Skip to content

Instantly share code, notes, and snippets.

Show Gist options
  • Select an option

  • Save avelinoapheris/6e6fee1fd13cc62ef180de5a118b5388 to your computer and use it in GitHub Desktop.

Select an option

Save avelinoapheris/6e6fee1fd13cc62ef180de5a118b5388 to your computer and use it in GitHub Desktop.
OpenADMET Blind Challenge Structure Track

PXR Challenge: Method Summary (ver228)

Track: Structure Prediction · LDDT-PLI: 0.5725 · BiSyRMSD: 3.41 · Coverage: 1.00

Method

We generate a large ensemble of candidate poses with ESMFold2 and select one pose per fragment by a physics-based interaction energy rather than by model confidence.

1. Ensemble generation. ESMFold2, 85 independent random seeds with 5 diffusion samples each (about 425 poses per fragment). All fragments use the same PXR protein (293 aa) and a single precomputed MSA. No templates, no fragment-specific finetuning.

2. Pose scoring. Each candidate pose is protonated (protein hydrogens with PDBFixer at pH 7.4, ligand hydrogens with RDKit from the CIF bond orders) and the hydrogens are relaxed against frozen heavy atoms, so the predicted heavy-atom pose is preserved. A single-point evaluation with the OrbMol-v2 machine-learned interatomic potential then gives an interaction energy, E(complex) − E(protein) − E(ligand), and per-atom forces.

3. Selection. For each fragment we discard poses whose maximum atomic force exceeds 20 eV/Å (broken geometry or clashes), then keep the surviving pose with the lowest interaction energy. If every pose fails the force gate, the least-strained pose is kept, so every fragment is covered.

Model confidence (ipTM) is recorded but not used to filter or rank.

Models and software

  • ESMFold2 (structure prediction)
  • OrbMol-v2 (interaction-energy scoring)
  • PDBFixer + OpenMM/AMBER14 (protein hydrogens), RDKit + MMFF (ligand hydrogens)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment