中文

Direction analysis

Anew Labs

ByteDance's AI pharmaceutical team, spun out in 2026. Known publicly for Protenix, an open-source AlphaFold3 reproduction; internally it wires structure prediction, generative design, affinity and R&D reasoning into one closed-loop engine, with IL-17 and IL-4R in the pipeline.

Last verified AI4S.fyi's own reading of the public record; its sources are listed at the end

← Anew Labs organization page

The claim

Wiring structure prediction, design, affinity and decision reasoning into one engine driven by experimental feedback can substantially cut the experimental search a candidate takes.

What has to hold for this route to work.

The real bottleneck. The team treats the handoffs between two experiments as the constraint: who fixes the epitope, who cuts thousands of candidates to what a bench can absorb, which residues change once measurements return. Those steps are written into one workflow and handed to an agent dispatched in conversation.

What is established. Open-source structure prediction (three preprints, downloadable weights, externally runnable) and one antibody design loop that ran twice with measured SPR data.

What is still a claim. That a closed-loop engine accelerates drug discovery: one undisclosed target, one modality, with functional and developability evaluation listed as next steps and no IND or trial registration.

The chosen product abstraction. Outward, an in-house pipeline; inward, a research interface driven in natural language, closer to a research assistant for scientists than to a fully automated lab.

How the loop runs

Operating sequenceTarget and epitope selectionGenerative designStructure and affinity evaluationWet-lab measurement (SPR)Feedback-driven optimization[S1][S2]

The four components of AnewDDE [S1][S2]
Structure: AnewFoldAntibody-antigen complexes, protein-ligand interactions, blind pocket identification and molecular-glue ternary complexes
Quantitative: AnewAffinityBinding-affinity prediction approaching FEP accuracy, with a reliability estimate attached to each prediction
Design: AnewDesignDiffusion-based generation, liability filtering and clustering, and affinity maturation guided by experimental feedback
Decision: AnewMindA hundred-billion-parameter model for ADMET and developability analysis, candidate ranking and long-horizon R&D reasoning
Interface: AnewScienceThe scientist states the objective in natural language; the agent dispatches the four layers and asks for confirmation at each fork

Current public boundary. Structure prediction, candidate generation and binding measurement; functional activity, developability and any clinical milestone are not yet public.

  • Experiments are run by people; the agent turns decisions into runnable compute.
  • The public campaign covers one undisclosed target and one modality.
  • Of the components, only the earlier Protenix series is externally runnable.

What has been demonstrated

ResultEvidence typeSources
Three structure-prediction preprints authored by ByteDance Seed, with downloadable weightsPublic artifact[S5][S6][S7]
A nanobody campaign: 16 of 150 clones at 1.8-8.1 nM by SPRCompany-reported[S1][S2]
AnewFold's Top-1 success rates on three antibody-antigen collections and on protein-ligand complexesCompany-reported[S1]
AnewMind's ADMET evaluations and the internal PharmBench benchmark (200 subquestions, 17 models)Company-reported[S1][S3]
Invited podium talks on orally bioavailable pan-IL-17 small molecules at AAI and FOCISCompany-reported[S8]
A $290M spin-off round at a $1.5B valuationIndependent third party[S9]

Where the evidence stops

StepStatus
Open, externally runnable structure prediction●
Generative design producing measurable candidates●
Experimental feedback entering the next design round◐
Functional activity and developability data○
Independent third-party replication○
An IND filing or clinical milestone○

Where the money comes from

Who pays. An in-house pipeline first, with the platform partly open and partly internal; compute supplied by Volcano Engine

For what. Structure prediction, candidate generation and filtering, affinity prediction, R&D decision reasoning

What becomes a durable asset. The team sits in an unusual spot: public standing in the structure-prediction community through Protenix, and a pipeline of its own. Since the spin-off, what it publishes has been shifting from model weights to technical reports.

What can currently be confirmed: three structure-prediction preprints authored by ByteDance Seed with open weights (Protenix, v1, v2); one technical report wiring four models into a closed-loop engine (AnewDDE); one nanobody campaign against an undisclosed internal target with measured SPR data; a set of IL-17 small-molecule results presented in invited conference talks; and a $290M spin-off round. The record has an unusual shape: the earliest evidence is the most externally checkable, and the newest, fullest evidence is internal and readable only through a technical report.

Scope: This page tracks the public record only: the company's own releases and technical reports, preprints and conference information, and third-party reporting. No inside information. It records how far the public material goes, not how good the company is.

Overview

DirectionWhat can be confirmedLatest evidence
1. Structure predictionThree Protenix preprints, open weights, externally runnablePreprint, 2026-04-11
2. Antibody design16 of 150 clones at 1.8-8.1 nM by SPRCompany research page, 2026-09-15
3. Affinity and samplingAnewAffinity ranks better than its comparator on 199 ligandsTechnical report, 2026-09-17
4. The decision layerA 100B-parameter model; PharmBench holds 200 subquestionsCompany research page, 2026-09-15
5. PipelineIL-17, IL-4R and two undisclosed targets; IL-17 has conference dataConference, 2026-06-09
6. Organization and capital$290M at $1.5B, ByteDance keeping 56%Reuters reporting, 2026-09-16

1. Structure prediction: the half that is public

What can be confirmed: In January 2025, ByteDance released Protenix, a full reproduction of AlphaFold3 covering protein-ligand, protein-protein and protein-nucleic-acid prediction, with open weights and code. Protenix-v1, in February 2026, describes itself as the first fully open-source structure-prediction model to beat AlphaFold3 while holding the same training-data cutoff, model size and inference budget, and shows consistent gains from increased sampling at inference time. Protenix-v2, in April 2026, reports antibody-antigen success rates up to 13 points above v1, with five seeds beating the previous thousand-seed results, and extends the system into biomolecular design.

All three are authored by ByteDance Seed, readable on bioRxiv, with downloadable weights. This is the most externally checkable stretch of the whole record.

Its relationship to AnewFold needs stating. The structure component inside AnewDDE is called AnewFold, and the report gives its antibody-antigen Top-1 success rates as 62.3% on AF3-AB, 60.8% on PXMeter-AB and 76.2% on FoldBench-AB, plus 77.4% on FoldBench protein-ligand, with Protenix-v2 and AlphaFold3 among the comparators. In other words, one of the things it is measured against is the team's own earlier open model, and AnewFold itself is not publicly runnable today.

Boundary: Structure and interface prediction. Confidence: high for the Protenix series (preprints, open weights, outside users); medium for AnewFold's margin (self-evaluated, model not public).

Open threads: whether AnewFold is released or submitted for review as Protenix was; how much of the margin survives when a third party rebuilds the antibody-antigen comparison on sets of its own choosing.

2. Antibody design: a campaign that ran twice

What can be confirmed: AnewDesign ran a nanobody design campaign against an undisclosed internal protein target. It starts from published VHH binders and a public crystal structure: the agent searches literature and structures, proposes candidate epitopes, then launches diffusion-based generation with structure-prediction-based evaluation. Filtering removes CDR motifs carrying deamidation, fragmentation, isomerization and N-linked glycosylation liabilities, then clusters what remains to keep diversity.

Fifty clones went to the bench in the first round; SPR found 7 binders between 46 and 390 nM, a 14% hit rate. Those measurements and developability labels became reference data for targeted optimization, with a partial-diffusion algorithm holding validated regions fixed. A hundred sequences in the second round returned 16 candidates at 1.8-8.1 nM. Across both rounds: 150 clones, 16 at single-digit nanomolar, 10.7%.

The second round is the part worth recording. A first-round hit rate like that is not unusual in antibody discovery; the dividing line is whether measured data can be turned into targeted improvement. Affinity improved by roughly an order of magnitude between rounds, which means the measurements re-entered the design. The positive control in the same figure sits at 1.7 nM, putting the optimized candidates in the range of an existing reference molecule.

Boundary: Binding. The company lists functional and developability evaluation as next steps. Confidence: medium (measured data, but an undisclosed target and company reporting).

Open threads: a campaign on a second target; function and developability data for these clones; disclosure of target difficulty, since a target with a solved structure and published binders to start from is not the same as a hard one.

3. Affinity and sampling: the line aimed at FEP

What can be confirmed: AnewAffinity aims for accuracy approaching free-energy perturbation (FEP) at a speed that suits large-scale screening, with an estimate of each prediction's own reliability attached. The technical report gives a comparison on 199 ligands: a pairwise ΔΔG RMSE of 1.057 kcal/mol, a weighted Pearson R² of 0.553 and a Spearman of 0.724, against 1.016, 0.486 and 0.620 for Boltz-2.

The two metrics point different ways. On absolute error Boltz-2 is slightly better; on rank correlation AnewAffinity is clearly better. In a real project, ordering is usually what decides which compound gets made first, so the honest reading of these numbers is "better for prioritization" rather than "more accurate".

The same line has two earlier public pieces (AnewSampling, a dynamics-aware sampling engine, and AnewOmni, an all-atom generative model, both from March 2026) and an invited talk at the Alchemistry free-energy workshop in Barcelona in May 2026, titled Learning the Thermodynamic Landscape of Protein-Ligand Interactions at Scale.

Boundary: Computational prediction. Confidence: medium (internal benchmark against public comparators, run by the team).

Open threads: AnewAffinity against real FEP on the same set of prospective projects; whether the workshop talk becomes a paper.

4. The decision layer: AnewMind and PharmBench

What can be confirmed: AnewMind is an internal model at the hundred-billion-parameter scale, post-trained on all parameters, that handles ADMET and developability analysis, relative ranking of candidates and long-horizon R&D reasoning. The company reports it in the top five across all four aggregate ADMET and developability evaluations, first on cyclic-peptide permeability and near the lead on antibody developability, while holding up on general benchmarks such as MMLU-Pro and GPQA Diamond.

To evaluate longer-horizon decisions, pharmaceutical R&D experts built an internal benchmark, PharmBench, around real cases: project context and experimental evidence are revealed progressively to see whether a model can integrate them into an R&D judgment. It holds 200 subquestions across seven capability dimensions and compares 17 models.

The value and the limit of a benchmark like this are the same fact. It measures judgment in realistic R&D situations, which is exactly what public benchmarks lack, and it was designed and run by one company, and is not public, so no one outside can recompute it.

Boundary: Analysis and ranking recommendations. Confidence: medium-low (internal benchmark, not public).

Open threads: whether PharmBench is released; how often AnewMind's judgment actually changed a project decision rather than being recorded alongside it.

5. Pipeline: four targets, two of them undisclosed

What can be confirmed: The company's site lists four pipeline entries (IL-17AA/AF/FF, IL-4R and two undisclosed targets) across stages from Exploratory to IND Enabling.

IL-17 is the one with public detail. At the American Association of Immunologists (AAI) annual meeting in April 2026, the team gave an invited podium talk on orally bioavailable pan-IL-17 small-molecule inhibitors designed by its generative platform, described as the first reported molecules to engage the IL-17F/F dimer strongly. In vivo data followed at FOCIS in June, which the company describes as biologics-like activity.

Where a conference talk sits as evidence. Invited podium talks are selected by a meeting's scientific committee, which is not the same as self-publication, and also not the same as a peer-reviewed paper. No compound structure, dose, animal-model detail or trial registration is public.

Boundary: Preclinical. Confidence: medium (committee-reviewed talks plus company description); low for anything clinical.

Open threads: a paper or patent disclosure on the IL-17 program; a first IND filing; any primary information on IL-4R or the two undisclosed targets.

6. Organization and capital: spun out of ByteDance

What can be confirmed: Reuters reported that in September 2026 Anew Labs closed $290M at a $1.5B valuation, with ByteDance retaining 56%. HSG, IDG Capital and Hillhouse led, with 5Y Capital co-leading and Gaorong, Primavera, Boyu and the state-backed Shanghai Future Industries Fund participating; SBP Group came in as a strategic investor. Roughly 50 core staff moved to the new company, and compute is still supplied by ByteDance's Volcano Engine. The team is led by Kai Liu, with offices in Shanghai, Singapore and San Jose.

The spin-off changed the shape of the record. The team's first public identity was as ByteDance Seed structure-prediction researchers whose output was open models. Since the split, the newest work arrives as a technical report, with AnewFold and AnewMind staying internal. That is not unique to this company, but it makes the question of which parts remain externally checkable unusually clear here.

Boundary: Capital and organization. Confidence: high (multiple outlets citing Reuters, details consistent), but this is evidence about capital, not about molecules.

What is not there

  • No peer-reviewed paper for AnewDDE. The September document is a company technical report.
  • AnewFold, AnewAffinity and AnewMind are not publicly runnable. What outsiders can run is the earlier Protenix series.
  • The antibody campaign's target is undisclosed, so task difficulty cannot be judged from outside.
  • No functional or developability data; the report lists both as next evaluation steps.
  • No IND, trial registration or named drug candidate.
  • PharmBench is not public, and neither is the list or configuration of the 17 compared models.

Method and scope

Material comes from the company's site and research pages, the AnewDDE technical report, Crossref metadata and abstracts for the three Protenix preprints, and Reuters reporting relayed by The Next Web. Every number quoted above was read at its source; the technical report was parsed and checked against the research pages.

Company-reported performance is labelled as such throughout. The distinction this page tries to hold is between what outsiders can run (the Protenix series), what the company measured itself (the AnewFold, AnewAffinity and AnewMind benchmarks), what has measured experimental data behind it (the SPR results in the antibody campaign), and what currently has only a name (the two undisclosed targets).

Primary sources

Evidence timeline

DateEventNew evidenceWhat it does not proveSources
The AnewDDE technical reportFour components wired into one closed-loop engine: AnewFold for structure, AnewAffinity for affinity, AnewDesign for generation and AnewMind for R&D reasoning, with benchmark results on antibody-antigen interfaces and protein-ligand complexes.The benchmarks were run by the team, and neither AnewFold nor AnewMind is publicly runnable today.TECH REPORT · COMPANY-REPORTED ↗
AnewDesign: a nanobody campaign with measured dataAgainst one undisclosed internal protein target: 7 binders at 46-390 nM from 50 clones in the first round, then 16 clones at 1.8-8.1 nM after feedback-guided optimization: 150 clones and 10.7% across both.SPR measures binding; functional activity, developability and the difficulty of the target remain undisclosed.RESEARCH PAGE · COMPANY-REPORTED ↗
Spin-off and a $290M roundReuters reports a $1.5B valuation with ByteDance keeping 56%, led by HSG, IDG Capital and Hillhouse, roughly 50 core staff moving across, and compute still supplied by Volcano Engine.Can the loop repeat on other targets and other modalities, and carry a candidate to IND?REPORTING · REUTERS ↗
The Protenix-v2 preprintAntibody-antigen success rates gain up to 13 points over v1, five seeds beat the previous thousand-seed results, and the system extends into biomolecular design.Can the loop repeat on other targets and other modalities, and carry a candidate to IND?PREPRINT · PUBLIC ARTIFACT ↗
First oral pan-IL-17 small molecule presented at AAIThe company says these are the first small molecules to engage the IL-17F/F dimer strongly, with in vivo data added at FOCIS in June; both were invited podium talks.A conference talk is not a peer-reviewed paper, and no compound structure, dose or trial registration is public.CONFERENCE · COMPANY-REPORTED ↗
The Protenix-v1 preprintDescribed as the first fully open-source structure-prediction model to beat AlphaFold3 under the same data cutoff, model size and inference budget, with consistent gains from inference-time scaling.Can the loop repeat on other targets and other modalities, and carry a candidate to IND?PREPRINT · PUBLIC ARTIFACT ↗
Protenix: an open-source AlphaFold3 reproductionByteDance published and open-sourced a full AlphaFold3 reproduction covering protein-ligand, protein-protein and protein-nucleic-acid prediction. It is where the team's standing in the community starts.Can the loop repeat on other targets and other modalities, and carry a candidate to IND?PREPRINT · PUBLIC ARTIFACT ↗

On evidence. A CEO claim is not a benchmark; a benchmark is not a physical experiment; a physical experiment is not manufacturing scale; a third-party test is not qualification; qualification is not deployment. Every line keeps its source and says what it supports and what it cannot show.

Sources

  1. AnewDDE: An Agentic Drug Discovery Engine (technical report) · Company technology report
  2. AnewDesign: An Integrated Workflow for Antibody Design and Optimization · Company research page
  3. AnewMind: A Scientific Reasoning LLM for Drug Discovery Decisions · Company research page
  4. Anew Labs: platform, pipelines and teams · Company
  5. Protenix-v2: Broadening the Reach of Structure Prediction and Biomolecular Design · Preprint
  6. Protenix-v1: Toward High-Accuracy Open-Source Biomolecular Structure Prediction · Preprint
  7. Protenix - Advancing Structure Prediction Through a Comprehensive AlphaFold3 Reproduction · Preprint
  8. 2026 AAI Oral Presentation: First-in-Class Small-Molecule Pan-IL-17 Inhibitor · Conference
  9. ByteDance's AI drug unit Anew Labs raises $290M at $1.5B valuation · Third party reporting

← Anew Labs organization page