Predicting Off-Target Binding Profiles With Confidence Using Conformal Prediction.

Area of Science:

• Computational chemistry
• Pharmacology
• Machine learning

Background:

• Ligand-based models in drug discovery can predict off-target interactions and adverse effects.
• Current methods often lack confidence measures, providing only point predictions.
• Reliable prediction of off-target interactions is crucial for early hazard assessment.

Purpose of the Study:

• To develop a methodology using Conformal Prediction for predicting off-target interactions with confidence measures.
• To create reproducible scientific workflows for model training and pre-processing.
• To provide accessible models via web and OpenAPI interfaces for drug discovery applications.

Main Methods:

• Utilized Conformal Prediction for off-target interaction prediction.
• Trained models on 31 targets from the ExCAPE-DB dataset.
• Employed signature molecular descriptors and support vector machines.
• Ensured reproducibility through openly available GitHub workflows.

Main Results:

• Generated predictions with confidence p-values for each class, addressing the lack of confidence in traditional methods.
• Demonstrated the methodology's utility with compounds from DrugBank.
• Made models publicly available through online interfaces.

Conclusions:

• Conformal Prediction offers a robust approach to predicting off-target interactions with associated confidence levels.
• The developed methodology enhances early drug safety assessment and facilitates compound screening.
• Openly available workflows and interfaces promote reproducibility and accessibility in drug discovery research.