Uncertainty-aware genomic deep learning with knowledge distillation.

Area of Science:

• Genomics
• Machine Learning
• Computational Biology

Background:

• Deep neural networks (DNNs) are powerful for regulatory genomics prediction.
• Challenges exist in ensuring DNN prediction reliability and interpretability.
• Understanding model decision-making is crucial for biological insights.

Purpose of the Study:

• Introduce DEGU (Distilling Ensembles for Genomic Uncertainty-aware models) to enhance DNN robustness and explainability.
• Integrate ensemble learning and knowledge distillation for improved genomic predictions.
• Provide calibrated uncertainty estimates for trustworthy deep learning applications in genomics.

Main Methods:

• DEGU distills an ensemble of DNNs into a single model.
• Captures both average predictions and variability (epistemic uncertainty).
• Includes an auxiliary task for estimating data-based (aleatoric) uncertainty.

Main Results:

• DEGU models inherit ensemble performance benefits in a single model.
• Improved generalization to out-of-distribution sequences.
• Consistent explanations of cis-regulatory mechanisms via attribution analysis.
• Calibrated uncertainty estimates with conformal prediction coverage guarantees.

Conclusions:

• DEGU enables robust and explainable deep learning in genomics.
• Enhances reliability and interpretability of predictive models.
• Facilitates trustworthy applications of AI in biological research.