Open Protein Structure Prediction Models Now Rival AlphaFold 3

Open models catch up with AlphaFold

For most of the past two years, the best tools for predicting the 3D shape of biological molecules came with conditions. AlphaFold 3 set the benchmark, but it restricts commercial use. Open alternatives existed, but they were generally a step behind. Releases throughout 2026 have largely removed that tradeoff. Several fully open models now report accuracy that matches or beats AlphaFold 3, and each has been released under licenses that let companies build products on it.

ByteDance moved first. In February its Seed team published Protenix-v1, which it describes as the first fully open-source structure predictor to outperform AlphaFold 3 under the same training-data cutoff, model size and inference budget.11 The release covers both the code and the model weights under the Apache 2.0 license. The model predicts all-atom structures for complexes that combine proteins, DNA, RNA and small-molecule ligands.13 The team also released a second version trained on data through June 2025, aimed specifically at drug discovery.11

In late May, the Chan Zuckerberg Initiative's Biohub went further. It released ESMC, ESMFold2 and a new ESM Atlas as what it calls an open "world model" of protein biology.14 The atlas contains more than one billion predicted structures. That is roughly 800 million more entries than the AlphaFold Database.12 Biohub says ESMFold2 outperforms AlphaFold 3 at predicting how antibodies bind their antigens, which is one of the most commercially important problems in the field.14

The trend continued after that. In August, OpenBind released OpenBind-0, a model built on OpenFold3 and tuned for proteins bound to small-molecule ligands. Its training data, code, weights and training recipes are all public under Apache 2.0.18 In September, a Korean academic group released AtlasFold under the MIT License. It reports antibody-antigen performance comparable to AlphaFold 3 and ESMFold2.16 Packaging has also changed: OpenFold3 is now offered as a ready-to-deploy microservice by both NVIDIA and AMD.20

What "accessible to all" really means

What matters here is less any single benchmark than the licensing. Coverage of ESMFold2 stresses that it is fully open-source with no commercial restrictions. Observers say this could make it much easier for both universities and companies to adopt.15 OpenFold3's developers made the same point by contrast: AlphaFold 3 limits commercial use, while OpenFold3 uses the permissive Apache 2.0 license.20

In practice, a small biotech can now run, fine-tune and ship a structure predictor at roughly AlphaFold 3's level without negotiating access. OpenBind says users can retrain its model on mixtures that include their own proprietary data. It notes this matters because reliable fine-tuning usually requires keeping examples from the original training set.18 A closed model cannot offer that kind of customization.

The benchmark claims deserve caution, because they come from the developers themselves. They also do not fully agree. Biohub's own comparison puts ESMFold2 ahead of both AlphaFold 3 and Protenix-v1 on protein-protein and antibody-antigen docking.14 The Protenix paper says AlphaFold 3 still leads on protein-ligand and protein-DNA docking, even as Protenix-v1 wins elsewhere.16 The ByteDance team also built PXMeter, a toolkit that tests several models on the same tasks, and its results show that how a benchmark dataset is designed can change which model comes out on top.13

My reading is that "beats AlphaFold 3" is now a claim about specific tasks, not the whole field. The more useful conclusion is that the open models are now close enough to the best closed one that the remaining gap matters less than having access at all.

Independent testing supports that view, with limits. A June study in Frontiers compared open models for screening nanobody binders. Measures from Boltz-2 performed best at identifying true binders, but the authors found that current predictors do not reliably generalize to antigens they have not seen before.19 Better structure prediction is not the same thing as reliable binding prediction.

From prediction to design

The newest open releases are being used to design proteins, not just predict their shapes. Biohub reports using ESMFold2 to design antibodies and binders against targets linked to cancer and immune conditions. In lab tests, a high share of those designs worked as the model predicted.15 Biohub describes some of these designs as binders with therapeutic-level affinity.14 Protenix-v1 was also tested as a filter for mini-binder designs and achieved the best ranking scores on most targets.16

Gene editing shows where this could lead. Profluent's OpenCRISPR-1 was generated by protein language models trained on CRISPR-Cas sequences. It was the first AI-designed editor to precisely edit the human genome, and the work was published in Nature in 2025.2 Profluent's models generated millions of candidate sequences. The team used AlphaFold2 to predict structures for 5,000 of them, then synthesized and tested 209 in human cells.1 That workflow already uses structure prediction as a filter between generating a design and testing it in the lab.

The results were strong. OpenCRISPR-1 matched the widely used SpCas9 enzyme on on-target editing, with median indel rates of 55.7% versus 48.3%. It cut off-target editing by about 95%.1 It also worked as a base editor, changing single DNA letters at 35% to 60% efficiency across three target sites.2 Its off-target sites were a subset of those seen with SpCas9, which suggests it does not create new cutting patterns.2

Some coverage gets the details wrong. A July 2026 article cites an 85% reduction in off-target activity and less than 40% sequence identity to natural proteins. It also dates the Nature paper to July 16, 2026.7 The paper itself appeared in July 2025.2 Earlier reporting put the off-target reduction at roughly 95%, with 403 mutations relative to SpCas9 and 182 relative to the nearest natural protein.1 The paper also notes that combining three natural Cas9s reproduces about 98.3% of OpenCRISPR-1's sequence.2 Readers should treat "fully synthetic" as a description of how it was made, not proof that it is entirely unlike natural proteins.

OpenCRISPR shows both sides of the openness story. Its sequence is free to license for research and commercial use.3 The license excludes some uses, including human germline editing.9 But Profluent has kept the AI model that created the editor in-house.5 In other words, the company shared the product, not the system that designs it.

Drug discovery: the walls that remain

The pull between open and closed approaches is clearest in drug discovery. Isomorphic Labs, the DeepMind spinout, released a technical report in February on its proprietary Drug Design Engine, IsoDDE.21 The company says it roughly doubles AlphaFold 3's accuracy on the hardest ligand-binding cases: about 50% versus 23.3% for targets with little similarity to the training data.24 Isomorphic is aiming for first-in-human trials by the end of 2026, a year later than it originally predicted.21 As of October, the company's most advanced program was still described as expecting trials by year-end.27

This creates an uneven picture. Basic structure prediction is quickly becoming a free commodity, while the leading firms are keeping their advantage in proprietary drug-design engines built on top of it. Open models like OpenBind-0 are trying to close that gap for small-molecule work and plan updates within about six months of the newest public structural data.18

The main limit is downstream, and here the coverage disagrees most. Counts of AI-derived drugs in the clinic range from fewer than 12 AlphaFold-derived candidates25 to at least 23 compounds where structure prediction played a significant role28 to more than 200 AI-discovered drugs in development.29 The differences mostly come from definitions, since there is no standard way to disclose AI's role in a drug's discovery.28 The more sober analyses agree on one point. Average timelines from target identification to filing an investigational new drug application have not changed much. Time saved in early discovery has gone into experimental validation, and AI does not speed up toxicity and drug-behavior testing, trial execution or regulatory review.28 One analysis also notes that no AI-discovered drug has yet won FDA approval.30

The bottom line

Open access to accurate structure prediction is now a reality in 2026. It is backed by permissive licenses, released weights and support from chipmakers. Its biggest effect is likely to be wider participation: academic labs, startups and researchers outside the large pharmaceutical companies can now work with models at the level of the best closed system. The coverage does not support the larger claim that this will directly speed up getting drugs to patients. Lab testing, clinical trials and regulatory review remain the slowest steps. For now, the open releases give more researchers the tools to start; they do not shorten the process that follows.