Correcting gradient-based interpretations of deep neural networks for genomics.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Deep neural networks (DNNs) are increasingly used for analyzing high-throughput functional genomics data.
• Post hoc attribution methods aim to interpret DNNs by highlighting important features, such as nucleotides in DNA sequences.
• Current attribution methods often produce noisy results, making it difficult to interpret the learned patterns in genomic data.

Purpose of the Study:

• To identify and address a source of noise in attribution maps generated by DNNs for DNA sequence analysis.
• To improve the interpretability and reliability of DNNs in the field of regulatory genomics.

Main Methods:

• Investigated the impact of one-hot encoding in DNA on DNN attribution methods.
• Developed and applied a statistical correction to reduce attribution noise.
• Evaluated the effectiveness of the correction across various genomic DNNs.

Main Results:

• Identified a novel noise source in DNN attribution maps stemming from the handling of one-hot encoded DNA.
• Demonstrated that this noise is a common issue in genomic DNNs.
• Showcased that the proposed statistical correction significantly reduces attribution noise, leading to more reliable attribution maps.

Conclusions:

• The developed statistical correction effectively mitigates noise in DNN attribution maps for genomic data.
• This method enhances the interpretability of DNNs, facilitating meaningful insights in regulatory genomics.
• The findings offer a promising advancement for applying deep learning in genomic sequence analysis.