Equidock-Diff
A protein-ligand docking system built with PyTorch and geometric learning, achieving 1.20 A mean aligned RMSD on a 20-complex evaluation panel.
Problem
Protein-ligand docking pipelines are sensitive to reproducibility and evaluation consistency, which makes model comparison difficult across runs.
Approach
I built a docking workflow around geometric deep learning with reproducible training, sampling, and evaluation pipelines. The system supports checkpoint reuse and run tracking to enable controlled iteration.
Outcome
- Achieved 1.20 A mean aligned RMSD across a 20-complex evaluation panel.
- Established repeatable experiment workflows to compare modeling choices with less variance.
- Improved confidence in model performance reporting through standardized evaluation runs.