Uncertainty-aware genomic deep learning with knowledge distillation.

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Deep neural networks (DNNs) show promise in regulatory genomics but face challenges in prediction reliability and interpretability.
• Understanding the factors driving DNN predictions is crucial for biological insights and model validation.

Purpose of the Study:

• To introduce DEGU (Distilling Ensembles for Genomic Uncertainty-aware models), a method enhancing DNN robustness and explainability.
• To provide reliable and interpretable predictions for regulatory genomics tasks using deep learning.

Main Methods:

• DEGU integrates ensemble learning and knowledge distillation to create a single, uncertainty-aware DNN.
• It captures both predictive consensus and variability (epistemic uncertainty) from an ensemble.
• An optional task models data variability (aleatoric uncertainty) from experimental replicates.

Main Results:

• DEGU-trained models achieve ensemble performance in a single model, improving generalization to out-of-distribution sequences.
• Attribution analysis reveals more consistent explanations of cis-regulatory mechanisms.
• Models provide calibrated uncertainty estimates with conformal prediction offering coverage guarantees.

Conclusions:

• DEGU enhances the trustworthiness of deep learning applications in genomics research.
• The method offers robust, explainable, and uncertainty-aware predictions for functional genomic tasks.
• DEGU facilitates reliable decision-making in genomics by quantifying prediction uncertainty.