Oyster: A Neural Network for Modeling Genomic Sequences with Exact Position-Specific k-mer Contributions

Oyster is a glass-box neural network that learns position-specific k-mer contributions while ignoring k-mer associations, enabling exact extraction of k-mer effects from model weights.

• Why it matters: k-mer associations are combinatorial and can obscure model interpretation, so it is unclear when ignoring them still yields faithful, human-understandable insights.
• What they did: Oyster was compared against models that include k-mer associations in two case studies—protein-DNA interactions and enhancer activity—using extracted position-specific k-mer contributions from weights.
• The result: In both case studies, association-aware models performed better, but Oyster predictions were in strong agreement, acting as faithful proxies and producing concepts aligned with literature plus novel ones.