BioMADE: Predicting Torsades de Pointes from molecular structures through biologically informed representations.

Area of Science:

• Computational toxicology
• Pharmacology
• Machine learning in drug discovery

Background:

• Drug-induced arrhythmias, especially Torsades de Pointes (TdP), are a significant safety concern in drug development.
• Current in silico methods often lack sufficient biological context, relying on chemical properties or complex simulations.
• Machine learning (ML) offers potential for analyzing complex datasets to predict cardiotoxicity.

Purpose of the Study:

• To introduce BioMADE, a novel ML framework for predicting TdP risk.
• To leverage small-molecule-protein activity profiles for TdP prediction without extensive mechanistic data.
• To develop a biology-informed ML approach for enhanced drug safety assessment.

Main Methods:

• Utilized ChEMBL activity data to train gene-specific ML models.
• Constructed a latent biological embedding (BioMADE embedding) for molecules using arrhythmia-relevant genes.
• Employed a support vector machine classifier with BioMADE embeddings to predict TdP risk.

Main Results:

• BioMADE embeddings showed superior classification performance for biological elements (e.g., ATC3 class) compared to Molformer and MACCS.
• BioMADE achieved an AUROC of 0.89 in internal validation for TdP prediction.
• BioMADE demonstrated competitive performance against state-of-the-art models, reaching an AUROC of 0.74 on an external dataset.

Conclusions:

• BioMADE offers a scalable, biology-informed, and generalizable method for predicting drug-induced toxicities.
• Integrating protein activity profiles is crucial for accurate adverse drug reaction prediction.
• This framework highlights the importance of human biology in toxicology modeling, complementing chemical descriptors.