Leveraging conformal prediction to annotate enzyme function space with limited false positives.

Area of Science:

• Biomedical research
• Computational biology
• Machine learning applications

Background:

• Machine learning (ML) models are vital for prioritizing candidates in drug discovery and biological research.
• Overconfident ML predictions can lead to numerous false positives, hindering experimental validation.
• Quantifying prediction uncertainty and controlling the false discovery rate (FDR) are crucial for reliable biological discovery.

Purpose of the Study:

• To develop a machine learning (ML) framework, CPEC, for FDR-controlled biological discovery.
• To address the issue of overconfident ML predictions and reduce false positives in experimental validations.
• To provide a tool for prioritizing biological hypotheses with guaranteed FDR control.

Main Methods:

• CPEC integrates a deep learning model with conformal prediction, a statistical method.
• Conformal prediction offers rigorous statistical guarantees for ML model predictions.
• The framework was evaluated using enzyme function annotation as a case study.

Main Results:

• CPEC demonstrated reliable FDR control, ensuring predictions do not exceed user-specified levels.
• The framework achieved comparable or better prediction performance at a lower FDR than existing methods.
• CPEC provided accurate predictions for enzymes under-represented in training data.

Conclusions:

• CPEC is an effective ML framework for FDR-controlled biological discovery.
• The tool is valuable for applications requiring high validation yield rates within limited experimental budgets.
• CPEC enhances the reliability of ML-guided biological discovery, particularly in enzyme function annotation.