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 using experimentally resolved co-crystal structures, but real use cases often rely on computed inputs (docked or predicted complexes). To quantify this benchmark-to-deployment mismatch, we compared the CASF-2016 performance of five reproducible PLBAP pipelines across crystal structures, GNINA docking into holo/apo/AlphaFold3-predicted receptors, and AlphaFold3 co-folding. Critically, access to an experimentally resolved apo receptor conformer provided only marginal benefit over AlphaFold3-predicted receptor structures. AlphaFold3 co-folding was competitive with, and for some models significantly better than, rigid-receptor docking into the apo conformer (Holm p ≤ 0.006). Multipose averaging showed source-dependent effects and did not recover near-crystal performance for any perturbed input. Interaction-level profiling revealed distinct shifts in protein-ligand interaction distributions that help explain the observed decrease in predictive performance. These findings provide practical guidance for two audiences: (1) model developers should consider reporting performance on benchmarks extending beyond crystal structures to more appropriately reflect deployment settings, and (2) end users should expect performance differences depending on structure generation method and pose selection.