Rationalizing Predictions of Isoform-Selective Phosphoinositide 3-Kinase Inhibitors Using MolAnchor Analysis

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Artificial Intelligence in Drug Discovery

Background:

• Explaining machine learning model predictions is crucial for their adoption in drug discovery.
• Phosphoinositide 3-kinase (PI3K) inhibitors are important therapeutic agents, but achieving isoform selectivity is challenging.

Purpose of the Study:

• To develop and validate a novel methodology, MolAnchor, for generating chemically intuitive explanations of machine learning predictions for PI3K inhibitor isoform selectivity.
• To identify specific structural fragments responsible for predicting inhibitor selectivity.

Main Methods:

• Generation of a test system for predicting PI3K inhibitor isoform selectivity.
• Systematic analysis of correct predictions using the MolAnchor methodology, based on explainable artificial intelligence "anchors" concept.
• Comparison of MolAnchor explanations with feature importance values from other methods.

Main Results:

• MolAnchor successfully identified well-defined structural fragments, often a single substructure, responsible for predicting isoform selectivity in most cases.
• Distinct recurrent substructures were found for inhibitors with different isoform selectivities.
• MolAnchor explanations demonstrated superior interpretability compared to feature importance values.
• Two recurrent substructures were directly linked to PI3K isoform selectivity, suggesting a causal relationship.

Conclusions:

• The MolAnchor methodology provides chemically intuitive and interpretable explanations for machine learning predictions in drug discovery.
• Identifying specific substructures linked to selectivity can guide the design of more selective PI3K inhibitors.
• This approach enhances the integration of predictive modeling in drug discovery projects by elucidating the basis of compound activity and selectivity.