EvoIF

EvoIF is a compact model for protein fitness prediction: given a protein and a set of mutations, it scores how those mutations are likely to affect function. This is a central task in variant-effect prediction and protein engineering, and strong zero-shot predictors typically lean on either evolutionary information from homologous sequences or on structure-aware models. EvoIF's contribution is to combine complementary evolutionary signals efficiently, matching much larger systems while using a tiny fraction of their training data.

The method rests on a conceptual framing: it interprets evolution as an implicit reward-optimization process and masked language modeling as a form of inverse reinforcement learning that recovers those evolutionary preferences. Concretely, EvoIF fuses two complementary sources of signal, within-family profiles built from retrieved homologs, and cross-family structural-evolutionary constraints distilled from the logits of an inverse-folding model, through a compact transition block that calibrates the combined probabilities for log-odds scoring of mutations.

EvoIF was introduced in October 2025 by Xiaoran Jiao, Jigang Fan, Shengdong Lin, Zhanming Liang, Weian Mao, Chenchen Jing, Hao Chen, and Chunhua Shen, a collaboration led from Zhejiang University and Peking University with additional partners. It is described in an arXiv preprint released under a CC BY-NC-ND 4.0 license; the authors state code will be released at the linked repository, which is not yet public.

#Key Features

• Complementary evolutionary signals: EvoIF fuses within-family profiles from retrieved homologs with cross-family constraints distilled from inverse-folding logits, which ablations show are complementary sources of information.
• Inverse-reinforcement-learning framing: The model interprets evolution as implicit reward optimization and masked language modeling as inverse RL, motivating how it extracts evolutionary preferences.
• Frozen foundation backbones: EvoIF builds on a frozen ESM-2-650M protein language model and a frozen ProteinMPNN inverse-folding model, with a Geometric Vector Perceptron network processing structure.
• Extreme data efficiency: With 76M parameters, it reaches state-of-the-art quality using only about 0.15% of the training data required by recent large-scale models.
• Robust across conditions: Ablations confirm consistent improvements across function types, MSA depths, taxa, and mutation depths.

#Technical Details

EvoIF is a lightweight 76M-parameter model that fuses sequence-structure representations with evolutionary profiles through a compact transition block, producing calibrated probabilities used for log-odds mutation scoring. It relies on frozen backbones, an ESM-2-650M protein language model and a ProteinMPNN inverse-folding model, with a Geometric Vector Perceptron network handling structure, and is trained on 30,948 experimental protein structures from CATH v4.3.0. On ProteinGym (217 mutational assays covering more than 2.5 million mutants), the MSA-free EvoIF reaches a Spearman correlation of 0.489 and the MSA-enabled EvoIF-MSA reaches 0.519, surpassing ESM-2-650M (0.414), SaProt (0.457), and TranceptEVE L (0.456), and matching or exceeding the much larger VenusREM (0.518) and AIDO Protein-RAG 16B (0.518) while using a small fraction of their training data.

#Applications

EvoIF is aimed at protein engineers and variant-effect researchers who need fast, accurate zero-shot fitness scores without training on labeled assay data. Because it builds on frozen public backbones and trains on a modest structure set, it is inexpensive to run and reproduce, making it practical for ranking candidate mutations, guiding directed-evolution campaigns, or prioritizing variants for experimental testing. Once the authors release their code, groups will be able to apply it to their own targets and integrate its scores into design pipelines.

#Impact

EvoIF demonstrates that carefully combining within-family and cross-family evolutionary signals can rival billion-parameter systems on ProteinGym at a fraction of the compute and data, a useful counterpoint to the trend of scaling protein models ever larger. Its framing of masked language modeling as inverse reinforcement learning offers a principled lens on why evolutionary and inverse-folding signals are complementary. As a preprint, its benchmark results await peer review, and its non-commercial CC BY-NC-ND license restricts commercial reuse.