dock-postprocess is an open-source structure-based drug design toolkit for standardizing docking results and carrying them through restrained OpenMM minimization, pose QC, protein-ligand interaction analysis, ligand conformational strain analysis, and integrated design prioritization.

The workflow is designed for practical docking postprocessing where preserving the docked binding geometry is important.

It supports conventional protein-ligand complexes as well as multichain receptors such as molecular-glue and PROTAC ternary complexes.

All code is on GitHub https://github.com/averysader/dock-postprocess

Mac mini has either M5 Pro chip or M6 chip

M5 Pro

• 15‑core CPU with 5 super cores and 10 performance cores, 16‑core GPU with Neural Accelerators, 16‑core Neural Engine, 307GB/s memory bandwidth
  Configurable up to 64GB unified memory
• 18‑core CPU with 6 super cores and 12 performance cores, 20‑core GPU with Neural Accelerators, 16‑core Neural Engine, 307GB/s memory bandwidth
  Configurable up to 64GB unified memory

M6

• 12‑core CPU with 2 super cores, 4 performance cores and 6 efficiency cores, 12‑core GPU with Neural Accelerators, Dual 16‑core Neural Engine, up to 170GB/s memory bandwidth
  Configurable up to 32GB unified memory

Mac Studio

Mac Studio has either M5 max or M5 Ultra chips

M5 max

• 18‑core CPU with 6 super cores and 12 performance cores, 32‑core GPU with Neural Accelerators, 16‑core Neural Engine, 460GB/s memory bandwidth
  36GB unified memory
• 18‑core CPU with 6 super cores and 12 performance cores, 40‑core GPU with Neural Accelerators, 16‑core Neural Engine, 614GB/s memory bandwidth
  Configurable up to 128GB unified memory

M5 Ultra

• 30‑core CPU with 10 super cores and 20 performance cores, 64‑core GPU with Neural Accelerators, 32‑core Neural Engine, 1.2TB/s memory bandwidth
  Configurable up to 256GB unified memory
• 36‑core CPU with 12 super cores and 24 performance cores, 80‑core GPU with Neural Accelerators, 32‑core Neural Engine, 1.2TB/s memory bandwidth
  Configurable up to 512GB unified memory

All come with a variety of storage options.

I usually go to Arstechnica for a review

The devices’ popularity for production inference took off after macOS 26.2 shipped last December. According to Apple’s release notes, 26.2 enabled “low-latency communication between Thunderbolt 5 hosts for use cases including distributed AI inference using MLX.” Thunderbolt 5 is a very fast wired data connection, and MLX is an open source array framework designed to help machine learning workflows take full advantage of the M-series chips’ unified memory.

This is baseline model a single task CheMeleon model trained on pEC50 data curated from ChEMBL for PXR. It is a no split model, meaning it has been trained with no data allocated to validation and test sets and with just a training set of 1.0.

Instructions for installing and running the pxr-chemeleon-v1 model, which predicts PXR bioactivity for a set of compounds, using OpenADMET’s Anvil framework (openadmet-models).

Prerequisites

Before you start, make sure the following are installed on your computer:

• git — used to download the model and code
• git lfs — a git extension needed to download the (large) model files
• conda or mamba — used to create an isolated Python environment

You can check whether you already have these by running:

git --version
git lfs --version
conda --version   # or: mamba --version

If any of these commands fail, install the missing tool before continuing.

Git is automatically installed with Xcode Command Line Tools. If Git is not installed, entering git in the Terminal will trigger a Mac dialog asking if you would like to Install Xcode Command Line Tools. Click “Install” to begin the download and installation process.

Use Homebrew to install git lfs,

brew install git-lfs

Instructions for installing conda https://docs.conda.io/projects/conda/en/latest/user-guide/install/macos.html

Installation

Open a terminal and work through the following steps in order, running each
command one at a time and waiting for it to finish before moving to the next.

Enable Git LFS, this only needs to be done once.

git lfs install

Download the model

I created a folder called PXR and moved into it

cd /Users/chrisswain/Projects/PXR

Then cloned the model into the folder

git clone https://github.com/OpenADMET/openadmet-models

cd openadmet-models/

Then created the conda environment

conda env create -f devtools/conda-envs/openadmet-models.yaml

Then activated the environment

conda activate openadmet-models
pip install -e .

I then ran a calculation using full paths to files provided.

openadmet-models % openadmet predict \
    --input-path /Users/chrisswain/Projects/PXR/pxr-chemeleon-v1/compounds_for_inference.csv \
    --input-col OPENADMET_CANONICAL_SMILES \
    --model-dir  /Users/chrisswain/Projects/PXR/pxr-chemeleon-v1/anvil_training/ \
    --output-csv  /Users/chrisswain/Projects/PXR/pxr-chemeleon-v1/predictions.csv \
    --accelerator cpu
    
INFO     Finished prediction                                                                             
INFO     Predictions saved to /Users/chrisswain/Projects/PXR/pxr-chemeleon-v1/predictions.csv  

To run your own prediction you need to edit the path to the input file

–input-path /Users/chrisswain/Projects/PXR/pxr-chemeleon-v1/compounds_for_inference.csv

and the name of the column containing the SMILES

–input-col OPENADMET_CANONICAL_SMILES

If you want to use the gpu edit

–accelerator cpu

If you are also interested in docking ligands into PXR this PYMOL session is a great start. https://macinchem.org/2026/08/07/openadmet-pxr-challenge-pymol-session-file/

There are now three pipelines

1. Molecular Clustering — Morgan FP → Tanimoto similarity → Butina clustering
2. 3D Conformer Generation — DG (4D) → ETK (3D) → MMFF94 optimization
3. PM6_D semi-empirical SCF — full d-orbital NDDO (S/P/Cl/Br/I) with PM6-D3H4 corrections

Full details are on GitHub https://github.com/guillaume-osmo/mlxmolkit

What’s new

Semi-empirical SCF on Apple Silicon — 7 methods (RM1, AM1, PM3, PM6, PM6_SP, PM6_D, AM1*) plus PM6-D3H4 post-SCF corrections — bit-exact to PYSEQM for PM6_D, with no PYSEQM/PyTorch runtime dependency. Every entry point is covered by tests/test_{methods_api,pm6_d_native,pm6_d3h4,pyseqm_port,rm1_scf}.py (83 tests total).

Requires macOS with Apple Silicon (M1/M2/M3/M4). RDKit is needed for molecular input:

conda install -c conda-forge rdkit
pip install mlxmolkit-rdkit

Cambridge Cheminformatics Network Meeting
Hybrid Mode – at the CCDC on Union Road, Cambridge and Online (via Zoom)
Direct Zoom registration: https://cam-ac-uk.zoom.us/meeting/register/NBe00sdHRTOIC1pWXqPcyg
For further details please visit: https://www.c-inf.net
Special Edition – ‘AI in Chemistry Warm-Up Event’, just the evening before!

Programme

Symmetries & Efficient Representations for Generative Structural Biology Models
Alvaro Prat, Oxford University

Automated Extraction of Experimental Datasets from the Literature: Are Agents the Answer?
Ben Honore, Imperial College London

ANNalog: Generation of MedChem-Similar Molecules
Dave Dai, Queen Mary University of London

CheMeleon is Robust but Un-improvable: Where to Next?
Jackson Burns, Massachusetts Institute of Technology (MIT)

RSC CICAG Quantum Computing in Chemistry: Current Capabilities and the Road to Utility
Thursday 19 November 2026, Burlington House, London.

Confirmed Speakers
Matthias Degroote – Boehringer Ingelheim – Quantum computing for drug design

Maria-Andrea Filip – University of Cambridge

Chiara Leadbeater – University of Cambridge – First-Quantised Electron Scattering Simulations

Sabrina Maniscalco – Algorithmiq

Ivan Rungger – NPL

Submit abstracts here

Registration here

Applications are particularly welcomed from early and mid-career UK researchers.

You must be based at a UK research organisation eligible for UK Research and Innovation (UKRI) funding.

The full economic cost (FEC) of your project can be up to £602,500. UKRI will fund 80% of the FEC (£482,000).

Projects must start on 1 February 2027 and last for up to 24 months.

https://www.ukri.org/opportunity/next-generation-artificial-intelligence-ai-explainable-ai/?utm_medium=email&utm_source=govdelivery

The latest of these challenges is the OpenADMET PXR Blind Challenge. The challenge is in two parts an Activity Dataset of over 11,000 molecules that can be used to produce machine learning models. In addition there is a Structure dataset containing 184 new X-ray structures with small molecules bound together with In addition, 68 structures from the PDB.

The pregnane X receptor (hPXR) is the major determinant of CYP3A gene regulation by drugs and other xenobiotics. In addition, PXR mediates induction of P450s 2B6, 2C8/9, and 3A4, as well as the drug transporters MDR1, organic anion transporting polypeptide C, bile salt export protein, and multidrug resistance-associated protein 2.

The binding site is large and hydrophobic with several important hydrogen bonding interactions. There may be multiple binding conformations. Similar to CYP3A pharmacophore, many (but not all) CYP3A substrates/inhibitors are also CYP3A inducers.  As shown below many structural classes can be accommodated by the receptor

Whilst 184 X-ray structures of the receptor with ligand bound are available, navigating between the structures is a significant challenge. Fortunately, Manish Sud has done much of the heavy lifting for you and has generated a PyMOL session file named OpenADMET-PXR-Crystal-Structures-Aligned.pse.zip (258M). It provides a ligand centric hierarchical views to visualize the data. The B factor visualization is also available in the PyMOL session file.

The uncompressed file is fairly large (1.25 GB) so will take a little while to load but once loaded it provides a means to easily compare structures. This is a fantastic resource.

There are issues with this system, the same file type e.g. .sdf can be opened and viewed by a number of applications, and a given application can open multiple different file types, and different apps might also provide QuickLook extensions. Whilst you can turn off a particular QuickLook extension in System preferences, this eliminates QuickLook views for every file type registered for that extension. I’ve not found a way to assign a particular file type to a particular QuickLook extension.

There is also the issue of file types, macOS uses a combination of Uniform Type Identifier (UTI) and file extensions. Unfortunately this system means that you can have a single file type with multiple entries in the database of UTI.

The day after the review was posted I got a comment on BlueSky

Recently found QuickLookDoctor for setting QuickLook extension preferences. markedapp.com/ql/doctor/

I quickly downloaded it and it is an invaluable tool for setting which Quick Look extension is used for which file type.

https://raymol.io

A Metal rendering engine replaces OpenGL, adding real-time shadow mapping, ambient occlusion, and order-independent transparency at interactive frame rates. On supported Apple Silicon, hardware ray tracing is available for higher-quality lighting. The direct-download and Homebrew builds also include the built-in MCP server, so Claude can drive Raymol. The Mac App Store build leaves that out because App Store sandboxing doesn’t permit it — the engine, rendering, and every other feature are identical.

https://prolif.readthedocs.io/en/stable

It can be installed using Conda

conda install -c conda-forge prolif