TrustAffinity: accurate, reliable and scalable out-of-distribution protein-ligand binding affinity prediction using trustworthy deep learning

Polypharmacology has emerged as a new paradigm to discover novel therapeutics for unmet medical needs. Accurate, reliable and scalable predictions of protein-ligand binding affinity across multiple proteins are essential for polypharmacology. Machine learning is a promising tool for multi-target binding affinity predictions, often formulated as a multi-modal regression problem. Despite considerable efforts, three challenges remain: out-of-distribution (OOD) generalizations for compounds with new chemical scaffolds, uncertainty quantification of OOD predictions, and scalability to billions of compounds, which structure-based methods fail to achieve. To address aforementioned challenges, we propose a new model-agnostic anomaly detection-based uncertainty quantification method, embedding Mahalanobis Outlier Scoring and Anomaly Identification via Clustering (eMOSAIC). eMOSAIC uniquely quantifies distribution similarities or differences between the multi-modal representation of known cases and that of a new unseen one. We apply eMOSAIC to a multi-modal deep neural network model for multi-target ligand binding affinity predictions, leveraging a pre-trained strucrture-informed large protein language model. We extensively validate eMOSAIC in OOD settings, showing that it significantly outperforms state-of-the-art sequence-based deep learning and structure-based protein-ligand docking (PLD) methods by a large margin as well as existing uncertainty quantification methods. This finding highlights eMOSAICs potential for real-world polypharmacology and other applications.