An intrinsically interpretable neural network architecture for sequence-to-function learning.

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Sequence-based deep learning models predict genomic functions but lack interpretability.
• Current methods require computationally intensive post hoc analyses for model interpretation.
• Highly parameterized models often obscure internal mechanics.

Purpose of the Study:

• Introduce a novel deep learning architecture, the totally interpretable sequence-to-function model (tiSFM).
• Enhance model performance and parameter efficiency compared to standard convolutional models.
• Enable intrinsic interpretation of model parameters in relation to sequence motifs.

Main Methods:

• Developed the tiSFM deep learning architecture.
• Applied tiSFM to analyze open chromatin measurements across hematopoietic cell types.
• Compared tiSFM performance against a state-of-the-art convolutional neural network.

Main Results:

• tiSFM demonstrated superior performance over a tailored convolutional neural network on open chromatin data.
• Identified context-specific transcription factor activities crucial for hematopoietic differentiation (e.g., Pax5, Ebf1, Rorc).
• tiSFM parameters provided biologically meaningful interpretations for predicting epigenetic state changes during development.

Conclusions:

• tiSFM offers an interpretable alternative to standard deep learning models in genomics.
• The model's interpretable parameters facilitate biological insights into sequence-function relationships.
• tiSFM is effective for complex tasks like predicting developmental epigenetic transitions.