Deciphering RNA splicing logic with interpretable machine learning.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Machine learning, especially neural networks, accelerates scientific discovery but often lacks interpretability.
• Current neural networks cannot explain their predictions, hindering scientific understanding.
• RNA splicing is crucial for converting genomic information into functional products.

Purpose of the Study:

• To present an interpretable-by-design neural network model for RNA splicing analysis.
• To achieve predictive accuracy comparable to state-of-the-art models while providing insights.
• To advance scientific discovery through interpretable machine learning.

Main Methods:

• Developed an interpretable neural network architecture.
• Trained the model on large datasets for RNA splicing prediction.
• Introduced a visualization tool to trace the model's decision-making process.

Main Results:

• The interpretable model achieved predictive accuracy on par with state-of-the-art methods.
• A novel visualization allowed tracing predictions from input sequences to output.
• The model identified previously uncharacterized components of RNA splicing logic.

Conclusions:

• Interpretable machine learning can provide significant scientific insights.
• The developed model successfully elucidates RNA splicing mechanisms.
• This approach advances scientific discovery by making complex models understandable.