Mitigating Molecular Aggregation in Drug Discovery With Predictive Insights From Explainable AI.

Area of Science:

• Artificial Intelligence
• Drug Discovery
• Computational Chemistry

Background:

• Small colloidally aggregating molecules (SCAMs) cause false positives in high-throughput screening (HTS).
• Identifying and mitigating SCAMs is crucial for efficient drug discovery.
• Current methods struggle with chemically counter-intuitive aggregation properties.

Purpose of the Study:

• To apply a novel explainable AI (xAI) model, MEGAN, for SCAM identification.
• To leverage xAI for understanding and designing molecules with altered aggregation properties.
• To reduce false positives in drug discovery screening pipelines.

Main Methods:

• Development and application of a multi-channel graph attention network (MEGAN).
• Utilizing explainable AI (xAI) for molecular property classification.
• Generating molecular counterfactuals based on xAI insights.
• Experimental validation of MEGAN predictions and counterfactuals.

Main Results:

• MEGAN successfully identified SCAMs, addressing a key challenge in drug discovery.
• xAI insights enabled the design of alternative compounds with modified aggregation behavior.
• Experimental validation confirmed the model's predictive accuracy and the utility of counterfactuals.
• Demonstrated the ability to alter aggregation properties via minor structural modifications.

Conclusions:

• MEGAN, an xAI model, effectively identifies SCAMs and reduces false positives in HTS.
• xAI provides valuable insights for designing improved drug candidates.
• Integrating this method accelerates the discovery of viable lead molecules.