Using Attribution Sequence Alignment to Interpret Deep Learning Models for miRNA Binding Site Prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• MicroRNAs (miRNAs) are key regulators of gene expression at the post-transcriptional level.
• miRNA binding to target transcripts is mediated by the Argonaute protein family.
• Accurate prediction of miRNA target sites is crucial for understanding gene regulation, but current methods face challenges in interpretability.

Purpose of the Study:

• To introduce a novel interpretation technique, attribution sequence alignment, for deep learning-based miRNA target prediction models.
• To provide a human-readable visual representation of miRNA:target interactions.
• To demonstrate the utility of this method for biological data analysis and model interpretation.

Main Methods:

• Developed attribution sequence alignment, a technique for interpreting deep learning models on a 2D representation of miRNA and target sequences.
• Applied the method to visualize and analyze miRNA:target interactions.
• Demonstrated applications in clustering experimental data and refining predicted binding sites on long transcripts.

Main Results:

• Attribution sequence alignment provides an interpretable, visual output for complex miRNA target prediction models.
• The method successfully clustered experimental data into distinct binding classes.
• It effectively narrowed down potential miRNA binding sites within long transcript sequences.

Conclusions:

• Attribution sequence alignment enhances the interpretability of deep learning models for miRNA target prediction.
• This technique facilitates the exploration of biological insights derived from neural networks.
• The method is versatile, applicable to various neural network models and extensible to other interaction domains like protein-protein interactions.