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

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning in drug discovery

Background:

• Structure-based deep learning models for protein-ligand binding affinity prediction (PLBAP) are typically benchmarked using crystal structures.
• Real-world applications often utilize computed protein-ligand complexes (docked or predicted).

Purpose of the Study:

• To quantify the performance gap between PLBAP models benchmarked on crystal structures versus those using computed inputs.
• To evaluate the impact of different structure generation methods on PLBAP performance.

Main Methods:

• Compared five reproducible PLBAP pipelines using CASF-2016.
• Evaluated performance on crystal structures, GNINA docking (holo/apo/AlphaFold3-predicted receptors), and AlphaFold3 co-folding.
• Analyzed multipose averaging and interaction-level profiling.

Main Results:

• PLBAP performance decreased significantly with computed inputs compared to crystal structures.
• AlphaFold3 co-folding was competitive with, and sometimes superior to, rigid-receptor docking.
• Experimentally resolved apo receptor conformers offered minimal advantage over AlphaFold3-predicted structures.
• Multipse averaging did not restore crystal-level performance for perturbed inputs.
• Interaction-level profiling revealed shifts explaining performance degradation.

Conclusions:

• A significant benchmark-to-deployment mismatch exists for PLBAP models.
• Model developers should report performance on diverse structural inputs beyond crystal structures.
• End users must consider the impact of structure generation methods and pose selection on prediction accuracy.