OpenBind opens the binding-data bottleneck — and why general LLMs fail at drug discovery
The UK's OpenBind initiative takes aim at the structural data gap that's quietly limited AI drug discovery since AlphaFold. Also: the FDA's AI-triage inspection pilot, Novo's enterprise OpenAI bet, and Insilico's empirical case against general-purpose models.
The gap between AI ambition and the data that grounds it runs through every story this week. The UK's OpenBind initiative is trying to build a public binding-data resource that could do for drug-ligand interactions what the Protein Data Bank did for structure; Insilico is publishing evidence that general foundation models simply fail at drug discovery without domain data; and Novo Nordisk is betting an enterprise-wide OpenAI partnership can pull it out of a bad year. Meanwhile the FDA keeps threading AI into how it inspects factories, qualifies tools, and phases out animal testing. The common thread: whoever controls the right data — and the regulatory pathway around it — sets the terms.
Lead — OpenBind releases a public binding-data resource
InfrastructureThe UK's OpenBind initiative released 800 protein-drug binding measurements and a free predictive model (OpenBind v1), positioning it as proof that an automated chemistry plus high-throughput X-ray crystallography pipeline works at industrial scale. Led by Diamond Light Source and funded by the Department for Science, Innovation and Technology, it chains synthesis, binding assays, crystallography, and structure determination into one continuous, AI-ready data factory. The bet is explicit: just as the Protein Data Bank enabled AlphaFold, OpenBind aims to become the equivalent open resource for drug-ligand binding rather than leaving that data proprietary. It hit 800 measurements in seven months — previously years of work — with future tranches targeting COVID-19, malaria, and dengue. The strategic implication is sharper than the current dataset size: companies whose moat was simply owning binding data now face the disruption proprietary mapping firms met when Google Maps arrived, while those with genuine architectural advantages keep them.
The FDA sends AI-picked inspectors for one-day factory screens
RegulatoryThe FDA launched a pilot using AI to flag low-risk drug manufacturing facilities for abbreviated, one-day screening assessments, announced May 6 with 46 visits already completed. The model scores facilities on product type, inspection history, and operational characteristics; when inspectors find significant issues they escalate to full multi-day inspections, though most assessments so far returned "No Action Indicated." The pilot spans biologics, medical products, human and animal foods, and clinical research facilities through September 30, 2026. For quality teams, this makes AI triage a new variable in inspection probability and intensity — and because the risk model learns from every visit, clean records and well-documented systems increasingly translate into a lighter inspection burden. The FDA frames it as a "force multiplier" to stretch strained inspection capacity across more facilities per year.
Novo Nordisk puts OpenAI at the center of its recovery
DealsNovo Nordisk announced an enterprise-wide partnership with OpenAI spanning drug discovery, manufacturing, and commercial operations — billed as the broadest-scope pharma-AI deal yet, covering the full value chain rather than a single function. CEO Mike Doustdar framed it as positioning Novo to lead the next era of healthcare; financial terms were undisclosed. The timing reads as strategic signaling during a rough stretch: Novo's stock fell 40% in 2025, it replaced its CEO, and it faces intense oral GLP-1 competition from Lilly's orforglipron. Enterprise AI typically delivers early wins in administrative functions while R&D evidence takes years, so the partnership's credibility will hinge on measurable R&D acceleration in forthcoming quarterly updates, not press releases.
Insilico argues general models don't work for drug discovery
PapersInsilico Medicine's MMAI (Multi-task AI) Gym for Science — trained on 120 billion tokens of drug discovery data across 1,000+ benchmarks — showed up to 10x gains over leading general-purpose foundation models, which failed on 75–95% of challenges. The resulting compact model, LFM2-2.6B-MMAI v0.2.1, reached state-of-the-art performance on multiple tasks despite its small size, and the methodology paper was accepted at ICLR 2026. MMAI applies supervised fine-tuning and reinforcement learning with domain-specific reward signals to adapt general models to drug discovery. The empirical takeaway is that "just use GPT-4" is not a strategy; advantage sits in domain fine-tuning, proprietary benchmarks, and closed-loop improvement. Notably, Insilico is positioning itself as a model-builder rather than only a drug developer — a distinction that shapes licensing value and how pharma should evaluate AI vendors.
CellCentric raises $220M for an oral epigenetic myeloma drug
FundingCellCentric closed a $220 million Series D to advance inobrodib, an oral, first-in-class p300/CBP inhibitor, in multiple myeloma, with Pfizer joining as co-investor alongside Venrock, RA Capital, HBM Healthcare, and Sofinnova. The Cambridge/Boston biotech targets the p300/CBP epigenetic complex — a mechanism entirely distinct from myeloma standards like proteasome inhibitors, IMiDs, CD38 antibodies, CAR-T, and bispecifics — and Pfizer's participation signals acquisition optionality. In one of oncology's most crowded indications, a novel oral mechanism drawing a $220M round with Big Pharma at the table signals both investor conviction in differentiated biology and appetite for oral options in a largely injectable field. For AI drug discovery firms working in oncology, CellCentric's trajectory sets a valuation benchmark for genuinely novel mechanisms at clinical stage, regardless of how they were discovered.
Boehringer's survodutide posts Wegovy-like Phase III weight loss
PapersBoehringer Ingelheim's survodutide, a GLP-1/glucagon dual agonist, delivered 16.6% weight loss versus 3.2% for placebo in Phase III, with analysts calling the results "Wegovy-like" and noting hints of a muscle-preservation benefit. Landing at competitive efficacy, it pushes the obesity arms race beyond the current two-player dynamic and beyond pure GLP-1 mechanisms.
FDA moves to phase out preclinical animal testing
RegulatoryThe FDA announced a formal push to end animal testing in preclinical development, endorsing AI models and organ-on-chip technologies as replacements. Experts quickly flagged the gap between announcement and implementation, cautioning that celebrating progress on a decade-long timeline is not the same as progress today.
Benchling report shows the AI adoption gap widening
Industry dataBenchling's 2026 Biotech AI Report found that of the 50% of biotech firms actively using AI in 2025, those users reported faster time-to-target, and 42% saw improved hit rates. The divergence between AI-fluent and AI-aspiring organizations is widening faster than most leadership teams realize, with the other half yet to document measurable benefit.
FDA qualifies its first AI drug development tool
RegulatoryThe FDA completed qualification of its first AI drug development tool under the 21st Century Cures Act, establishing a formal regulatory pathway for AI tools used in drug development decisions. The move draws a clear line between internal AI use and regulatory acceptance of AI-informed submissions within defined evidentiary standards.
The Onion Desk
Clinical AI company Aidoc raised a $150 million Series E led by Goldman Sachs to flag concerning CT findings "approximately four minutes before the radiologist would have noticed them," touting 97% sensitivity across 11 acute indications and 110 million cases across 2,000 hospitals — as clinicians describe the alerts as "timely" and "I had already put in the order." Elsewhere, a PitchBook report thrilled investors with 80–90% Phase I success rates for AI-native biotechs versus a 40–65% industry average — a "paradigm-shifting" finding built on all of ten datapoints, prompting conference organizers to book seventeen AI drug discovery panels for Q3. And the FDA issued a formal Request for Information asking the public to explain how to actually run faster clinical trials, clarifying that the agency "has the vision and the press release but would find additional perspectives helpful"; Commissioner Makary cited proof-of-concept trials with AstraZeneca and Amgen while comments stay open until May 29.