1. Key takeaways from the month
Between Sep 24 and Oct 23, 2026, AI drug discovery produced one Phase 3 start, a run of discovery-stage deals with small cash components, one open-weight model and two financings that closed. Five conclusions follow; each points to the chapter that holds the evidence, and every status is judged against Oct 24, 2026.
1. The clinical news is early-phase, and no AI-originated drug read out at pivotal stage. Insilico said on Sept 30 that rentosertib "has officially entered Phase III", and the registry entry NCT07687459 shows a start on Sept 9 with 320 patients planned, read second-hand. Recursion's REC-4881 reached a 47% median polyp reduction in a Phase 1b/2 study, and REC-1245 had two confirmed partial responses, with registrational work targeted for 2H 2027. Only five of 129 AI-enabled assets in trials had reached Phase 3 by March 1, and Isomorphic had no IND as of Sept 30. Insilico added only preclinical candidates, ISM1354 on Sept 30 and ISM9546 on Oct 13, its 34th and 35th. [Chapter 2]
2. Pharma paid modest cash for access to AI platforms. Upfront money was 2.5-3.6% of headline value where both figures were disclosed: $40M of about $1.5B for AstraZeneca and insitro, $32M of $888M for Insilico and Servier, and $30M of $1.2B for Generate and Zealand, the last reported only. AI Proteins and Astellas is worth more than $400M in upfront and milestones combined, with the upfront undisclosed. The one disclosed milestone was Sanofi's $10M option payment to Recursion, taking receipts to $144M. Non-AI licensing paid far more cash: Sanofi paid Regeneron $1B upfront on Oct 1, against upfronts of $30-40M in the AI deals. [Chapter 3]
3. The month's real model news is open, but its evidence is the authors' own. NVIDIA and the University of Washington released RFD4-Proteina with BSD 3-Clause weights and 435.6M parameters. Its lab result, 158 of 188 TrkA designs binding (84%), comes from a preprint that is not peer reviewed. Insilico's inClinico Agent reported AUROC 0.828, on a benchmark Insilico built and ran itself. The other design results are preprints, and the structure-prediction item was a third-party method for Boltz-2 rather than a new release. [Chapter 4]
4. Money went to listed companies; the largest private number is only a report. Recursion closed a $200M offering of 61,540,000 shares at $3.25 on Oct 16, and Iambic priced a Nasdaq IPO of about $150M at $16. Isomorphic Labs is reported to be in early talks at a valuation of at least $40B, with nothing closed and no round size or investors reported. Xaira raised nothing and said it is funded until it has molecules in the clinic. Recursion's runway now extends into 2029. [Chapter 5]
5. Regulation is present only as a dependency, not as a new rule. No new FDA or EMA instrument appears in the month's material. Recursion's registrational trial awaits the outcome of FDA discussions, and "AI-discovered" is not a regulatory category. The evidence gaps a reviewer would see are in Chapters 2 and 4: small early datasets and vendor-run validation. [Chapter 6]
1.1 Candidates and deals at a glance
| Company | Candidate or deal | Stage or value | Date | Status (Oct 24) | Source |
|---|---|---|---|---|---|
| Insilico Medicine | Rentosertib, IPF (NCT07687459) | Phase 3; 320 planned | Start 2026-09-09 | Recruiting (second-hand registry reading) | The Signal |
| Recursion | REC-4881, FAP | Phase 1b/2; 47% median polyp reduction | Cutoff Sept 10; reported Oct 15, 2026 | Registrational trial targeted 2H 2027 | Recursion |
| AstraZeneca and insitro | MASH and cardiomyopathies collaboration | $40M upfront; up to about $1.46B milestones | Oct 15, 2026 | Signed | insitro |
| AI Proteins and Astellas | Miniprotein collaboration | More than $400M upfront and milestones combined | Oct 21, 2026 | Signed; upfront undisclosed | AI Proteins |
| Insilico and Servier | Oncology collaboration | Up to $888M; $32M upfront and near-term | Milestone Oct 20, 2026 | First milestone reached; payment undisclosed | Insilico |
| Generate Biomedicines and Zealand | Obesity protein collaboration | Up to $1.2B; $30M upfront | Oct 13-14, 2026 | Reported | RSM via BioSpace |
| Recursion | Follow-on offering | $200M; 61,540,000 shares at $3.25 | Announced Oct 15; closed Oct 16, 2026 | Confirmed | Wilson Sonsini |
| Iambic Therapeutics | Nasdaq IPO | About $150M; 9.4M shares at $16 | Priced Oct 14, 2026 | Completed, per press | pharmaphorum |
| Isomorphic Labs | New funding round | Valuation of at least $40B | Reported Oct 8, 2026 | Reported; not closed | Bloomberg |
2. Clinical pipeline progress
Chapter 1 gives the month's conclusions; this chapter tests them against the trial record, one candidate at a time. Status is judged against 2026-10-24.
2.1 Candidates at a glance
| Item | Key figure | Date | Status | Source |
|---|---|---|---|---|
| Insilico rentosertib, IPF (NCT07687459) | Phase 3; 320 planned | Start 2026-09-09 | Recruiting (second-hand) | The Signal |
| Recursion REC-4881, FAP | 47% median polyp cut | Cutoff 09-10 | Phase 1b/2; pivotal start targeted 2H 2027 | Recursion |
| Recursion REC-1245 | 2 confirmed partial responses; disease control 11/12 | Cutoff 08-21 | Phase 1; 45 mg to expansion | Recursion |
| Sanofi-optioned candidate | Phase 1 planned 2027 | 2026-10-15 | Preclinical | Recursion |
| Insilico ISM1354, ISM9546 | 34th and 35th candidates | 09-30; 10-13 | Preclinical | Insilico |
| Isomorphic Labs | No IND or trial yet | As of 09-30 | Planned (end-2026) | BiopharmaTrend |
| Iambic IAM1363 | Phase 1 | 2026-10-16 | Reported | pharmaphorum |
| Xaira XA-1, XA-4 | Named; no timing | 2026-10-03 | Preclinical | Catalyst Brief |
2.2 Insilico: rentosertib enters Phase 3 while the pipeline stays preclinical
Rentosertib (ISM001-055) is an oral TNIK inhibitor for idiopathic pulmonary fibrosis (IPF). According to rentosertib.co.uk, Insilico's PandaOmics software picked the target and its Chemistry42 software generated the molecule. Insilico's Sept 30 release says rentosertib "has officially entered Phase III" but gives no date.
The start date conflicts across sources. A vocal.media essay says July 2026. rentosertib.co.uk says September 2026. The Signal's Oct 6 edition reads the registry as NCT07687459 (GENESIS-IPF-3), start 2026-09-09 marked actual, 320 estimated enrolment, 47 sites, recruiting. We follow Sept 9, the registry's actual date, which agrees with. Our registry copy is glossary-only. No company-stated readout date is in the material.
The only efficacy evidence is the Phase 2a trial (Nature Medicine, June 2025, before this window). In the highest-dose arm, lung function rose by 98.4 mL over 12 weeks, against a 20.3 mL fall on placebo, a gap of 98.4 + 20.3 = 118.7 mL. The groups were small, and liver-related adverse events raised concern (New Market Pitch).
Keep the claim narrow. Generate:Biomedicines began a Phase 3 of its AI-engineered anti-TSLP antibody GB-0895 on 3 December 2025, so Insilico's claim is the first drug with both an AI-identified target and an AI-designed molecule in Phase 3. "AI-discovered" is not a regulatory category.
Insilico's other news was preclinical. ISM1354, a GIPR antagonist for obesity, was nominated on Sept 30 as the 34th preclinical candidate (10th of 2026). Its non-GLP monkey study showed no significant toxicity and an estimated ~45-fold exposure-based safety margin; both are company data. ISM9546, an oral VAV1 molecular-glue degrader for autoimmune disease, followed on Oct 13 as the 35th (11th of 2026). Servier's first milestone on Oct 20 covered a preclinical oncology programme; both programmes have completed target nomination (terms in Chapter 3).
2.3 Recursion: two early datasets and one option
On Oct 15, 2026 Recursion reported updated data for REC-4881, a MEK1/2 inhibitor in familial adenomatous polyposis (FAP). The release says the opportunity came from phenotypic screening with Recursion's AI foundation models. In the Phase 1b/2 TUPELO study (4 mg daily, cutoff Sept 10), 87% of patients had a lower total polyp burden after 12 weeks, with a 47% median reduction. In the upper GI tract the median reduction was 55%, and two patients had at least 98%. Six of 12 improved by at least one Spigelman stage (50%).
Per the company, no treatment-related event exceeded Grade 3, and 21 of 26 patients (80.8%) had Grade 2 or lower. Four had Grade 3 events, and four stopped for adverse events, none for a Grade 3 event. The release reports no comparator. It targets a registrational trial in 2H 2027, pending FDA discussions, and more data come on Nov 2 (upcoming).
REC-1245, an RBM39 degrader, is in the Phase 1 dose-escalation part of DAHLIA (cutoff Aug 21). At doses of 45 mg or more there were two confirmed partial responses in endometrial cancer, one unconfirmed response in liver cancer and disease control in 11 of 12 patients (11 ÷ 12 = 91.7%). Recursion says the molecule was found within 18 months after screening 204 compounds. Endometrial expansion data are expected in 2H 2027.
Sanofi also optioned the first development candidate from their collaboration, an oral small molecule against a dual target in inflammatory disease. Phase 1 is planned for 2027, and Sanofi will fund and lead it; payment terms are in Chapter 3. Recursion will seek partners for REC-617 and REC-3565 and focus on five other programmes. That is reprioritisation, not discontinuation.
2.4 Who is not yet in the clinic, and what the record shows
Isomorphic Labs has named no candidate or target. BiopharmaTrend reports that as of Sept 30 it had announced no IND and no trial start. Its Sept 29 post shows one working concentration-response curve from an unnamed "Target D". The first-trial target slipped from end-2025 to end-2026 (Reuters via TFN). A Bloomberg-sourced report says no patient has been dosed (valuation talks: Chapter 5).
Iambic's IAM1363, an oral brain-penetrating HER2 inhibitor, is in Phase 1; the company is pitching a registrational trial for next year. Xaira named XA-1 (a cancer antibody built "from scratch in seven weeks") and XA-4, with XA-2 and XA-3 preclinical, and gave no efficacy data or clinic timing. Undated secondary context: Polaris lists GB-0895 in Phase 3 (SOLAIRIA-1 and -2) and Absci's ABS-201 in Phase 1/2a (NCT07317544).
No dated discontinuation fell inside the window. The failures in our material are older: Exscientia's DSP-1181 was discontinued, BenevolentAI's BEN-2293 missed its secondary objectives, and Recursion's REC-994 gave mixed Phase 2 results.
A Cancer Treatment Reviews analysis counted 129 AI-enabled assets from 69 companies in trials as of March 1, 2026. Observed success was 77.8% in Phase 1 and 19.5% in Phase 2, and only five assets had entered or completed Phase 3. A vocal.media essay puts Phase 2 success near 40%. We follow the peer-reviewed count.
The month's clinical news is early-phase: one Phase 3 start, two Phase 1-sized Recursion datasets, and a field where only five assets had reached Phase 3 by March. No AI-originated drug has a pivotal readout in this window.
3. Partnerships and deals
Chapter 2 showed that the month's clinical news came from a few candidates; the deals below show where pharma money is going before any clinical data exist. Every October deal in this chapter is a discovery-stage contract, and none includes a candidate in the clinic.
| Item | Key figure | Date | Status (as of 2026-10-24) | Source |
|---|---|---|---|---|
| AstraZeneca and insitro (MASH, cardiomyopathies) | $40M upfront; up to ~$1.46B milestones; single-digit royalties | Oct 15, 2026 | Signed | insitro; AstraZeneca |
| AI Proteins and Astellas (miniproteins) | More than $400M in upfront and milestone payments combined; upfront alone undisclosed; royalties | Oct 21, 2026 | Signed | AI Proteins (Business Wire via Yahoo Finance) |
| Insilico and Servier (oncology) | Up to $888M; $32M upfront and near-term; first milestone payment undisclosed | Milestone Oct 20, 2026; deal signed early 2026 | Milestone reached | Insilico |
| Generate Biomedicines and Zealand Pharma (obesity proteins) | Up to $1.2B; $30M upfront; up to $35M early preclinical milestones | Oct 13-14, 2026 | Reported (no company release in material) | RSM via BioSpace; The Pharma Letter |
| Recursion and Sanofi (immunology) | $10M milestone; $144M received to date; more than $300M still eligible | Oct 15, 2026 | Option exercised | Recursion |
| AstraZeneca and Oracle (trial AI) | Terms not disclosed; non-binding | Date not stated | Announced | pharmaphorum |
| Chai Discovery and GSK | Terms not stated | Date not stated | Reported | pharmaphorum |
| insitro models in Lilly TuneLab | No value stated | Oct 6, 2026 | Announced | Pharma Journal |
3.1 Headline value against cash: three new October contracts
Where both figures are disclosed (AstraZeneca-insitro, Insilico-Servier, Generate-Zealand), upfront cash is only 2.5-3.6% of the headline value. The rest is contingent on milestones that are years away.
AstraZeneca and insitro is the clearest case. On Oct 15, 2026 insitro announced a worldwide exclusive research collaboration for novel targets and therapeutics in MASH and cardiomyopathies: $40 million upfront, up to approximately $1.46 billion in discovery, development and commercial milestones, and single-digit royalties. AstraZeneca's own release confirms the same two numbers but lists only "development and commercial" milestones. insitro applies its Virtual Human platform to advance AstraZeneca-nominated targets to development candidate; AstraZeneca then leads clinical development, regulatory work and commercialisation. Neither release gives a total. PharmExec and pharmaphorum headline "$1.5 billion", which is the sum 40 + 1,460 = 1,500; PharmExec adds that the upfront is under 3% of it, and neither company disclosed the number of targets or payment schedule. Pharma Journal says Bloomberg reported up to $1.86 billion and that it found no basis for that in the release. We follow the releases.
AI Proteins and Astellas was announced on Oct 21, 2026 under a Research Collaboration and License Agreement with Astellas Innovation Management, an Astellas affiliate. AI Proteins is eligible for "over $400 million in upfront and milestone payments, plus royalties". It designs de novo miniproteins against multiple Astellas-selected targets. Astellas holds the exclusive worldwide licence and takes over development, regulatory work and commercialisation. Targets, indications and the upfront amount are not disclosed, so no upfront ratio can be computed. The $400 million is a ceiling, not revenue.
Generate and Zealand is the weakest-sourced of the three. RSM, citing BioSpace, reports up to $1.2 billion with $30 million upfront and up to $35 million in early preclinical milestones, tiered royalties, and a split in which the AI company leads research through candidate selection. RSM does not name the parties; The Pharma Letter's story headline names Zealand as the payer and Generate as the recipient. TechTimes dates the announcement to Oct 13, while The Pharma Letter lists it under Oct 14. We treat it as reported.
| Deal | Upfront | Headline | Upfront share (derived) | Source |
|---|---|---|---|---|
| AstraZeneca and insitro | $40M | $1.5B (40 + 1,460) | 40 ÷ 1,500 ≈ 2.7% | insitro; PharmExec |
| Insilico and Servier | $32M (with near-term R&D payments) | $888M | 32 ÷ 888 ≈ 3.6% | Insilico |
| Generate and Zealand | $30M | $1.2B | 30 ÷ 1,200 = 2.5% | RSM via BioSpace |
The Generate headline may or may not include the upfront; the source does not say, so the 2.5% is indicative. A BioBucks tracker also warns that undisclosed upfronts are becoming the norm, and that a ceiling with no cash figure "tells a reader nothing about what was actually committed".
3.2 Milestones paid and options exercised
On Oct 20, 2026 Insilico (HKEX: 3696) announced the first R&D milestone in its oncology collaboration with Servier. Both programs have completed target nomination, and the early drug discovery milestone was met for the first, triggering an undisclosed payment from Servier. The deal carries up to $888 million, including $32 million upfront and near-term R&D payments. Insilico leads AI-driven discovery and early development, and Servier shares R&D costs and takes over clinical validation and commercialisation. Peptide & Protein reports signing on Jan 4, 2026, so the milestone came about 9.5 months in, consistent with Insilico's "within 10 months". Neither target nor tumour type is named.
Insilico also states, citing its 2026 interim report, that signed collaboration value in 2026 reached about $7.3 billion by mid-August, and nearly $11 billion cumulatively since 2021; first-half revenue was about $106 million (+287%). These are company claims of potential value, not cash.
The only milestone with a disclosed amount this month is Recursion and Sanofi. On Oct 15, 2026 Recursion announced that Sanofi optioned the first development candidate from their collaboration, triggering $10 million. Total upfront and milestone payments received are now $144 million, with more than $300 million in further milestones and double-digit royalties still possible. The collaboration has reached six progress-based milestones across five programs. The candidate's clinical plan is in Chapter 2; the same-day $200 million offering is in Chapter 5.
3.3 The wider partner map: Novo, Lilly, NVIDIA and the licensing backdrop
Several big names produced little dealable news in the window. Novo Nordisk has no new AI deal in this material. Its Anthropic collaboration was dated Sep 16 and its Orbis Medicines macrocycle pact (up to $1.4B) Sep 17, both before the window. Novo told reporters on Sep 22 it has more than 30 strategic AI partnerships (reported by Axios, via Pharmatica). Lilly appears through its platform: insitro said on Oct 6 that its models for predicting how small molecules behave in the body are available in Lilly's TuneLab, with no financial terms stated. Lilly and NVIDIA's joint AI lab, up to $1 billion over five years, dates from January 2026, so it is context only. In the window, NVIDIA is linked to Danaher on autonomous AI labs (date and terms not stated). GSK moved Chai Discovery from technology evaluation to a full partnership, with terms and date not stated. AstraZeneca and Oracle signed a non-binding agreement on AI and real-world data for trials, starting with hypertension in Abu Dhabi; terms are undisclosed.
Non-AI licensing ran far larger. BioBucks counted 157 BD&L transactions in 2026 as of Oct 2, with Merck paying SciBrunch $400M upfront (Sep 28), Novo paying Hengrui $300M upfront (Sep 29, up to $2.6B) and Sanofi paying Regeneron $1B upfront (Oct 1, up to $8B). Against those, the AI deals' upfronts of $30-40M are small. Pharma appears to pay modest cash for AI discovery access and large cash for assets that already exist.
4. Models and platforms
Chapter 3 covered what pharma pays to access AI platforms; this chapter looks at the models and agents behind them, and at who has actually tested them. The month's one open-weight release was RFD4-Proteina. Nearly every benchmark number below was produced by the model's own authors or vendor.
| Item | Key figure | Date | Status | Source |
|---|---|---|---|---|
| NVIDIA / UW IPD RFD4-Proteina | 435.6M parameters, BSD-3-Clause weights | Weights Oct 19; announced Oct 21, 2026 | Open, available | NVIDIA newsroom; Unite.AI |
| RFD4-Proteina, human TrkA binders | 158 of 188 designs bound = 84% | Oct 21, 2026 | Authors' preprint, not peer reviewed | R&D World |
| UW RFdiffusion3 (Science) | Active enzymes in zero-shot 96-construct campaigns | Oct 15, 2026 | Peer-reviewed, first release | Science |
| UW Prompteina | 9 binders; best 107 nM | arXiv Oct 9, 2026 | Preprint; NeurIPS 2026 accepted | Tech Times |
| Princeton T-REX | 2.43× vs tree search; 2.48× vs best generator | Sep 24, 2026 | Preprint; open source | bioRxiv |
| ConforFlux on Boltz-2 | +3 to +17 points per-state success | v2 Oct 19, 2026 | Preprint; NeurIPS 2026 | bioRxiv |
| Claude binder-design eval | Opus 5.5 82.6%, Sonnet 5.5 82.3% | Oct 23, 2026 | Vendor self-report | BenchLM |
| Schrödinger Bunsen | ~80% of HPC modeling jobs run by agent | Launched end of July 2026 | Early access | BiopharmaTrend |
| Insilico inClinico Agent | AUROC 0.828 (efficacy), 0.809 (safety) | Oct 22, 2026 | Launched, enterprise clients | Insilico release |
| DaltonTx antibody platform | 87,000 structures per hour | Oct 1, 2026 | Launched | DaltonTx via GlobeNewswire |
4.1 RFD4-Proteina: the month's open model
NVIDIA announced RFD4-Proteina on Oct 21, 2026 at GTC Berlin as a BioNeMo open model built with the University of Washington Institute for Protein Design (IPD). Status: available. The model generates sequence and all-atom structure together. Its model card dates the weights to Oct 19 and Diffusers support to Oct 20, and the NIM listing on build.nvidia.com and NGC is also dated Oct 20.
| Release fact | Value | Source |
|---|---|---|
| Licence | BSD 3-Clause for model, weights and code; optional third-party components carry separate terms | R&D World |
| Size | 435.6M parameters; 1.74 GB in FP32 (435.6M × 4 bytes ≈ 1.74 GB) | Hugging Face card via Unite.AI |
| Speed, one H200 | 250-token design in 5.0 s (BF16), 2.8 s after regional compilation | Hugging Face card via Unite.AI |
| Speed vs RFdiffusion3 | 200 residues in 1.35 s vs 5.12 s, i.e. 5.12 ÷ 1.35 ≈ 3.8×; peak memory 8.50 GB to 1.36 GB (about 6.3× less) | NVIDIA benchmark via R&D World |
| Distribution | Hugging Face, GitHub, NVIDIA hosted service; partners include Amazon Bio Discovery, Anthropic's Claude Science, Benchling, Databricks, Google Cloud, Nebius, OpenAI, Tamarind Bio | NVIDIA newsroom |
Who evaluated it. The authors did. NVIDIA says the model beat the evaluated baselines on four computational suites: protein binding, small-molecule binding, enzyme scaffolding and DNA binding. According to R&D World's reading of the 116-page preprint, which has not been peer reviewed, it produced 5.0 to 10.4× as many distinct passing structural clusters per GPU-hour as the next-best method, with AlphaFold3 in the evaluation loop. GEN reports the developers' summary as five- to ten-fold higher success rates and speed. These agree. The 3.8× figure against RFdiffusion3 is a narrower, single-baseline speed test.
The laboratory results also come from the authors:
- TrkA binders: 158 of 188 designs bound with affinity below 1 µM, pooled over two campaigns. 158 ÷ 188 = 84%. The preprint compares this with 18 of 60 (30%) for an August study that used Claude Mythos Preview to coordinate public tools, so the comparison is the authors' own and the two campaigns differ in design.
- Pesticide-degrading enzymes: 89 of 288 designs broke down paraoxon (31%).
- Lenacapavir synthesis step: 4 of 67 designs (6%) formed the desired amide. The authors call it, to their knowledge, the first de novo enzyme proof of concept for an active-pharmaceutical-ingredient synthesis step.
RFD4-Proteina is the only open-weight, permissively licensed model in this month's material, but its evidence is still the authors' own preprint and no independent benchmark appears.
4.2 Other design research: who tested what
RFdiffusion3 appeared in Science on Oct 15, 2026 (first release). The authors report active de novo phosphotriesterases and zinc proteases specific for transthyretin in zero-shot, 96-construct campaigns. Licence not stated.
Prompteina accepts design goals written in plain English. According to Tech Times, it was posted to arXiv on Oct 9, 2026 and accepted at NeurIPS 2026. Its authors report nine de novo binders to an unnamed rare-disease target, the best at 107 nM. They also report a lead on at least 13 of 20 targets and a 14% gain in interface confidence over the next-best baseline. These are the authors' computational metrics via a secondary outlet, so treat them as unverified. Licence not stated.
T-REX (Princeton, with MIT) uses LLM agents to decide where a binder-design campaign spends compute. The bioRxiv version posted Sep 24, 2026 reports the most structurally distinct hits on all seven targets under matched compute: a geometric mean of 2.43× over a non-agentic tree-search controller and 2.48× over the strongest generator-only baseline per target. Code is open source on GitHub, licence not stated.
Claude on binder design. BenchLM's page dated Oct 23, 2026 lists Claude Opus 5.5 at 82.6% and Claude Sonnet 5.5 at 82.3% on a 15-target composite binder-design task (mean over five runs per target). BenchLM says most rows are provider self-reports, labels the table display-only and does not rank on it. New Market Pitch reports 17 validated scorpion-toxin binders from 211 Claude-assisted designs (17 ÷ 211 ≈ 8%) in October 2026.
Structure prediction. The month's material has no new AlphaFold, OpenFold or Boltz release. The one item is ConforFlux, a third-party, training-free method for Boltz-2. Its Oct 19 bioRxiv version reports +3 to +17 percentage points of per-state success over default Boltz-2 on four conformational-change categories, with higher clashscores, and gains on OpenFold3 and Protenix. Boltz is also named among companies building agents on NVIDIA BioNeMo. For Isomorphic, IsoDDE was announced on Feb 10, 2026 and remains proprietary. The company claims more than twice AlphaFold 3's accuracy on a hard protein-ligand benchmark. On Sep 29 its president described a "Target D" run in which one molecule gave a working concentration-response curve; BiopharmaTrend notes that no IND had been announced as of Sep 30, 2026 (see Chapter 2).
4.3 Agentic platforms: launches with vendor-run validation
Schrödinger Bunsen is an agentic co-scientist that plans and runs the company's physics-based simulations. BiopharmaTrend says it launched at the end of July, before this window. Status: early access with selected customers, including a strategic software agreement with Bristol Myers Squibb. Full commercial release is expected before the end of 2026, with NVIDIA and Google Cloud supplying compute. Chief scientific officer Robert Abel says about 80% of modeling jobs on Schrödinger's HPC now run through Bunsen, and the scientist reviews the plan first. This is a company statement about internal use, not a benchmark. The material shows no October launch.
Insilico inClinico Agent (Oct 22, 2026) is a trial-intelligence agent, not a molecule designer. It uses a database of more than 100,000 trials and two gradient-boosted models. The release reports these figures:
| inClinico Agent metric (Insilico) | Value |
|---|---|
| Efficacy endpoint model, held-out trials from 2021 on | AUROC 0.828 (n = 2,347) |
| Safety/toxicity endpoint model | AUROC 0.809 (n = 1,485) |
| 939 trials completed in 2025 | AUROC 0.805, accuracy 0.734 |
| Prospective ClinBench | 3,311 ongoing Phase I–III trials, completing by Dec 31, 2027 |
| Predicted success: five frontier LLMs vs agent | 75–82% vs about 50% (53% historical rate) |
"Held-out" describes the data, not the evaluator: Insilico built the benchmark, ran the comparison and published it on medRxiv. ClinBench has no outcomes yet, so it tests nothing until the trials finish. The agent is sold to industry partners and enterprise clients.
DaltonTx (Oct 1, 2026) added antibody discovery to its Dalton platform. It says it folded the 2.6 million paired OAS sequences at 87,000 structures per hour. That is 2,600,000 ÷ 87,000 ≈ 30 hours. The release gives throughput, not accuracy.
Judgement. The month's real technical news is open, composable design tools (RFD4-Proteina, T-REX) that agents can call. Commercial agents (Bunsen, inClinico, Dalton) are launches whose validation is internal. For developers, the BSD-3 weights are testable today. For buyers, ask for an external run before treating any of these numbers as comparable.
5. Funding
Chapter 4 covered the models and platforms that investors are pricing; this chapter follows the money from Sep 24 to Oct 23, 2026 and separates what has closed from what is only reported.
| Item | Key figure | Date | Status (as of 2026-10-24) | Source |
|---|---|---|---|---|
| Isomorphic Labs, new round | Valuation of at least $40B, possibly up to $50B; size and investors not reported | Oct 8, 2026 | Reported, in early talks, not closed | Bloomberg; TFN |
| Recursion, follow-on offering | $200M gross; 61,540,000 shares at $3.25 | Announced Oct 15; closed Oct 16, 2026 | Confirmed (company release, law firm) | Recursion via GlobeNewswire; Wilson Sonsini |
| Iambic Therapeutics, Nasdaq IPO | About $150M; 9.4M shares at $16; valuation above $800M | Priced Oct 14; trading from Oct 15, 2026 | Completed, per press | pharmaphorum; Endpoints |
| Biossil | $153M at a $1B valuation, led by OpenAI's fund | Week of Sep 29 to Oct 5, 2026 (exact date not stated) | Reported, newsletter only | The Signal |
| Xaira Therapeutics | No new round; $1B raised at the 2024 launch | Oct 3, 2026 update | No financing in the window | The Catalyst Brief |
5.1 Isomorphic Labs: a valuation target, not a round
On Oct 8, 2026 Bloomberg reported that Alphabet's Isomorphic Labs is in early talks to raise at a valuation of at least $40 billion, with one source saying it could reach $50 billion. Status: reported. Nothing has closed. TFN notes that neither the round size nor the investors have been reported, and that terms could change. Isomorphic did not respond to a request for comment, and no company statement exists in the material.
The previous round is better documented. TFN says Isomorphic closed a $2.1 billion round on May 12, 2026, led by Thrive Capital, with Alphabet and GV returning and MGX, Temasek, CapitalG and the UK Sovereign AI Fund joining. That followed a $600 million round in March 2025, so external funding totals 2.1 + 0.6 = $2.7 billion, a figure TFN and Crypto Briefing share.
The sources disagree on the earlier valuation. Crypto Briefing says it was "approximately $17.5 billion". TFN and Startup Fortune both say the May round's valuation was undisclosed. New Market Pitch gives an implied $12–20 billion (estimate). We follow "undisclosed" and treat $17.5 billion as an unverified secondary figure. On that figure, $40B ÷ $17.5B ≈ 2.3×. On the New Market Pitch range, the step-up is 2× to 3.3×.
A $40 billion target values Isomorphic at roughly 10 times Chai Discovery's confirmed $3.8 billion, before any dosed patient. It is a negotiating position until documents are signed.
The 10× comparison is TFN's: $40B ÷ $3.8B ≈ 10.5, and $40B ÷ $0.8B = 50 against Basecamp Research. Several outlets note that Isomorphic has not dosed a patient, and that first-in-human trials are planned for the end of 2026 (see Chapter 2). Partner terms are in Chapter 3.
5.2 Public-market raises: Recursion and Iambic
Recursion (Nasdaq: RXRX) announced a $200 million underwritten offering of common stock on Oct 15, 2026, in the same release as REC-4881 and REC-1245 data and a Sanofi option (see Chapters 2 and 3). Buyers named by the company are Janus Henderson Investors, Invus, Longaeva Partners, ARK Invest and Driehaus Capital Management. The release calls it "oversubscribed" but gives no demand figure, so that word is the company's description only.
Pricing and closing are confirmed. StockTitan reports 61,540,000 Class A shares at $3.25, with Leerink Partners, Goldman Sachs, Morgan Stanley and Evercore as joint bookrunners. Check: 61.54M × $3.25 = $200.0M. Wilson Sonsini, Recursion's counsel, states that the offering closed on Oct 16, 2026 at about $200 million gross, before discounts and expenses. Recursion says its expected cash runway now extends into 2029, including these proceeds. The material truncates the cash balance, so we cannot say how many months the $200 million adds. StockTitan records RXRX closing +11.23% on Oct 15.
Iambic Therapeutics, a Nvidia-backed AI oncology company, priced a $150 million IPO on the evening of Wednesday Oct 14, 2026 and rose on its first trading day (Endpoints, paywalled after the first lines). According to pharmaphorum, it sold 9.4 million shares at $16, the midpoint of its range, for a valuation above $800 million. The arithmetic checks out: 9.4M × $16 = $150.4M. The proceeds go mainly to IAM1363, an oral HER2 inhibitor in Phase 1, and Iambic is pitching a registrational trial for next year. Shares traded about 13% above the IPO price, at a little over $18. That quote is a snapshot taken when pharmaphorum published on Oct 16.
5.3 Private-market context and what did not happen
Only one private round in the window appears in the material, and it is thinly sourced. The Signal newsletter reports that OpenAI's fund led a $153 million round at a $1 billion valuation for Biossil, whose thesis is that part of pharma's failure rate comes from poor patient matching. The round is reported as the newsletter's own item, with links to The Globe and Mail and BetaKit that we cannot verify. We treat it as reported.
The closest AI-native rounds sit just outside the window, so they are context only:
- Basecamp Research raised a $140 million Series C at an $800 million valuation on Sep 23, 2026, one day before the window opens (Reuters via TFN).
- Chai Discovery raised $400 million at $3.8 billion in a Series C announced Jul 14, led by Index Ventures, with total funding above $600 million.
- ByteDance spin-off Anew Labs reportedly closed a $290 million first round at about $1.5 billion. The exact date is not stated. 36Kr, published Sep 26, places it a few days before the Sep 18 policy plan, and the investors named are HSG, IDG Capital and others.
Xaira did not raise. On Oct 3, 2026 it disclosed two candidates, XA-1 and XA-4, and CEO Marc Tessier-Lavigne said it was capitalized so as not to return to the markets until it had molecules in the clinic. He would consider raising sooner if investor interest justified it. The Catalyst Brief notes that Xaira did not disclose its cash balance. We found no new Generate:Biomedicines or Chai round in the window. BiopharmaTrend says only that Generate raised $400 million on Nasdaq, with no date stated.
New Market Pitch (last updated Sep 24, 2026) ranks valuations from one dataset, so these values can be compared directly:
| Company | Valuation | Basis in that dataset |
|---|---|---|
| Isomorphic Labs | $12.0B–$20.0B | Implied from raise (estimate) |
| XtalPi | $4.9B | Public market cap |
| Insilico Medicine | $4.7B | Public market cap |
| Chai Discovery | $3.8B | Announced private round |
| Xaira Therapeutics | $2.7B–$3.5B | Proxy-based estimate |
| Recursion | $2.2B | Public market cap |
| Generate:Biomedicines | $2.2B | Public market cap |
The same source warns that capital raised is a weak shortcut for valuation. Eikon has raised more than $1.1 billion yet trades far below that level. Judgement: this month's money went to a listed company with clinical data (Recursion) and a listed oncology company with a Phase 1 asset (Iambic). The largest private price signal, Isomorphic, has no closed round behind it.
6. Regulatory discussion
Chapter 5 showed that the month's money followed listed companies with clinical data; this chapter asks what regulators and reviewers would make of the same record. The earlier chapters contain no new FDA or EMA instrument, no credibility framework and no guidance document for AI in drug development, so none is described here. What they do contain are regulatory touchpoints: registry entries, pending agency discussions, the kind of evidence offered, and who produced it. This chapter uses only those, and it is a reading of the record, not legal or medical advice.
6.1 What the record says about evidence
Trials are judged as trials, not as AI. Insilico's own framing is that rentosertib is the first drug with both an AI-identified target and an AI-designed molecule in Phase 3, and the source adds that "AI-discovered" is not a regulatory category. Generate:Biomedicines began a Phase 3 of GB-0895 on Dec 3, 2025, before this window. The registry record for rentosertib, NCT07687459, shows a start on Sept 9, 2026 and 320 planned patients, but our reading is second-hand and the sources disagree on the start month. The only efficacy evidence remains the Phase 2a study, with small groups and liver-related adverse events that raised concern.
Agency discussions gate the next step. Recursion targets a registrational trial for REC-4881 in 2H 2027, pending the outcome of FDA discussions. That is a plan, not an agreement. Its safety statement is a company statement: no treatment-related event above Grade 3, and four patients stopped for adverse events. The release reports no comparator, so the dataset cannot show how much of the effect is the drug. Recursion has also said it will seek partners for REC-617 and REC-3565, which is reprioritisation, not discontinuation. No dated discontinuation fell inside the window.
Base rates matter more than labels. A peer-reviewed count put Phase 1 success for AI-enabled assets at 77.8% and Phase 2 at 19.5% as of March 1, 2026, with five of 129 assets in or past Phase 3. A reviewer reading a new Phase 1 dataset should start from those rates, not from the platform's claims.
Model evidence is mostly vendor-run. RFD4-Proteina's laboratory results come from a preprint that is not peer reviewed. Insilico built and ran its own inClinico benchmark, and its prospective ClinBench set of 3,311 trials cannot test anything until the trials finish by Dec 31, 2027. BenchLM labels its Claude binder-design table display-only.
The month produced rules of evidence only by implication: registry entries, pending FDA discussions and unreplicated, author-run model results. Decisions should rest on the dated trial record and on independent checks, not on whether a company calls a molecule AI-discovered.
6.2 What changed and what it implies, by reader
| Reader | What changed | Implication | Evidence |
|---|---|---|---|
| Developers | A Phase 3 start (rentosertib) and a Phase 1b/2 dataset with a registrational plan (REC-4881) arrived in one month | Expect to be judged on trial endpoints and base rates; plan for FDA discussions before any registrational start | — |
| Pharma partners | New deals paid 2.5-3.6% of headline value upfront where disclosed, and Astellas did not disclose its upfront | Treat headline values as ceilings; ask for disclosed upfronts, milestone triggers and who runs validation | — |
| Investors | Recursion's offering closed with data in hand; Isomorphic's $40B is only reported | Separate closed financings from reports, and clinical data from platform claims | — |
| Model users and buyers | An open-weight model and several agents launched, with benchmarks run by their authors or vendors | Ask for an external run before treating any figure as comparable | — |
6.3 Open questions the record cannot answer
Four gaps recur. The material gives no company-stated readout date for rentosertib. It gives no round size or investors for Isomorphic. It does not say how many months Recursion's $200 million adds to its runway beyond "into 2029". And no independent benchmark of RFD4-Proteina appears. Each is a place where a later disclosure, such as a registry update or a peer-reviewed paper, would change the picture. Two further caveats apply to the deal evidence. Where upfronts are undisclosed, as with AI Proteins and Astellas, no upfront share can be computed, and a BioBucks tracker warns that a ceiling with no cash figure says little about what was committed. Where media headlines differ, as with the $1.5 billion AstraZeneca-insitro headline (40 + 1,460) and Bloomberg's unexplained $1.86 billion, the company releases are the better guide.
6.4 Key dates
Dates already past on Oct 24, 2026 are marked "In effect"; later ones are "Upcoming".
| Date | Event | State on Oct 24, 2026 | Source |
|---|---|---|---|
| Dec 3, 2025 | Generate:Biomedicines begins GB-0895 Phase 3 | In effect | rentosertib.co.uk |
| Sept 9, 2026 | Rentosertib Phase 3 start (registry, marked actual) | In effect | The Signal |
| Oct 15, 2026 | Recursion reports REC-4881 and REC-1245 data, Sanofi option, $200M offering | In effect | Recursion |
| Oct 16, 2026 | Recursion offering closes | In effect | Wilson Sonsini |
| Oct 20, 2026 | Servier first milestone with Insilico | In effect | Insilico |
| Oct 21, 2026 | RFD4-Proteina announced; AI Proteins and Astellas sign | In effect | NVIDIA newsroom; AI Proteins |
| Nov 2, 2026 | More REC-4881 data at the CGA-IGC meeting | Upcoming | Recursion |
| End of 2026 | Isomorphic first-in-human target; Schrödinger Bunsen commercial release expected | Upcoming | TFN; BiopharmaTrend |
| 2027 | Sanofi-optioned candidate Phase 1 planned; Iambic registrational trial pitched | Upcoming | Recursion; pharmaphorum |
| 2H 2027 | REC-4881 registrational trial targeted, pending FDA discussions; REC-1245 endometrial expansion data | Upcoming | Recursion |
| Dec 31, 2027 | ClinBench trials due to complete | Upcoming | Insilico |
7. References
This document draws on the following 57 sources, listed in the order they are first cited.
- Breaking the Safety Barrier: Insilico Medicine Nominates ISM1354, the Next-Generation Small Molecule GIPR Antagonist — insilico.com
- The Signal: Bio/Health — Edition #8 — October 6, 2026 • Buttondown — buttondown.com
- Recursion Pharmaceuticals, Inc. - Recursion Announces $200 Million Oversubscribed Follow-On Offering, Clinical Data Updates, and Partnership Milestone — ir.recursion.com
- Artificial intelligence–enabled drug discovery: A growing but largely early-stage clinical pipeline — sciencedirect.com
- The $2.7B Question: How Do We Know If Isomorphic's AI Works? | BiopharmaTrend — biopharmatrend.com
- AI-driven Molecular Glue Innovation: Insilico Nominates ISM9546, an Oral, Peripheral-Restricted VAV1 Degrader as PCC for Autoimmune Diseases — insilico.com
- insitro Announces a Strategic Collaboration with AstraZeneca for AI-enabled Discovery of Novel Therapeutics for MASH and Cardiomyopathies - insitro — insitro.com
- Insilico Medicine and Servier Achieve First Milestone in AI-Driven R&D Collaboration, Triggering An undisclosed Milestone Payment — insilico.com
- 5 things to know in life sciences: Week of Oct. 12, 2026 — realeconomy.rsmus.com
- AI Proteins Announces Research Collaboration and License Agreement With Astellas Pharma to Discover and Develop Novel Miniprotein-Based Therapeutics Valued at Over $400 Million — finance.yahoo.com
- Recursion Pharmaceuticals, Inc. - Recursion Announces Sanofi Options Potential First-in-Class Immunology Development Candidate from Their Collaboration — ir.recursion.com
- Biotech BD&L Tracker 2026: 157 Licensing & Partnering Deals — biobucks.co
- Healthcare and Life Sciences Leaders Build on NVIDIA to Advance Drug Discovery and Expand Patient Care | NVIDIA Newsroom — nvidianews.nvidia.com
- NVIDIA Debuts Open Protein Design Model RFD4-Proteina with UW Medicine – Unite.AI — unite.ai
- RFD4-Proteina: NVIDIA and UW's open protein design model — rdworldonline.com
- Insilico Medicine introduces inClinico Agent: An agentic AI system designed to accelerate clinical trial intelligence and predict trial outcomes — insilico.com
- Agentic campaign control for high-throughput de novo ... — biorxiv.org
- ConforFlux: Particle-Guided Trunk Repulsion for Diverse ... — biorxiv.org
- Wilson Sonsini Advises Recursion on $200 Million Follow-On Offering | Wilson Sonsini — wsgr.com
- Recursion prices about $200M stock offering at $3.25/share | RXRX Stock News — stocktitan.net
- AI in drug discovery & development: News watch, October 2026 | pharmaphorum — pharmaphorum.com
- Watch Google DeepMind's AI Drug Spinoff, Isomorphic Labs, Seeks Funding at Valuation of $40 Billion - Bloomberg — bloomberg.com
- DeepMind spinout Isomorphic Labs eyes $40B valuation less than five months after $2.1B round: Report — TFN — techfundingnews.com
- Xaira Reveals XA-1 And XA-4, Framing X-Design As A Drug-Ready Antibody Engine | The Catalyst Brief — thecatalystbrief.com
- How AI was used to discover rentosertib | Rentosertib UK — rentosertib.co.uk
- Has AI discovered a real drug? – New Market Pitch — newmarketpitch.com
- AI Can Design a Drug. It Still Can’t Prove It Works. | 01 — vocal.media
- Study Details | NCT07687459 | Study Evaluation Rentosertib (INS018_055) Administered Orally in Patients With Idiopathic Pulmonary Fibrosis (IPF) | ClinicalTrials.gov — clinicaltrials.gov
- Insilico Medicine and Servier Achieve First Milestone in AI-Driven R&D Collaboration, Triggering An undisclosed Milestone Payment - BioSpace — biospace.com
- Recursion Announces $200 Million Oversubscribed Follow-On — globenewswire.com
- Isomorphic Labs reportedly seeks funding at a valuation of at least $40 billion — cryptobriefing.com
- AI Protein Design Market Size, Share & Forecast 2034 — polarismarketresearch.com
- AstraZeneca announces new strategic collaboration with insitro to use advanced AI systems to develop transformative therapeutics for difficult to treat cardiometabolic diseases — astrazeneca.com
- AI Proteins signs miniprotein collaboration with Astellas worth over $400 million — thepharmaletter.com
- AZ and Oracle team up on AI platform for clinical trials | pharmaphorum — pharmaphorum.com
- AstraZeneca & insitro MASH Deal: $40m AI Collaboration | Pharma Journal — pharma-journal.com
- AstraZeneca Taps insitro AI for MASH in Deal Worth Up to $1.5 Billion — pharmexec.com
- AI Proteins, Astellas Strike $400M+ Deal for Miniprotein Drug Discovery - Pharmacally — pharmacally.com
- Plain-English AI Produces Validated Drug-Candidate Proteins Against Rare Disease Target — techtimes.com
- Servier Pays Insilico First Milestone in $888M AI Deal — peptideandprotein.com
- Novo Nordisk account brief | Life Sciences Radar | Ortent — ortent.co
- Alliances - BioSpace — biospace.com
- Novo Bets on AI Drug Discovery as GLP-1 Patents Near Expiry | Pharmatica — pharmatica.io
- Big Pharma Turns to AI to Find New Drugs - Pharma Journalist — pharmajournalist.com
- Atomic functional site scaffolding in biopolymers with RFdiffusion3 | Science — science.org
- De novo protein-binder design Leaderboard — October 2026 — benchlm.ai
- When Agents Run Computations: Schrödinger’s Robert Abel On Where The Human Still Decides | BiopharmaTrend — biopharmatrend.com
- DaltonTx, the decision engine for drug discovery, launches — globenewswire.com
- Nvidia, UW Institute for Protein Design Unveil Open-Source Protein Design AI Model — genengnews.com
- What Is Isomorphic Labs? Alphabet AI Drug Discovery Company — layer3labs.io
- Updated: Iambic rises after $150M IPO for AI-driven cancer pipeline — endpoints.news
- Alphabet's Isomorphic Labs seeks valuation of at least $40 billion in new funding round — ca.finance.yahoo.com
- Isomorphic Labs Funding Talks Put a $40 Billion Valuation Ahead of Clinical Proof — remio.ai
- Isomorphic Labs seeks funding at a valuation of at least $40 billion - Startup Fortune — startupfortune.com
- Top AI in Drug Discovery Startups by Valuation (2026) – New Market Pitch — newmarketpitch.com
- AI Drug Discovery: On the Eve of Full Value Realization in Modern Pharmaceutical Innovation — eu.36kr.com
- 13 AI Drug Discovery Companies to Watch in 2026 — techlifesci.com