Trustworthy Protein–Ligand Binding Affinity Prediction via Reliability-Aware Multi-Engine Fusion

Computational drug discovery relies on docking engines to predict binding affinity, yet these engines often provide conflicting results without indicating which prediction is trustworthy. This paper addresses the lack of interpretable confidence measures in multi-engine docking by introducing RELIABLE-BA, a framework designed to quantify uncertainty and identify the most reliable predictions for any given protein-ligand complex.

RELIABLE-BA treats each docking engine as an 'evidential expert' that outputs a Normal-Inverse-Gamma (NIG) distribution, which naturally decomposes uncertainty into aleatoric (data noise) and epistemic (model knowledge) components. The model introduces a novel reliability network that learns to weight each engine's contribution based on the specific molecular context of the protein-ligand pair. These reliability-scaled experts are then aggregated using a Mixture of Normal-Inverse-Gamma (MoNIG) framework, which produces a final, calibrated prediction that accounts for both individual engine uncertainty and disagreement between engines.

Experiments on the PDBBind and BDB2020+ benchmarks demonstrate that RELIABLE-BA achieves competitive point prediction accuracy while significantly improving uncertainty calibration compared to standard consensus or ensemble methods. By using the model's uncertainty estimates to filter for high-confidence predictions, the authors observed a reduction in mean absolute error (MAE) of up to 25.7%. Further validation on the SARS-CoV-2 Mpro dataset and 5HT2A receptor confirms that the framework maintains its performance on clinically relevant targets, providing a principled way to prioritize compounds for experimental validation.