AI for Bio · Model

HCAbLM

Nona Biosciences language model trained only on fully human heavy-chain-only antibodies, aimed at developability

AntibodiesPharmaChina, United StatesActive中文

Overview

NameHCAbLM · aliases: HCAbLM
Full nameNeither the preprint nor the press release gives an expanded name
Ambiguous names (not auto-linked)None
FamilyNot a family entry. A single model; the group's separate generative HCAb model is a different model and is not merged
OrganizationNona Biosciences, HBM Holdings (Harbour BioMed)
Organization typePharma
CountryChina, United States
Model classAntibody LM (developability)
DomainAntibodies
First release · paper revision
Basis: bioRxiv preprint v1 posted source
Other milestones
  • First public evidence: · bioRxiv preprint v1 source
  • Preprint v1: · bioRxiv preprint v1 source
  • Official announcement: · Company press release source
Latest update · no later event
Basis: Only the v1 preprint exists
No later sourced event found (as of 2026-11-16); no new version, code or weights found source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · First released 2026-09; no later version yet
OpennessCode: Unclear
Weights: Unclear
The press release and preprint page do not say whether weights or code are public
Benchmarks reported0 counted · none of the 44
Wet-lab validationNone recorded
Unverified fields† open_status.code, open_status.weights, training_data.a_public
As of2026-11-17

What it does

A 366M-parameter language model that learns only the sequence grammar of fully human heavy-chain-only antibodies (HCAbs); the preprint says HCAbs differ from conventional human VH mainly in CDR1/2, CDR3 architecture and parts of the framework. Frozen representations are transferred to SEC purity, HIC behaviour and thermal stability. It is not a structure or design model, unlike IgGM or RFantibody; its comparators are general and antibody language models such as ESM3.

Benchmarks reported

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

How it evaluates itself: Sequence distribution analysis; Frozen-embedding transfer; Press-release baseline claim

Other evaluations (outside the 44)

  • Sequence distribution analysis (Own task set) Self-reported: Compares HCAb and conventional human VH sequences; shifts concentrate in CDRs and some framework regions
  • Frozen-embedding transfer to developability (Own task set) Self-reported: Predicts SEC purity, HIC behaviour and thermal stability with grouped internal validation and retrospective cross-project evaluation, all at the developer
  • Press-release baseline claim (Comparison against baselines only) Self-reported: Claims to beat IgLM, AbLang and a 6B-parameter ESM model on SEC and HIC; no numbers and the dataset is unnamed
  • FLAb and GDPa1 (Public benchmark not among the 44) Self-reported: Not pairings: the FLAb2 model list does not include it, and the public evaluation in the press release is unnamed and cannot be taken as GDPa1

Developers report: a 366M model trained on 31.8M fully human HCAb sequences represents developability better than larger general or antibody language models.

Independent evidence: No independent evidence found. Only the press release and preprint exist; no third-party evaluation found

Training data

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

ClassValueDetail
a. Public databasesNot disclosedWhether public data were mixed in is not disclosed source
b. Private or licensed dataNo source
c. Own wet-lab data used for trainingYesImmune repertoires from the company's transgenic mouse platform: 73 immunised mice, 31.8M fully human HCAb sequences source
d. Wet lab used only for validationYesSEC, HIC and thermal-stability measurements are used for transfer evaluation source

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: Nona Biosciences was set up in 2022-11 as a wholly owned subsidiary of HBM Holdings (Harbour BioMed, HKEX 02142), offering discovery and preclinical services up to IND; at launch the group said its transgenic mouse platform had been used by over 50 companies and institutions.
  • Funding: No separate funding for Nona; the group reported about $158M total 2025 revenue (annual results release), and the 2025-03 AstraZeneca deal included a $105M equity investment.
  • Products and partnerships: Products include H2L2 and HCAb transgenic mouse platforms and the Hu-mAtrIx AI discovery platform; in 2025 the group had deals with AstraZeneca, Otsuka, Pfizer, Bristol Myers Squibb and others; a separate generative HCAb model was announced in 2025-10.
  • Related models: IgGM / IgGM-FM, RFantibody, ESM3

Reading notes

  1. The only antibody language model on the site trained on a private transgenic-mouse repertoire; the data are distinctive but there is no results table to compare.
  2. The ESM comparator is named only in press-release wording; the actual model should be checked in the preprint.
  3. FLAb2 and GDPa1 are the public tests it most needs; worth following up.
  4. Weight availability is unstated, so it cannot be reproduced externally.

Caveats

  1. Comparison claims come from a press release with no numbers and an unnamed dataset.
  2. All developability evaluation was done inside the developer.
  3. None of the four added antibody benchmarks on the site has a public score for it.

Sources

  1. bioRxiv preprint (preprint)
  2. Company press release (press)
  3. Nona launch press release (press)
  4. Group 2025 annual results release (press)
  5. AstraZeneca collaboration press release (press)
  6. FLAb2 paper (third-party paper)

Corrections

None recorded.

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-17.