Multimodal out-of-distribution individual uncertainty quantification enhances binding affinity prediction for...

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Polypharmacology, using single drugs to target multiple proteins, offers potential for unmet medical needs.
• Accurate, reliable, and scalable prediction of protein-ligand binding affinity is essential for polypharmacology.
• Current machine learning methods face challenges in predicting binding affinity for novel compounds, quantifying uncertainty, and scaling to large compound libraries.

Purpose of the Study:

• To address the limitations of existing methods in multitarget binding affinity prediction.
• To introduce a model-agnostic uncertainty quantification method for improved prediction robustness.
• To enable scalable and accurate predictions for polypharmacology applications.

Main Methods:

• Developed embedding Mahalanobis Outlier Scoring and Anomaly Identification via Clustering (eMOSAIC), a model-agnostic anomaly detection-based uncertainty quantification method.
• eMOSAIC quantifies individual prediction uncertainty by measuring the divergence between known and unseen data representations.
• Integrated eMOSAIC with a multimodal deep neural network and a structure-informed protein language model for multitarget binding affinity prediction.

Main Results:

• eMOSAIC demonstrated superior performance in out-of-distribution settings compared to state-of-the-art methods.
• The method effectively quantifies prediction uncertainty on a compound-by-compound basis.
• The integrated model showed significant improvements in predicting protein-ligand binding affinity.

Conclusions:

• eMOSAIC overcomes key challenges in multitarget binding affinity prediction, including generalization to novel compounds and uncertainty quantification.
• The proposed method offers a scalable solution for predicting binding affinities across large compound libraries.
• eMOSAIC holds significant potential for advancing polypharmacology and other drug discovery applications requiring robust predictions.