Additive baselines furnish no evidence for epistasis learning by MULTI-evolve

As machine learning (ML) continues to expand its role in biology, researchers are increasingly seeking ways to capture complex phenomena that govern how proteins function and evolve. One such phenomenon is epistasis, where the effect of one mutation on a protein’s activity depends on the presence of other mutations. For example, if two sites are in contact in the folded 3D protein structure, the effects of mutations to alternative amino acids at these sites are often interdependent.

MULTI-evolve, the Arc Institute’s “model-guided, universal, targeted installation of multimutants” (DOI: 10.1126/science.aea1820), is an ML-based workflow developed to address this challenge. Combining protein language models with neural-network-based epistatic modelling, the method aims to identify improved protein variants from a relatively small set of single- and double-mutant measurements and to predict the behaviour of higher-order mutants. Most notably, the approach claims a universal ability to learn epistatic interactions across diverse protein systems, a characteristic reflected in the “U” (“universal”) in the MULTI-evolve acronym.

Although the paper sets its sights high, it lacks one key element: straightforward comparisons with a simple additive baseline model to answer the crux of the question: did the model learn, and made predictions based of, epistatic patterns?

The authors of the preprint highlighted here, “Additive baselines furnish no evidence for epistasis learning by MULTI-evolve”, (https://doi.org/10.64898/2026.04.23.719915) address this fundamental question and, following their own reanalysis of the datasets used in MULTI-evolve, found that the model is in fact unable to learn epistasis but instead also uses a basic additive model for its predictions. This point is especially important when looking at higher‑order mutant predictions, where closer analysis shows that the model consistently collapses to an effectively additive behaviour.

Building on this observation, the authors argue that claims about sophisticated phenomena such as epistasis learning should always be preceded by carefully designed baseline experiments. This is intended not only to strengthen scientific rigor, but also to promote a fair and standardised review culture and to make results more accessible to non‑specialist readers. In addition, they call for broader interdisciplinary training of early‑career researchers, so that complex research questions can be approached from multiple methodological perspectives without sacrificing analytical stringency.

For a more in‑depth and lively exploration of the study, including the authors’ own take on their analysis and on where machine learning in protein engineering and epistasis modelling is headed, Gian Marco Visani and William S. DeWitt discuss this preprint in a dedicated spotLight episode.

Tags: additive models, bioinformatics, computational biology, epistasis, machine learning, model validation, multi-evolve, protein engineering, protein evolution, scientific rigor