AI for Bio · Model

EVOLVEpro

Few-shot active-learning wrapper for protein engineering, about a dozen variants per round

ProteinsAcademic labUnited StatesMaintained only中文

Overview

NameEVOLVEpro · aliases: EVOLVEpro, EvolvePro
Full nameSources use EVOLVEpro only; no expanded name
Ambiguous names (not auto-linked)None
FamilyNot a family entry. A single method; no family boundary needed
OrganizationBrigham and Women's Hospital / Harvard Medical School, Mass General Brigham Gene and Cell Therapy Institute, Beth Israel Deaconess Medical Center, MIT
Organization typeAcademic lab
CountryUnited States
Model classDesign (active learning on pLM embeddings)
DomainProteins (also: Antibodies)
First release† partly verified · preprint
Basis: bioRxiv preprint (date carried from the existing note, not re-fetched this round)
Note: The preprint's first posting date was not re-verified this round source
Other milestones
  • First public evidence: † · bioRxiv preprint (date carried from the existing note, not re-fetched this round) source
  • Preprint v1: † · bioRxiv preprint (date carried from the existing note, not re-fetched this round) source
  • Journal online: · Science journal publication (Crossref issued date) source
Latest update · journal
Basis: Science journal publication (DOI 10.1126/science.adr6006)
Journal version online; repository commits through 2025-06-11 are maintenance. source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · Latest substantive event is the 2025-01 journal paper; later commits are maintenance
OpennessCode: Open
Weights: Unclear
Two public repositories; it has no weights of its own and relies on third-party pLM weights
Benchmarks reported0 counted · none of the 44
Wet-lab validation2 recorded campaigns
Unverified fields† wetlab_validation[1], first_release, open_status.weights
As of2026-11-16

What it does

Uses frozen large protein-language-model embeddings as features and fits a light regressor to each round of measured activity, then picks the next batch of mutants; it pretrains nothing itself. Unlike zero-shot models such as ESM3, it depends on experimental data for each target protein, so it fits settings that already have an activity assay.

Benchmarks reported

None of the 44 benchmarks. This model is on the zero-pairing list.

How it evaluates itself: In silico DMS active-learning comparison; REGN10987 antibody engineering; PsaCas12f, PE2, Bxb1, T7 RNAP engineering

Other evaluations (outside the 44)

  • In silico DMS active-learning comparison (Own task set) Self-reported: Compares eight representations on 12 DMS datasets; frozen pLM embeddings rank best and beat zero-shot prediction (self-reported)
  • REGN10987 antibody engineering (Wet-lab validation) Self-reported: Five active-learning rounds; the hit rate rose and the final product reached nanomolar affinity (self-reported)
  • PsaCas12f, PE2, Bxb1 and T7 RNAP engineering (Wet-lab validation) Self-reported: 4-8 rounds of about a dozen single mutants each; the reported fold gains are inconsistent and unverified

Developers report: Developers report that few-shot active learning on pLM embeddings improved five different proteins within 4-8 rounds of about a dozen mutants each (self-reported).

Independent evidence: No independent evidence found. No third-party replication found; the authors disclose patents and company roles

Training data

Group in Figure 7: Own wet-lab training data.

ClassValueDetail
a. Public databasesYesFrozen ESM-2 15B weights pretrained on public sequence data; 12 DMS sets only for in silico grid search source
b. Private or licensed dataNoThe paper names no private or licensed data source
c. Own wet-lab data used for trainingYesPer-protein active-learning rounds of about 11-16 measured variants train the top regressor source
d. Wet lab used only for validationNoThe same experiments are training input and application result, not validation only source

Wet-lab validation

Every rate carries its numerator and denominator. Claims that could not be verified are shown as unverified, without numbers; different paper versions are separate entries.

REGN10987 antibody engineering (self-reported) (PMC preprint version)
Targets 1 · Designs tested — · Hits — · Success hit rate rose from 9.1% to 54.5% over five rounds; ELISA binding improved 63% · Best affinity IC50 11.9 nM

Denominator basis: Five active-learning rounds of about 11-16 variants each; the paper gives percentages, not numerators and denominators
Self-reported with no independent replication; endpoints must be named precisely (a reduced-immunogenicity by-product is not protein activity) and the figures come from an early version

Self-reported · † · verified · source
Enzyme and protein engineering set (unverified) (PMC preprint version)
Targets — · Designs tested — · Hits — · Success — · Best affinity —

Denominator basis: 4-8 rounds of about 12-16 single mutants each; the abstract and the discussion disagree on the fold-improvement range, so no figures are given (unverified)
Unverified: figures must be shown per endpoint and never as one fold-change range; no independent replication

Self-reported · † · †partly verified · source

Background

  • Organization: An academic team across Harvard-affiliated hospitals (including a gene and cell therapy institute and a virology and vaccine center) and MIT, with collaborators at the University of Tokyo and the Inamori Research Institute; the work spans genome editing, CRISPR diagnostics and protein engineering.
  • Funding: The acknowledgements list several NIH grants plus foundation and venture funding; the Japanese co-authors are supported by AMED, JSPS and JST.
  • Products and partnerships: No product and no pharma partnership; the authors have filed patents on the method without giving numbers. It builds on a third-party open pLM with no affiliation to that model's publisher.
  • Related models: ESM3

Reading notes

  1. These are per-protein experimental results and cannot be lined up against zero-shot model scores.
  2. Each fold improvement starts from a different baseline, so they cannot be pooled into one range.
  3. There is no score relation to ProteinGym: its 12 DMS sets are only a data source.
  4. The tiny per-round budget is the point: with an assay in hand, active learning is more practical than zero-shot.

Caveats

  1. Zero pairings: no site benchmark entry exists.
  2. The enzyme and protein engineering set's fold gains are inconsistent and marked unverified.
  3. All results are developer-reported; no independent replication was found.
  4. Its own wet-lab data feeds the training loop, so it is not comparable to validation-only models.

Sources

  1. Science journal paper (journal) ·
  2. Preprint full text (PMC) (preprint)
  3. Official repository (full data) (repository)
  4. Official repository (deployable model) (repository)

Corrections

1 correction
FieldWasNowBasisFound
wetlab_validationA single fold-change range summarised the gains across all engineering casesThe abstract and discussion ranges disagree, so figures are shown per endpoint and no single range is given (unverified)source2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.