pipette
ESEspañol

Embedding kernels for sequence-function relationships

W. Carabba, C. Marti-Gomez, D. M. McCandlish, S. Petti

Preprint

In the authors' words

A central challenge in biology is understanding how variation in biological sequences (DNA, RNA, and protein) impacts function. Recent advances in high-throughput techniques have enabled the measurement of such sequence-function relationships on an unprecedented scale. However, analyzing these complex datasets requires models that are both flexible and mathematically tractable. We consider Gaussian process regression, a Bayesian method that specifies a Gaussian prior distribution over sequence-function relationships. The prior is defined by a kernel function that encodes the expected covariance between pairs of sequences. We study a family of embedding kernels in which the covariance depends on the distance between sequences in a learned embedding space, making these kernels both expressive and interpretable. We show that linear embedding kernels generalize previously studied classes of kernels and can more flexibly express variability in the expected similarity between mutational effects across genetic backgrounds, allowing the resulting priors to encode more complex biological mechanisms such as RNA base-pairing. We implement these methods using a GPU-accelerated Gaussian process framework, allowing us to learn linear embeddings and make predictions with datasets containing hundreds of thousands of measurements. We find that our embedding kernels match or improve the predictive performance of other kernels across various high-throughput sequence-function datasets, including protein datasets ranging from short motifs to full-length proteins. Furthermore, the learned embedding spaces provide insight into biological mechanisms underlying sequence-function relationships.

Main resultThe abstract does not state a limitation.

Appeared: Wednesday, November 18. bioRxiv. Preprint, not yet peer-reviewed.

DOI: 10.64898/2026.11.15.766687