ProBEiT combines a protein’s amino-acid sequence with a tokenised structural description. Its peer-reviewed Communications Biology paper, published on 26 October, tests whether this improves predictions of protein behaviour.[1]
Why combinations need their own evidence
A mutation’s effect can depend on which other mutations are present. This dependence, called epistasis, makes assembling useful changes harder than adding a list of individual improvements. A well-characterised example comes from a 2019 experiment on fluorescent proteins: researchers constructed all 8,192 combinations linking two variants separated by thirteen amino-acid substitutions.[2]
One substitution reduced brightness in one genetic background but increased it when another substitution was present. Adding a third change almost eliminated that interaction. The same molecular edit therefore acquired a different effect as its surroundings changed. The experiment measured brightness through cell sorting and sequencing, revealing interactions beyond simple pairs.[2]
The arithmetic grows quickly. In an illustrative design with twenty positions and two choices at each, there are 1,048,576 combinations. Testing each change separately would sample only a tiny fraction of those possibilities. Useful prediction needs information about combinations, including cases where a promising individual change becomes unhelpful in company.[2]
A measured heat challenge
The model keeps separate processing units for the amino-acid sequence and the structural description. A shared attention step lets each position draw information from both. The input type determines which unit processes it.[1]
For glycerol-3-phosphate oxidase, the source data list 39 engineered variants and a wild-type control. Researchers measured activity remaining after 30 minutes at 46°C. Candidates came from multiple models and were selected by the experimental team; the ranking comparison was retrospective.[1]
ProBEiT’s predicted scores had a Spearman rank correlation of 0.436 with measured residual activity, against 0.269 for fine-tuned ESM2. The moderate correlation leaves considerable disagreement between predicted and measured order.[1]
Sources & further reading
- Efficient structural information integration into protein language models with implicit mixture-of-experts ProBEiT Communications Biology · 26 Oct 2026 · Accessed 28 Oct 2026
- Learning the pattern of epistasis linking genotype and phenotype in a protein Nature Communications · 16 Sep 2019 · Accessed 28 Oct 2026