Roche counts a tracked AI or computational contribution in 40% of its pipeline decisions and grades how much each mattered.
Europe’s medicines regulator is weighing generative AI in drug manufacturing under a documented control strategy, with no method yet for measuring hallucination.
Novo Nordisk and Anthropic announce a partnership to test Claude Science in drug discovery.
An AstraZeneca-authored paper uses machine learning to identify lung cancer patients who may benefit more from tremelimumab.
Xaira reports AI-designed antibodies produced binders on a receptor target where library screens failed.
Use Case · September 28, 2026
Roche counts AI or computation in 40% of pipeline decisions
Roche reports that 40% of its pipeline decisions from the fourth quarter of 2025 through the second quarter of 2026 had a tracked AI or computational contribution. Coverage reached 44% in the second quarter. The company presented the figures at its Pharma Day in London on September 28, 2026.
Roche grades each contribution: 41% fundamental or critical, meaning it materially influenced the decision, 48% supportive, meaning it meaningfully improved confidence in the decision, and 10% informative, meaning it informed thinking. Roche also cites 18 months from lead identification to clinical candidate, which it describes as twice as fast and with reduced immunogenicity risk, for its lab-in-the-loop approach, where AI is embedded in each design iteration from target discovery to the clinic. Its target-assessment AI agent, TargetNexus, is on track to be part of 80% of research portfolio decisions by the fourth quarter.
What it means for you. If your board asks how much AI is changing research and development (R&D), Roche has given you a way to answer: count the portfolio decisions AI touched and grade how much each one mattered, using definitions your teams can apply the same way.
EMA considers allowing generative AI in GMP with documented controls
The European Medicines Agency (EMA) published a report on October 2, 2026 from a multistakeholder workshop on Annex 22, the planned AI annex to the EU good manufacturing practice (GMP) guide. The workshop ran June 30 and July 1. The current draft says generative AI, large language models (LLMs), and other probabilistic or adaptive models should not be used in critical GMP applications. Consultation comments supported allowing them, and the drafting group has discussed doing so, provided they meet the annex’s requirements and are supported by a “fully documented, robust, risk-based control strategy.”
Concerns remain about what that strategy needs. Regulators asked whether hallucination, fabrication, and sycophancy are ordinary statistical uncertainty or a different kind of failure, and the discussion set no method for estimating confidence intervals or hallucination rates. The report records that in adaptive or generative AI, inability to detect a failure may itself be a central risk factor. It also says that if a third-party model or application programming interface (API) cannot provide the outputs, rationale, logs, or evidence needed for control, the manufacturer cannot measure and justify the system adequately. Accountability stays with qualified or responsible persons and cannot be delegated to AI, and human oversight cannot compensate for inadequate validation. The drafting group will review the input as it revises Annex 22.
What it means for you. If your manufacturing or quality teams want generative AI in GMP workflows, the annex may move from a ban toward a conditional yes, though the text is not final. Ask your suppliers now for the logs, outputs, and change records the report describes, and confirm who among your qualified persons would carry accountability.
Novo Nordisk and Anthropic partner to test Claude Science in drug discovery
Novo Nordisk and Anthropic announced a collaboration on September 16, 2026 to help Novo discover and develop new medicines faster. Novo will test Claude Science, Anthropic’s tool for scientific reasoning in research and development, in specific R&D workflows and apply it to scientific challenges in biological reasoning. It will also use Anthropic’s frontier models to strengthen AI-driven software engineering across the company.
The companies say they will jointly address drug discovery challenges that Novo’s scientists identify. The collaboration includes “robust data governance and human oversight,” which Novo says keeps AI use in line with its ethical and compliance standards.
What it means for you. Your peers are choosing frontier-model partners for both research and software engineering. If your board asks where you stand, prepare two answers: which workflows you would test first, and what data governance and human oversight you would require, since Novo names both as part of the deal.
Machine learning signature points to lung cancer patients who may benefit more
AstraZeneca-affiliated authors report in Clinical Cancer Research, published October 1, 2026, an analysis of the POSEIDON trial in metastatic non-small cell lung cancer (NSCLC). Machine learning models analyzed the trial data and produced genetic signatures to identify patients who may derive greater overall survival benefit from adding tremelimumab to durvalumab and chemotherapy.
For patients with nonsquamous tumors, the top-ranked 50% reached a hazard ratio of 0.56 (95% confidence interval, 0.33–0.97). The authors conclude that such approaches could be used in the future to improve precision in tailoring therapies.
What it means for you. If your clinical teams hold completed-trial data, this is a concrete pattern: use machine learning to look for the patients who benefit most, then plan prospective validation before building a strategy on it.
Xaira says AI found antibody binders where library screens failed
Xaira reports that its X-Design AI, which uses public and proprietary data and works zero-shot, designed VHH antibodies de novo for two of its programs and two reference targets. The company posted the results on October 2, 2026.
For XA-1, an oncology program, X-Design produced 18 confirmed binders from 182 designed sequences after quality control, a 10% binder rate, in seven weeks. For XA-4, a G protein-coupled receptor (GPCR), Xaira says two naive library screens produced no functional binders and llama immunization was unsuccessful. X-Design gave seven confirmed binders from 92 selected candidates, and the program is in lead optimization. On two reference GPCRs, CXCR4 and APJ, hit rates were 35% (32 of 91) and 29% (26 of 91). The work is preclinical, and Xaira says a systematic assessment of X-Design will follow in a future post.
What it means for you. If you are weighing AI antibody-design partners, these are the numbers to ask every vendor for: binder rate, time to a confirmed binder, and results on targets you choose. Xaira’s figures give you a benchmark for those conversations.
Ginkgo Datapoints and Apheris will build a 10,000-antibody dataset and train models on members’ proprietary data without exposing raw sequences, with initial access projected for early 2027.
