An equivariant Bayesian convolutional network predicts recombination hotspots and accurately resolves binding motifs.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Convolutional neural networks (CNNs) excel with abundant data but struggle with limited, noisy biological sequence data.
• Predicting phenotypes from DNA sequences requires methods that optimally utilize limited training data and inherent sequence structure.

Purpose of the Study:

• To develop a robust deep learning model for analyzing DNA sequences with limited and noisy data.
• To improve prediction consistency and accuracy in genomic sequence analysis.

Main Methods:

• Integration of equivariant networks with Bayesian dropout for enhanced CNNs.
• Implementation of a model with exact reverse-complement symmetry for DNA sequence analysis.
• Application of the model to predict recombination hotspots and identify protein binding motifs.

Main Results:

• The novel model demonstrates improved prediction consistency and accuracy over standard CNNs and motif finders.
• Accurate prediction of recombination hotspots from DNA sequence data.
• Identification of previously unobserved binding motifs for the PRDM9 protein, validated by experimental assays.

Conclusions:

• The developed model offers a powerful approach for analyzing complex genomic data, particularly in low-signal or limited-data scenarios.
• This method advances the identification of functional elements in DNA sequences, such as protein binding sites.