ExplaiNN: interpretable and transparent neural networks for genomics.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Deep learning models, particularly Convolutional Neural Networks (CNNs), show great promise in genomic sequence analysis.
• However, a major limitation of these powerful models is their inherent lack of interpretability.
• This opacity hinders their widespread adoption and validation by domain experts in biology.

Purpose of the Study:

• To introduce ExplaiNN, a novel framework that integrates the predictive power of CNNs with the transparency of linear models.
• To enable interpretable predictions for key genomic tasks, thereby bridging the gap between deep learning and biological understanding.
• To provide both global (cell state) and local (sequence-specific) biological insights from genomic data.

Main Methods:

• ExplaiNN combines Convolutional Neural Networks (CNNs) for feature extraction with interpretable linear model components.
• The framework is designed for predicting transcription factor (TF) binding, chromatin accessibility, and de novo motif discovery.
• It supports integration with pre-trained models and annotated position weight matrices as a plug-and-play platform.

Main Results:

• ExplaiNN achieves predictive performance comparable to existing state-of-the-art methods on genomic tasks.
• The model provides transparent predictions, offering clear biological insights at both global and local sequence levels.
• Demonstrated ability to predict TF binding, chromatin accessibility, and discover de novo motifs with high accuracy.

Conclusions:

• ExplaiNN successfully addresses the interpretability challenge in deep learning for genomics.
• The tool enhances biological understanding by providing transparent, actionable insights into genomic sequence function.
• ExplaiNN is poised to accelerate the application of deep learning methods by genomic domain experts.