Difficulty Replicating Antibody Optimization Results

[benyaminjami]

I've been trying to replicate the antibody optimization results detailed in the paper, strictly adhering to the provided instructions and ensuring the activation of OpenMM relaxation. However, my attempts to reproduce the expected outcomes have not been successful. According to the paper, the average ddG calculated by GNN predictor should be approximately -7, but my results consistently show an average ddG of around -4, even when num_residue_changes is set to 0.

[kxz18]

Sorry for the delay! I checked the codes and found that I accidentally changed the logic of optimization when I was refactoring the codes for release. I've updated the repo with the corrected codes, and now it should work by directly following the instructions in README:

GPU=0 bash scripts/test/optimize_test.sh checkpoints/cdrh3_opt.ckpt checkpoints/cdrh3_ddg_predictor.ckpt ./all_data/SKEMPI/test.json 0 50

Sorry for my carelessness again!