Colloquium "Interpretable machine learning models for higher-order epistasis in sequence-to-function data"

20 October 18:10 – 19:30

Speaker: Rick Farouni, HSE University

Title: Interpretable machine learning models for higher-order epistasis in sequence-to-function data.

Abstract: This talk reviews interpretable models for epistasis in sequence-to-function data across protein biophysics, statistical genetics and machine learning. I clarify mathematical connections across selected formulations and examine how measurement scale, reference sequence or background distribution, and prior affect the interpretation of their results. An exploratory benchmark on simulated landscapes finds that the best-performing estimator depends on the generating landscape. The central message is that models should be evaluated both on how accurately they predict and on how faithfully they recover the underlying interactions.

Place: Zoom