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

Abstract

Genomics data are sequences, and exploration in these data involves mapping these sequences to function and understanding that mapping. Analogous to a sentence composed of words, a sequence's function is understood through its constituent subsequences, or k-mers. The presence of k-mers, their positions, and their relationships with other k-mers (k-mer associations) all affect sequence function. Understanding the sequence function requires determining the contributions of each of its k-mers and the impacts of their positions and k-mer associations. k-mer associations are inherently combinatorial and may be hierarchical. Although these associations may be impactful for biological phenomena, k-mer associations cannot be understood by humans and severely obfuscate model analyses, yielding generalizations, simplifications, and approximations. Ignoring k-mer associations, however, yields concepts that are human-understandable and exact. To determine whether ignoring k-mer associations is appropriate, models that learn or ignore k-mer associations can be compared. Oyster is a novel NN that learns k-mer composition and k-mer positions while ignoring k-mer associations, the first with these abilities to our knowledge. Oyster is a glass-box model as the exact position-specific contributions of k-mers are directly extracted from the model weights. Oyster is compared with models that consider k-mer associations using two case studies spanning two important genomics topics: protein-DNA interactions and enhancer activity. Across these case studies it was found that while models that consider k-mer associations perform better than Oyster models, the predictions of the Oyster models were in strong agreement, serving as faithful proxies. Human-understandable concepts were extracted from the Oyster models revealing concepts that align with the literature as well as novel ones.