Pharma’s private structures “supercharge” protein AI—the bottleneck is the data wall

In a 14 September 2026 Nature news piece, Ewen Callaway frames proprietary pharma data as the next fuel for protein AI: a system trained on more than 20,000 industry structures beats AlphaFold-class models limited to public repositories. Most of the article sits behind a paywall for non-subscribers; the same window, infrastructure firm Apheris published federated fine-tuning results with Columbia’s AlQuraishi lab and five drugmakers—the experiment Nature’s headline is pointing at.

The falsifiable claim should not stop at “secret data is magic.” Narrower and harder: on protein–ligand co-folding that drug discovery actually needs, public PDB coverage of drug-like complexes is thin; keeping private crystal structures behind firewalls while exchanging only model parameters is already enough to lift interface and pose accuracy on industry-held benchmarks above public models—without returning weights to open science.

The scientific intuition Callaway relays is not new: for drug discovery, AlphaFold-class tools need examples of how proteins interact with drug-like molecules, not another round of general folding PR. Apheris’s public numbers quantify it. PDB entries that include an approved or investigational drug sit around the low tens of thousands, with fewer added after AlphaFold 3’s training cutoff. Five companies contributed 20,167 experimentally determined private structures, ran synchronous federated averaging without sharing coordinates, and finished in under ten weeks.

The evaluation set is 1,056 partner-held-out structures. Federated model AISB-1-Fed raised high-quality interfaces (PL-lDDT ≥ 0.8) from 35.6% for public OpenFold3 Preview 2 to 52.1%, and correct ligand poses (bisyRMSD ≤ 2 Å) from 28.9% to 46.8%; public Boltz-2 sat near 40.9% and 36.5%. A same-recipe control fine-tuned only on public PDB isolates the private-data effect. Nature’s 15 September clarification fixed AbbVie’s location (North Chicago) and did not walk back the performance story.

Evidence quality needs layers. First, Nature as secondary reporting confirms the news fact—private pharma data, super-public models, named participants including AbbVie—and situates it beside Škrinjar et al. on co-folding generalization in Nature Structural & Molecular Biology. Second, the percentage points that can be checked come from the consortium operator’s technical blog, not a peer-reviewed main paper; the alliance says a journal submission is planned while only aggregate metrics are public now.

Third, the limits are also public: the test set comes from the same five firms, not a fully external blind set; cross-partner train–test ligand similarity cannot be measured without moving structures; only structure metrics are released, not affinity or binder classification; weights and data stay closed. The steelman is fair: if the goal is faster in-house virtual screening and lead optimization, closed federation can deliver internally auditable gains without solving open-source politics first.

The strongest steelman for “supercharge” admits that public co-folding already looks strong on sets like PoseBusters, while medicinal chemistry needs accuracy on drug-like ligands and industrial targets outside the training distribution—exactly where dense proprietary series live. Federated learning routes around the legal wall; it does not pretend firms will upload crystal coordinates to the public PDB.

The steelman’s hard edge is clear: without external independent benchmarks and peer review, outsiders cannot rule out optimistic bias from partner-held-out industrial distributions; privacy and attack-test claims are operator-stated and costly to reproduce; if public science only gets headline-level lifts, the gap between pharma AI and academic AI widens. Nature’s paywall itself makes Callaway’s quotes and caveats harder for non-subscribers to verify.

For R&D leadership, the near-term variables are whether the federated model enters real design loops and whether the next initiative adds affinity and activity data as the network has sketched. For open-science and policy readers, the signal is that the marginal bottleneck in drug AI has shifted from “do we have a Transformer?” to “who can lawfully mobilize industrial structural assets.” For investors, that discounts “yet another public co-folder” narratives and raises the pricing power of data governance and federation infrastructure.

For Chinese and other pharma observers, what travels is not the word “secret” but contracts, threat models, and a synchronous federation recipe—conditional on having enough private structures and compute. With only public PDB, the same training cookbook will not reproduce AISB-1-Fed’s held-out numbers.

Over the next six months, watch two trackable artifacts: whether the consortium pushes results through peer review with external or chemically broader benchmarks, and whether any weight or distilled license reaches a wider industrial or academic circle. If review and external tests hold, Nature’s “supercharge” hardens from news rhetoric into a citable methods claim; if the lift flashes only on partner-held sets, the accurate reading is that private-data federation proved the economic value of a data moat—not that the public scientific frontier of protein AI has already been rewritten.

【No.2】