Genomic functions arise from nucleotide sequences and their overlapping k-mers – subsequences whose contributions depend on their composition, position and associations. Understanding these contributions requires computing a k-mer contribution function that may or may not consider k-mer associations. Neural networks that model associations yield powerful predictors but are notoriously hard to interpret; conversely, models that ignore associations deliver exact, position-specific contributions yet might underperform. We introduce Oyster, the first convolutional architecture that can be toggled between Exact (ignoring associations) and non-Exact modes. Exact Oysters yield closed-form k-mer contributions directly from their weights, without post-hoc attribution. By letting users choose between interpretability and complexity, Oyster provides a unified framework for transparent, high-performance sequence-to-function modeling. We apply Oyster to predict intensities of YY1-DNA interactions in human K562 cells from 500-nt DNA windows and intensities of eleven histone post-translational modifications. Exact and non-Exact variants achieved statistically indistinguishable performance, highlighting that k-mer associations are not necessarily important for all biological phenomena. Modelling of YY1-DNA interactions is dependent on the YY1 motif which is expected but is also dependent on several histone post-translational modifications including H3K9ac and H2AFZ.
The Hidden Power of Normalization: Exponential Capacity Control in Deep Neural Networks
arXiv:2511.00958v1 Announce Type: cross Abstract: Normalization methods are fundamental components of modern deep neural networks (DNNs). Empirically, they are known to stabilize optimization dynamics and