AI for Pharma

A New York Times report (read here in syndicated form) says Tabula conducted the lab experiments and produced the phage cocktail that reached a patient in October, while Radical Numerics supplied some of the genome designs; Radical acknowledges Tabula's contribution and cites contractual confidentiality for not saying more. Attribution is disputed, so the share of the clinical milestone that rests on Radical's Omnii-1 model, is not established

Beacon models led more than 270 participants in predicting inhibition potency (pIC50) against CYP3A4, CYP2D6, CYP2C9 and CYP1A2 for 750 compounds whose measurements were withheld until the challenge closed, with NVIDIA third and Recursion fourth. A second independent blind-benchmark win for the ADMET vendor after the OpenADMET-ExpansionRx challenge, on the endpoint that drives drug-interaction risk; the ranking is reported by Inductive and no numerical scores are published in the release.

Supplementary Phase 2 data presented at SITC 2026: neoantigen-specific T cells detected directly ex vivo, mainly CD4+ with CD8+ responses in about half of patients, and a significant association (p<0.001) between the AI-Immunology platform's predictions and measured T-cell responses; 12 of 16 patients had objective responses over two years. Every patient also received pembrolizumab for two years and there is no control arm, so the findings do not isolate the vaccine's clinical effect. Evaxion develops its own vaccines rather than selling the platform.

A 100-task benchmark built with Ligo that tests whether an AI agent can interpret raw wet-lab evidence rather than answer from the literature. OpenAI reports GPT-6 Astra earning 66.7% of available partial credit against 43.8% for GPT-5.6 Sol at the same reasoning setting. The results are vendor-reported and cover only OpenAI's own models, so this measures laboratory-data interpretation rather than comparing vendors.

Kos, a $125M New York and Athens life-sciences investor, took part through its inaugural Kos BioVentures fund in the Sutter Hill-led Series A that Fathom had already announced; Kos's own cheque size is not disclosed. Fathom's Microcosmos platform pairs physics-based (quantum-chemistry) simulation with AI to model protein motion and drug binding, but the release describes advancing an internal pipeline rather than selling the platform, so it is a drug developer rather than a vendor.

Growth has slowed sharply from the 13% of fiscal 2026: revenue of $174.5M was up 4% year on year, net income fell to $34.7M from $62.1M and adjusted EBITDA to $88.9M (50.9% margin) from $100.8M, with full-year guidance of $678–683M. The one AI figure is distribution rather than usage, more than 190 health systems have purchased the AI suite (Ask, formerly DoxGPT); no active-prescriber count was reported this quarter. A current, audited read on the physician-engagement channel that OpenEvidence's valuation is priced against.

Jonathan Heeney's Laboratory of Viral Zoonotics and its 2017 spin-out DIOSynVax used an AI tool over global surveillance sequences to design 'super-antigens' shared across H5 variants; the chosen candidate gave complete protection in mice and ferrets against variants that have infected humans, where about half the animals survived with the WHO's stockpiled vaccine (Cell Reports Medicine, Nov 2026). Preclinical only, human trials are the stated next step, and DIOSynVax develops its own vaccines rather than selling design tools.

Roche will use Chai's models and platform in a number of discovery programs across its therapeutic areas, Chai's seventh named big-pharma partner in 2026 after Lilly, Pfizer, Novartis, argenx, BMS and GSK. Financial terms, program count, modalities and which Chai models are in scope are undisclosed; the Chai-3 'high target success rate' language in the release carries no figures.

A Nasdaq-listed biospecimen marketplace (ISPC) buys an AI agent that identifies disease-associated proteins and a model that monitors disease trends: $750k cash, $2.5M in restricted stock and up to $1.25M in milestones due only on delivery and acceptance of the products, so integration and acceptance are still outstanding. No product names, validation data or revenue disclosed; the release also flags Nasdaq-listing and going-concern risks.

The UK government's venture fund joins Twin Path, Odyssey, Nauta and others, with angels Patrick Collison and Chris Gibson, behind a model that rebuilds quantum chemistry for GPUs instead of learning from past experimental data. The '100x cheaper' calculation claim and the drug-discovery benefits are company statements framed as aspirations in the release, with no benchmark or peer-reviewed result; the round size is given only as eight figures.

Anthropic's life-sciences lab describes using Claude alongside its researchers to investigate vault particles, a large and poorly understood cellular protein complex. The work is not peer-reviewed and the team documented model errors that needed correction, so it is a worked example of the model-plus-wet-lab loop Anthropic is building, not a validated result.

UCB will use Cradle's protein-design platform in selected discovery projects from hit generation through preclinical readiness, the first top-tier pharma named publicly for Cradle's antibody work. Financial terms and outcomes undisclosed; a collaboration announcement, not a result.

Financing for a supplier of human iPSC-derived cell models used in discovery and safety testing, the wet-lab reagent layer under the virtual-cell and ADMET models on this map rather than an AI vendor. The $25M is a single round, not total funding (the company release names M&G Investments and Abcam founder Jonathan Milner as backers and gives no total raised); Source upgraded 4 Nov 2026 from the BioXconomy report to the company announcement.

Reports stronger performance than selected baselines on molecular-property prediction and antibody-design tasks. The antibody results are computational proxies with no wet-lab validation, and the comparisons are the developer's own.

Adds an agent-assisted option for molecular analysis and design inside Google's Antigravity agent environment, a frontier lab packaging cheminformatics tooling as agent skills, the same move as Claude Science's skills and connectors. Google's own demonstration produced computational proposals that still require experimental validation; no partner or customer results.

Joins Signant's applications, study data and workflows for sites, sponsors and CROs, with SHAI DataAssist, a plain-language query tool over study datasets. Ships as standard with Signant solutions rather than as a separate product, an IRT and eCOA incumbent bundling an AI layer into existing licences. Launch announcement; no customer results or metrics.

Post-trained from GLM-5.3 on chemistry, materials and life-sciences data plus the company's own lab records and agent trajectories; weights on Hugging Face and ModelScope, API access, and use inside its Mira AI-scientist platform. Another open alternative under the commercial science-model floor. The benchmark comparisons against GPT-6-Astra and Kimi-K3 are the developer's internal evaluations, not independent.

Exclusive option rights over a decade could take Fresenius from 29 to 66 biosimilar molecules, with WuXi Biologics paid fee-for-service per candidate selected; Fresenius chooses which programs advance and is not obliged to develop all of them. No AI content, and WuXi Biologics (HKEX: 2269) is a separate listed company from WuXi AppTec, the WuXi on this map; included as a read on biologics CDMO scale.

A module in the Optichannel 360 platform that scores how a biopharma brand's therapies surface in AI answer engines and turns the gaps into prioritised content fixes, commercial teams starting to treat AI answers as a channel to be managed. Launch announcement only; no independent performance outcomes.

$411.5M Q3 revenue, $108.1M from data and applications, ~$500M of pharma licensing bookings year to date, and a BMS oncology-and-Alzheimer's data licence worth up to $192.5M over five years. The only audited quarterly read on what pharma pays for real-world data.

Adds a clinical-stage asset to Biossil's development portfolio; the $350,000 upfront is distinct from up to $91M in contingent milestones.

Claude-powered analysis extends an incumbent instrument vendor into AI-assisted interpretation of its own single-cell data: researchers ask questions of Chromium data in plain language and an agent plans, runs and documents the analysis.

The biggest jump of any question in the Q3 survey. Sentiment, not outcomes, worth reading against the flat Phase II data.

The second data-room integration for Prudentia in 2026 after Ideals. Diligence agents that run inside the room the deal already uses, within its permissions and without training on the data, remove the main objection BD teams have had to AI in due diligence.

Not a function-level pilot: compound screening, target ID and trial design across oncology, internal medicine and immunology, with Edison scientists forward-deployed into Pfizer teams. The first pipeline-wide AI-scientist deployment at a top-5 pharma on this map.

Cellarity, Etiome and Empress Therapeutics combined into one AI drug-discovery company under former Adaptimmune CEO Adrian Rawcliffe, consolidation rather than new formation, and a broader target-to-small-molecule competitor. Already on the watchlist.

Financing for a laboratory-robotics vendor competing for the physical execution layer of AI-driven experiments.

100M single-cell RNA profiles plus 11,000+ protein profiles, with GSK, Pfizer and Sanofi as foundational partners. Pre-competitive data infrastructure rather than a product.

Will acquire 10–15 shelved drug programs over 3–5 years, each in a separate entity, advanced on one AI platform, AI-native company as asset acquirer rather than discoverer. A fund commitment, not a $600M corporate financing round.

AI foundation models to flag treatment-related tissue changes for pathologist review, the first named AI-native deal in the Nonclinical Safety node. Terms and performance undisclosed.

Partner-reported: 35,000 antibody designs in a month (Divergence), a CAR-T component 70% shorter with 10x receptor expression (Generate), 89% faster manufacturing-doc review (Enveda), trial-plan assessment at BMS. Mostly preclinical, no independent replication, but it widened Claude's footprint on this map from 4 functions to 10 and makes Anthropic a drug developer.

Accompanying preprint reports experimental screening, making it a candidate competitor to existing antibody-design platforms; the commercial offering remains unclear.

Exclusive worldwide licence to de novo miniprotein candidates against multiple Astellas targets; over $400M in upfront and milestones plus royalties. AI Proteins' second $400M-class pharma deal after BMS.

The oldest AI medical-writing vendor reports scale in trials rather than a published time saving; investor and amount undisclosed. First news item on the Medical Writing node, created 2 November from the Pfizer workforce audit.

Ex-OpenAI science lead Kevin Weil emerges from stealth at an $865M valuation, targeting the synthesis step pharma outsources to Chinese CROs. No product yet.

Radical Numerics' Omnii-1 generated 700,000 genomes; 16 proved viable and a cocktail reached one patient under compassionate use. No peer review and no disclosed outcome.

Low double-digit-million upfront against ~$1.1B in milestones for tissue-specific lipid nanoparticles, and the same ~1–2% upfront ratio as insitro's AstraZeneca deal.

An open protein-design model combining programmable design with search, distribution via BioNeMo rather than licensing revenue.

Brings an agentic operating layer spanning multiple stages of drug development. The frontier labs are buying the workflow layer rather than competing with it.

Two Nature papers on RFdiffusion3, atomic-level control around catalytic cores, with active enzymes reported on a first design attempt.

Wraps 50+ design tools behind one API, cutting tokens 86% and GPU cost 64%. Ships skills for Claude Code. Free via RosettaCommons.

MASH and cardiomyopathies. Partner-funded target discovery: AZ nominates targets and leads all development. The upfront is 2.7% of the headline.

Chemistry-native AI for formulation R&D. Every named customer is chemicals or consumer goods, not pharma; the pharma relevance is by analogy.

Agents run entire workflows, design sequences, order from a CRO, trigger lab work, analyse results, turning an ELN vendor into a lab-orchestration competitor.

CTMS, eTMF, Study Startup and Site Connect across IQVIA's studies. The two largest incumbents consolidating rather than competing.

Purpose-built models to predict clinical trial outcomes, 'run the trial before running the trial'.

Paris-based, data-engine first: generates its own pre-training datasets to predict how human cells respond to treatment.

Turns a trial brief into protocol and submission package checked against current guidance. Claims 82% of avoidable protocol errors caught.

Cortellis, DRG and BioWorld change hands for $600M against $389.8M revenue, roughly 1.5x, on a business declining 6.9% a year.

Valuation above $800M. Investors increasingly judge AI-discovery companies on clinical data rather than platform claims.

Chai's sixth big-pharma partner in 2026, after Lilly, Pfizer, Novartis, argenx and BMS.

Used by roughly 40% of US physicians, ~$300M annualised revenue from pharma advertising. Valuation up 15x in about 18 months.

Foundation models for toxicologic pathology image analysis. Funding undisclosed; the AstraZeneca collaboration followed a month later.

Advanced AI across targeted areas of clinical development, a named large-pharma deployment for the agentic orchestration layer.

Natural-product chemical space rather than de novo design; three candidates already in human testing.

Alongside a new in-house life-sciences lab and a verification programme for vetted labs, unusual vertical integration for a model vendor.

Third named deal in a year after Sanofi (three-year licence, Oct 2025) and an extended Merck agreement (Dec 2025); a Hypothesis Generation Agent followed on 10 September.

Time to regulatory first draft down ~97%, from about 100 hours to three.

Syneos is deploying the platform across its therapeutic-area, medical advisory, real-world evidence and proposal teams, an enterprise-wide CRO rollout rather than a pilot, and a route into pharma that runs through a services provider rather than direct sales.

The commercial-analytics incumbent answering AI with headcount rather than models, a read on how much of pharma's AI budget sits in Commercial Analytics, not R&D.

An agentic co-scientist that executes validated physics methods rather than retrieving and summarising, an explicit shot at general-purpose AI.

Four production capabilities in ICON's Orbis platform, site intelligence, enrolment-risk prediction, protocol optimisation, client access inside Claude, plus Claude Science for clinical teams. A frontier lab entering clinical operations through the largest listed CRO.

Chai-3 reports binding at therapeutic affinities in about half of attempts, roughly double Chai-2.

One of very few agentic science platforms with peer-reviewed validation. 10,000+ labs, 15,000+ scientists in nine months.

An AI workbench for scientists with 60+ skills and connectors, running locally or on a lab's own HPC so proprietary data never leaves.

Alnylam built custom agents on Marmot for its commercial organisation from August 2025 and reports reporting cycles cut from months to hours; now extending across the enterprise.

The largest anonymised quality assessment of preprints to date. The finding that 12.9% of top-scoring papers landed in lower-ranked journals is an argument that journal rank is a weak proxy, which is the premise the product is built on.

240+ healthcare organisations and 13,000+ sites in 20+ countries; 52M patients added in a year, mostly outside the US. A conversational AI interface shipped in January.

Redpoint-led round takes total funding to $125M and ~50 customers a year after commercial launch, the regulatory-documentation category now has a near-unicorn.

Transformation Capital leads; Alexion named as a customer; claims MLR review rounds cut from 3.2 to 1.2 per asset.

Target discovery, validation and translational biology, synthesising experimental, clinical and biomarker data.

A consultancy shipping agentic software for medical affairs (KOL identification, field insight) and MLR content review, straight into the territory of Sorcero and Solstice.

Claims 50% shorter site selection, 60% shorter IND/NDA authoring (Weave) and 20–30% gains in safety and medical writing, a CRO packaging AI-native vendors as its own offering.

Sanofi Ventures joins; total raised $35M for a privacy-focused pharma commercialisation platform.

Led by Thrive Capital. Still no patient dosed; first clinical trials expected by end of 2026.

Ten niche oncology and immunology targets, same query and schema across Gosset, Claude, GPT, Gemini and Perplexity. The useful detail is that exposing the curated index as an MCP tool closed most of the gap, which argues the moat is the data rather than the model.

A hyperscaler selling a science application rather than infrastructure. 40-plus biological foundation models, an agent layer, and integrated wet labs at Ginkgo, Twist and A-Alpha Bio, launched with an MSK case that went from target to 46 confirmed binders in weeks. The customers it ships to already hold their data and compute on AWS.

AI tools and agents for workflow automation and sponsor oversight across clinical pharmacology, clinical development and device trials; part of the post-Labcorp turnaround pitch.

Consultancy plus agent vendor for trial coordination, onboarding and adherence; BCG's own benchmark says few biopharmas yet see sustained AI value.

The reference AI-native in commercial pharma absorbed into a commercial-services company; terms undisclosed. The decision layer gets bundled with orchestration rather than sold alone.

Takes total funding to $27M. A distributor's venture arm leading a diligence-software round is a signal that the buyer for this category is broadening beyond pharma BD to anyone evaluating life-science assets.

NewSpring-led round brings total to $59M; claims a third of the top 30 pharma as customers.

OpenAI models embedded across the Accelerator Drug Development offering, with the stated aim of cutting clinical-trial cycle time and killing unlikely therapies earlier.