AI for Bio · Model

TerraBind

Protein-ligand structure and affinity foundation model trading diffusion for speed

Small moleculesAI-native biotechUnited StatesProduct only, no public artifact中文

Overview

NameTerraBind · aliases: TerraBind, EMMI Predict
Full nameThe paper gives no expanded name; the commercial product name is listed as an alias
Ambiguous names (not auto-linked)None
FamilyNot a family entry. Only one version (arXiv v1); no family members
OrganizationTerray Therapeutics
Organization typeAI-native biotech
CountryUnited States
Model classCo-folding / affinity prediction (diffusion-free)
DomainSmall molecules (also: Structure & co-folding)
First release · paper revision
Basis: arXiv v1 submitted (the only version)
Note: The company press release came four days later and is not used as first-release evidence source
Other milestones
  • First public evidence: · arXiv v1 source
  • Preprint v1: · arXiv v1 source
  • Official announcement: † · Commercial product launch press release (a promotional event, not a new version) source
Latest update† partly verified · no later event
Basis: Commercial product launch press release
The latest sourced event is a press release for the commercial product launch; no later version, code or weights were found and the arXiv record is still v1. The event is promotional, not a new model version
Note: Must display as unverified: a press release is not a model version event source
The most recent sourced event, not a claim that nothing changed since.
StatusProduct only, no public artifact · No public code or weights; offered as a commercial product
OpennessCode: Closed
Weights: Closed
The paper gives no code or weights link, so it is treated as not open; it is offered under a commercial product name
Benchmarks reported2 counted (modified protocol 1, retrospective 1)
Wet-lab validation1 recorded campaign
Unverified fields† latest_update, release_milestones.official_announcement, background.funding, background.products_partnerships
As of2026-11-16

What it does

Uses a smaller pairformer trunk with a diffusion-free optimisation scheme to output structure and affinity directly, plus a per-inference uncertainty estimate; parameter comparisons hold only at the trunk level and must not be set against the baseline's full model size. Nesso-1 explicitly follows this slimming route, while AQAffinity builds on another open co-folding model; the comparison baseline is Boltz-2.

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. 2 of 2 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
05 PoseBustersDeveloper report, modified protocol—Self-reportedMerged into a chart with two other benchmarks; no number for this benchmark alone †partly verified · evidence
01 CASPRetrospective dataset use—Third-partyRetrospective CASP16 affinity evaluation · evidence

Other evaluations (outside the 44)

  • Two structural sets (off-site) (Public benchmark not among the 44) Self-reported: Merged into the same chart with two other public structural sets; pose accuracy self-reported as comparable to an earlier baseline
  • Affinity evaluation on 18 internal targets (Own task set) Self-reported: Per-target correlation comparison on developer-owned data that outsiders cannot recompute
  • End-to-end speed comparison (Comparison against baselines only) Self-reported: A speed-up measured at a fixed GPU, token length and sample count per complex; the conclusion depends on that setup
  • Role of a high-throughput activity dataset (Comparison against baselines only) Self-reported: In the paper that dataset is only a source of affinity training data, not a reported evaluation, so it cannot be set against the baseline's same-named result

Developers report: a much smaller trunk with diffusion-free optimisation matches an earlier baseline's pose accuracy, runs more than an order of magnitude faster end to end, and correlates better than the comparator on a released affinity set and on internal targets.

Independent evidence: No independent evidence found. With no code or weights, no third-party evaluation was found; the claimed advantage partly rests on developer-owned target data

Training data

Group in Figure 7: Public data + wet lab for validation only.

ClassValueDetail
a. Public databasesYesPublic structure and predicted-structure databases, public complex sets, with affinity from several public activity databases plus around a million synthetic decoys source
b. Private or licensed dataNoNone for the base model; nine proprietary crystal structures across two targets were used only in a separate fine-tuning experiment source
c. Own wet-lab data used for trainingNoThe paper states the base model uses public data only source
d. Wet lab used only for validationYesExisting biochemical and cell assay data for 18 internal targets used for evaluation only, not new wet-lab work 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.

Proprietary evaluation data (internal target assays) (arXiv v1)
Targets 18 · Designs tested — · Hits — · Success better than the comparator on 15 of 18 targets (developer-reported) · Best affinity —

Denominator basis: 18 internal targets with more than 95 IC50 points each; nine proprietary crystals across two targets used only in a fine-tuning experiment
The data come from ongoing projects, are uncurated and span several chemical series; the total compound count is not given
Executor: The data belong to the developer; not an independent evaluation and not a new wet-lab campaign

Self-reported · · verified · source

Background

  • Organization: A US AI small-molecule drug-discovery company founded in 2018 (headquartered in California; the paper's address differs from the city in earlier coverage), running its own pipeline alongside pharma collaborations; its platform includes a data engine the company says produces a billion-scale set of small-molecule binding measurements with ultra-miniaturised chips.
  • Funding: A $60M Series A in 2022 led by one venture firm; a $120M Series B announced 2024-10-17 co-led by a new investor and a chipmaker's venture arm with several other participants; the company reports over $200M raised in total.
  • Products and partnerships: Collaborations include a large pharma company (2023-12, multi-target), a life-science research company, a transcription-factor-focused biotech (2024-10) and another large pharma company (2024-12); public materials mention upfronts, milestones and tiered royalties with no amounts disclosed.
  • Related models: Boltz-2, Nesso-1, AQAffinity

Reading notes

  1. A substantial part of the claimed affinity advantage comes from developer-owned targets that outsiders cannot reproduce.
  2. The speed claim depends on a specific hardware and sampling setup and need not hold elsewhere.
  3. PoseBusters appears only in a merged bar chart with no standalone value, so it cannot be compared with same-named figures elsewhere.
  4. Other teams are following the diffusion-free slimming route, suggesting affinity tasks may not need full atom-level diffusion.

Caveats

  1. The latest update is a company press release (promotional), not a model version, and must display as unverified.
  2. No public code or weights, so the results cannot be independently recomputed.
  3. Parameter comparisons hold only at the trunk level.
  4. The CASP entry is retrospective dataset use and does not count toward reported benchmarks.

Sources

  1. arXiv 2602.07735 v1 (preprint)
  2. Paper full text (HTML) (preprint)
  3. Commercial launch press release (press)
  4. Company product page (official site)

Corrections

2 corrections
FieldWasNowBasisFound
latest_update2026-02-12 recorded as a model updateSame date, but it is a press release and the kind is no later sourced event; must be marked unverifiedsource2026-11-16
other_evaluationsReports evaluation results on a high-throughput activity datasetIn the paper this dataset is only a training-data source, with no evaluation resultssource2026-11-16

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