Skip to content
    AI Pharma Pulse · Oct 9, 2026

    Friday, October 9th

    Open on LinkedIn

    Sign in to LinkedIn, then select Subscribe.

    Foundational AI Updates

    Google’s MedGemma medical models get a peer-reviewed write-up in Nature Medicine On October 6, Nature Medicine published an open-access paper on MedGemma, Google’s collection of open medical vision-language models built on Gemma 3 that read both images and text. The paper also introduces MedSigLIP, a medically tuned vision encoder derived from SigLIP that powers MedGemma’s image understanding. On out-of-distribution tasks, the authors report gains over the base Gemma 3 models of 2.6–10% in medical image question answering, 15.5–18.1% in chest X-ray finding classification, and 10.8% in agentic evaluations. They also report that fine-tuning MedGemma can be more effective than fine-tuning the base Gemma 3 model for medical tasks, particularly when training data is limited.

    Why it matters: Teams building tools for safety case intake, medical writing, or imaging endpoints now have an open, peer-reviewed medical model to start from instead of a black box. Because it is open, sponsors and vendors can document it, validate it, and lock a version, which is what an inspection-ready, risk-based credibility case for a defined context of use needs.

    Source: Nature Medicine — An open vision-language model for diverse medical applications https://www.nature.com/articles/s41591-026-04626-w

    Drug Development & AI

    GSK expands work with Chai Discovery after Chai’s designs passed GSK’s own wet-lab tests On October 9, Chai Discovery announced a collaboration giving GSK access to its platform, including its protein folding and design models, across GSK’s pipeline. In the evaluation, Chai’s models generated designs to GSK targets zero-shot, without target-specific training, and GSK’s internal wet-lab testing found binders to all tested targets. GSK’s SVP R&D Technologies, Christopher Austin, said the technology complements GSK’s own internal models, which it continuously retrains against its own wet-lab data.

    Why it matters: A large pharma company expanded the relationship only after testing the designs in its own lab. Expect AI vendors to be judged the same way: by whether results hold up in experiments, with the evidence on file before a sponsor relies on the tool.

    Source: BioSpace / Business Wire — Chai Discovery Announces Collaboration with GSK Following Successful Wet-Lab Evaluation https://www.biospace.com/press-releases/chai-discovery-announces-collaboration-with-gsk-following-successful-wet-lab-evaluation

    Insitro-built in vivo PK models become available to biotechs in Lilly TuneLab On October 6, insitro announced that the machine learning models it built with Eli Lilly to predict in vitro and, critically, in vivo properties of small molecules can now be used by certain biotech companies in Lilly TuneLab. The models were trained on multi-species preclinical data Lilly generated over decades from hundreds of thousands of unique molecules, and insitro positions them as an alternative to the industry-standard method of extrapolating from in vitro assays. The collaboration was announced in September 2025, and the models will be updated as the dataset expands.

    Why it matters: Smaller sponsors are getting preclinical prediction tools that used to be available only to organizations with the scale to generate that much data. Model output is moving upstream of compound prioritization and animal-study decisions, so sponsors will need to record which model version they used and how they validated it for their chemistry.

    Source: insitro — insitro Designed Machine Learning Models for Small Molecule in vivo Pharmacokinetic Behavior Prediction Now Available in Lilly TuneLab https://www.insitro.com/news/insitro-designed-machine-learning-models-for-small-molecule-in-vivo-pharmacokinetic-behavior-prediction-now-available-in-lilly-tunelab/

    CHARM Therapeutics doses first patient with CHM-029, billed as the first co-folding-designed molecule in the clinic On October 7, London-based CHARM Therapeutics said it has begun dosing and fully enrolled the first cohort in a Phase I/II trial (NCT07751991) of CHM-029, a menin inhibitor for relapsed or refractory acute myeloid leukemia with an NPM1 mutation, KMT2A rearrangement, or NUP98 rearrangement. CHARM used its proprietary AI platform DragonFold to design the molecule to stay potent against known menin resistance mutations, and calls it the first molecule designed with a protein-ligand co-folding model to reach clinical studies (company claim). The company reports less than 24 months from first synthesis to first patient dosed and expects initial data in 2027.

    Why it matters: Structure-prediction models are moving from design tools to molecules in patients. The trial, not the model, now carries the evidence burden, and sponsors and partners should expect to see how AI-informed design rationale is documented alongside the preclinical package.

    Source: GlobeNewswire / CHARM Therapeutics — CHARM Therapeutics announces first patient dosed and first cohort fully enrolled in its Phase I/II clinical trial of CHM-029 for Acute Myeloid Leukemia https://www.globenewswire.com/news-release/2026/10/07/3376361/0/en/charm-therapeutics-announces-first-patient-dosed-and-first-cohort-fully-enrolled-in-its-phase-i-ii-clinical-trial-of-chm-029-for-acute-myeloid-leukemia.html

    Insilico’s AI-designed lung fibrosis drug rentosertib is now dosing in a Phase III trial Insilico Medicine said on September 10 that it dosed the first patient in GENESIS-IPF-3 (NCT07687459) at Peking Union Medical College Hospital, with Shanghai Pulmonary Hospital enrolling its first patient the same day; an October 8 AI Industry Today analysis put the milestone back in focus this week. Rentosertib is for idiopathic pulmonary fibrosis (IPF), and AI was used both to prioritize the target, TNIK, and to design the molecule. The 52-week randomized, double-blind, placebo-controlled trial plans to enroll 320 patients at 47 centers in China, with the annual rate of decline in forced vital capacity (FVC) as the primary endpoint. The earlier randomized Phase IIa study ran 12 weeks in 71 patients, and the drug remains investigational.

    Why it matters: This is one of the clearest tests yet of a drug where AI chose both the target and the molecule, and from here it is judged by the same late-stage standards as any other drug. Regulators will weigh the clinical results, not how the drug was discovered.

    Source: Insilico Medicine — Insilico Medicine Doses First Patient in GENESIS-IPF-3, the World’s First Phase III Trial of a Generative AI-Driven Innovative Drug https://insilico.com/news/isn1009261-insilico-medicine-doses-first-patient-genesis-ipf-3

    Regulatory, Policy & Industry Moves

    FDA’s drug center stands up an office that combines AI, real-world evidence, and trial modernization CDER’s Office of Innovation and Clinical Trial Modernization (OICTM) became operational October 1 inside the Office of Medical Policy, with dedicated divisions for artificial intelligence and real-world evidence analytics. Marie Bradley announced on LinkedIn that she is acting director of the Division of Real-World Evidence Analytics. This week, Cecilia Almeida, Associate Director for Regulatory Policy in the Office of Medical Policy, confirmed the office publicly, and Clinical Trial Vanguard’s October 6 analysis argued that placing AI and RWE under one roof reflects a view that they are converging into a single methodological challenge. FDAWeb noted that the July 29 Federal Register notice identified the office and its divisions but did not describe their functions, staffing, or regulatory responsibilities.

    Why it matters: FDA put its AI division next to the people who review real-world evidence and trial methods, which suggests AI-derived results will be reviewed as scientific evidence rather than as software. Defending them is a job for biostatistics, epidemiology, and regulatory teams, not just IT. Expect questions about context of use, model credibility, and documentation that lets a regulator reconstruct a decision, and note that the office has yet to publish what it will actually do.

    Source: Clinical Trial Vanguard — The FDA Created an AI Office. Now Comes the Hard Part. (with FDAWeb — New CDER Office Focused on AI, Real-World Evidence) https://www.clinicaltrialvanguard.com/clinical-bellwether/the-fda-created-an-ai-office-now-comes-the-hard-part/ https://www.fdaweb.com/new-cder-office-focused-on-ai-real-world-evidence/

    Originally published on LinkedIn. Read the original

    The Pulses

    Stay close to what’s changing.

    AI Vendor Pulse and AI Pharma Pulse, published on LinkedIn.

    About 1,070 newsletter subscribers · October 2026

    AI Vendor Pulse

    Mondays

    Open on LinkedIn

    AI Pharma Pulse

    Fridays

    Open on LinkedIn

    To subscribe, sign in to LinkedIn and select Subscribe on the newsletter page.