Drug Discovery Has A Data Problem

The easiest way to tell the Anthropic drug-discovery story is to reach for DeepMind.

John Jumper, one of AlphaFold’s creators and a 2024 Nobel laureate, is leaving Google DeepMind for Anthropic, and just a few days ago, Claude Science launched. Anthropic has been moving deeper into life sciences and healthcare. If you want a clear direction from them, there it is. The frontier AI lab hires the AlphaFold guy, wraps Claude around scientific workflows, and starts walking toward drug discovery.

It is a very good story, but it is also missing the part that made AlphaFold possible.

AlphaFold was not some mystic magic sprayed over biology. It was a brilliant model derived from a rare, unusually clean, unusually public scientific problem. Protein structures have been deposited for decades into the Protein Data Bank, a public archive established in 1971. The National Science Foundation put it plainly after the 2024 Nobel: the PDB provided the training library for AlphaFold [6].

I want you to remember that because this is exactly what most drug discovery initiatives currently lack.

There is no Protein Data Bank for whether a new drug works in humans. There is no single, canonical, decades-deep public archive for target validity, toxicity, ADMET, dosing, off-target effects, human efficacy, and all the messy reasons a beautiful molecule fails.

Those data exist, but they are scattered, unstructured, expensive to generate, and often locked inside companies that paid dearly to make them.

That is the asymmetry everyone keeps trying to talk around. Models are increasingly abundant, but the raw experimental data are still scarce.

Don’t get me wrong here, Claude Science looks to be an incredible tool. Anthropic describes it as an AI workbench that brings databases, code, compute, figures, manuscripts, and scientific workflows into one environment. It includes more than 60 curated skills and connectors, including tools for genomics, single-cell analysis, proteomics, structural biology, and cheminformatics.

That is not nothing. For a scientist drowning in file formats, broken pipelines, and cross-database grunt work, it may be genuinely valuable.

But it is also not the DeepMind move. TechCrunch reported the key line from Anthropic directly: Claude Science is not a new AI model and not a more capable model for biology.

It runs the same Claude models already available today. The same piece noted the strategic gap. Google DeepMind owns foundational science models like AlphaFold and AlphaGenome; Anthropic can call them as tools.

So the question is not whether Anthropic has hired brilliant people; it obviously has. The question is whether more model brilliance solves the bottleneck that drug discovery actually faces.

DeepMind’s own drug arm is a great example of that reality check. Isomorphic Labs was founded as a DeepMind spinout and built on the AlphaFold lineage. Yet in January 2026, Demis Hassabis said Isomorphic now expects its first clinical trials by the end of 2026, after previously aiming for the end of 2025.

That delay is not really that surprising; this is drug discovery after all, it would be strange if timelines were actually met.

Once you leave structure prediction, the clean public-data regime ends. The model can suggest all sorts of candidates, and the workflow can be accelerated. The Nobel laureate can improve the science, but a drug still has to survive assays, animals, manufacturing, regulators, dosing, safety, and people.

Anthropic may still build something important here. The most interesting version is not that it becomes another DeepMind. It is that Claude becomes the operating layer through which pharma companies use their own proprietary data and validated tools. That could be a real business.

But then the moat belongs to whoever owns the data. Drug discovery will be decided by the data the model cannot fake.

Inductive Bio plugged its models into Anthropic's new science platform the same day the platform launched. On June 30, hours after Anthropic unveiled Claude Science, Inductive Bio shipped a connector that lets scientists query its ADMET prediction models, the ones that placed first among 370-plus entries in the OpenADMET-ExpansionRx blind challenge, from inside Claude in plain language. It is the first outside tool to move into the workbench on day one; whether that becomes the pattern (specialist models competing to live inside frontier assistants) or stays a one-off is not yet shown. [Link]

A biotech announced positive "Phase III" results for an AI-designed program that has never been in a clinical trial. NorthStrive Biosciences, a PMGC Holdings subsidiary, and partner Yuva Biosciences reported that four small molecules picked by Yuva's MitoNova AI raised ANT1, a muscle-energy protein, by as much as 50%, but the test was an in-vitro assay in primary human skeletal-muscle cells, and "Phase III" here means the third stage of the companies' internal discovery program, not a clinical Phase 3. The next step they describe, confirmatory testing in a more mature muscle model, is still preclinical. [Link]

Insilico signed its third nine-figure-plus AI discovery pact of the year and still has no drug past mid-stage trials. On July 1 Takeda agreed to pay Insilico about $60 million upfront and near-term against a deal that could reach roughly $600 million in milestones plus royalties, for candidates from Insilico's Pharma.AI platform, nine days after Insilico's up-to-$2.5-billion neuroimmune deal with SK Biopharmaceuticals on June 22. The upfront is real money; the headline totals are contingent maximums that pay out only if candidates advance, and Insilico has not yet carried an AI-discovered molecule through Phase 3 or to approval. [Link]