Bridging predictive reliability and explainability: a multi-representation deep learning framework for chemical space...

Area of Science:

• Computational Chemistry
• Immunology
• Machine Learning

Background:

• Virtual screening (VS) requires integrating chemical and biological data for effective decision-making.
• Coordinated interactions between molecular representations, learning algorithms, and explainability are crucial for complex bioassays.

Purpose of the Study:

• To establish a benchmarking framework for VS of immune-related bioassays.
• To enable consistent, predictive, and interpretable modeling by integrating molecular representations, ML/DL algorithms, and explainability.
• To derive latent concepts for hypothesis generation in immune target-ligand interactions.

Main Methods:

• A wide range of Machine Learning (ML) and Deep Learning (DL) models were evaluated across descriptors, images, and graph-based molecular representations.
• Benchmarking focused on predictive reliability and model interpretability.
• Concept Whitening was combined with the AttentiveFP architecture for an integrated explainable DL framework.

Main Results:

• Support Vector Machines (SVM) showed the strongest overall performance among classical ML models.
• The AttentiveFP architecture outperformed other DL models.
• Identified context-specific optimal VS networks and derived latent learned concepts.

Conclusions:

• The developed framework provides a consistent approach for VS of immune targets.
• The integrated explainable DL framework maintains predictive reliability while enabling concept alignment and separation.
• The study highlights potential for model ensembling and reveals relationships between data characteristics and model behavior.