Dynamics-aware-FeatureDock
Structure-based machine learning has ridden one wave: AlphaFold made static structure cheap and universal, and models like FeatureDock, Boltz-2, and DiffDock-L now predict binding from a single frozen pose.
Core trick
Adds Dyna-1 predicted per-residue conformational-exchange values as an extra pocket-feature channel in FeatureDock's ligand-occupancy model.
Tags
Description
Structure-based machine learning has ridden one wave: AlphaFold made static structure cheap and universal, and models like FeatureDock, Boltz-2, and DiffDock-L now predict binding from a single frozen pose. But proteins are not statues. Much of what determines whether a drug binds — and how tightly — lives in µs–ms conformational motion that no static structure captures. We believe dynamics is the next dimension, and the next leap in AI for protein design will come from models that can see it.
We investigated whether ground-state dynamics information improves ML models of small-molecule–protein binding. Using Claude Science, we integrated Dyna-1, which predicts per-residue conformational exchange from any protein structure with no experiment required, into FeatureDock, adding a dynamics channel to its pocket features. As hypothesized, pose prediction improved most in flexible regions, exactly where a single static structure fails.
This matters because it is a concrete step toward making dynamics a reusable, experiment-free ground-truth signal for machine learning. The same prior could sharpen co-folding and docking engines like Boltz-2 and DiffDock-L, and give protein-design pipelines a way to engineer motion on purpose, turning dynamics from an after-the-fact surprise into an optimizable design objective across countless AI-protein models.
README excerpt
# Dyna-1-FeatureDock **Augmenting FeatureDock with µs–ms protein dynamics for pocket-aware docking.** Built for the *Built with Claude: Life Sciences* hackathon (July 7–13, 2026) --- ## What this is Take [FeatureDock](https://doi.org/10.1038/s44386-025-00005-6) — a transformer that predicts where a ligand sits in a protein pocket — and make it dynamics-aware: teach it not just the static shape of the pocket but how the pocket moves, so it docks more reliably in flexible pockets where a single crystal structure misleads it. FeatureDock, as published, reads one static structure and predicts a ligand-occupancy probability envelope over grid points in the pocket. That works when the pocket is rigid; it struggles when the pocket breathes on the µs–ms timescale — exactly the motion a crystal structure cannot show. We append a per-residue µs–ms dynamics channel — predicted by [Dyna-1](https://www.biorxiv.org/content/10.1101/2025.03.19.642801v1) — to the FEATURE tensor at every grid point: **6×80 → 6×81**. Retraining on HPC was done with the help and written with Claude Science! --- ## Background: why dynamics, and why not existing dynamics-aware models Existing dynamics-aware binding models such as Boltz-2 are trained on NMR ensembles and short MD trajectories. Neither is ground-truth dynamics: <img src="docs/assets/abl_src_02_not_ground_truth.png" width="640"/> - **NMR ensembles** are structural snapshots consistent with experimental restraints, not validated exchange k