Reproduce “A mechanistically interpretable neural network for regulatory genomics” on OpenResearch

About this paper

Alex M. Tseng, Gokcen Eraslan, Tommaso Biancalani, Gabriele Scalia

An architecture for regulatory genomics, ARGMINN, directly encodes DNA sequence motifs and their combinatorial syntax within its structure, enabling transparent motif discovery and the interpretable design of novel functional sequences. The model demonstrates enhanced motif identification and instance calling accuracy, and can generate functional DNA sequences validated by independent models.

Key insights

• Mechanistic interpretability can be intrinsically designed into deep learning architectures for genomics by forcing specific network components (e.g., first-layer filters) to learn distinct biological concepts (motifs).
• A custom regularization strategy applied to convolutional filters can effectively prevent redundancy and ensure each filter specializes in learning a unique and biologically relevant DNA motif.
• A modified attention mechanism, operating with an explicit memory stream, enables the direct and transparent identification of motif instances and their learned syntactical relationships, bypassing the need for complex post-hoc interpretation methods.

Summary by alphaXiv.