New spotLights conversation on machine learning, epistasis and scientific rigor
16 September 2026
Machine learning is increasingly shaping protein engineering, but how can we tell whether models truly capture complex biological phenomena such as epistasis, where the effect of one mutation depends on the presence of others?
In our latest spotLights episode hosted by Leonie Brüne, Gian Marco Visani and William S. DeWitt discuss their preprint, “Additive baselines furnish no evidence for epistasis learning by MULTI-evolve.” The study revisits the recently published MULTI-evolve framework and asks a fundamental question: does the model genuinely learn epistatic interactions, or can its predictions be explained by a much simpler additive model?
Beyond the specific case of MULTI-evolve, the conversation explores the importance of robust baseline comparisons in machine learning, scientific rigor in interdisciplinary research and the future of ML-driven protein engineering.
Listen to the episode below and discover the authors’ perspective on what it takes to make meaningful claims about biological complexity in the age of AI and machine learning.