Learning sequence-function relationships with scalable, interpretable Gaussian processes.

Area of Science:

• Genetics and Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Understanding genotype-phenotype relationships is crucial in genetics but complicated by epistasis (context-dependent mutational effects).
• High-throughput phenotyping generates large datasets, but standard models struggle with generalizability and interpretability.
• Deep neural networks offer flexibility but lack interpretability and uncertainty quantification.

Purpose of the Study:

• To introduce a novel family of interpretable Gaussian process models for sequence-function relationships.
• To capture epistasis using flexible prior distributions that generalize fitness landscape models.
• To provide scalable and interpretable methods for exploring complex genetic interactions.

Main Methods:

• Developed interpretable Gaussian process models with flexible prior distributions to model epistasis.
• Incorporated site-, allele-, and mutation-specific factors to quantify epistatic effects.
• Utilized GPU acceleration for scalability to large datasets (protein, RNA, genome-wide SNP).

Main Results:

• Achieved superior predictive performance on large biological sequence datasets.
• Generated interpretable model parameters that recover known genetic features.
• Uncovered novel epistatic interactions, providing new insights into the genotype-phenotype map.

Conclusions:

• The developed Gaussian process models offer a scalable and interpretable approach to studying sequence-function relationships.
• These models effectively capture epistasis and provide deeper insights into the genotype-phenotype map.
• The methods are applicable across diverse biological systems, including DNA, RNA, and protein sequences.