Living with lupus nephritis means constantly adapting to a body that can be unpredictable. A life-threatening manifestation of lupus, lupus nephritis, occurs when the immune system attacks the kidneys, causing irreversible damage if left untreated. From varying symptoms, like inflammation, fatigue and pain, to the ongoing need for medical care, this disease can disrupt routines, limit activities, and impact every moment of the day for those living with it.

Behind the list of symptoms and statistics are real people finding strength in the face of a serious autoimmune disease. Raena, a person living with the disease, shares how she manages symptoms, cares for herself, and finds moments of joy each day, proving that life with a chronic condition is about more than survival; it’s about finding ways to thrive.

Starting the day with lupus nephritis

For Raena, managing lupus nephritis has taught her to listen closely to her body and adapt her routine to honor what it needs. While no two mornings are exactly alike, each day begins with a check-in with her body.

Some mornings bring relief. Joints feel at ease, sleep was restful, and tasks like washing up or applying sunscreen feel manageable. On these days, Raena moves through her routine with confidence.

Challenging mornings are a different story. When joint pain and swelling hit Raena as soon as she wakes up, she draws on her toolkit of coping strategies. She requires extra time in bed to prevent dizzy spells that can occur. She keeps an eye on the weather, aware that temperatures that are too hot or too cold can trigger a flare.

When the simple act of eating feels challenging, she sips on a nutritional supplement drink to keep her energy up for the day ahead. Through intentionality and recognizing her body’s signals, Raena has reclaimed agency in her daily life.

Navigating the day’s challenges

Approaching challenges as the day unfolds requires ongoing adaptation and resilience. Throughout her workday, Raena has developed smart strategies in response to her body’s needs. When her fingers cramp or swell, she listens to this cue to slow down, stretch, and adjust. When brain fog sets in, Raena relies on sticky notes and digital reminders to keep track of tasks.

Over time, she’s fine-tuned her diet, choosing hydrating foods that ease her joint flares and steering clear of known triggers. She’s also learned to stay vigilant about sun exposure, which can spark rashes and itching, by wearing sunscreen and covering up while she enjoys time outdoors.

Despite the challenges of her chronic illness, Raena cherishes quality time with family. Parenting while managing lupus nephritis has taught Raena to show up for her daughter in sustainable ways. She celebrates and encourages her daughter’s active life and prioritizes being truly present whenever flares subside. She stays connected through phone calls and leans on her extended family for support, something she has learned to do over time.

Setting the stage for a restful night

As evening settles in, Raena has cultivated her own rituals for winding down. When sleep doesn’t come easy due to pain, she uses the quiet hours to tidy up and organize her space – small tasks that give her a sense of calm and accomplishment. On the pain-free nights, she finds comfort in positive thoughts and settles into bed with gratitude. For Raena, the night is a time to reset and care for her body and mind, setting the stage for whatever the next day may bring.

Reflections on living with lupus nephritis

Raena's daily experiences show that living with lupus nephritis means staying adaptable, self-aware, and open to support. Listening to her body and what it needs is key. Simple routines help create a sense of stability. Above all, Raena emphasizes the importance of reaching out for help and building a network of support from family and friends.

As research continues to advance, stories like Raena’s highlight both the resilience of people living with lupus nephritis and the importance of continued innovation to ease the burden of this complex disease. Genentech is committed to improving patients' lives through science and advancing innovative therapies for lupus nephritis and other immune-mediated diseases.

In retinal care, communication shapes how patients understand their diagnosis, navigate treatment decisions, and stay engaged in care over time. This article highlights five practical principles to strengthen communication with patients and care partners in retinal care to improve access—offering perspectives patients, caregivers, and healthcare professionals can apply across the care journey. Drawing on insights shared during the Genentech and Prevent Blindness panel Removing Communication Barriers in Ophthalmology, these principles reflect a shared goal across the ophthalmology community: making communication a core part of delivering effective, patient-centered care.

Start with Listening

Before education comes understanding. Patients rarely arrive thinking only about imaging results. They may be worried about losing independence, navigating work and family responsibilities, or managing long-term treatment. Those concerns aren’t always visible unless invited into the conversation.

Julie Owens, a patient advocate living with thyroid eye disease, has described how overwhelming a diagnosis can feel—even for someone experienced in navigating healthcare. Information alone isn't enough. Patients need clarity, context, and space to process. A few simple focused questions (e.g., “What concerns you most right now?”) can reshape how a patient experiences the plan that follows.

Translate Complexity—Medical and Financial

Retinal care involves complex terminology, and so does insurance. For many patients, both can feel intimidating. As leaders at Prevent Blindness often note, health literacy is not about intelligence. It is about familiarity. When medical terms or referral pathways aren’t explained simply, patients may leave uncertain about next steps. Financial terms like “deductible” or “coinsurance” can create similar barriers, sometimes showing up as delayed scheduling rather than direct questions.

Plain language, clear summaries, and teach-back (e.g., “Can you tell me how you understand the plan?”) help ensure information is not just delivered, but understood. When complexity is translated thoughtfully, patients are better equipped to make informed decisions.

Putting Shared Decision-Making Into Practice

In retinal care, where treatments may be procedural and ongoing, patients can feel there is little room for dialogue. Yet even within clinical guidelines, there’s space to address concerns and preferences.

Shared decision-making means finding a sense of alignment—that patients understand the rationale, feel heard, and leave with a clear sense of partnership. In high-volume practices, operational support such as care teams, scribes, and structured follow-up materials can help protect the time needed for those conversations.

Support the Whole Care Team, Including Care Partners

For many individuals with retinal disease, care partners play a very important role. They coordinate appointments, assist with transportation, navigate digital portals, and reinforce instructions after visits. Yet their contributions aren’t always acknowledged.

Supporting care partners strengthens the care experience and is part of providing comprehensive care. Brief check-ins can signal inclusion and respect. Advocacy organizations, including Prevent Blindness, offer programs like the ASPECT Patient Engagement Program that helps equip patients and care partners with tools to communicate effectively and navigate systems with greater confidence. When care partners feel supported, patients are more likely to remain engaged in long-term treatment.

Use Technology Thoughtfully

Digital tools and emerging technologies, including AI, are increasingly part of healthcare. Their impact depends less on the tool, and more on how it’s introduced and contextualized. Technology can enhance communication by organizing information, visualizing disease progression, or reducing administrative burdens but it can also create friction if patients are unsure about how to use it.

Thoughtful integration requires attention to accessibility, comfort level, and clarity. In some cases, hesitation may reflect visual challenges. In others, it may reflect unfamiliarity with digital systems. Distinguishing between the two allows practices to respond appropriately. The goal isn’t adoption for its own sake. It’s alignment—ensuring innovation supports the patient-provider relationship.

Looking Ahead

Advances in retinal science continue to evolve, offering new possibilities for treatment and care delivery. As those advancements progress, communication will remain central to how patients experience them. When providers listen with intention, simplify complexity, encourage shared decision making, support care partners, and introduce technology thoughtfully, the result is more than understanding—it’s building greater trust.

Organizations such as Genentech and Prevent Blindness continue to explore how collaboration across the ophthalmology community can strengthen patient-centered communication.

Because in retinal care, communication isn’t an accessory to treatment. It is foundational to it.

We’re kicking off season seven with an in-depth look at the wild world of ophthalmology, where scientists are pushing the boundaries of what’s possible for vision treatments. In this episode, co-host Danielle Mandikian is joined by guests Deepak Lamba, Director and Distinguished Scientist, Ophthalmology, and Susie Crowell, Senior Director, Project Team Leader, to explore the latest in the field of ophthalmology since our season five episode on eye diseases. Together, they dive into the bizarre biology of how our eyes actually work, unpack how and why eye diseases develop, and discuss how transformative new therapies could one day restore sight.

If you would prefer to read a transcript of this episode, please click here.

SUBSCRIBE BELOW TO CATCH EACH EPISODE

If you want to learn more about the groundbreaking science happening in our labs, click here. To learn more about the jobs in our research and early development group, click here.

Transcript of Two Scientists Walk Into A Bar: “The Wild World of Eyeballs” with Deepak Lamba and Susie Crowell

Maria: I'm Maria Wilson.

Danielle: And I'm Danielle Mandikian.

Maria: And we are scientists. We. Love. Science.

Danielle: Yeah, we do. So, when we aren't doing it, the next best thing is to talk about science! And what's really awesome is that we're surrounded by some of the most brilliant minds in research!

Maria: We are going to step away from the labs today to talk to other scientists about the cool stuff they are thinking about, working on and imagining . . .

Danielle: . . . as well as how some of these discoveries just might lead to new medicines. So, grab your favorite drink, get ready to unlock your science brain and join us for Two Scientists Walk into a Bar…

Maria: The show for scientists, science geeks, and the people who love them!

Danielle: We're giving everyone a pop quiz. Can you spell ophthalmology?

Employee: O-P-T-H-T-H. . . wait sorry.

Employee: No, I don't even know what that is. [Laughs]

Employee: O-P-T-H- . . .

Employee: It's like O-P-T-H-A-M-O-L-O-G-Y. No?

Danielle: That was the fastest anyone has spelled it.

Employee: Was it right?

Danielle: I don't know, but it was fast.

Employee: It's wrong, yeah. [Laughs] Okay.

Employee: O-P-T-H-A-L-M . . .

Employee: O-P-H-T-H-A-L-M-O-L-O-G-Y.

Danielle: Winner! Are y'all in ophtha?

Employee: I'm the one that discovered it was spelled wrong in the database! [Laughs]

Danielle: Hi, everyone! So this season, we're concentrating on groundbreaking discoveries. And since discovery starts with vision, we're gonna take a deep dive into what's really pushing the boundaries in ophthalmology. Back in season six, we explored where vision research is focused. See what I did there? But today, we're gonna learn about the wild ways that research is pushing for transformative therapeutic strategies. I'm excited to introduce and hang out with cutting-edge scientists who live and breathe vision. Susie Crowell, a senior director and project team leader – as well as an old friend. And Deepak Lamba, director and distinguished scientist. Welcome to the bar!

Susie: Thanks for having us.

Deepak: Thank you.

Danielle: So, like in our last eye episode, we like to start off with basic understanding. In that one we covered how does the eye actually work in the biology? But I thought it'd be fun to get meta with it. Susie, how do we even perceive vision to begin with?

Susie: It's so much wackier than I ever thought about before I started working in this space.

Danielle: Tell me more.

Susie: How eyeballs work is just wild. I mean, basically all of our senses are just like taking whatever it is in our world and turning it into electrical impulses that our brain deciphers, and we're just so good at pattern recognition that we're like, "hey, what if we paint it purple," or whatever it happens to be. And it's a really complicated sense where we're taking light in through, you know, the window, the pupil in the front of our eye – it hits our retina and then gets, you know, transmitted through the optic nerve coax cable into the visual cortex in the brain where the brain's like, "I'm gonna make meaning out of this." But like, the number of things that have to happen there, it's really complicated! [Laughs] And it's kind of crazy when we talk about trying to make medicine that addresses things in the visual realm because like how do you measure if it works when we could have, like, a really bizarre philosophical discussion about whether I see the same thing that you see?

Deepak: Yeah. It's complicated. It's not like light goes in, reaches the brain, and you see something.

Susie: Yeah.

Deepak: There's different ranges of light. In fact, our light-sensing cells – the light-sensing photoreceptor cells – these cells have a weird way. So they're normally active all the time. So when they sense light, they stop responding to light, and that's how our brain senses. It's called hyperpolarization. So when light hits these light-sensing cells, suddenly the channels close down, and that causes our secondary order neurons to sense light.

Danielle: Right, so that makes me think of glaucoma – right? – where there's a problem with these neurons even communicating to begin with. How do we even deal with that? How do we get them to start working again?

Deepak: Glaucoma is way more challenging because then that's the one and only connection between your eye and your brain.

Susie: That's like the cord is gone.

Deepak: And so you can't just join the frayed ends together. You actually have to have the new cells make a new cord all the way. And that is the reason it's not been very successful so far – because the cues that are there in the microenvironment to tell you, "okay, this is your route to the brain," go away once you are fully developed.

Danielle: What about for other parts of the eye that are needed for vision? Are there other places where there is damage that actually leads to pathology?

Deepak: Yes. And there are different sort of initiators of damage which can cause different disease. So like, there are people who have inherited degenerations, for example. People like – who have retinitis pigmentosa, which is this broad disorder with different mutations in different genes which are involved in light sensing are affected, and so that's the cause. So that causes abnormal proteins to accumulate inside the cell, and the cell is like, "I'm done, I can't deal with it." They start dying. And so you lose those cells. At least early in the disease, the rest of the circuit is still there, because it's a local circuit – only one cell right next to it is gone. So potentially, you may be able to replace it. But the big challenge with all of it is our brain and our eyes are so smart that they feed in the missing information. So essentially, you can lose close to 80 percent of your cells and not realize you have vision loss.

Danielle: That's super scary.

Deepak: It is insane.

Susie: Cause your brain is like photoshopping it for you.

Deepak: Exactly.

Susie: This is my thing, because I'm missing like a huge chunk. So we're sitting in a triangle right now. We're all about three feet from each other. When I'm looking at you, Danielle, I can't see Deepak at all, even though he's in a part of my peripheral vision that most people would be able to. But there's just like a weird, you know, navy blue smear. [Laughs] Like it's like an AI generative, whatever. And it's my brain trying to be like, "well, I know this was there a minute ago when she looked that way, so I guess I'm gonna fill it in."

Deepak: Yeah. The brain fits in the information, but it often fakes it. And so there are people, like people who have advanced diseases, where they don't realize it, but then they hit a pole, and –

Susie: Yeah.

Deepak: – then it's like, how did I miss that it was right in front?

Susie: Well, then for some of these that are degenerating slowly over time, perceiving that change without having, kind of, you know, tests or something that's making you notice it is really difficult. So when I first started in this space, I thought a lot about – so neovascular age-related macular degeneration where blood vessels are growing in parts of your retina they're not supposed to. And that causes leak of fluid, and the fluid, you know, messes up the optics essentially – can't see. And there was a lot of talk about how patients don't comply with their treatments.

Danielle: Yeah. [Laughs]

Susie: And if you're also, you know, elderly and have somewhat impaired vision, it's not like you can just drive yourself there. It's a big problem. This isn't compliance. [Laughs] I get very upset!

Deepak: Exactly.

Susie: Because my experience of losing vision slowly versus abruptly – I have experience with both of those – and it's a completely different situation. The slow loss, I only know about it because I'm obsessed with my visual field –

Danielle: Yeah.

Susie: – and I check it all the time.

Karen: Hey, Danielle.

Danielle: Hey, Karen.

Wellington: Hey, Danielle.

Danielle: Hey, Wellington.

Karen: Danielle, how do you explain why compliance is so important for people who have eye diseases and are taking treatments? How is the field addressing this issue?

Danielle: Oh man, that's like a huge area. So first of all, for compliance, the medications only work if they're at the right concentration in the target area for the right amount of time. So, if you want to get to a certain concentration, you have to have multiple routes of administration. But it's really uncomfortable, and the patient population that has to get it is already at a disadvantage because they can't see well, so they need to get someone to take them to their appointments. They're usually a higher-aged population. So there's a lot of hurdles to actually getting treatment. So to address that, there's a lot of work that's done to try to do what are called, like, long-acting delivery systems, or something that helps it so that you don't have to keep giving as many doses. And if you don't have to do that as often, it makes it easier for people to actually follow through. That's exactly what Susie and I worked on together!

Danielle: Most of the time whenever I talk to folks about diseases in the eye, the most common culprit is problems with vascularization and VEGF. Deepak, could you tell us a little bit more about this?

Deepak: So basically, in some of these vascular diseases, what happens is there's a cue from the dying cells which tells them we need more food. And that cue to the vascular cells is VEGF. VEGF stands for vascular endothelial growth factor. And so this factor tells them, "oh, we need more blood with us to grow in because we need more food." Unfortunately, it reaches a point where these cells are not controlled anymore. And so they just start migrating and invading into the eye and then become leaky because they are newer cells – the body's not used to having new blood vessels in older ages. And so what our therapies are is to try to block that to see if we can sop up all of that extra VEGF in that milieu so that your vessels now stop growing in. And that sucks up all the extra fluid.

Danielle: So like a VEGF therapy is really something to sop up all the VEGF.

Deepak: Pretty much.

Susie: Because it's like – the cells are like "I'm hungry!"

Danielle: [Laughs]

Susie: Your body's like, "panic!" And it overbuilds and freaks out and it's leaky.

Deepak: Exactly.

Susie: And then we want to calm it down.

Deepak: Yeah, damp it down. And so that affects that part. But again, as you can imagine, also part of that is if you're doing that and your cells are growing into the wrong spaces, leaking – that drives inflammation. And so those are the things that are missing right now. So we can target the VEGF really well, the vascular supply. What we are missing is what's around it. You're trying to stamp down the inflammation that's happening in that space. Also, these new blood vessels cause other defects which cause cells to become more fibrotic. And so that's another signal that develops in these patients where there's local fibrosis. And if there's a thick scar in the back of the eye, then the eye is not gonna function well. And so those are the areas that we need to still try to figure out – how do we get to those biologies? And these are all side effects. Like the main part of the disease is still happening that we still need to get to.

Susie: I think it's been about 20 years since anti-VEGFs really came onto the market in full form. And that was a massive gamechanger, as I think you guys have previously discussed, where patients weren't just, you know – we weren't just slowing disease down. They actually got vision back as we like closed down the new blood vessel factory. [Laughs]

Danielle: Yeah. [Laughs]

Susie: And that fluid would go away. And that's been huge, and I think there's been a lot of iteration on the anti-VEGFs in the field. There's been a little bit of new biology introduced in this neovascular AMD and then the other kind of leaky diseases, so diabetic eye diseases.

Danielle: Mm-hmm.

Susie: But there hasn't been another one of those like massive, transformative inflection points since then. There's been kind of chipping away.

Danielle: Well, that's like the alley-oop for y'all, right? VEGF is for the vascular-related kind of pathologies. But what about the ones that y'all were talking about where photoreceptors have been damaged and are missing? Are those gonna be in pathologies completely independent of those VEGF? Or like, how would you even address those, maybe is a better question?

Susie: Both glaucoma and, like, geographic atrophy.

Deepak: Yeah. So basically, even when talking about just neovascular AMD, macular degeneration – that's the end stage of the disease. What leads to it is some of the earlier stages – intermediate AMD.

Danielle: Mm-hmm.

Deepak: And that can either lead to that or to GA, which is geographic atrophy. So it is like the area is dying, and the whole geographic area has atrophied. And that's where you're losing these light sensory photoreceptor cells, the support cells at the back of the eye, called the pigment epithelium. And so there are lots of efforts on seeing once you've lost these cells, can you replace them back? So there's a few things in the clinical trial stage right now of trying to replace either the pigment epithelial cells or photoreceptor cells. It's very, very challenging, as you can imagine, putting new cells in an environment which is undergoing active disease.

Danielle: Mm-hmm.

Deepak: So, that's one area. The other area is, can we save what's being lost? Can we go in early, try to hit some of those pathologies early? So that's that. And then on the other side is glaucoma. Glaucoma is a disease where, as I had mentioned earlier, the main cell going from the eye to the vein – the main connection – those cells are dying. Currently, all the treatments are to try to reduce the pressures. Often, this is connected to pressure in the front of the eye.

Danielle: Mm-hmm.

Deepak: And we have lots of drops and things like that for it, but they never get to the main problem of the disease, which is these cells dying at the back. And so that's another big area where we need to figure out how to start saving them.

Danielle: Maria was super excited that we were gonna talk to y'all. And she really wanted me to ask you to give a little bit more detail about how cell therapies are actually gonna potentially play a role in this area. Because I mean, this is such like a new place to be thinking about it. Normally for cell therapies, I think most of us hear about it in the oncology space, but not in ophtha.

Susie: There's some really cool stuff going on here and a lot of different approaches. So Deepak's mentioned a number of cell types that are at play here. You've got your nerve fiber that goes to the brain. You have the pigmented epithelial cells. I think of those as like the garbage dump and grocery store workers in your eye.

Danielle: [Laughs]

Susie: They're taking care of the photoreceptor layer, which is like the business center. And then, you know, there's a few other cell types in there as well. But let's focus on maybe those three. There's some cell therapy approaches where they're basically like – alright, in geographic atrophy where a lot of those pigmented epithelial cells have dropped out and you just have a big desert –

Danielle: [Laughs]

Susie: – food desert, they're like, okay, let's just put a slurry of, you know, a mix of RPE cells in there, fresh ones, and see if they'll, you know, take root and hang out. So that's one approach. There's another where they're like, okay, let's culture, you know, a film of RPE – it's the retinal pigmented epithelial cells – and then put the entire, like a fancy band-aid, just put that whole thing in there and like unfurl it.

Danielle: Oh, almost like a tissue graft kind of?

Susie: Yeah.

Deepak: Mm-hmm.

Danielle: Dope.

Susie: Yeah, it's like really cool.

Danielle: Yeah. [Laughs]

Susie: And then there's other ideas where it's like, alright, you lose your RPE layer, and you know, what happens when you have a food desert? People move away. The photoreceptors, actually in this case they die, so it's a little bit darker.

Danielle: Yeah. [Laughs]

Susie: So then you have two problems, you know? You need to reestablish the food and garbage capacity, and then you need the people to come back – the photoreceptors in this case. So let's put a mix of both in there and see if they work themselves out. And will they do that if you mix them together into like a cell replacement smoothie, or is it you have to do one first? Deepak's like "no, Susie."

Deepak: [Laughs]

Susie: Do you do like one layer first and then the other? Or do you need to do the tissue graft? There's a lot of, like, geometry and architecture and gross smoothies, I guess.

Danielle: Cell smoothies are super popular right now.

Deepak: [Laughs]

Susie: That's so gross.

Danielle: I know.

Susie: I regret saying that.

Danielle: I know, but everyone and their mom is all about cell smoothies and seeing like how to rearrange each other.

Susie: It's crazy. Like how do these guys go where they're supposed to?

Deepak: [Laughs] So yeah. So everything she said is exactly how it is. Like we have people who are trying to figure out how we can put either on a single-cell layer.

Danielle: Mm-hmm.

Deepak: The pigment epithelial cells, those grocery slash garbage spaces – very complicated to put it together – but that's exactly what they do. They are the main cells, which because your light sensing cells actually have no vascular supply. So they actually rely –

Danielle: Mmmm.

Deepak: – on these pigmented epithelial cells.

Susie: Well and they're so busy. These are like the most metabolically active cells in your body.

Deepak: Absolutely.

Danielle: I didn't even think about it. I don't know how many retinas I've stared at, but yeah.

Deepak: So they actually have to bring in the food, get rid of all the dead stuff. And what makes it even more complicated is our light-sensing photoreceptors have this weird adaptation at the tip called their outer segment. And they shed 10 percent of that every night – every night when you go to sleep.

Danielle: Did you know that?

Susie: So I did technically know that, but like I don't really get it, you know? Like . . .

Danielle: Why have we not talked about that before?

Susie: It's just . . . like it's a crazy system.

Deepak: It's insane. So at night, they just have to go crazy because they have to chew it all up, package it all together, and get rid of what needs to be gotten rid of. But also put things back so the photoreceptors can keep making more.

Danielle: Yeah.

Deepak: Because they rely on the same cells.

Susie: Like little sidebar – the other thing that happens while you're sleeping is that the rest of your brain is like, "oh, the visual cortex isn't busy. What if we just put that to work for other things?" So your brain is like fighting to take that space away and, like, give it to other senses.

Danielle: Like temporarily hijack it?

Susie: No, it wants to take it forever. And then you wake up and it's like, never mind, let's give it back. [Laughs]

Danielle: Shut up. [Laughs]

Susie: It's crazy! It's crazy. The whole thing is nuts. I can't believe it works. [Laughs]

Karen: Okay Danielle, I think I missed this in my schooling, but can you talk more about this smoothie metaphor, please?

Danielle: Yeah, I mean, you know what, you can tell who trained in biochemistry versus cell biology whenever you say cell smoothie by how hard they laugh, slash grimace. [Laughs] But like, you know, basically whenever we have like, you know, systems that we're trying to create in vitro, it's like a mixture of cells. So sometimes the way you get that mixture of cells is you go to an organ and you dissociate the cells and it – I mean, you don't put it in a blender, cause that's like when you go the biochemist route, but for the cell biologist in the room, we do tricks to like dissociate the cells and it just literally looks like a strawberry smoothie. That's what everyone's insides look like. Fun fact. Hope that doesn't ruin your morning routine or protein shake.

Wellington: Okay, non-scientist here – that's a little gross.

Danielle: [Laughs] Yeah, but it's totally utilitarian, right? So the whole point of this is that for whatever end application, we need to get to cells, right? But if we're harvesting them from whole organs, we have to be able to separate them and either purify certain cell populations, culture the ones we want, whatever your game is – you've got to have them separated and that requires a smoothie!

Danielle: I was really shocked to hear people even talk about optogenetics as a form of treatment. I mean, I'm familiar with that from neuroscience where we're trying to use it as a tool to understand circuitry. How is this even an option for a therapy?

Deepak: So optogenetics is absolutely fascinating. So it's the field where you're trying to make non-light-sensing cells see light. And as you point out in neuroscience, it's cool – like can you make certain parts of your brain respond and see how it's connected? But in the eye, because you can actually do it in a live human person, maybe you can use it therapeutically. And so there's a lot of excitement in that area on seeing if once you've lost your light-sensing cells, can you give that ability back to the second order neurons, which are these bipolar cells, or the ganglion cells, axon cells connecting to the brain. Of course, you lose some resolution by doing that because there are about 200 million photoreceptors in our human eye, and you're down to a few thousand of the inner retinal neurons. But as I said, there's so much redundancy in our vision that there is a potential that these few cells could still give you some useful vision.

Susie: I saw at a conference a number of years ago – not that many years ago –

Deepak: Mm-hmm.

Susie: – you know, I think some of the early testing of this kind of approach is – it's occurring in people who are fully blind.

Deepak: Yes.

Susie: No light sensing whatsoever. And there was a woman that they tested this on who had no light sensation. And then they did this, and then she was able to basically locate a pen on a table –

Deepak: Yes.

Susie: – where it was like high contrast, good light conditions, but like that is completely insane! Zero light sensation to, like, I can find an object. And like, when we're talking about being able to – I mean, the way that sight underpins your ability to participate in society . . .

Danielle: Yeah.

Susie: It's not to say that you can't when you're fully blind, but we don't make it easy.

Danielle: Yeah. [Laughs]

Susie: And to be able to find an object –

Danielle: Mm-hmm.

Susie: – or even just detect, you know, a little bit, is – it's . . .

Deepak: Huge.

Susie: Like science fiction, a) of all, and b) just so massive an improvement for somebody's ability to like, navigate or interact.

Deepak: Yeah. And the field has moved quite a long ways. Initially, these optogenetic proteins people were putting were like in very small dynamic ranges, very small wave lengths of light that responded. But the field has moved so much now that the current clinical trials can actually work where you don't need special goggles to see things like that. They could work in normal light levels and within the normal light wavelengths. Of course, you won't get color vision back, but you would – if you can even, as Susie pointed out, make sense, not hit a wall right in front of you because you couldn't see it – those are huge improvements in these patients.

Susie: And like as a person who, despite still being – you know, I have good visual acuity on the eye chart, but like I run into things a lot. [Laughs] I get a lot of concussions.

Danielle: [Laughs]

Susie: Knowing that like there may be – you know, we can't fix all the problems, but like there is a possibility for giving back some light sensation, it gives me some kind of like – I don't know . . .

Danielle: Gets you in the feels?

Susie: It really does. It really does.

Danielle: It's a – you know, it's the same kind of thing, too, where you look at, like, how people have been kind of taking on like a different strategy for prosthetics.

Susie: Mm-hmm.

Danielle: It's that same thing of like, how are you mapping into the brain and just restoring this? It's such a trip.

Susie: Oh, it's crazy. There was another podcast I listened to where they like hijack other senses.

Danielle: Uh-huh.

Susie: And Deepak knows what I'm gonna say. There's – okay, there's this one episode where they like basically convert, you know, there's a sensor that converts information about the spatial environment into electrical impulses, and they basically attach it wherever. But this one, it was to the tongue so that you could see through your tongue. And it works!

Deepak: The other science fiction thing that's happening is this chip technology. So there was this recent study that just came out – it's a company based on work here at Stanford – of an electronic chip that has at least rescued some vision in patients. It's like a chip with like a hundred electrodes, you put where your light-sensing photoreceptors are.

Danielle: Mm-hmm.

Deepak: And the chip is wirelessly connected to, say a camera that's on – that you just wear as a part of your clothes.

Danielle: Mm-hmm.

Deepak: And that sends a signal to the eye.

Danielle: What a trip.

Deepak: So that was like one of those early things that I was working on on prosthesis, and there was one company that made a little bit of a leap in that space. But, again, it was so difficult to get these chips to work, and this updated version of the chip, the data looks really interesting.

Danielle: Are they like carbon nanotubes?

Deepak: Something like that. These are micro electrodes. They're spatially depolarized based on which area. And so originally when these were made, they were connected by wires to the –

Danielle: Mm-hmm.

Deepak: – to a sensor on the front of the eye, and that had to connect to a camera, which was in a box that you would have to hang around and walk. But doing those early studies were very fascinating. And there's videos of people who received some of these early chips. They could actually walk on the road because they could make out those edges of the roads –

Danielle: Yeah.

Deepak: – not fall down, see when to stop. And now it has moved to this new version of the chip, which is actually even more better adaptable.

Danielle: Well, so this brings me up to the next thing. Thinking forward into the future, what are some things that you are the most hopeful for? You think we're gonna be like – just fast-forward 25 years.

Susie: Oh my god, I really want a glaucoma drug that's not an eyedrop. Like desperately. Like glaucoma – Deepak kind of mentioned it earlier – but like what you do when you get glaucoma is you put eyedrops that sting into your eyes between one and three times a day. And that might help, but it might not. It's certainly not gonna feel like it's helping.

Danielle: [Laughs]

Susie: And same kind of deal where we're talking about like, oh, they're not complying with their treatment regimen. And I'm like, well, it hurts, and I can't tell if it's helping me.

Deepak: And they can't see, and they're trying to put these drops in there.

Susie: So I just – I want like neuroprotective agents. So the eyedrops, they're usually intraocular pressure lowering, which technically helps protect your neurons if there's too much pressure on them.

Deepak: Mm-hmm.

Susie: But something that directly protects and supports the health of the neurons so they aren't dying, I would just be thrilled if we managed that.

Deepak: For me, it's cell therapies.

Susie: Yeah.

Deepak: I'm super excited about it. It's gonna be a challenging right there, but I'm very hopeful – that's the only way we are actually gonna rescue.

Susie: Right.

Deepak: Because we can do things to slow things down. But there are lots of patients who are – reach that stage that you have to figure out a way to rescue them. And to me, that's the ultimate rescue.

Susie: I have another one!

Danielle: What you got? [Laughs]

Susie: I'm so – 25 years – we're on the cusp of changing the way that we assess vision. So like the way that we do clinical trials now, by and large for ophthalmology, is best corrected visual acuity, which is literally the eyechart. How many letters can you read? We talk about how the trial went in terms of how many letters you could read between, you know, the group that was treated and the group that wasn't.

Danielle: Does the photoshop thing that you were talking about screw that up?

Susie: Yes. My best corrected visual acuity is amazing. And again, triangle – can't see Deepak.

Danielle: [Laughs]

Susie: Like I have – that's not considered clinically significant. I'm missing over half of the vision in my right eye – it's significant! But it doesn't register on that eyechart. So there's other things like visual field, which there's, you know, long-established ways of measuring, but we're starting to see things like using virtual reality headsets, which as a patient I can tell you, it's – the test feels very similar, and they compare very well to the kind of gold standard way of doing this – the Humphrey Field Analyzer. But I get to sit on my couch, and I can do it, you know, more often. It's so much more accessible. And that also gives us the ability in clinical trial design to actually take more measurements, which gives us more sensitivity. And there's – you can use them to do other things completely, like mazes.

Deepak: Like playing a game. [Laughs]

Susie: Or playing a game. My kids like it, yeah. [Laughs]

Danielle: And I wanna come over and play that, yeah.

Susie: There's a lot of cool stuff with visual endpoints. And that's just like the tip of the iceberg. There's other kinds of ways to measure sensitivity. And just the idea of being able to measure vision in ways that are clinically meaningful and actually get at the, you know, full – I mean, this is such a robust sense. We talked about how it's hard to know when it's changing. Being able to actually say, look, you have intermediate AMD. And even though you can't tell or you haven't noticed yet, like maybe it's just that your anxiety is a little higher and you don't know why.

Deepak: Mm-hmm.

Susie: But it's actually because you can't see as well. [Laughs]

Danielle: [Laughs]

Susie: And it's just something we haven't been able to measure.

Deepak: Yeah it's the small changes. Like you lose depth perception to some extent.

Susie: Mm-hmm.

Deepak: All of that you really cannot measure.

Susie: Nighttime vision.

Deepak: Nighttime vision, yeah.

Susie: And like suddenly you don't want to go out anymore, and you're like, oh, I just, you know, don't really like it. And it's like, no, it's because you can't see as well.

Danielle: What is your call to arms to the scientists out there? What breakthroughs need to happen in order to actually make that 25-year plan realistic? Call to arms. Not a to-do list. [Laughs]

Deepak: [Laughs]

Susie: Yeah. Now I'm thinking about my to-do list. It's a mixture of, you know, bold science where we're, you know, as individuals, as companies, as a field we're willing to take smart risks on what we're trying. And really synthesizing what we understand across what different experiments are being done in the field and then making small steps along the way at the same time.

Deepak: That's what I was going to say. It'll be incremental. Don't expect it to be like – hit the jackpot the first time. Because we'd never do it. But using the knowledge that we're gaining to get that incremental success. There will be obviously a fairly large number of failures, but just accepting a failure and saying this is not gonna work is not gonna get us there. So we need to just be diligent at it. Fund the right type of experiments in the lab and those experiments in the clinic. And I think that's how we're gonna achieve what we need to get to those patients eventually. Because it's the patients in the end.

Danielle: So like that, I'm – at my core, agree with y'all. Make those smart risks. Take the incremental steps. But the other thing that sometimes whenever I'm talking to early-in-career scientists is sometimes surprising about what scientific backgrounds need to be at the table.

Susie: Mm-hmm.

Danielle: So cell biologists. PK scientists. Neuroscientists.

Susie: Yeah, I mean, there's a lot – a lot of what is necessary to move the needle on these different, you know, potential transformations is super cross functional, by which I mean we need a variety of backgrounds at the table. This isn't something that's simply a, you know, I don't know . . .

Deepak: Single lab solving thing.

Susie: Yeah.

Deepak: It's always – there has to be a team behind. Depending on what you're trying to get. If you're looking at, say, optogenetics, you require how do you deliver it? So there's the wireless people that need to work. Then there's the protein people to figure out how these proteins are gonna respond to light? What's the right range we need to work on? There's people who . . .

Danielle: You need structural biologists.

Deepak: Structural biologists.

Susie: We need people with like engineering backgrounds.

Deepak: Sure.

Susie: And we need people who can communicate between all of these different –

Deepak: Yeah.

Susie: – backgrounds, because a lot of us don't naturally understand what each other is saying.

Deepak: Yeah.

Danielle: Yeah, right? [Laughs] That – everyone laughs when you bring stuff up like that because it's so painfully true.

Susie: It's super true, so.

Deepak: And then if you're talking about like AMD or wet AMD, there's a – now you're looking at a vascular disease. You have the inflammation disease. Now you – not only that, we want a protein scientist who can connect those things together to make these bi-specifics, multi-specifics.

Susie: Mm-hmm.

Deepak: Hit multiple biology simultaneously – that's what multi-specifics are supposed to do. Reach the site of action – that's where the PKPD people come. It's like no, you made this amazing molecule – it's so big that it just won't even reach where you want it to go. [Laughs]

Susie: Bad news. [Laughs]

Deepak: And it's like, oh man, that's awesome. [Laughs]

Danielle: That's our backstory. That's our origin story.

Susie: Yeah.

Deepak: Yeah. So it's one of those challenges. I think it's a team approach, and the more people talk to each other, I think we – that's the way to get there.

Susie: Thinking about optogenetics in particular. There are so many components to that that you need to get to work. And how do you think about developing a therapy when you have to fix, you know, 17 different things at once? And the answer is that you can't fix them all at once. You need to make progress on them individually if you can. And so having multiple kind of opportunities within one organization that inform on different parts of that equation can be super helpful. Or looking, you know, across the field at what are we learning from this attempt and this experiment over here, non-clinically or clinically? And I'm thinking about there's been a lot of work on using gene therapy approaches for what's called bio-factory approaches where you're basically, you know, injecting these into the eye and asking the cells to make a little something extra – maybe an anti-VEGF, for instance. And that's where a lot of the, like, work in AAV or gene therapy in ophthalmology has been – not all of it. But using that to understand, you know, how much are we getting in there and like actually getting it to work and do the thing we're asking it to do can be really helpful when you're trying to figure out how to do something completely different, like optogenetics. But maybe you're using the same way in, the same door. And then you're basically kind of simplifying that question a little bit by solving for something else in another location. And I'd like to see, particularly when I think about this from kind of a sort of organizational framework, how do the different things that we're pursuing inform each other and, you know, move us forward together instead of just thinking of them individually? So . . .

Deepak: Mm, yeah. No, these cross-group learnings are so critical as we want to move this forward. Yeah.

Danielle: That's my favorite thing about being a scientist. [Laughs]

Susie: It's so fun.

Deepak: Yeah, it's so much fun.

Danielle: I have to say, this has been a lot of fun. It's really great to meet you, Deepak, and it's so great to see you again, Susie.

Susie: Thanks.

Danielle: This was a lot of fun! Thank you so much for joining us.

Susie: Thank you for having us, yeah.

Deepak: Yeah, thank you for having us. This was awesome.

Karen: Danielle, really cool episode. Eye-opening, if you will.

Danielle: [Laughs]

Karen: What do you think is the next biggest breakthrough in ophthalmology, given your experience in the field?

Danielle: Probably the thing that I know the least about, which would be regenerative cell therapies. I think that's so amazing, so cool. And I really think that that can get at some logistics that some of our classic pathways just can't even touch, right? Because most of the time we're trying to sop up a protein that's not supposed to be there or something like that. But having a therapy that can actually replace damaged cells – like, that's groundbreaking.

Wellington: It was incredible to hear about some of the technologies that take even non-light-sensitive cells and make them quote unquote see. Do you think we're gonna get to a place where we can really ameliorate the vascular diseases, or do you think we can actually go from zero and restore somebody's vision that never had it?

Danielle: So I'd say both. We have strategies that are really great at dealing with the vascularization, we just need to improve maybe some delivery and make it easier for patients. But also, I think this idea of expanding how people can process vision, like that's something that's so new that we haven't been able to touch ever before. I mean, that's really where it's at – I think that's cool.

And that's our show! Thanks so much for listening. If you haven't already, rate our podcast wherever you listen. It'll help new people find us. And make sure to subscribe! If you have any questions about the show, you can contact us at podcast@gene.com. And now for me, it's back to stalking cells!

The name Two Scientists Walk Into A Bar is under license and used with permission from the Fleet Science Center

For many people, conversations about weight come with years, and sometimes decades, of judgment or uncomfortable experiences, including in healthcare settings. Too often, obesity has been framed as a personal failing rather than what science clearly shows it to be: a chronic, multi-faceted disease shaped by biology, environment, and access to care.

The words used to talk about obesity matter because they shape how people approach and experience care — and whether conversations with healthcare professionals feel respectful, productive, and grounded in science rather than judgment. When language reflects biology instead of blame, it creates space for more meaningful dialogue and more effective care.

Scientists like Maria Wilson, PhD, gRED director and cardiovascular and metabolism research head for Genentech, are working to make the biology of obesity easier to understand so people can make informed choices about their health.

“Everyone has heard about obesity and GLP-1s in the last few years, but it’s important to journey deeper than the news cycle,” says Dr. Wilson. “Understanding the biology behind weight gain can help us feel more informed and support better conversations with clinicians, so people living with obesity can understand their bodies without judgment,”

What Obesity Is — and Isn't
Obesity is a chronic disease shaped by biology, not a lack of willpower

Obesity is a chronic disease – not an issue of willpower, as it has long been portrayed. It stems from complex interactions between genetics, neurobiology, eating behaviors, access to healthy foods and the broader environment.¹ These signals operate largely outside conscious control, helping explain why hunger and fullness can feel misaligned — and why managing obesity is not a question of willpower.

To fully understand obesity as a chronic disease, it’s important to look at the biological systems that influence hunger, fullness, and how the body regulates energy. When hunger signals don’t seem to match food requirements, it isn’t imagined. Some people who live with obesity report a constant internal talk track about food and meals, which can feel all-consuming with no off switch. Behind that “food noise” is a complex network of biological signals that regulate appetite.

One part is within the digestive system, which releases hormones called incretins when we eat. The two main incretins are Glucagon-like Peptide-1 (GLP-1) and Gastric Inhibitory Polypeptide (GIP), which support communications from the gut to the brain. GLP-1 and GLP-1/GIP therapies are designed to work with these biological pathways by addressing underlying drivers of appetite and metabolism.^2,3,4  Natural incretins have subtle effects on appetite and satiety.

Therapies mimic this natural effect by providing superpowered versions of these hormones that more powerfully influence food intake.

Other hormones that impact weight management are produced in the pancreas. These include insulin, amylin and glucagon. Insulin is the main regulator of blood sugar, lowering blood sugar levels, typically after eating to prevent them from rising too high.⁵ Glucagon is a multi-functional hormone that counteracts insulin to prevent blood sugar from dropping too while also impacting how the body uses energy, in turn contributing to weight management.⁶ Amylin helps slow how quickly food leaves the stomach and promotes feelings of fullness.⁷ Together, these hormones work in a constant coordinated system to regulate blood sugar and energy balance.

History helps explain why these hormones do what they do. Dr. Wilson notes, “When food was scarce for our ancestors, the body developed alarm systems to prevent too much weight loss. Today, food is readily available, but those biological signals remain. Understanding them allows scientists to develop approaches that work with human biology rather than against it.”

How Doctors Assess Health
And why the conversation is evolving

Understanding obesity through this biological lens shifts the focus from “lifestyle choices” to care grounded in evidence. This shift can fundamentally change the tone of a doctor’s visit.

While obesity is a commonly used term, scientists often use more precise language to describe what’s happening in the body. One example is adiposity, which refers to how much fat tissue the body stores. Fat cells are not inherently harmful - they serve important roles in insulation, energy storage, and hormone signaling.

However, when adiposity becomes excessive, adipose tissue can send altered signals that affect appetite, metabolism, and how energy is stored and used. Understanding this helps explain why hunger, fullness, and energy balance may feel out of sync for some people.^8,9 “Adipose, also known as fat, shouldn’t be a dirty word,” Dr. Wilson adds. “Humans need fat cells for their body to work well. The goal of weight management should be finding adiposity balance, as having too much — or too little — can cause malfunctions in the body.”

Increasingly, patients and providers are moving away from a narrow focus on the scale toward whole-body health — including metabolic risk denoted by lipid panels, related conditions like hypertension and diabetes, and overall wellbeing.¹⁰  While Body Mass Index (BMI) is a useful population-level tool, it does not account for fat distribution, muscle mass, or underlying metabolic health. Two people with the same BMI can have very different cardiometabolic risk profiles.

Clinicians may also assess comorbidities — additional health conditions such as type 2 diabetes, hypertension, or heart disease — as well as markers like blood pressure, cholesterol levels, and blood glucose. Metabolic syndrome refers to a cluster of risk factors that increase the likelihood of cardiovascular disease and stroke.

These terms are common in medical conversations. But when discussed in isolation, they can unintentionally narrow the focus to a number on the scale rather than the whole person. When language broadens in this way, it can help patients feel more empowered to ask questions, seek clarity, and advocate for discussions that are centered on health outcomes - not just weight.

This is where individual biology comes in: understanding the physiological drivers behind obesity gives patients and clinicians a fuller, more nuanced picture of health, supporting conversations that are productive, science-based and judgement-free.

A Chronic Journey
Why empathy, understanding, and science must evolve together

As scientific understanding advances, so does the language used to describe obesity. One emerging framework is adiposity-based chronic disease (ABCD) - a term that reflects how excess adiposity can drive downstream health conditions, from mechanical strain on joints to metabolic dysfunction that increases cardiovascular and renal risk.

This approach shifts the focus from weight alone to the broader health impacts of adipose tissue and reinforces that obesity is a chronic, biologically mediated disease requiring long-term management.

No one chooses the biology they are born with. People living with obesity should not have to navigate a chronic disease alone or feel judged for factors beyond their control. When conversations with healthcare professionals are respectful, productive, and grounded in science rather than stigma, patients are more likely to engage openly and explore treatment options. Obesity is a recurring disease that requires scientific advancements and empathy to address. There is more to be done in how we treat this condition, and at Genentech, teams are working to advance more personalized and sustainable approaches to weight management – with a focus on treating the whole person, not just a number on the scale.

Changing how we talk about obesity is one meaningful step toward evolving how it’s understood, treated and experienced.

¹ World Health Organization. Obesity and overweight. Available from: https://www.who.int/news-room/fact- sheets/detail/obesity-and-overweight. Accessed Jan 21, 2026.
² GoodRx. What Are Incretin Mimetics, and How Do They Affect Weight Loss, Blood Sugar, and Type 2 Diabetes? Available from: https://www.goodrx.com/conditions/diabetes-type-2/what-are-incretins. Accessed Jan 21, 2026.
³ World Health Organization. Obesity: GLP-1 therapies. Available from: https://www.who.int/news-room/questions- and-answers/item/obesity-glp-1-therapies. Accessed Jan 21, 2026
⁴ WebMD. Can You Boost GLP-1 Naturally? Available from: https://www.webmd.com/obesity/features/natural-glp1- boosters. Accessed Jan 21, 2026.
⁵ Cleveland Clinic. Insulin. Available from: https://my.clevelandclinic.org/health/body/22601-insulin. Accessed Feb 11, 2026.
⁶ Cleveland Clinic. Glucagon. Available from:https://my.clevelandclinic.org/health/articles/22283- glucagon. Accessed Feb 11, 2026.
⁷ Boyle, C.N., Lutz, T.A., & Le Foll, C. Amylin – Its role in the homeostatic and hedonic control of eating and recent developments of amylin analogs to treat obesity. Molecular Metabolism. 2018;8:203–210. doi:10.1016/j.molmet.2017.11.009.
⁸ Bays, et al. Obesity, adiposity, and dyslipidemia: A consensus statement from the National Lipid Association. Journal of Clinical Lipidology. Volume 7, Issue 4p304-383July-August, 2013.
⁹ Cleveland Clinic. Adipose Tissue (Body Fat). Available from: https://my.clevelandclinic.org/health/body/24052- adipose-tissue-body-fat. Accessed Jan 21, 2026.
¹⁰ Volger, et al. Patients' Preferred Terms for Describing their Excess Weight: Discussing Obesity in Clinical Practice. Obesity (Silver Spring). 2011 Jul 14;20(1):147–150. doi: 10.1038/oby.2011.217.

Your favorite science podcast is back! To kick off season seven of Two Scientists Walk Into a Bar, hosts Maria Wilson and Danielle Mandikian sit down to celebrate a decade of the podcast and look back at our previous season on unmet medical needs. But what’s next? Get ready for a season that’s going bigger than ever before. To mark the 50th anniversary of Genentech, Maria and Danielle will draw a thread from the seminal breakthroughs of the last 50 years in biotechnology to the incredible innovations that will shape the next decade and beyond.

If you would prefer to read a transcript of this episode, please click here.

SUBSCRIBE BELOW TO CATCH EACH EPISODE

If you want to learn more about the groundbreaking science happening in our labs, click here. To learn more about the jobs in our research and early development group, click here.

Transcript of Two Scientists Walk Into A Bar: “Season Seven Teaser” with Maria Wilson and Danielle Mandikian

Danielle: Hi, everyone! Welcome to the teaser episode for Two Scientists Walk into a Bar. I'm your host, Danielle Mandikian.

Maria: And I'm your other host, Maria Wilson.

Danielle: It is so good to see you! It's wild that we don't get together more often. You know, we had such a great season. What were some of the episodes that you really enjoyed?

Maria: It was a great season. I really enjoyed – in all of the episodes – how we talked about unmet patient need and how science is really advancing to meet needs of patients. I think particularly, I enjoyed my conversation with Daniel Lafkas about mucus. I mean, I learned a lot about mucus, which was, you know, things everybody should know. I also really enjoyed the conversation we had about digital twins with Iraj Hosseini. I took some of that actually back to my own work – I really learned a lot from that conversation. And I really enjoyed talking with my friend Manu Chakravarthy about sort of my area of the work that I do in my day job around solutions for patients living with obesity. What about you?

Danielle: Oh, man. It was a – that was a big season. The unmet need theme was amazing. And it also kind of, like, reinvigorated how I was thinking about my work as well. But some of the talks that really stood out to me were the big swings that we're taking. Like Aviv Regev came on, and she talked a lot about how we're utilizing AI to really kind of transform the development and identification of new therapies, and that was absolutely exciting. And you know – because I love cells, I'm a cell girl – we also got to do an interview with Todd McDevitt about cell therapies, and that was so cool. I really liked that episode because he really kind of laid out the landscape of what everyone is actually working on, and it was kind of like a call to arms on really getting together to solve it. Gives me goosebumps! [Laughs] So anyways, it was a lot of fun and I'm really kind of just thinking about what's next.

Maria: Oh that's a really good point! What should we do? Well, it is the 10th year of the podcast, despite being season seven, because we did have that global pandemic, and also it is the 50th anniversary of Genentech and the advent of the whole biotechnology industry. So I think we should go really big.

Danielle: I think so too – go big and really kind of bring in all of the groundbreaking work that we've done in the past and how it's linked to where we are now. I think this is going to be cool.

Maria: I love that. We'll draw a thread from the past through to the next 10 years and beyond.

Danielle: Everyone loves the underdog story in science and seeing how the innovation meets it. So, let's get to it.

Maria: Let's get to it.

Danielle: Yeah. My timer is going off, so I gotta run. But it was really good to see you.

Maria: It was great catching up with you, Danielle.

Danielle: You too. I'll see you soon!

Maria: One more thing before you go, please don't forget to like or subscribe to Two Scientists Walk Into a Bar wherever you get your podcasts. See you soon!

The name Two Scientists Walk Into A Bar is under license and used with permission from the Fleet Science Center

Sara Mostafavi discovered her interest in science through what she describes as a “continuum of influences.” As the oldest of three daughters, she remembers trailing behind her father as he fixed things around their home, absorbing his problem-solving instincts. She loved puzzles, patterns, and understanding how the world worked.

A different kind of influence came from her paternal grandfather, a general practitioner who served as the only doctor in a remote mountain village in Iran for five decades – embodying a life rooted in purpose and service. On her mother’s side, generations of people who worked the land fostered her appreciation for nature, careful observation, patience, and a deep sense of responsibility for the well-being of their community.

She spent her first 16 years in Tehran, where students are channeled early into specific educational tracks – math, science, or the arts – and she was firmly on the math path. That sense of certainty shifted when her family moved to Canada. Suddenly, the range of possible directions felt wide open. For the first time, she had the space to reflect on what kind of impact she wanted to make, opening a field of opportunities that would ultimately shape her path.

Initially drawn to medicine for its human focus, she enrolled at Queen’s University, majoring in life sciences on the pre-med track. But she soon realized she missed the intellectual rigor of mathematics. With no prior exposure to computer science, she enrolled in an introductory course while searching for direction – and it changed everything. “It felt like the kind of puzzle I’d always loved,” she recalled

To pursue computer science, Sara transferred to the University of Toronto, where she graduated with a double major in biology and computer science. She made this shift at a moment when computational biology – blending machine learning with high-throughput biology – was just beginning to take shape. She later pursued a doctoral degree in computer science, developing computational models that infer the function of previously uncharacterized genes.

Sara’s postdoctoral work took her first to Stanford’s computer science department, where she worked on early large-scale human cohort studies using genome-wide association approaches to better understand the genetic basis of disease. As her postdoc mentor, Daphne Koller, took a leave to co-found Coursera, Sara seized the opportunity to spend a year at Harvard Medical School with Christophe Benoist, working on large-scale immunology projects. Immersion in immunology proved pivotal. “It gave me a different way to think – like a biologist,” she recalls. “Looking for anomalies and contradictions in the data, rather than optimizing models around average behavior.”

Over the following decade, as a faculty member at the University of British Columbia and later at the University of Washington, Sara helped advance AI-driven genomics as the field matured from early promise to practical possibility. She was increasingly drawn to problems that required scale – both in the aggregation of diverse biological data and in the precision engineering needed to train large models effectively. In her lab at UW, she built a research program focused on genomic AI: developing predictive models of gene expression and function directly from DNA sequence. The work addressed a foundational question in biology – how variation in the genome gives rise to variation in function and disease – and offered a path toward causal understanding. As data resources and modeling approaches matured, so did her conviction that these tools were ready to move beyond theory and begin delivering real impact.

BRIDGING ACADEMIA AND INDUSTRY

Sara had long admired Genentech not only for its scientific innovation, but also for its rigor, integrity, and commitment to building a culture that supports its people while enabling sustained scientific progress. For much of her career, she viewed AI in biology as a field still in active formation – one that required substantial foundational research before it could consistently add value beyond well-designed statistical approaches. That conviction led her to focus on advancing the underlying science in academia, which she saw as a uniquely open environment for exploration: shaping how biological questions are framed computationally, what data are generated and interrogated, and how models are built to reflect underlying biological mechanisms.

As the field matured, she began to see those efforts converge. “For a long time, I didn’t think AI tools were mature enough to trust for high-stakes biological decision-making,” she recalls. “But as data improved, it became clear that well-designed AI and machine learning methods could uncover mechanistic insight that was simply out of reach for simpler approaches.”

Conversations with Aviv Regev, Head of gRED, and John Marioni, Head of Genentech’s Computational Sciences Center of Excellence (CS COE), reinforced how intentionally Genentech was investing in this convergence – creating the conditions for AI-driven biology to move from exploration to real therapeutic impact.

A year ago, Sara made the move from academia to industry – not as a departure from discovery, but because her own work at the intersection of AI, biology, and biomedicine had reached a point where it could be tested in environments where decisions carry real consequences. “Academia gave me the freedom to explore across disciplines and to sit with difficult questions,” she says. “But I reached a point where I wanted to apply that understanding in settings where insight and decision-making are tightly coupled.”

Today, as Vice President and Global Head of AI Biology and Translation, Sara leads a multidisciplinary team of AI scientists, engineers, statisticians, and computational biologists in Genentech’s CS CoE. Together, they are developing foundation models and predictive tools to understand the consequences of genetic perturbations, model tissue and molecular dynamics, and characterize patient trajectories – work aimed at accelerating research and improving decision-making across the drug-development pipeline.

Looking ahead, Sara views AI not as a replacement for scientific intuition, but as a way to deepen it. “The goal is to bring mechanistic and predictive layers into how we study biology so we can understand complex systems more precisely, ask better questions, design smarter studies, and make more confident decisions faster,” she explains. “I came to Genentech because I believe in what’s possible – I want us to be at the forefront of shaping how AI can make a meaningful difference in patients’ lives.”

LEADING WITH INTENTION

For Sara, the most meaningful moments in her career have come from working with and mentoring others. “Watching someone move from grappling with ambiguity to developing intellectual independence – learning not just the science, but the art of doing science – nothing compares to that,” she says. “It’s about learning how to frame questions, decide what matters, and make progress when there’s no clear roadmap.” At Genentech, she sees that same process unfolding within her teams.

“I’m really proud of how we’re moving from largely independent computational work toward tightly integrated team science,” she reflects. “The problems we’re tackling to improve the discovery of new medicines are inherently complex – they require many forms of expertise. What’s special about Genentech is that it creates the space to do deep science together, in a way that advances each discipline through the collaboration itself.”

That kind of collaboration doesn’t happen overnight. “We all speak different scientific languages, and learning to work through that ambiguity – where our conceptual frameworks, definitions, and intuitions about the core of the same problem don’t map neatly across disciplines – is part of the work itself,” she says.

Sara’s approach to mentorship is shaped by the same philosophy – and by the mentors who guided her early in her career. “When you’re deep in a complex scientific problem, it can feel lonely,” she says. “You’re making decisions constantly, without knowing for sure if they’re right.” She has always valued mentors who are willing to sit with that uncertainty – not to micromanage, but to help clarify thinking and build confidence in one’s own reasoning; mentors who invest in the scientific and personal growth of their mentees, rather than focusing solely on the outcome of any single project.

Her advice to young scientists follows the same approach “Do good work and be persistent,” she says. “Focus deeply, finish what you start, and reflect on what you learn. Every question – even the ones that don’t pan out – teaches you something about how to do science better, and learning how to tell that story is part of our science.”

ENDURANCE AND PERSISTENCE

Outside of work, Sara’s life is full of motion . . . literally. She has an energetic nine-year-old son, whose boundless energy keeps her moving, whether through competitive chess matches, impromptu soccer games or wrestling sessions. And for as long as she can remember, Sara has been a dedicated long-distance runner.

Running, she says, has shaped her scientific mindset – teaching endurance and the importance of staying in forward motion. “Running is like meditation for me,” she explains. “It’s one of the few times my mind goes completely quiet, and I can explore new ideas. The first 10 minutes are always the hardest. But once you get past that barrier, everything becomes easier – and that’s how starting a new research project feels.”

Wherever she travels, she brings her running shoes. Her favorite runs are through cities – moving between buildings, absorbing the energy of people, and finding rhythm in the flow around her. She has also completed multiple marathons, drawn to the discipline and challenge of pushing herself over long distances.

For Sara, running, science, and leadership are guided by the same principles: keep moving forward, stay focused, and trust that persistence – step by step, iteration after iteration – leads to breakthroughs that matter.

From the very beginning, Genentech pioneered a new way of doing science, combining rigorous research, cross-disciplinary collaboration, and the courage to follow the science—wherever it leads—to deliver medicines for patients. Here, bold thinking isn’t encouraged. It’s expected. It’s mandated.

We’ve come this far with the support of South San Francisco leaders, the workforce, and community members who help create a vibrant city where innovation, opportunity, and quality of life go hand in hand. Progress like this takes allies who believe in reimagining what’s possible, and we feel lucky to have them here in South City.

Today, we’re investing in the next 50 years. By thoughtfully transforming our campus, we are reinforcing a long-term commitment to South San Francisco, supporting thousands of local jobs, strengthening the regional economy, and ensuring we remain ready to lead the next revolution in human health for decades to come — not by maintaining the status quo, but by building the next era of biotech leadership.

New state-of-the-art research facilities are being designed to foster work across scientific disciplines, integrate AI technologies, and support the next generation of discovery — while advancing Genentech’s ambitious sustainability goals through resilient infrastructure built to adapt to the impacts of global climate change.

The goal isn’t just better labs. It’s bigger leaps.

These investments will empower Genentech’s scientists and engineers to collaborate more seamlessly, glean new insights for target and drug discovery, and answer fundamental questions about human disease biology. By modernizing our campus within its existing footprint, we are creating a more connected, future-ready research environment — one that is more collaborative, sustainable, and resilient — built for bold thinking, breakthrough science, and the next era of biotechnology innovation. We’re here to build what lasts, not what’s easy.

The gRED Center is a research hub and world-class facility of approximately 1.25 million square feet, designed to create flexible, sustainable, modern labs and collaborative spaces that enable our scientists to pioneer transformative therapies for patients.

In the 1970s, the biggest revolution in biology was molecular biology. When Genentech was founded in 1976, it brought recombinant DNA technology to drug discovery, changing the world. Now, we’re leading another revolution, powered by AI and large-scale data. We’re building the gRED Center to integrate human biology, massively scaled lab methods, and advanced computational science in one place, so breakthroughs move faster and farther.

Genentech draws on the best scientific and technical talent from the Bay Area, leveraging AI and large-scale lab data to pursue and accelerate groundbreaking science. And we know the next era of medicines requires more diverse research at every level, from teams, data, exploration, and experimentation—made possible by interpretable, predictive, and generative tools like AI and machine learning.

The gRED Center represents the first major step in a multi-decade campus evolution. The project reinforces our belief that the future of medicine will be discovered where science, technology, and human ingenuity meet, and that future will be built right here in South San Francisco. Construction on the new facility is estimated to occur between 2027 and 2033.

At the heart of this initiative is our belief that a healthier, more hopeful future should be within reach for everyone, and that means ensuring patients have access to the medicines they need. With a revitalized campus, our scientists, data experts, and engineers will work in world-class facilities that will enable us to think bigger, push beyond the status quo, and reimagine how we address the world’s most urgent health challenges. By accelerating drug discovery and development from early research through clinical trials, we are strengthening our ability to deliver life-changing medicines to patients worldwide.

Our breakthroughs begin with the bold thinking happening today. We are inventing what’s next in science, technology, and patient care, confirming our role as the trailblazers who will shape a healthier, more hopeful future for all.

During construction, it will create more than 1,500 on-site jobs. In total, the project is expected to support approximately 4,000 direct and indirect jobs in South San Francisco and 4,500 jobs across San Mateo County during peak construction years.

Beyond construction, Genentech’s presence delivers sustained economic impact. Every 100 jobs at Genentech’s South San Francisco campus supports an additional 167 jobs across San Mateo County, reflecting the company’s role as a major driver of regional employment.

Our local operations generate significant economic value, supporting $3.1 billion in total labor income for South San Francisco workers and $5 billion in countywide wages that flow through local businesses, services, and neighborhoods. In addition, long-term investment in the campus significantly increases local tax revenue, helping fund essential city services and infrastructure that benefit the entire community. That’s momentum you can feel on Grand Ave, not just in the lab.

Together, these impacts underscore a simple belief that innovation thrives when communities thrive, and Genentech’s continued investment in South San Francisco is designed to support both for generations to come.

Since 2017, Genentech has invested $65 million in local communities through charitable giving and employee volunteerism. As part of our long-term Master Plan, we are also investing more than $250 million in housing, transportation, and community programs, supporting infrastructure and services that help the city thrive.

Over the past five years, Genentech has contributed $10 million to the creation of affordable housing in South San Francisco, and those funds are already making an impact.

One million dollars launched the Hello Housing Accessory Dwelling Unit (ADU) Pilot Program, which has resulted in 28 completed units to date, with 8 more currently in development. The remaining $9 million has been leveraged by affordable housing developers to produce more than 270 additional affordable homes.

By partnering with the City to expand housing options, we are helping ensure that the people who power innovation here, from teachers and first responders to researchers and service workers, can continue to call South San Francisco home.

Our investment in community extends to the next generation as well. Through Futurelab, our STEM education program, Genentech supports 8,000 K–12 students in the South San Francisco Unified School District, helping spark curiosity and open pathways to careers in science and technology. We’re building the next wave of trailblazers, future disruptors with lab coats.

Economic opportunity also means supporting local entrepreneurs. Through Genentech Goes to Town, we have invested more than $3.6 million in over 200 small businesses in South San Francisco, helping strengthen the city’s economic backbone.

Together, these partnerships reflect a long-standing belief that innovation is strongest when rooted in community and that our success is inseparable from South San Francisco's. This is how we are building a future that belongs to everyone.

Together, we can continue to lead the industry for the next 50 years through our deep curiosity, teamwork, perseverance, and the bold, trailblazing spirit that leads to life-changing scientific breakthroughs for all patients.

As we celebrate our 50th anniversary in 2026, we are looking ahead with purpose. The next revolution in human health, and the next 50 years of biotechnology breakthroughs, will be shaped in South San Francisco by the scientists, partners, and community that have defined our success from the beginning. We’re not here to repeat history. We’re here to rewrite it.

We are proud to remain rooted in our hometown, committed to the people who make this city thrive, and focused on building a future where bold ideas, breakthrough science and shared prosperity move forward together for generations to come.

A future powered by disruption, driven by purpose, and built for what’s next.

This article is reprinted with permission from Endpoints News. The original published article can be found here.

Genentech’s shift from one of the largest drug benefit plans to a privately-held entity is expected to save the Roche subsidiary tens of millions of dollars, the company’s CEO said.

Genentech CEO Ashley Magargee exclusively told Endpoints News that the move could save about $70 million through 2028. Genentech is at least the second large pharma to sign onto Rightway, a pharmacy benefit startup founded in 2017.

Magargee declined to specify which pharmacy benefit manager Genentech left, saying only it was one of the three largest — Optum Rx, CVS Caremark or Express Scripts. A Genentech spokesperson said the company signed a three-and-a-half-year contract with Rightway, which began in mid-2025. The switch applies to roughly 25,000 employees and dependents spanning both Genentech and Roche Diagnostics.

The company was reporting a roughly 13% annual increase in its pharmacy benefit costs, Magargee said, and projects $150 million in spend last year. That rise is now expected to flatline this year with Rightway, she added.

A Genentech employee first disclosed the move on a Rightway webinar last August. CVS, in an emailed statement, said that it was not the pharma company’s previous PBM.

Competitive search

Magargee said Genentech had a competitive year-long search for a new PBM, selecting nine companies for further evaluation before landing on Rightway. She acknowledged that even as a pharma company enmeshed in the US healthcare system, it was hard to untangle from one of the largest PBMs.

“It took us a while to actually review our own medicines for our own employees and what they rely on,” Magargee said. “And that’s when it became really elucidating how much the traditional model just wasn’t serving us or our employees.”

Magargee was sold on Rightway’s “modular” benefits model. Employees can now see how much they pay for different prescriptions, and what the coverage would be, depending on the source.

Genentech also constructed an evidence-based formulary with Rightway, versus the standard model based on rebates. That means that employees pay lower copays for medicines with better data. The obesity drugs are an example in which Genentech asked for lower out-of-pocket costs for employees, citing their strong evidence. Roche and Genentech drugs are not higher on their own formulary, though Magargee said she wants employees to access drugs that they work on.

Eli Lilly CEO David Ricks disclosed on a podcast last year that it too was moving to Rightway. Both moves come as drugmakers and the federal government continue to wage a political fight against traditional PBMs and their pricing practices.

Congress has been working on PBM reform for years, but has struggled to pass legislation despite bipartisan support. Magargee said her executive position allowed her to seek out better transparency through the marketplace.

“My personal point of view on it was what can we do as an employer, an ultimate payer, to see whether this model is working well for us or working against some of our goals,” she said.

Editor’s note: This story was updated after a Genentech spokesperson clarified that pharmacy benefits cover 25,000 employees and dependents, not 25,000 employees, and also to include a comment from CVS.

At Roche and Genentech, we are using artificial intelligence (AI) and machine learning (ML) to transform how we discover, develop, and deliver medicines. More than add-ons or a collection of tools, these technologies comprise a dynamic ecosystem that powers, and is embedded into, every aspect of our work. Integrating AI “end-to-end” has helped us transform the entire journey of making medicines – from identifying new targets and turbo-charging drug design, to devising smarter clinical trials, to enabling agile, efficient manufacturing – with each step informing the others.

In building an AI-powered ecosystem, we’ve advanced the development of new molecules more rapidly or impactfully than would ever have been possible before. Our scientists have also created molecules that would have been challenging to design without the addition of AI. The secret isn’t just better technology, it’s how our scientists are partnering with AI to amplify their expertise and solve harder problems faster – and how we are collaborating with other AI leaders – to redefine how R&D is done.

Turbocharging Our Discovery Engine

Every day matters to patients. That’s why we've implemented AI across our early research and development teams to accelerate and improve some of the most difficult and time-consuming steps of R&D: identifying new disease targets and discovering new ways to address those targets.

Traditional approaches to target discovery have been lengthy, labor-intensive, and risky because biology is incredibly complex – on average, it takes more than 10 years to move a candidate from discovery through Phase III trials. Instead of spending months or years poring over massive biological datasets, our scientists can now leverage autonomous agents equipped with multiple AI biology models to help them zero in on the most promising disease targets identified in a fraction of the usual time. These same AI-driven approaches are helping us more efficiently select indications for our therapies, allowing programs to move forward with stronger evidence and greater precision.

AI is also having a dramatic effect on drug discovery. Historically, only one or two novel molecules out of 100 make it through clinical development. Today, all of our antibody programs and ~90% of eligible small molecule programs integrate AI to streamline discovery, make design more predictive, and increase the likelihood of success. For example, in oncology, AI tools recently helped our scientists redesign a molecule to reduce unwanted immunogenicity, speeding its transition into development. In another oncology project, our researchers designed a specialized molecule, called a “degrader,” 25% faster and with a structure that wouldn’t have been achieved without AI. In immunology, we used AI to deliver a backup small molecule with a distinct scaffold for an important disease target in just seven months, versus more than two years for the original lead developed before AI.

The foundation of Genentech's R&D strategy centers on the “lab in the loop,” where data from the lab and clinic feed AI models designed by our researchers to identify trends, make predictions, propose new experiments, and generate new molecular designs. After scientists conduct experiments based on these predictions and designs, they input the results back into the model, thereby improving its performance. It's an iterative cycle of data, computation, experimentation, and discovery.

“The beauty of the ‘lab in the loop’ is that it augments scientists, rather than replacing them,” says Aviv Regev, Head of Genentech Research and Early Development (gRED). “AI allows us to iterate quickly and at scale, freeing our teams to ask bigger questions – and answer them better and faster than ever.”

To operationalize this vision and scale AI across our R&D organization, we launched the Computational Sciences Center of Excellence (CSCoE) in 2025. The CSCoE, which includes computational scientists from both Genentech and Roche, will ensure all future AI projects serve the breadth of our research and early development activities and the sharing of data, models, tools, insights, and best practices across the teams for our entire research and early development portfolio.

Boosting the Efficiency of Clinical Development

After we've identified promising candidates in research, AI helps us make smarter decisions about which molecules to advance into clinical development and how to pursue and execute clinical trials. Historically, up to 80% of clinical trials face delays, and 20-30% of patients typically drop out for various reasons. Success is never certain, and many drugs fail to demonstrate effectiveness in the middle and late phases of development. These delays and setbacks not only increase the cost of R&D but also make patients wait longer for potentially life-changing treatments.

To overcome these barriers, we are using AI to make clinical trials faster, smarter, and more efficient. We’ve established a “clinic in the loop,” which leverages a wide variety of computational approaches to inform clinical trial design and decision making. For example, AI-powered simulations enable us to model different trial scenarios, test trade-offs, and identify the best strategies to reduce patient dropout rates and risks. By optimizing factors like which patients to enroll and how to measure treatment success, AI helps us ensure our studies are not only scientifically robust but also more efficient. We are also using AI to analyze patient data in real time, which helps us select the best locations and strategies for running clinical trials and cuts planning time from weeks to minutes.

Beyond improving operations, AI is enhancing how we monitor and evaluate the effectiveness of investigational treatments during clinical trials. For example, we have developed AI tools to analyze complex medical images, such as scans used to track disease progression. These tools can analyze vast amounts of data that humans cannot and interpret data much faster – reducing the time experts spend reviewing images by over 90% – while providing more detailed and consistent insights. This is especially important in diseases like ulcerative colitis, where AI can assess treatment effectiveness with more precision, helping researchers make faster, data-driven decisions, and in ophthalmology, where AI can analyze retinal scans to detect diseases early and help inform treatment to preserve people’s vision.

In areas where we have deep experience and rich data, like oncology, we are creating a “clinic in the loop” to transform clinical development. First, we train ML models on past trial data to predict outcomes like tumor response and survival. Then we validate those predictions with data from new studies. Once the models are validated, we use them in simulations to inform future clinical trial design and make more confident decisions about how to proceed when trials read out. The result is smarter trials, faster learning, and a clearer path to getting effective and safe medicines to patients.

“Artificial intelligence has become a true force-multiplier across our development engine, including in clinical trials,” says Venkat Sethuraman, Senior Vice President and Global Head of Data Science and Analytics at Roche and Genentech. “AI-driven tools help us do everything from generating endpoints more quickly to monitoring treatment response more precisely in trials. They give us real-time insights and cut months – and potentially years – off the traditional timeline, ensuring we can bring life-saving medicines to the people who need them faster and more reliably.”

Tackling Production at Scale

Novel medicines can only help patients if we can produce them consistently and in sufficient quantities to meet treatment needs. AI is helping us assess “manufacturability” from the earliest stages of drug discovery. By predicting problems before we move a potential medicine into development, AI models enable our scientists to choose molecules – even in early research stages – that can be readily produced at scale, illustrating the importance of end-to-end integration of the technology.

As we advance a medicine through late-stage development to commercial manufacturing, AI is transforming our ability to improve efficiency, quality, and scalability of manufacturing through applications that identify risks, predict problems with equipment, monitor quality, and more. AI is also enabling us to build smarter factories through the use of data-powered simulations called digital twins, autonomous robots, and tools that personalize production and train our work force – ultimately reducing costs and boosting productivity.

“Even before a medicine is approved and commercialized, we use AI to simulate, design, and optimize manufacturing processes that are resilient, compliant, and ready to scale to deliver higher value for our patients,” says Daniele Iacovelli, Head of Data and Digital for Roche’s Global Manufacturing Network. “Our data scientists have developed AI tools that can identify production risks by analyzing historical and real-time data to predict potential issues in product quality, process variability, and regulatory compliance.”

Aligned with our sustainability goals, we’re also integrating AI into manufacturing to improve quality control or product consistency while reducing waste. Through advanced computer vision and ML, we are empowering our experts to detect defects and inconsistencies with enhanced precision and speed. These insights significantly boost quality, minimize waste, and strengthen the trust associated with every produced batch. Furthermore, agentic AI farms and automation are poised to transform the entire lifecycle of our medicines – from development and manufacturing to supply – ultimately enriching the experience for both our patients and our employees.

“In manufacturing, our ambition with AI goes beyond efficiency,” says Iacovelli. “It is about building a future-ready supply network that patients can always rely on: resilient by design, intelligent by default, and able to deliver the medicines people need, exactly when they need them.”

Boosting AI-Driven R&D through Collaboration

No single company can unlock the full potential of AI alone. At Roche and Genentech, our strength lies in collaboration within, across, and outside our organizations.

“AI only delivers real impact when diverse expertise comes together, including biology, chemistry, clinical insight, and cutting-edge computational power,” said John Marioni, Senior Vice President at gRED and Head of the Computational Science Center of Excellence. “That’s why collaborations with AI innovators are essential to accelerate and scale what we can achieve in drug discovery and development.”

Our strategic collaboration with NVIDIA, which began in 2023, accelerated our ability to rapidly predict the structure of molecules, reduce training time for our biological AI models, and streamline image analysis – compressing some of the slowest steps of drug discovery, enhancing our ability to discover molecules, and boosting the efficiency of our “lab in the loop.” For example, our Equifold model, trained in part with the NVIDIA BioNeMo platform, was our first model to predict the dynamic conformational ensembles of antibodies, thereby turning a GPU (Graphics Processing Unit)-hours-long calculation of surface properties into seconds and enabling such calculations to scale to early research.

Through this collaboration, we are also leveraging AI across our global manufacturing network to guarantee smooth and efficient plant operations. For example, we’re constructing an “omniverse” – digital twins that accurately model the flow of material, processes, and personnel – to optimize plant design and operations with unprecedented speed and cost efficiency.

Our partnership expanded in 2026 to launch a large-scale AI factory to further enable our end-to-end AI approach. Roche’s combined on-premise and cloud infrastructure now exceeds 3,500 GPUs, making this the largest announced GPU footprint available to a pharmaceutical company. Working together with NVIDIA, we’re generating more high-quality data, training smarter AI models, and speeding up the progress of new medicines through our pipeline.

“Our ongoing collaboration with Genentech proves that AI can revolutionize R&D,” said Rory Kelleher, Senior Director of Business Development for Life Sciences at NVIDIA. “By unifying accelerated computing, foundation models, and real-world scientific workflows, we are helping to build a new engine for R&D – one that can dramatically expand the scale, speed, and scope of what scientists can discover.”

We are also collaborating with data and AI innovators to accelerate breakthroughs in difficult-to-treat disease states like neurodegenerative disorders. We are partnering with Recursion, which creates maps of human biology, and together we identify novel disease targets and better understand the drivers of disease. Recently, we collaborated on a microglial map built from approximately 46 million images that links 17,000 genes to the unique characteristics of microglial cells – important immune cells involved in neurodegeneration. The microglial map allows scientists to jointly use AI to uncover pathways that traditional methods miss, addressing long-standing bottlenecks in discovering new treatments for neurodegenerative diseases.

Beyond the scientific research itself, Genentech is using AI to reinvent the way lab experiments are conducted through our collaboration with Medra, pioneering the next era of lab research by linking data, AI, and physical lab work. Like a self-driving vehicle, Medra's physical AI platform leverages advances in AI models and general purpose robotics to help our scientists reinvent research through the "lab in the loop." Medra's platform integrates with Genentech’s internal ML infrastructure and its laboratory information management system, providing a unique continuously learning environment to rapidly take predictions, run experiments, and iteratively optimize scientific experiments, giving our scientists more freedom to generate and test new hypotheses.

While ML has made a tremendous impact on process improvement, it doesn’t eliminate some of the most tedious non-clinical aspects of research. Scientists spend a vast amount of time on procedural tasks, like sourcing data and sifting through journals. That’s where the gRED Research Agent and our collaboration with Amazon Web Services comes in. This technology eliminates many of the tedious manual aspects of research by combing through vast data sets and cross-referencing scientific journals. It can potentially help automate more than 43,000 hours of manual biomarker validation across therapeutic areas each year, giving our scientists more time to bring new medicines to patients faster.

Leading the Next Biotech Revolution

Fifty years ago, Genentech pioneered the biotechnology industry, and that same spirit guides our work today. Just as recombinant DNA and monoclonal antibodies once enabled us to reshape what was possible for patients, AI is helping us lead the next revolution in biotechnology. By combining human ingenuity with powerful AI systems across research, development, and manufacturing, we are reimagining R&D from end-to-end and solving problems that were once beyond our reach. For Genentech, the AI revolution in biotechnology promises to help us deliver more – and better – medicines faster. For more patients than ever before, it means the chance to live longer, healthier lives.

In our season six finale, we dive deeper into how artificial intelligence (AI) is shaping the future of drug discovery and scientific research. With remarkable scale and speed, AI models parse through complex datasets and confirm or generate hypotheses, which can help scientists accelerate R&D. In this episode, co-host Danielle Mandikian welcomes Aviv Regev, Head of gRED, and Jure Leskovec, Professor of Computer Science at Stanford University, to talk about foundation models and autonomous agents. Together, they explore the opportunities and challenges of applying AI in drug discovery, including balancing innovation with scientific rigor and the evolving role of scientists. They also discuss how AI is reshaping the future of research — from building more biologically meaningful models to advancing agent-based systems and lab automation.

If you would prefer to read a transcript of this episode, please click here.

SUBSCRIBE BELOW TO CATCH EACH EPISODE

If you want to learn more about the groundbreaking science happening in our labs, click here. To learn more about the jobs in our research and early development group, click here.

Transcript of Two Scientists Walk Into A Bar: “Foundation Models and Agents” with Aviv Regev and Jure Leskovec

Maria: I’m Maria Wilson.

Danielle: And I’m Danielle Mandikian.

Maria: And we are scientists. We. Love. Science.

Danielle: Yeah, we do. So, when we aren’t doing it, the next best thing is to talk about science! And what’s really awesome is we’re surrounded by some of the most brilliant minds in research!

Maria: We are going to step away from the labs today to talk to other scientists about the cool stuff they are thinking about, working on and imagining . . .

Danielle: . . . as well as how some of these discoveries just might lead to new medicines. So, grab your favorite drink, get ready to unlock your science brain and join us for Two Scientists Walk into a Bar…

Maria: The show for scientists, science geeks, and the people who love them!

Danielle: Do you know what an AI agent is, and can you define it?

Employee responses:

Can I define an AI agent? It's like the guy who's like running things in the background, right?

It's often a kind of, like, LLM core, which exists in an environment where it has some amount of memory and access your tools that it can use to execute tasks.

An AI agent? No, not me. [Laughs]

I think it's for help or troubleshooting, or like customer support, maybe?

I think an AI agent helps aggregate all of the information that is out there and summarize it to the question that you're asking so that you, yourself, do not have to go individually search the different sources.

Danielle: Hi everyone, welcome to the show! I'm Danielle Mandikian, and believe it or not, we're already on the final episode of the season. And what we're gonna do is round up our discussions about unmet need with our old friend, artificial intelligence. Last season, we got the opportunity to talk to a lot of scientists about how AI is utilized in their research. And if you haven't checked out those episodes yet, queue it up! But today what we're gonna do is take a deeper dive into foundation models and the emergence of autonomous agents. And what we'll do is we'll talk about the capabilities of these tools, how they're actually accelerating discovery in both academia and industry, and then we'll also talk about what this really means for the future of scientific research. So, to explore this, I'm really excited to welcome back computational biologist and AI expert Aviv Regev. Welcome!

Aviv: Thank you for having me.

Danielle: And making history on the pod, we have our first ever external guest. So I would like to welcome to the bar a computer science professor from Stanford University, Jure Leskovec. Welcome.

Jure: Yeah. Thank you for having me.

Danielle: All right, so can one of y'all explain to me what is a foundation model? And are there different types?

Jure: So it's a large neural network model that has been pretrained on a lot of data, and generally, it's been pretrained in what is called a self-supervised way, right? So an example of this would be large language models are trained over long pieces of text or long pieces of computer code, and they're essentially predicting the next word, or they are predicting the next character or the next token, right? And the exciting thing is as you pretrain these large models this way, they basically are able to kind of learn that domain and as you make these models bigger, they're starting to get capabilities that you did not put directly into the model. So now on the biology side, we also see the emergence of these large, let's say pretrained models. You can think of them at several different, let’s say, scales of resolution. You can think of molecular foundation models – basically you get a sequence of amino acids of the protein, they will give you a representation, a very compact representation of the – of that protein. And then, you know, now on top of that, you can build folding models, you can do a lot of different things – but now you have a representation of that molecule. Then you have foundation models at the single-cell level that are trained now on a lot of single-cell RNA seq data to give in our compact representation of the cell. And then you could also start thinking about foundation models for spatial data or for tissues – so collections of cells.

Danielle: What makes these tools so powerful in our research?

Aviv: So, what makes them powerful in the research is kind of at multiple levels. The first one is, you have to think about how people used to use models. Say you're a biologist. You do an experiment, you collect some data, and you in fact train your model on the data you collected, and you ask it questions that you kind of predefine to yourself. Usually, you would train it actually in the context of that question and within the context of the data that you collected. And that's it. And when you're done with that, you're kind of gonna let it go. It is quite likely no one will use that model again. It was just a representation of the data that you already collected, and you asked it a relatively narrow question. And the question that you predefined is the one based on which you actually build the model. In the foundation model world, you kind of turn everything a different way. So first of all, it's not just the data you collected – it's a lot of data from a lot of places across many, many, many different contexts, many of which you can't even decide for yourself whether they're even relevant or not. But it turns out when you show, during training, when the model gets retrained on this much broader variety of things, it picks up much more essential pieces of understanding and information kind of about the world. So if it is, for example, a model over cells, in the traditional world – I don't know – I am interested in the gut, in a disease like inflammatory bowel disease, or IBD. So, I'm going to take a dataset, I'm going to collect a dataset of cells from patients with IBD, I'm going to do some analysis on them and ask them IBD questions. And you might think that's the best thing I can do because I'm only looking at IBD. But it turns out that you understand better the language of cells as a model if you go to see, first of all, many other things that happen in the gut. It's not just IBD. There's healthy people. There's patients with celiac disease. There is colorectal cancer. It's not just your colon, it could also be your small intestine and your esophagus. It turns out, it's even better if you also let it see cells from the skin and from the lung and from the kidney and from the brain, and maybe also from many, many other organisms throughout evolution. You learn the language of cells better, and now you can give better answers – both about IBD and about plenty of other things. And in fact, we're not limited. It turns out maybe whatever's happening in IBD is also important when you have a lung disease or a skin disorder or colorectal cancer. So it gives you this broader context, which is something we usually appreciate in a scientist. We say, “Wow, this scientist knows so much.” It's because they were exposed to a lot of different things. They draw all these connections – similar thing for the model.

Danielle: Can you help me understand what are some of the limitations or challenges that we face currently with building these models?

Jure: Yes. So, I think there is a lot, right? It is a super fast-moving field, and we are learning a lot as we go. I think right now we do not, for example, understand too well what is the best data to use, how much data we need, how this data needs to be, in many times, collected, sampled, and so on, right? And I would say similarly, like even if you think about large language models, we kind of train them on the entire internet just because it's there, right? And that is already great, but in the post-training phases, we kind of collect special datasets that allow, kind of – that give models special abilities, like writing code, doing mathematical reasoning, and so on. So I think if – now in, let's say, in biology, we follow that trajectory, I think we need to invent what are these post-training phases, and what kind of data we need to collect and what kind of tasks we need post training that will really give these models new abilities. I think another challenge is that, you know, biology is a discovery-based science – so it's not about repeating the same thing over and over again, but it's kind of discovering and generalizing to something new. And once it's been discovered, we don't need to go there anymore, we want to go somewhere else. And I think that's fundamentally harder than, you know, asking ChatGPT when Napoleon was born or something. Like it's something we already know, and it's just kind of recalling. So I think this kind of notion of generalizing to somewhere else where you were not trained, right? You were trained on one type of data, but really the questions we want to ask is about the area where there is no data, because they don't know that yet, right? So, understanding that generalizability and developing the technology that allows us to generalize well and kind of predict well the things that haven't yet happened or that we just don't know yet, I think is the key benefit of these models, and giving them that ability, I think, is also a big open question. How much of that we can do? Where can we do it? How can we do it? Under what conditions, and so on.

Aviv: And on the data side, we don't even know whether the kind of style of data we have, its capacity, what has already been done – is that a lot or a little? We don't really know. We don't know if we had a hundred times or a thousand times more, would that be transformative and actually change what can happen with models, or in fact it's going to be totally the same? It's like that's what's – what there is. I think there's a belief that more will be better, but we don't really know because we kind of have to learn the rules of the game as we play it. But I think there's two things that are especially important in science. One is that in science, we also have, for certain areas, some fundamental understanding of the laws or constraints of the phenomenon. So for example, if you work with molecules, there's principles from physics and from chemistry that we actually understand. Is it actually beneficial to introduce these constraints and kind of let the model only operate in a world that's, like, physically and chemically legitimate? Or is it better to just show it a lot of physics and chemistry, it will learn it all on its own and learn the right thing? That's an open question. We don't know the answer to it. However, it's kind of an empirical question. You can try and see how things perform. Similarly, when you're in biology – so molecules really matter to us a lot because we try to make great molecules. We want the model actually to generate these new general things, these new things that the model didn't see because those – that could be the next medicine, for example. Similarly, when you try to understand biology – like find what might be the gene that causes disease – then you need causality. It's not enough that there is a description of what we already know. You want the model somehow to be able to tell you like what if – to answer “what if” questions. If I did this, would that be beneficial? Or, what might be the reason or the cause that I'm observing this phenomenon? And for that, again, you need the model either to be interpretable in those terms, or actually to be built originally within the framework where this is baked in. And historically again, the first generations of these models, this was not part of them. And we don't know whether this would actually make them better, or it's something that we should just let them, kind of, see a lot and, somehow, they figure it all out and we just come after the fact and get the explanation to it. But these are super important for biology, for chemistry, for the impact of AI on science and on medicine.

Wellington: Hi Danielle!

Danielle: Hey, Wellington. Hey, Karen, welcome back! Did you miss us?

Karen: I really missed you guys. Although, I was at a bar in Italy, as you mentioned, and enjoying my Aperol Spritz when the last episode dropped.

Wellington: So jealous. [Laughs]

Karen: [Laughs] But it's great to be back. Jure and Aviv were mentioning that it's a really big challenge to not know the right amount or type of data to train these models. So what gives us the confidence to trust the answers that we get from these models?

Danielle: You know, I don't know that we have a set answer for that. I think it's a very iterative process and that's kind of why people are always talking about needing to keep scientists in the loop to kind of keep evaluating the progress and then keep refining as the model needs to evolve.

Wellington: So to follow up on Karen's question, when you're talking about iterations, is this the scientific way of saying you're double checking your work?

Danielle: Yeah, triple check. So something that we always do as scientists is this idea of cross-validation, and that's where we take expected results or truths that we know that we can count on and cross-check it against the predictions that we would get out of a given system.

Danielle: One of the things that y'all had brought up that I think is also very interesting is, you know, when you were talking about these foundational models and the prediction, it's kind of like this kind of figuring out what's the next part in a sequence, like linearly or spatially or whatever. But then, you also talked about building the models so that they could be thinking about the hypothesis or the reason. So is there another component about reasoning or logic that is the next phase of these types of foundational models?

Jure: So I think yes. Like, I agree with that. But basically, I think right now, maybe where we started was like we have these special biological entities, we have molecules, we have cells and so on, so we need to build representations for them. And these representations need to be enriched with all the knowledge of biology we have. And now, as Aviv was nicely saying, now that, if you have this enriched presentation, you actually need less data to build an accurate model on top of it than to directly build the model on the raw data, right? So if you say, okay, I have some gut cells, it's actually beneficial to push them through the foundation model, get a kind of biologically meaningful, enriched representation and then analyze that, than to analyze the raw data. Because the raw data is so high dimensional, so noisy, it's very hard to see stuff there to build models, tests on that. So you want to push that through the foundation model, it – the noise gets taken away, biology gets added in, you have now this small, compact representation on which you can run analysis much more reliably, right? So that, I think, is the first step in this foundation model.

Aviv: Again, for a listener who might be a scientist but not a computer scientist, this is actually what the scientist's brain does a lot. You show them something, it's kind of messy, it's kind of unclear, but they can see inside it already. And that is because, in fact, their brain was trained with a lot of prior scientific knowledge, so they know to pay attention to these things, and they know that these things are less material, and they know, let’s say, these genes and their expression, they mean something already, or these chemical groups in a small molecule mean something already. That's actually part of what our brains do. They do a lot of other things, but they do that. And they do that on the foundation of a lot of knowledge that they were already exposed to. So the foundation model is that. It's kind of this exposure to a lot of preexisting data, information, knowledge – whatever you would want to call it – so that when you bring another piece of information through, you don't start from scratch. It's not like, you know, if a kid who's like, I don't know, 10 or 13 years old and never yet studied something, and you show them something, they won't see the same thing as if you showed it to the same individual 10 years later after they just got through a lot of information. That's the –

Jure: To a trained brain, in a sense.

Aviv: That's the trained brain. So that is actually what Jure is describing. He's describing it for the computer scientists. But we actually have a lived experience of that as human beings. That's actually what happens to us. It's not the only thing. Our brains can do a lot of other super cool stuff that foundation models cannot yet do, but this piece – what the foundation model has an edge over, say, any of us as humans – is that they just digested a huge amount of stuff. And any human in their lifetime can only digest a certain amount of stuff. So that's a distinction. Back to you because that was just level one. [Laughs]

Jure: Exactly. Thank you for translating that. I think it's great. So now that we kind of understand these entities, now we want to do – we want to go kind of a level up and do this kind of human-level reasoning, right? And that's what the large language models who have read the entire internet, have read all the biological papers, are now able to do, right? They're able to reason over these representations, over these entities that we can encode with foundation models at a high level of obstruction that gets us then more towards this kind of, can we have now a virtual scientist? Can we talk to this, let's call it agent or whatever, that is then responding back to me, informed by my data, as well as informed by all the biological papers that can be found on the internet.

Danielle: I've heard the term agents before, but I don't really understand it. What is it?

Jure: Yeah. So what we think of an agent it's an autonomous system that is able to take actions, let’s say in the real world, get feedback from those actions, reason about those, and informed by that experience, take new actions. So if you want to have an agent, an agent has to be able to take actions. That's the key, right? Like software by itself that just outputs something on the screen, that's not enough. It has to do more. It has to be able to make connections, execute something, move something, push something. That's what an agent is. It's something that should – that is autonomous and it's able to make an action autonomously, get some result out of that actioning, and then plan what the next action should be.

Danielle: Aviv, could you give us an example of what utilizing an agent would actually mean with these foundation models, especially in the context of drug discovery or drug development?

Aviv: So I'll try. So let's start with the computational side, and then I'll turn to the actual physical, three-dimensional world, because agents can operate in that world in principle as well.

Danielle: Okay.

Aviv: So today, for example, or until recently, if you had, say, a dataset came in, okay, you wanted to analyze it and draw some scientific understanding from it. Say it came from – we started with inflammatory bowel disease – I'll just stick with it. You got some dataset from inflammatory bowel disease patients, and you wanted to analyze it and understand something about what is going on in them. Maybe they were treated with a medicine. Maybe you're just trying to figure out targets, new genes that you would want to go after, and so on. And so, you would usually have a computational biologist, a person who’s trained in application of computational methods to biological things, and they would apply a whole world of tools to this – those tools were previously developed and they analyze data – and you took the dataset and you ran this thing and you looked at the result and you're like, eh, not exactly. I don't exactly understand what's going on there. I'm going to try a little differently like this, and then I might move to this tool, and then I might move to that tool, and I will do this and I will do that. Now I'm gonna go read some papers because I need to understand what I'm seeing there. Okay, I read the papers, I have some thoughts, I'm gonna do – that's what the human does, right?

Danielle: Mm-hmm.

Aviv: Same idea. Replace the word human with autonomous agent, and it's very similar. You take an agent, which is, of course, appropriately coded and trained, et cetera. And it has the following abilities. First of all, it's equipped with a large language model itself, so you can talk to it, like we talk to, you know, you talk to ChatGPT. So you tell it: “Here's a dataset in IBD, I want you to analyze it for me.” That might be actually all you will tell it. Maybe you will say more, maybe that's all you will say. Like you hand it, you know, to your friend who sits next to you. And what can they do? They can actually run – the agent can run all of these analysis tools, the same tools, and can go to the literature because it has an LLM of the literature, so it can kind of read and think – “think,” in quotes, something along those lines. And you often would do something similar with the agent to what you do when you train a person. You actually show them some loose templates of how you work. And now what would the agent be able to do? They're able to receive your typed in instruction or spoken instruction. Then they have access to all these tools. They have the ability to launch the tool. They have the ability to receive back the results from the tool. And they can look at it – “look,” in quotes – and reason over it in their way of reasoning. And they have memory – they remember that before that, they did this step, and before that, they did that step, all of those steps, right? So they have context and memory. And they just go and do these things. They run this and they look at that and they do this and that. You can stop them, and they can also stop on their own and say, "Here's what I have for you. Like it?" And you can say, "Yeah, I do," or "I don't” – you know? – in the same way that you interact with other scientists essentially, and they become a sidekick for you. So that is in the virtual world because all this agent did is run code and look at results. But in fact, you can couple that thing to instruments. If the instruments are automatable, you can also, in principle, allow your agent to go and spin off an experiment. If it can be automated, you can spin off an experiment. And the agent can go and say, okay, now maybe it's in chemistry, now not in IBD in biology, I want the synthesis to happen. I have some automated system for synthesis, and I want the molecules back, and I'm gonna ship them off to that robot to measure their properties and then tell me what the proper – tell me what the measurements were, and I'll decide on the next step. That can also be an agent.

Danielle: What about you, Jure? Any examples?

Jure: So the question is kind of – it’s a simple question to define. So you want someone to help you design experiments, right? So you have a cell, maybe you have some tumor cells, and you're like saying, okay, what gene should I perturb for these cells to die or to express a given phenotype, right? So you have 20,000 things you can choose from and select – perturbing each one separately in the lab is too expensive, takes too long and all that. So what you would like to do is maybe you would like to select 64 or 32 or, you know, some –

Danielle: Something manageable. [Laughs]

Jure: – round – some round number for a computer scientist. And then you would try those, see what happens, and then you'd like to suggest the next batch and try it out and keep going like that, right? So, we built an agent, where basically you come in and say, okay, I have this type of cells, this is the phenotype I would like to express, suggest to me what genes should I perturb. Okay? And then you can go in the lab, around them, you go back and you say, okay, I did these genes, these are the results, and it will suggest the next set of experiments. Okay? So when we were developing this, it just didn't work. It was so terrible. I cannot tell you how terrible it was.

Danielle: I'm convinced that's good luck though. [Laughs]

Jure: Right? That's right. So then my student Yusuf had a brilliant idea – he went back to the LLM and he said: “Look, these are the genes that are the ground truth to perturb the cell type to get the phenotype you want. So what question should I ask you to give me this set of genes?" And you know what the thing responded? The thing said: "This is like you asking me to generate a random set of English words." And we were like, what is going on?

Danielle: How is it that funny? [Laughs]

Jure: So, we go into the original dataset, and what we found out is that the gene dictionary was off by one. So basically, all the genes were random. That's why it didn't work!

Danielle: Oh my gosh.

Jure: And now, show me a human who can look at a set of 54 letter number codes and say, oh yeah, these are random, they have nothing in common. Versus, oh yeah, these all belong to this pathway, I know it all. There is no such human. But the LLM has that knowledge. It said, look, sorry, this is a totally random set. We went back, we checked the data – it was off by one, so everything was garbled. We corrected it, and it started working beautifully.

Danielle: I love it.

Jure: Right, so it's just like, wow. Right? So that's one concrete example. And the beautiful thing is you can also ask it “Why? What's the reasoning behind it?” – point you to papers. You can really kind of work with it as a companion to design these experiments. And it works amazingly well. Better than current technologies that kind of run – learn out of the data because the current – this kind of bayesian optimization methods they are called – they learn from scratch. But the LLMs have read the “entire internet”-worth of biology papers, so they come in with all this prior knowledge that helps early on, and then, actually, they learn from the experimental results to suggest other later stages. So they both have a tremendous amount of background knowledge, as well as are able to take in kind of the evidence to suggest the next step.

Aviv: This requires, actually, two kinds of changes from the side that does the experiments. First of all, there is a human change. The humans have to change the way they think about – right? – now they have an agent. How do you use it? How do you sync with it? How do you engage with it? And it's really important because all of us, in some way or another, grew up in the world of calculator access. Meaning we actually believe that when you ask a computational device a question, you get a correct answer, right? You put in the thing in the calculator – you don't actually double check that it made a correct calculation. But the AI world is not like that. In one way, it is way richer, right? It read the literature, it did this, it did that. But it's not necessarily right. It's a lot more similar to a human. When you work with another person and you ask a really difficult calculation question, you're not gonna be certain that they're giving you a right answer. But they are going to be way more interesting than a calculator. So that's one big shift. The other is, experiments have to change.

Danielle: What do you mean by that?

Aviv: So when you do an experiment in a world like that, you might not want to design your experiments anymore the same way. For example, you might actually do a whole slew of experiments that are not for any particular question just because you think they make the foundation model better. Or, instead of, when you do screens – which are the kinds of experiments that Jure actually described their screens – historically, the goal was to do as big of a screen as possible up front. And why is that? Because – first of all, it's painful to do them. You know that better than I do.

Danielle: [Laughs] Yeah.

Aviv: But also, because then you will have all the results. You will read through all of them, and you will figure things out. But it's a fallacy because you actually can never get all the results in biology. You could always have given a different condition to the cell. And, yeah, maybe you don't count every gene at once – you can do all of them, but you can't do all their combinations. These are things that you can never do all of them. But if I can work with an agent, maybe I can do the best set of experiments overall if I let it kind of see things gradually – I'll do a batch, let it think for a moment. Come back, do another batch. Come back, do another batch. And I have to stay out of the way for a while, because if I intervene too much, I might actually kind of break the magic.

Danielle: Yeah.

Aviv: So that's one other thing that is a big shift in how you might think about experimental biology. And once you start thinking like that, you might want different styles of experiments being done and then invent some new lab techniques to make it possible, or the right kind of instrument to make it possible. It changes the calculation for the wet lab – not just for the people who do computational work.

Karen: Danielle, Jure’s example of the really cheeky agent response about that garbled data was hilarious, and it stood out as a perfect example of how agents can solve problems that a human might miss. So where else do you think using agents might be helpful in the lab?

Danielle: I mean, we're not there yet, but I would absolutely love it if I could get a cheeky flag down whenever I mishandle samples, you know? Or, I like – you know, I'm still like a wet bench scientist, meaning, like, I'm in the trenches digging through cells, pulling stuff out from a biochemistry perspective, and, like, analyzing it across like 40 million different machines. It'd be phenomenal if I can have someone step in and be like, uh-uh girl, you mislabeled that. That's not even what makes sense to analyze at this point. That would save me a lot of tears. [Laughs]

Danielle: Alright – representing wet lab scientists everywhere – whenever we hear this amazing capabilities of utilizing foundation models and agents to really speed up the prediction of what's a good target? How are we going to bind it efficiently? You know, sometimes that leaves some of us standing around being like, shoot, what are we supposed to do? Because historically, that's been us in the trenches, really kind of going through at this really slow pace of trial and error trying to find this out. So what does this look like in the future in terms of how foundation models and agents can actually interplay with scientists like me?

Aviv: So let me start by reducing the fear, and I'll start with a real example of an agent that was trained to do something known as spatial analysis for a spatial transcriptomics, okay? That is a pretty involved set of these steps and so on. There's a set of templates – people have done it before, but every time you need to do it, you do it from scratch. So we built an agent like that, but we also competed it against humans – I believe about 10 different people. Basically, we gave them the same tasks, and you can – on tasks that you can measure how accurate you are, how performant you are, and so on. And it turns out that, yes, an agent can be better than some of the humans. But it's not better than the best human. But what is especially important is that when you give the agent to a human, the human exceeds themselves and the agent, which means this is actually a tool that empowers you as a lab scientist or as a computational biologist or as a computational chemist, et cetera, et cetera. Because it lets you, first of all, take a lot of drudgery out of your work. Yeah, the first time you run spatial transcriptomics analysis, it's super interesting. The hundredth time, it's like, oh my god I have to do this again. Right?

Danielle: Yeah. [Laughs]

Aviv: But there's no replacement. Well, it turns out, you can figure it out when it's interesting, figuring out how to do something like that. But then, you have something that takes a lot of the drudgery out, and once in a while would show you something you wouldn't think about while you have a chance to think your own thoughts and kind of work with them together. Which, in fact, we love working with other humans for that reason. We exchange ideas. They do it like this, we do things like that, something better comes out. That is actually the world with agents. It's not a world without the scientists, it's a world where the scientist becomes a lot more empowered. It's like everyone all of a sudden has a whole bunch of interns working with them –

Danielle: Right.

Aviv: – rather than working alone. And it's amazing to have interns around. And of course, every intern is much more powered, because every intern now has an army of high school students working with them. You can kind of continue that analogy –

Danielle: Yeah, yeah, yeah.

Aviv: – as far as you want to go.

Danielle: But that's a great way to look at it, because then you have this kind of broader picture and you give the intern the guidance – like start this, let me know if you find anything interesting.

Aviv: Yeah.

Jure: And what is very interesting is you can always come back and ask it why? So you can kind of talk to it so it can explain why did it do what it done, why does it think this is a good next step and so on. So I think it is a really good kind of collaborator and brainstormer as well as, you know, kind of the robot that can automate certain mundane tasks that we as humans perhaps don't want to be doing. And I think that's very, as Aviv said, it's like very liberating because each one of us is expert in one area, but we are kind of not experts in other areas. And now if we have this kind of generalist agent, they can kind of supplement you or complement you in some sense in these areas where you are not the ultimate expert, right? And that kind of speeds up the scientific progress by a lot.

Danielle: How is this impacting science, research and development, especially from the perspective of industry versus academia?

Jure: I can quickly give maybe two vignettes here. So, we've developed this kind of agent that helps lab scientists accomplish their tasks. And the beautiful thing about this agent is that it has access to, let's say, 100 different, you know, ontologies and datasets and about 150 different kind of advanced bioinformatics tools that it can use to get its tasks done. And just two quick examples. So, one success was a wet lab at Stanford. They were trying to design a new cloning protocol, and amazingly, they kind of put in all the data into the agent and said, okay, this is what I want to do, give me that plasmid map [Laughs]. You know, I'm zero expert on this, right? But in the end, they got this cloning protocol of about, you know, at what temperature to put the cells into what kind of oven, for how long. I was just amazed. You know, I know nothing about it – so, what that cloning protocol was? And they tried it out, and it worked. And they said, look, it would take a super senior postdoc several days, weeks to come up with this. Now it took 10 minutes. Another example was a multiomics lab at Stanford School of Medicine who had continuous glucose monitoring data, physical activity data, as well as food intake data. And they just dropped, I don’t know, 500 different Excel spreadsheet files of activity of these different subjects and said, generate some good hypotheses out of this. And, you know, the thing would analyze the data and come up with a hypothesis how, I know, some temperature arises after consuming food or something like that, right? And the scientist in the lab was like, look, it would take me probably two weeks of work, two hours a day, three hours a day – and this thing did it in half an hour. So we see a lot of, I would say, enablement at this day-to-day level. Where we are going with this is actually to close the loop and say, can these systems automatically verify hypothesis and come up with experiments that try to falsify a hypothesis? And if hypothesis is not falsified, then can they come up with the next experiment and the next experiment, and this way, really try to discover knowledge. And if you think about this, this would be super amazing because we have these huge amounts of data that we as humans just don't have time to look through. And of course, we can kind of correlate everything with everything, but then we get millions of hypotheses and we don't know which ones are promising or not. So now if you could use some kind of agent that would prescreen, pre-pretest some of the hypotheses and come back and say, look, here are the ones that look interesting that I couldn't falsify, so likely they are true – you humans may want to look at this. That would be a huge enabler in, you know, the way we do science and how quickly the science progresses.

Aviv: It's always useful to remember how much of that process is filled with failure, right? In our industry, 90 percent of programs fail, either preclinically or clinically. That's horrible. Imagine we would make it better. Imagine we just moved from 10 percent success to 20 percent success or to 30 percent. That means, like, twice as many or three times as many benefits to humanity. That's a big deal.

Danielle: I always feel like I tell people that science is kind of like a bad gambling addiction. [Laughs] You know what I mean?

Aviv: [Laughs] Totally. But we want to change that, right?

Danielle: Yeah.

Aviv: The most critical step you actually have is deciding which targets you're gonna go after. Because if you go after the wrong target, you're gonna spend years doing many things at huge human effort, human ingenuity, time – at some point, you might go into patients. And if in the end that does not succeed, it's all because you chose the wrong target. So that's the first choice you make, but it's one that takes a long time sometimes to figure out. The other thing is that because each step is so complex – because there's so much reading and thinking and figuring out and so on – you often do it like in a sequential, in streaming mode. You try one, and if it doesn't work out, then you go and you try another one. And if that doesn't work out, you go and you try another one. What these tools allow us to do, what AI allows us to do more generally and agents allow us to do, and especially when they're coupled in a lab and we iterate – what we call the lab in a loop – when you do that, then you can actually look at a ton of things in parallel. Again, just like what Jure described. And yet, many, many of them are gonna fail. But you failed them fast, you failed them right away. You're not gonna waste all your time and effort on them. You're gonna focus your effort on where it's actually productive. That could be transformative for what we try to do for patients. Because without it, you're kind of doomed to always toiling one by one, one by one, through something that's enormous – that's, like, it feels unbounded, and it's pretty close to that. Then yeah, I mean, once in a while, we definitely are very, very lucky. But we'd rather be a lot more lucky than that. We don’t want it to feel like a gambling addiction. [Laughs]

Danielle: [Laughs] Yeah. Let's go 10 years into the future. You know, I've been hanging out with my favorite agent. We got through some cohorts. You know, it told me that my choices were probably not the best, sent me in the right direction. I went to bed, woke up to some results. My AI watch was like, hey, I've got the molecule, it's safe, whatever. Where do we go from there? Like how does this keep growing?

Aviv: I'll start by actually trying to imagine the physical world. I think especially for lab scientists, as you are one, labs haven't really changed in many, many decades. Actually, probably 100 years. The instruments changed dramatically in what they can do, but the lab environment didn't really. I'm not sure that would not – would remain true 10 years into the future. For example, it is very reasonable to assume there will be self-driving labs, like you have a self-driving car. Because, yes, in fact, that thing came to your watch. But it came from a lab that ran 24-7, and humans were busy with the long tail of creativity and the next really unimaginable thing. But the thing churned and churned and churned, and it constantly is connected together. And the lab might also have humanoid robots that have, you know, dexterity with things that look like human hands, so they can actually operate side by side. So that physical world is actually gonna get, you know, very embedded and embodied together with AI, and that has not yet happened as much to lab environments, but it's part of where the future is likely headed. Who knows? You know, it's hard to – not to, you know, prophesy, was given to the fools.

Danielle: No, I believe you though. I believe you.

Aviv: But still, that, I think, is an exciting side of it. The second part of it is I do think we will – in some ways, the scientific method will remain the scientific method, and in some ways, we might actually change the way that we think in science. And to me, that is the most exciting thing. And if you're a new person starting – if you're now an undergrad, or in your PhD or in your postdoc – that is what's gonna make your life, like, so fun. Because, you know, you still have a super flexible brain, and it's gonna shift a lot. So, the way people will do their science and think about their science might be quite different than the way that it is today. I’m not sure exactly in what way, but I think we will become much more adept at working at the true scale of science, which in chemistry – especially for drug discovery and in biology – is an enormous scale. Whereas before, we were limited, we were doomed to only work on small parts of it one at a time. Now all of a sudden, we would be able to carry because we would have our own brain and that brain next to us to carry it together.

Jure: Just to say, right, like basically, intelligence and reasoning are already kind of getting commoditized, right? So as time progresses, even more so. So I think that means that kind of the nature of work we do as humans is going to fundamentally change. Where I think that we are ready for the evolution – for this to really come to fruition – is the robotics, that is the next thing, right? Right now, kind of reasoning in the virtual in silica – that we can do quite well. But the interaction with the physical world, executing things in the physical world, is where still there is a huge gap. And maybe those gaps are actually bigger than we all think, right? But if you do things in the lab, you need to physically perform the experiment, right? So that, I think, is the next big innovation that needs to be done. Right now we basically have these agents, companions, they can reason, they are intelligent, they read the papers, they are super good as, kind of, collaborators – but I think it's the robotics piece, it's the next big thing.

Danielle: It's so exciting. And like, we had this episode about – Kim Homan came on and she was talking about different kinds of in vitro assays and using bioprinting and all this stuff, and like, none of this was even on the table like 10 years ago. I mean, it's so – I remember when liquid handlers came onto the scene, and we raced in our lab about who could do ELISAs fast – and it was like, you know, the old story with like, the steam-powered boat and stuff, and we were like living it in real time – it'd be like ah, you know? But it's all changed, now there's like whole rooms that automate certain kinds of assays. And so I do think that that's so realistic and plausible.

Aviv: Just like the models they're in, so does this amazing neural network known as the human brain – which is, by the way, amazingly energy efficient – learns. And as our circumstances change, the labs change, the computer science changes, AI comes about, and so on – our brains are gonna go with it, and they're gonna come up with all sorts of awesome stuff. And what's most exciting about thinking 10 years into the future is appreciating the fact there's gonna be stuff there that we're not imagining right now.

Danielle: This has been such a great conversation – so much fun. Thank you both for joining.

Jure: Yeah, thank you.

Aviv: Thank you.

Karen: Friends, I think this is one of my favorite episodes yet. We're going out with a bang for season six. A key take home that really stood out to me from this episode was that AI foundation models can't be – they can't do it all – and you really need the scientist working with these tools to be able to accelerate drug discovery. So my question to you, Danielle, is what is the most exciting thing you believe will come from this relationship that doesn't currently exist today?

Danielle: I really think the thing that's resonated the most with me is figuring out or imagining a situation where the scientists are freed up to do bigger questions that they're interested in. Like, I'm super fascinated by how biologics actually move throughout diseased tissue and it kind of sucks for me now because I can't dive that deep into it. But if we were able to free up certain time and resources to take care of finding a target, figuring out how to drug it, then it gives us a lot more space to dig deeper into some of these other questions and feed models that can even move it faster.

Wellington: Yeah, Danielle, I thought this was an incredible episode, and as you know, I've been asking, sort of, questions about data. Like, what are we going to do with all this data? And agents in this episode kind of answered that question for me. There's two things that are exciting to me. One is, as a scientist, it must be kind of exciting to think about things moving faster and getting answers to questions faster. And then I think about the translation to patients – how any improvement on how we target and discover will potentially translate to real gains from the patient side.

Danielle: Yeah, and I mean, absolutely. The gains on the patient side is what this is all about. But even just from like a scientific curiosity standpoint, the other thing that's really cool is getting novel insights that we just haven't come up with yet. And having a path forward and a new way to attack something is just priceless. So what makes this episode the perfect end to a really exciting season is that you can really start to see the light at the end of the tunnel, almost like a point in time where all of the work that's come together to get these types of models and agents out there is really gonna start to produce real results. And that's what we've all been waiting for.

Danielle: And that’s our show! Thanks so much for listening. If you haven't already, rate our podcast, wherever you listen – it’ll help new people find us. And make sure to subscribe. If you have any questions about the show, you can contact us at podcast@gene.com. And now for me it’s back to stalking cells!

The name Two Scientists Walk Into A Bar is under license and used with permission from the Fleet Science Center