Explaining Accurate Predictions of Multitarget Compounds with Machine Learning Models Derived for Individual Targets

Area of Science:

• Drug discovery and medicinal chemistry
• Computational chemistry and cheminformatics
• Machine learning in pharmacology

Background:

• Multitarget compounds (MT-CPDs) are vital for polypharmacology, enabling simultaneous interaction with multiple biological targets.
• Historically, MT-CPDs were discovered serendipitously, limiting their systematic development.
• Computational methods, including machine learning (ML), are increasingly used for designing and identifying MT-CPDs.

Purpose of the Study:

• To investigate the effectiveness of ML models in predicting MT-CPDs.
• To explore the unexpected finding that models developed for single-target compounds (ST-CPDs) could accurately predict MT-CPDs.
• To elucidate the underlying chemical rationale for this predictive capability using explainable ML.

Main Methods:

• Development and evaluation of machine learning models to distinguish MT-CPDs from ST-CPDs.
• Application of explainable ML techniques to analyze model predictions.
• Identification of structural feature subsets driving accurate predictions for both MT-CPDs and ST-CPDs.

Main Results:

• Accurate prediction of MT-CPDs was achieved using ML models originally derived for ST-CPDs.
• Explainable ML analysis revealed that specific structural features of MT-CPDs were key determinants for ST-CPD model predictions.
• A general relationship between feature subsets for MT-CPDs and ST-CPDs was uncovered.

Conclusions:

• ML models for ST-CPDs can effectively predict MT-CPDs, simplifying computational drug discovery workflows.
• Understanding feature subset relationships provides a chemically intuitive basis for MT-CPD prediction.
• This finding offers a more routine approach to designing and identifying MT-CPDs for polypharmacology.