DeepSIBA: chemical structure-based inference of biological alterations using deep learning

Area of Science:

• Computational chemistry
• Cheminformatics
• Pharmacology

Background:

• Predicting biological effects of chemical structures is crucial for in silico drug discovery.
• Current methods face challenges in accurately linking molecular structures to biological outcomes.

Purpose of the Study:

• To develop a deep learning model for predicting biological effects from chemical structures.
• To identify structurally dissimilar compounds with similar biological impacts.
• To infer signaling pathway signatures and influential substructures from chemical structures.

Main Methods:

• Utilized Siamese Graph Convolutional Neural Networks to map compound differences to biological effect alterations.
• Represented chemical structures as graphs for deep learning analysis.
• Employed deep ensembles for uncertainty estimation in predictions.

Main Results:

• The model accurately encoded molecular graph pairs and identified structurally dissimilar compounds affecting similar biological processes.
• Achieved high precision in predicting biological effects.
• Provided reliable predictions for novel chemical structures outside the training set.
• Successfully inferred signaling pathway signatures and influential substructures for anticancer drugs.

Conclusions:

• The developed deep learning model offers a powerful tool for in silico drug discovery.
• The approach enables accurate prediction of biological effects and identification of key molecular drivers.
• This method has significant potential for accelerating drug development and understanding drug mechanisms.