The protein-folding revolution has been running on a single public resource for five years. This week a group of drug companies showed what happens when you feed the models the data nobody has been allowed to see.
What Happened
On September 14, Nature reported results from the AI Structural Biology Network, a consortium of pharmaceutical companies that has spent the past year testing a straightforward premise: that the protein structures sitting in corporate vaults are worth more to AI models than anything left in the public record. They fine-tuned OpenFold3 — an open-source reimplementation of AlphaFold 3 — on 20,167 proprietary structures showing proteins bound to candidate drug molecules. The structures came from five companies, contributed in a way that kept each firm's data private from the others.
The results held up on a held-out test set of 1,056 protein–ligand structures. The consortium's model predicted more than half of them to a high level of accuracy. The publicly available version of OpenFold3, trained on public data alone, managed the same standard on about a third. Boltz-2, a competing open-source model, landed near 40 percent. The gain is not subtle, and it did not come from a new architecture — same model, more data.
"You add all this data, and you get a pretty big bump in performance," Mohammed AlQuraishi, the Columbia University computational biologist who took part in the work, told Nature.
Two caveats travel with the result and deserve to be stated up front. The study was published as a blog post and has not been through peer review; the group says it plans to submit a paper. And the model itself is not publicly available. What has been released is the finding, not the artifact.
Why It Matters
The Protein Data Bank holds more than 200,000 experimentally determined structures and has been the foundation of nearly everything in this field. AlphaFold 2 was trained on it, and that work earned a share of the 2024 Nobel Prize in Chemistry. For predicting what a protein looks like on its own, the PDB has been remarkably sufficient.
Drug discovery asks a different question. It needs to know how a protein behaves when a specific small molecule docks into it — the co-folding problem that AlphaFold 3 and its successors were built to address. And here the public record thins out fast. Paul Mortenson, who runs computational chemistry and informatics at Astex Pharmaceuticals in Cambridge, puts the number of PDB structures involving drug-like molecules at perhaps 10,000. Out of 200,000.
That scarcity has a measurable consequence. Work published this year in Nature Structural & Molecular Biology found that co-folding models degrade sharply when asked about molecules unlike anything in their training set — which is, inconveniently, the exact situation a medicinal chemist is in when exploring new chemical space. A model that only works on chemistry someone has already solved is a search engine, not a discovery tool.
Meanwhile, the missing structures exist. They were solved by X-ray crystallography and cryo-electron microscopy inside drug-discovery programmes and never deposited anywhere, because the compounds were proprietary. Nobody knows how many there are in total. Some researchers have guessed the private holdings may exceed the PDB itself.
Reaction
The detail that drew the most comment was not the headline number but a secondary comparison buried in the results: the pooled model also beat models trained on each individual company's data alone. Five separate silos, each fine-tuned on its own structures, lost to the combination.
John Karanicolas, who leads computational drug discovery at AbbVie, has made the point that the structures absent from the PDB are precisely the ones sitting in company archives — the gap in the public data maps almost exactly onto what industry has been quietly accumulating. The pooling result turns that observation into an argument with teeth. No single firm's vault is large enough to matter on its own. Five of them are.
That reframes a competitive question as a collective one, which is an uncomfortable place for an industry organized around exclusivity. The AISB structures stayed private throughout — the companies did not hand each other their crystallography. But the finding that the value lives in the aggregate rather than in any one archive undercuts the logic of hoarding, and several of the researchers involved have been explicit that the real conclusion points toward public data rather than better-organized private data.
AlQuraishi's framing is that the result strengthens the case for building comparable open datasets, which is a notably unselfish reading of an experiment his collaborators funded.
What's Next
There is already a project attempting exactly that. OpenBind, backed by up to £8 million in UK government funding, is generating protein–ligand structures for public release; it put out several hundred last month with thousands more in progress. Set against 20,167 proprietary structures it looks modest, but it is growing and, unlike the AISB model, anyone can use what it produces.
The near-term questions are practical. Peer review will test whether the benchmark comparisons are as clean as the blog post suggests, and held-out test sets drawn from the same five companies' pipelines may flatter a model more than truly independent chemistry would. Whether the consortium ever releases weights is a separate and probably harder question — the structures stayed private, but a model fine-tuned on them carries something of them forward, and no one has settled what that implies for compounds still in development.
The more consequential fork is who ends up owning the improvement. If the best co-folding models are the ones trained on data that cannot be published, academic groups get permanently second-tier tools in a field where the public-data tradition produced AlphaFold in the first place. If OpenBind and projects like it reach useful scale, the gap closes and everyone works from the same floor.
Closing Thoughts
What makes this study interesting is how boring the method is. No architectural innovation, no new training objective, no scaling milestone. Somebody took an existing open-source model, showed it structures it had never seen, and it got substantially better at the thing it was already trying to do. The bottleneck was never the model.
That is a pattern worth noticing, because it keeps recurring across applied machine learning and it keeps being less exciting than the alternative explanation. The field's attention flows toward architectures and parameter counts, which are legible and comparable and produce good charts. Data provenance is tedious by comparison, tangled up in legal agreements and institutional incentives, and it appears to be where a great deal of the remaining headroom actually sits.
There is also a quieter observation here about what the PDB was. It is a sixty-year habit of depositing your structures where strangers can use them — a norm, not a technology. It happened to produce the training corpus for one of the more significant scientific tools of the decade, which nobody planned. Whether anything comparable gets built for protein–ligand data is a question about institutions and funding, not about machine learning at all.
한글 요약
9월 14일 네이처 보도에 따르면, 제약회사들이 모인 AI 구조생물학 네트워크(AISB)가 자사 금고에 잠겨 있던 단백질 구조 데이터로 단백질 접힘 AI를 훈련시킨 결과를 공개했습니다. 알파폴드3의 오픈소스 재현판인 오픈폴드3에 다섯 개 회사가 제공한 20,167개의 비공개 구조를 추가 학습시켰고, 각 회사의 데이터는 서로에게 비공개로 유지됐습니다. 따로 떼어둔 1,056개 단백질-리간드 구조로 시험했을 때 절반 이상을 높은 정확도로 예측했습니다. 공개 데이터만 쓴 오픈폴드3는 3분의 1, 경쟁 모델 볼츠-2는 약 40% 수준이었습니다. 다만 이 연구는 블로그 글 형태로 공개돼 아직 동료 심사를 거치지 않았고, 모델도 공개되지 않았습니다.
왜 이런 차이가 나는지는 데이터 구성을 보면 분명합니다. 공개 저장소인 단백질 데이터 뱅크(PDB)에는 20만 개가 넘는 실험 구조가 있지만, 약물 후보 분자와 결합한 상태의 구조는 1만 개 정도에 불과하다고 아스텍스의 폴 모텐슨은 말합니다. 신약 개발에 정작 필요한 것은 이 결합 상태 데이터인데, 올해 발표된 연구는 이런 모델들이 학습 데이터와 크게 다른 분자를 만나면 정확도가 급격히 떨어진다는 점을 보여줬습니다. 제약사들이 X선 결정학과 극저온 전자현미경으로 만들어낸 구조는 대부분 공개된 적이 없고, 그 총량이 PDB보다 클 수 있다는 추정도 나옵니다.
가장 주목받은 대목은 따로 있습니다. 다섯 회사의 데이터를 합친 모델이 각 회사가 자기 데이터만으로 만든 모델보다 나았다는 점입니다. 어느 한 곳의 금고도 혼자서는 충분하지 않다는 뜻이어서, 데이터를 쥐고 있는 것보다 모으는 쪽이 낫다는 결론으로 이어집니다. 컬럼비아대 모하메드 알쿠라이시는 이 결과가 공개 데이터셋 구축의 근거를 강화한다고 봤습니다. 영국 정부가 최대 800만 파운드를 지원하는 오픈바인드(OpenBind)가 지난달 수백 개 구조를 공개했고 수천 개를 준비 중입니다. 새로운 모델 구조도, 더 큰 파라미터도 아닌 데이터가 병목이었다는 것이 이 연구의 핵심입니다.
참고: Nature – Ewen Callaway, Nature – AlphaFold is running out of data