# Recursion Posts Its First Phase 2 Win as Anthropic Muscles Into Pharma - AI Drug Discovery Weekly - Week of June 25–July 2, 2026 > Recursion's RX4881 hit Phase 2 proof of concept in FAP and moved into FDA pivotal-trial-design talks, the first hard clinical datapoint for an AI-native drug platform, just as Anthropic launched Claude Science and founders across the field argued the real moat is proprietary data and compute, not model IQ. Our synthesis for the week of June 25 to July 2, 2026. ## AI Drug Discovery Weekly ### Week of June 25–July 2, 2026: Recursion Posts Its First Phase 2 Win as Anthropic Muscles Into Pharma --- ### TL;DR - **A clinical proof point, finally.** Recursion's RX4881 hit Phase 2 proof-of-concept in familial adenomatous polyposis with "significant polyp burden reduction," and the company is now in FDA pivotal-trial-design talks, the clearest sign yet that an AI-native platform can put a real asset on the board, not just a slide. - **Big AI came for pharma this week.** Anthropic launched *Claude Science*, a model tuned for scientific and pharmaceutical research: the platform layer is now directly courting the same customers techbio sells to. - **The bottleneck is data and GPUs, not model IQ.** Multiple founders converged on the view that intelligence is not the binding constraint, proprietary biological data and compute allocation are. - **Spot check:** RXRX ~$3.80, SDGR ~$16.88, LLY ~$1,210. Recursion still trades near the low end of its 52-week range despite the readout. ### What's new **Recursion (RXRX) turned platform promise into a Phase 2 win.** On Opto Sessions this week, CFO/President Ben Taylor detailed that RX4881 reached Phase 2 proof of concept in FAP, a disease with no approved therapy, showing significant polyp reduction, with the company now working with the FDA on pivotal trial design and updates expected later in 2026. Taylor also walked through the business model that funds it: 7 milestone payments totaling more than $500M booked from Sanofi (five chemistry milestones) and Roche (two ~$30M neuroscience milestones), with the Sanofi framework allowing up to 15 programs at up to $343M each plus low-double-digit royalties ([Opto Sessions – Invest in the Next Big Idea](https://app.matterfact.com/podcasts/96fd15aa0a5d930990b5c938df50b516b52461209e48af31dd34f84c26151660), Jun 29, 2026). **Anthropic entered the arena.** Anthropic launched *Claude Science*, a version of its model tailored for scientific and pharmaceutical research, in what STAT framed as a broader push to become a platform in biology and life sciences ([STAT via TheFly](https://www.statnews.com/2026/06/30/anthropic-release-claude-science-ceo-dario-amodei/), Jun 30, 2026). Read alongside the founder commentary below, it sharpens the central question for the group: is the durable moat the model, or the data the model is trained on? **The FDA kept clearing plumbing.** Eli Lilly was named among the first participants in the FDA's new *PreCheck* pilot, which lets regulators review manufacturing facilities while they are still under construction ([CNBC via TheFly](https://www.cnbc.com/2026/06/29/eli-lilly-regeneron-in-fda-precheck-manufacturing-program.html), Jun 29, 2026), incrementally shortening the path from molecule to market that every discovery platform ultimately depends on. **On the frontier, the best generative-chemistry results came from private hands.** On Latent Space, Genesis Molecular AI (the rebranded Genesis Therapeutics) said its protein-ligand co-folding model, PEARL, hit sub-1Å RMSD versus the ~2Å industry standard and beat published open-source models on the external OpenBind (EVA-71A protease) challenge, a target none of the models had seen in training. Genesis is running a 24/7 "agentic" discovery platform, codenamed Sapphire ([Latent Space: The AI Engineer Podcast](https://app.matterfact.com/podcasts/47964319d96d672dbf84ee8875e71ccbac5d6459026cb2d6e4fb2f4a8f5b6191), Jul 1, 2026). In academia, UPenn's Cesar de la Fuente described mining the full 42,000-protein human proteome for antimicrobial peptides in ~60 minutes and a deep-learning model (APEX) that predicts activity from sequence alone, a reminder that the discovery frontier extends well beyond the listed names ([Editors in Conversation](https://app.matterfact.com/podcasts/d9f75f9f25ba1694d7368788bd5b8faa22e1ce6d395b352612a3d7184821aa98), Jul 1, 2026). ### The debate **Bulls point to compounding proof points.** Genesis's Evan Feinberg drew the direct analogy to coding agents: once the underlying models cross a quality threshold, agents amplify the value. "If your model is sitting at 1.8, 1.9 Angstrom RMSD, that's slop," he said: the prerequisite for a 24/7 agentic drug-hunting platform is models good enough that "medicinal chemists would actually want to make and not laugh at" the molecules ([Latent Space](https://app.matterfact.com/podcasts/47964319d96d672dbf84ee8875e71ccbac5d6459026cb2d6e4fb2f4a8f5b6191), Jul 1, 2026). Recursion's Taylor made the platform case in numbers: chemistry optimization cut development timelines from 4–5 years to 17 months using 90% fewer experimental chemistries, and he argues approved and pipeline drugs still address only 10–12% of the human genome ([Opto Sessions](https://app.matterfact.com/podcasts/96fd15aa0a5d930990b5c938df50b516b52461209e48af31dd34f84c26151660), Jun 29, 2026). **Skeptics say the constraint isn't intelligence.** The most consistent theme across founders this week is that models are not the bottleneck. Ambitious Bio's Elizabeth Hudson argued no amount of intelligence ("OpenAI, Anthropic, DeepMind, Microsoft") can solve drug toxicity prediction without complete cross-demographic molecular data, noting the largest human proteomic reference dataset still comes from just three European men ([Women in Tech Podcast](https://app.matterfact.com/podcasts/11cac2879b28f221558790c9c53d73af5609fcc7bfb65372f6d2804288609c2b), Jun 30, 2026). Ron Alfa, ex-Recursion and now CEO of Noetik, made the same point from the other side of a signed deal: he says LLMs alone can't crack cancer biology for lack of public-scale disease data, which is why Noetik spent ~2 years building a proprietary multimodal oncology dataset before licensing foundation models to GSK in what he calls "the first AI bio foundation model licensing deal" ([Pear Healthcare Playbook](https://app.matterfact.com/podcasts/1cd10c4f78f19bf3be2ca89a2d471fe7063ad6a5451d56894bdf09eb34d524af), Jun 30, 2026). Genesis's Sergey Edunov added the harder-nosed version: GPU allocation (LLM labs crowding out life sciences) is the field's key bottleneck ([Latent Space](https://app.matterfact.com/podcasts/47964319d96d672dbf84ee8875e71ccbac5d6459026cb2d6e4fb2f4a8f5b6191), Jul 1, 2026). **And a useful cold shower.** Nature's long read on the peptide craze is a reminder that novel ≠ efficacious: the peptide's own discoverer conceded "none of them are as effective as the GLP drugs. If they were, pharma would have developed them already" ([Nature Podcast](https://app.matterfact.com/podcasts/ccc6c2bf52be23d03eae8b395cb019e7bc551e02900351e2049c2bf8910f8473), Jun 29, 2026). The lesson transfers cleanly to AI-designed molecules: a benchmark win is not a clinical win. ### Stocks in play - **RXRX (~$3.80, near the low end of its $2.77–$7.18 52-week range, ~$1.7B mkt cap).** The RX4881 readout and pivotal-trial talks are the first hard datapoint the bull case has had in a while, yet the tape is not paying for it, a setup worth watching into the later-2026 update. - **SDGR (~$16.88, $10.95–$23.75 range, ~$1.3B mkt cap).** No fresh company-specific catalyst this week, but the Anthropic/Claude Science entry and the "data-not-models" debate cut to Schrödinger's physics-based-vs-learned positioning; quiet weeks are when the narrative gets reset. - **LLY (~$1,210, near its $1,238 52-week high, ~$1.14T mkt cap).** Named to FDA PreCheck; the AI-discovery angle for Lilly remains adjacent (obesity and manufacturing dominate its news flow), but it is the deep-pocketed buyer every techbio platform is ultimately selling into. ### Read-throughs - **Private frontier > public comps this week.** The most impressive generative-chemistry result (Genesis's PEARL) and the marquee licensing deal (Noetik-GSK) both sat in private hands, a reminder that the public AI-discovery basket (RXRX, SDGR, ABSI, RLAY) captures only part of where value is accruing. - **Anthropic's move pressures the "platform" thesis.** If frontier labs sell science-tuned models directly to pharma, the differentiated asset shifts decisively toward proprietary, hard-to-replicate biological data, favoring players with owned datasets (Recursion's 40+ PB, Noetik's multimodal oncology corpus) over those selling model access. - **Compute is now a biotech input cost.** Edunov's GPU-allocation warning means AI-discovery timelines are partly hostage to the same capex cycle driving the broader AI trade, a non-obvious macro linkage for a healthcare book. ### What changed vs last week Where the debate stands: the bull case just got its cleanest clinical datapoint of the cycle (Recursion RX4881), while the smartest skeptics have stopped arguing that the models don't work and started arguing that data and compute are the real moats. Next week we track whether Anthropic's Claude Science draws a pharma partner, any follow-through on Recursion's FDA pivotal-design talks, and whether the private-frontier lead (Genesis, Noetik) surfaces in public-comp commentary.