Computational methods for understanding genotype-phenotype maps, part 1
Cold Spring Harbor Laboratory (CSHL)
Abstract:
A fundamental question in biology is how variation in DNA, RNA, and protein sequences translates into quantitative phenotypic variation. Learning these genotype-to-phenotype relationships is challenging because sequence spaces are high-dimensional and the effects of mutations are often strongly context-dependent. Although high-throughput experiments can now measure phenotypes for large numbers of sequences, computational models are still needed both to predict unmeasured genotypes and to characterize the structure of the underlying landscape. Here, we introduce a conceptual and computational framework that connects theoretical results from the fitness-landscape literature with Bayesian inference for deciphering complex genotype-phenotype maps from empirical data.
In the first part of the talk (presented by Carlos Marti), we consider genotype-to-phenotype maps for which the complete sequence space remains computationally tractable, encompassing up to millions of genotypes. We define prior distributions over complete landscapes whose parameters directly control the expected values of statistics quantifying the magnitude and structure of genetic interactions. We then use Gaussian process regression under these priors to analyze high-throughput measurements of the Shine-Dalgarno sequence and a self-splicing intron, obtaining phenotypic predictions for every possible genotype. Finally, we investigate the global and local structure of the inferred landscapes using a combination of summary statistics, local quantitative descriptions and non-linear dimensionality reduction methods and reveal the molecular mechanisms underlying these complex maps.
In the second part of the talk (presented by Samantha Petti), we turn to genotype-to-phenotype maps whose sequence spaces are too large to enumerate or to generate predictions for every possible genotype. We introduce a new family of priors parametrized by the expected effects that individual mutations have on the predictability of other mutations. These parameters can be estimated from data using GPU-accelerated Gaussian processes and thus provide clues about how phenotype arises from genotype. Moreover, we derive tractable analytical expressions for computing a broad class of landscape statistics under these models that would ordinarily require predictions across the entire sequence space, and discuss how this new “kernel trick” can be used both to interpret the structure of the landscape and to navigate toward regions of sequence space with desirable properties. Together, these methods provide a practical toolkit for learning, interpreting, and exploring complex genotype-to-phenotype relationships from the rapidly growing collection of high-throughput datasets.
Biographies:
Carlos Martí is a postdoctoral researcher in the McCandlish Lab at Cold Spring Harbor Laboratory (CSHL). His research focuses on developing computational methods for analyzing empirical sequence–function relationships across diverse biological systems. Previously, he received his PhD from the Universidad Autónoma de Madrid, where he studied the regulation and evolution of alternative splicing using computational approaches, and an MSc in Bioinformatics from the Universitat Autònoma de Barcelona.
Sam Petti is an Assistant Professor of Mathematics at Tufts University and holds a secondary appointment in Computer Science. Her group is supported by the Burroughs-Wellcome Fund Careers at the Scientific Interface award. Previously she was an independent postdoctoral fellow at Harvard University’s NSF-Simons Center for Mathematical and Statistical Analysis of Biology. She received her Ph.D. from the multidisciplinary Algorithms, Combinatorics, and Optimization (ACO) program at Georgia Tech.