Two Critical Bugs Found in GPU Fleet Manager

Bug 1: Ligand Placement at Protein CoM (not binding site)

The holo MD script places ligands at the protein center of mass instead of using the DiffDock-predicted binding pose coordinates. For ROCK2/Fasudil this worked by coincidence (ATP pocket near CoM in PDB 2H9V). For LIMK2/bbb5, the ATP site is 33Å from CoM — the ligand started in the wrong place.

Fix: Read the docked pose SDF file and use those 3D coordinates as the ligand starting position. Fall back to CoM only if no docked pose exists.

Bug 2: HETATM Count Verification (includes waters)

The post-MD verification checks grep -c HETATM final_20ns.pdb which counts ~90,000 water molecule atoms. A ligand can dissociate completely and the check still passes.

Fix: After MD completion:

1. Extract non-water HETATM coordinates (ligand only)
2. Compute distance to nearest protein CA atom
3. Gate: closest CA must be < 10Å
4. Compute ligand RMSD vs initial position (gate: < 5Å relative to protein)

Impact

• All previous holo MDs with ligand placement at CoM need re-evaluation
• Fasudil/ROCK2 passed because CoM ≈ binding site (lucky)
• bbb5/LIMK2 failed because CoM ≠ binding site (exposed the bug)
• Fix prevents false positives in all future MD validations