AI for Bio · Model

Profluent-E1

Retrieval-augmented protein encoder for mutation-effect scoring and representation

ProteinsAI-native biotechUnited StatesMaintained onlySite entity page (Chinese site)中文

Overview

NameProfluent-E1 · aliases: Profluent-E1, E1-600m
Full nameThe paper refers to it as E1 and gives no expanded form
Ambiguous names (not auto-linked)E1
FamilyThree parameter sizes from one paper sharing training set and licence; the sibling generative model is excluded
  • E1-150M · †
  • E1-300M · †
  • E1-600M · †
OrganizationProfluent
Organization typeAI-native biotech
CountryUnited States
Model classProtein language model (retrieval-augmented encoder)
DomainProteins
First release† unverified · paper revision
Basis: Preprint v1, with weights posted at the same time
Note: The DOI request failed, so the date was not reverified this round source
Other milestones
  • First public evidence: † · Weights repository posted source
  • Preprint v1: † · Preprint v1 source
  • Code release: † · Code repository made public source
  • Weights release: † · Weights for three sizes made public source
Latest update · maintenance only
Basis: Repository changelog entry
The repository changelog records one bug fix in the MSA sampling code and nothing else; no new weights. source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · Only maintenance commits; no new version or weights
OpennessCode: Open · Apache-2.0
Weights: Open
Code under Apache-2.0; weights under a permissive licence with attribution, which the developer states is free for research and commercial use
Benchmarks reported3 counted (self 1, retrospective 2)
Wet-lab validationNone recorded
Unverified fields† first_release, release_milestones.paper_v1, release_milestones.code_release, release_milestones.weights_release, open_status.weights_license, training_data.b_private_licensed, benchmarks_reported
As of2026-11-16

What it does

An encoder that takes either a single sequence or retrieved homologues as context, used for mutation-effect scoring and representation. Its small size and retrieval mode set it apart from the sibling generative model ProGen3; PoET-2 works in the same in-context direction.

Benchmarks reported

Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 3 of 3 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
02 ProteinGymDeveloper report, official protocol—Self-reportedOnly a curve in the project README; no row on the official summary table †partly verified · evidence
01 CASPRetrospective dataset use—Third-partyCASP15 contact maps, numbers not retrieved †partly verified · evidence
15 CAMEORetrospective dataset use—Third-partyCAMEO targets used post hoc; curve only. †partly verified · evidence

Other evaluations (outside the 44)

  • Comparison with similar-size protein language models (Comparison against baselines only) Self-reported: The repository plots size against metric alongside two open encoders and two in-context models

Developers report: the single-sequence mode beats two common open encoders at similar size and the retrieval mode surpasses two in-context models; the public material gives curves, not values.

Independent evidence: No independent evidence found. No entry for the model was found in the official leaderboard file as of 2026-11-16, and no third-party evaluation was found

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesA genomic and metagenomic sequence set curated by the developer, with training reported at about 4 trillion tokens source
b. Private or licensed dataNot disclosedThe paper does not say whether the set contains licensed data source
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The developer is an AI-native protein and gene-editing company founded in 2022 in California, reported at about 20 people in 2024 with no current public headcount.
  • Funding: A $35M round in March 2024 and $106M on 19 November 2025, about $150M in total; the valuation is not officially disclosed, with media reporting it near $1B.
  • Products and partnerships: The line-up includes a generative protein language model, this encoder and an open gene-editing protein; on 28 April 2026 a multi-programme recombinase collaboration with a large US pharmaceutical company was announced, with milestones up to $2.25B and an undisclosed upfront.
  • Related models: ProGen3, PoET-2, ESM C

Reading notes

  1. Paper and repository give curves only, so claims of beating another model read as directional, not as checkable numbers.
  2. Retrieval-augmented and single-sequence models are not directly comparable on the same board; the mode must be matched first.
  3. The contact-map work reuses released targets from a continuous evaluation and one CASP round after the fact; it is not participation.
  4. The training set is a privately curated collection, so its composition cannot be reproduced externally.

Caveats

  1. The first-release date comes from the preprint; the DOI request failed this round, so it is marked unverified.
  2. The public material carries no per-model values, so no comparison can be rechecked.
  3. The official leaderboard file holds no entry for the model.
  4. Whether the training set contains licensed data is not disclosed.
  5. The latest event is a code bug fix, not a new version.

Sources

  1. Preprint 2025 (preprint)
  2. Preprint full text (preprint)
  3. Code repository (repository)
  4. Weights repository (600M) (repository)

Corrections

None recorded.

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