Overview
| Name | tFold · aliases: tFold, tFold-Ab, tFold-Ag |
|---|---|
| Full name | No separate expanded name |
| Ambiguous names (not auto-linked) | None |
| Family | Not a family entry. Single-model entry, not a family |
| Organization | Tencent AI for Life Sciences Lab |
| Organization type | Big-tech lab |
| Country | China |
| Model class | Antibody structure and design system |
| Domain | Antibodies |
| First release | · preprint Basis: bioRxiv v1 preprint (repository creation date not used) source |
| Other milestones | |
| Latest update | · journal Basis: Journal online publication of the design pipeline The design-pipeline paper was published; the last GitHub push is the 2025-03-14 packaging release source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Journal publication within 12 months |
| Openness | Code: Partly open Weights: On request GitHub states source and weights are provided but declares no standard open-source licence |
| Benchmarks reported | 0 counted · none of the 44 |
| Wet-lab validation | 1 recorded campaign |
| Unverified fields | None |
| As of | 2026-11-16 |
What it does
An antibody structure and design system: predicts antibody structure without MSA, predicts antibody-antigen complexes, and combines virtual screening with epitope-specific de novo design; the online service sits on the company's drug-discovery platform.
Benchmarks reported
None of the 44 benchmarks. This model is on the zero-pairing list.
How it evaluates itself: SAbDab-22H2-Ab structure test (n=169); SAbDab-22H2-AbAg (n=99); Virtual screening; De novo design, 4 targets
Other evaluations (outside the 44)
- Time-split antibody structure test (Own task set) Self-reported: 169 time-split antibody structures compared with general folding and antibody-specific models
- Antibody-antigen docking test (Own task set) Self-reported: Docking success on 99 complexes; alone it trails a general co-folding model, and improves with ensembling or epitope input
- Virtual screening (Own task set) Self-reported: Candidate pools for two targets ranked, with enrichment metrics reported by the developers
Developers report: Developers report MSA-free antibody structure prediction better than general folding models on antibody regions, and SPR-confirmed de novo binders on four targets
Independent evidence: No independent evidence found. No independent evaluation found
Training data
Group in Figure 7: Public data + wet lab for validation only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | Public antibody structure data (2022-07 time split) and public sequences for the language model |
| b. Private or licensed data | No | |
| c. Own wet-lab data used for training | Not disclosed | Own wet-lab training data not disclosed |
| d. Wet lab used only for validation | Yes | Phage display and SPR validated four targets |
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.
Denominator basis: 10,000 Fab pairs synthesised per target, with 10 Fc-fusion antibodies finally measured by SPR; success about 1.5%–4.4% with no per-target breakdown
Verified in the published paper, lifting the earlier unverified flag; the site's old "~82 nM" should read 81.6 nM. The preprint previously cited as a tFold source actually belongs to IgGM, whose figures must not be attributed here
Background
- Organization: An in-house lab in Shenzhen funded by the company, with no separate funding
- Funding: Not disclosed or not applicable
- Products and partnerships: The online service sits on the company's drug-discovery platform; the sister system IgGM shares authors
- Related models: IgGM / IgGM-FM, RFantibody, AlphaFold 3
Reading notes
- The site once listed the IgGM preprint as a source here; the figures must be kept apart
- Only the overall 1.5%–4.4% range exists, with no per-target numbers
- No results reported on any of the 44 benchmarks
Caveats
- No standard open-source licence is declared and the weight terms are unclear
- Structure and docking evaluations are self-built time-split sets
Sources
- PubMed/PMC record (journal)
- bioRxiv preprint (preprint)
- Crossref record 10.1038/s41467-025-67361-9 (journal)
- GitHub repository TencentAI4S/tfold (repository)
- Journal article DOI 10.1038/s41467-025-67361-9 (journal)
Corrections
2 corrections
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.