Robustness of Protein-Ligand Binding Affinity Prediction Models to Docked and Predicted Structures

Structure-based deep learning models for protein-ligand binding affinity prediction (PLBAP) are commonly benchmarked on experimentally resolved co-crystal structures, but real use-cases often rely on computationally generated inputs. This repository contains the code and aggregated results for a systematic study of how PLBAP performance changes when five reproducible pipelines are evaluated on:

Experimental co-crystal structures (Crystal)

GNINA rigid-receptor docking into holo co-crystal receptors (GNINA-Crystal)

GNINA rigid-receptor docking into apo receptors (GNINA-Apo)

GNINA rigid-receptor docking into AlphaFold3-predicted receptors (GNINA-AF3)

AlphaFold3 co-folding (AlphaFold3)

The five PLBAP models evaluated are Dynaformer, EGNA, EHIGN-PLA, GIGN, and OnionNet-2, benchmarked on the CASF-2016 dataset. Analyses address three questions: (1) how much does PLBAP performance degrade across structure sources, (2) does averaging predictions over multiple generated poses recover near-crystal performance, and (3) do protein-ligand interaction distributions shift in ways that explain performance differences.

Preprocessed structure files for each structure generation source are attached here in the prepped_structures.tar.gz archive.  This contains sub-archives for each structure generation method and pose (e.g., gnina-apo_best0001.tar.gz contains structure files of the best GNINA-Apo pose of each complex). The gnina-apo_best0001 archive, for example, contains one directory per complex (e.g., 1a30, 3ao4, etc). Each complex directory contains a protein file (e.g., 1a30_protein.pdb) and three ligand files (1a3_ligand.pdb, 1a30_ligand.sdf, and 1a30_ligand.mol2). These were used as inputs to the five afformentioned PLBAP models.

PLIP XML reports used in downstream analyses are attached in the plip_reports.tar.gz archive. This contains sub-archives with similar structure to the preprocessing archives mentioned above. However, the innermost contents will be raw XML files output by PLIP for each pose.

Associated code for structure processing, plip analyses, and more is provided at https://github.com/WoldringLabMSU/PLBAP_Robustness.