Our latest post
We set out to fine-tune OpenFold3 on PXR, a notoriously promiscuous target. Along the way we found that success does not come down to how much data you have, but on whether that data consistently shows the model what it needs to see.
Read more →
We set out to fine-tune OpenFold3 on PXR, a notoriously promiscuous target. Along the way we found that success does not come down to how much data you have, but on whether that data consistently shows the model what it needs to see.
The Heme-coming last dance begins: Interim leaderboard release & announcing a new Structure Track
OpenADMET is tackling the persistent data bottlenecks in predictive drug discovery. With major new backing, we are generating prospective ground-truth and structural data at scale. Together, we are turning ADMET from an unpredictable gamble into a rational design problem.
After the PXR blind challenge, we found that feeding a CheMeleon embedding and primary screen predictions into a tabular foundation model drove a top entry's accuracy. The combination cut our MAE from 0.50 to 0.43 and is now a first-class pipeline in openadmet-models.
When we launched the OpenADMET initiative, we set out with an ambitious roadmap: to slash assay costs, generate large-scale prospective datasets, and establish a
Small molecules remain one of the most practical tools for reaching patients in underserved regions. They can be formulated for ambient storage, manufactured at scale
Ninety-six semi-pure compounds span six 12-analog series, ranked by potency and yield-corrected. Five co-crystallized PXR leads map SAR into the pocket via scaffold-aligned conformers. Inspect complexes, toggle ligands, probe polar contacts, and track alignment quality alongside 2D structures.
Announcing the open-source release of our production-ready blind challenge infrastructure template, enabling anyone to spin up a molecular prediction competition in an afternoon.
Moving into the A and D of OpenADMET
Knowing where in chemical space your models will give reliable predictions is essential to getting actionable insights. But how do you figure out where your model is applicable?
Exploring the limits of ligand-based ML models in predicting PXR induction, based on key learnings from our recent blind challenge.
Octant details their robust, reproducible, and scalable assays to map the inhibition of the most pharmacologically relevant cyotochrome P450 isoforms for >10k molecules across chemical space.
We are excited to announce that OpenADMET has received new grant funding from Radial, the life sciences division of the Astera Institute.
Medicinal chemistry groups frequently rely on mental lists of structural alerts to predict CYP inhibition, yet these heuristics are rarely tested against real data. What happens when we evaluate these alerts using 7,800 experimental measurements?
We attempted to improve the CheMeleon foundation model. We were ultimately unsuccessful, but we learned valuable lessons about foundational training along the way.
We extend our previous active learning analysis to choose not just which compound to test, but which assay to run. Utilizing both primary screens and full dose-response measurements cuts the cost of finding actives by more than half, without costing the model any accuracy.
Our quarterly newsletter detailing our progress, goals and priorities.
Announcing the results of the PXR Blind Challenge — and the public release of the largest high-quality PXR induction dataset ever made available, with over 11,000 compounds.
How do you train robust deep learning models when most of the high-quality data is proprietary? Discover how to enrich multitask ADMET predictions using a public proxy for proprietary learnings.
Part 1 of a series on cofolding methods for ADMET targets: using structure prediction to model protein–ligand complexes for key Avoidome anti-targets like PXR and CYP3A4.
We're halfway through the PXR Blind Challenge: unblinding the first data, freezing the leaderboard, and stepping into Phase 2 — plus a look at the response so far and what's next.
With Octant, we use direct-to-biology high-throughput chemistry and standard-free quantification to explore PXR chemical space far beyond commercial libraries, part of OpenADMET's effort to map the Avoidome.
Synthesizing and assaying compounds is slow and costly. We explore whether active learning can use what a model already knows to guide which pEC50 data to acquire next.
Benchmarks are everywhere in ML — but when is a leaderboard gap real? Jon Swain on the statistics of comparing blind challenge entries and separating genuine performance from noise.
Improving 66 Pregnane X Receptor structures from the Protein Data Bank using modern refinement protocols to enhance ligand model quality.
A step-by-step guide to setting up a consumer AMD GPU workstation with ROCm to run the full openadmet-models stack, including dual-boot Linux setup.
Launch details for the PXR induction challenge, including the dataset, rules, and practical guidance for participants.
Lessons and reflections from the OpenADMET-ExpansionRx blind challenge — evaluating whether zero-shot ADMET predictions can be trusted in real drug discovery settings.
An introduction to the next blind challenge and why blinded datasets remain important for realistic benchmarking.
A look at OpenADMET's clearance modeling work and the broader challenge of generalizing across chemical space.
A behind-the-scenes look at the infrastructure and strategy required to generate useful ADMET datasets at scale.
Authors: Maria Castellanos, Hugo MacDermott-Opeskin, Jon Ainsley, Pat Walters DOI: 10.5281/zenodo.21784567 The ExpansionRx challenge, launched on October 27, 2025, closed two
Authors: Maria Castellanos, Hugo MacDermott-Opeskin DOI: 10.5281/zenodo.21784497 Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) properties can make or break the preclinical
OpenADMET's first public model release — predictive models for key ADMET properties, built on open data and designed for the drug discovery community.
Intermediate leaderboard release Authors: Maria Castellanos, Hugo MacDermott-Opeskin DOI: 10.5281/zenodo.21784394. Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) properties can make or
How do you know what you don't know? A deep dive into uncertainty quantification for ADMET predictions and why it matters for drug discovery decision-making.
Authors: Hugo MacDermott-Opeskin, Maria Castellanos DOI: 10.5281/zenodo.21784247 Additional details on the coming challenge Absorption, Distribution, Metabolism, Excretion and Toxicity–ADMET–properties
Authors: Hugo MacDermott-Opeskin, Maria Castellanos DOI:10.5281/zenodo.21784074 Small molecules continue to be the bricks and mortar of drug discovery globally, accounting
Author: Pat Walters, Chief Scientist, OpenADMET Originally published on Practical Cheminformatics. Starting today, September 15, 2025, I will assume a new role as Chief Scientist
Authors: James Fraser, Sriram Kosuri, John D. Chodera, Pat Walters DOI: 10.5281/zenodo.21783808 Small molecule therapies have always been and continue to be
Get the latest on blind challenges, models, and data delivered to your inbox.
