Medra launches an AI layer that plans and revises experiments for its robot lab, with a DARPA-funded project to turn written protocols into machine runs
What happened: A scientist sets a goal in chat; the AI Experimentalist drafts the plan, sends it to Medra's robots, reads the results and revises the protocol. Partners use it on site or through ML001, the 38,000 sq ft San Francisco lab Medra opened in April, where closed-beta work covers antibodies, protein engineering, gene editing and cell biology. It runs NVIDIA's Nemotron models and BioNeMo agent tools, among others.
Why it matters: Medra's bet is that testing, not ideas, is now the slow step, so the reasoning should sit next to the robots. Its one worked example is its own: an antibody binding assay cut from three days to about 14 hours by dropping a two-day cloning step and tuning the rest, with no data shown. DARPA's money says the government wants protocols turned into robot runs; it is not a result.
Bayer signs a drug-discovery collaboration with Iambic to use its Enchant and NeuralPLexer models on hard-to-drug targets
What happened: Bayer will discover small-molecule medicines using Iambic's platform, including its Enchant and NeuralPLexer models, aimed at targets that are hard to drug and at hit molecules that differ from known ones. Iambic gets an upfront payment plus milestones and royalties; no amounts, targets or number of programmes are given. Juergen Eckhardt, Bayer's head of business development and licensing, and Iambic chief executive Tom Miller spoke for the companies.
Why it matters: Iambic now has three pharma platform deals on the record this year, with Takeda in February, Bayer in June and AbbVie in September, and its IPO priced on 14 October. Bayer is paying for access to the models, a different arrangement from handing molecules over for milestones on a single asset.
Evidence and limits: Only the terms in the company's own release: no amount, no target, no number of programmes, and no result from the platform for Bayer yet.
Flex-Cat closes a catalyst-optimization loop and scales the result tenfold
What happened: The autonomous laboratory runs 680 experiments across three catalyst-optimization campaigns and transfers selected results from 2 mL discovery runs to a 20 mL reactor.
Why it matters: The important step is not autonomous search alone, but translation from miniaturized discovery experiments toward process-relevant conditions.
John Jumper leaves Google DeepMind to join Anthropic
What happened: John Jumper, who led AlphaFold and shared the 2024 Chemistry Nobel, said on X he is leaving Google DeepMind after nearly nine years to join Anthropic, after a break. Bloomberg, which broke the news, says he was a key member of Google's AI coding team.
Why it matters: The person who led AlphaFold is moving to a lab that has published its own science results this year, including protein binders Claude designed and a lab tested. His work at Anthropic is not yet public.
Insilico doses the first person in a Phase 1 trial of ISM8969, an NLRP3 inhibitor it designed with AI and co-owns with Hygtia Therapeutics
What happened: The single-centre, randomised, placebo-controlled trial in Australia will test single and multiple rising oral doses in about 80 healthy volunteers and 20 obese adults at risk of heart disease, for safety, tolerability and how the drug is absorbed and acts. ISM8969 is meant to reach the brain to treat chronic neuroinflammation, including Parkinson's disease. Under the deal, Hygtia, incubated by a Shenzhen fund and Fosun Pharma, holds half the global rights, and Insilico can receive up to $66M.
Why it matters: It adds a brain-penetrant inflammation drug to the Insilico molecules in human trials, under a deal in which a partner holds half the global rights. Insilico counts it as that deal's first clinical milestone.
Evidence and limits: Dosing a first volunteer shows the trial has started, not that the drug is safe or works; the 'AI-driven' design is the company's own description.
Noetik's TARIO-2 picked out likelier responders to Agenus's BOT+BAL immunotherapy from routine biopsy slides, looking back at 49 colorectal cancer patients in a trial
What happened: In a white paper with Agenus, Noetik's Ryan Dalton, Ron Alfa and colleagues use TARIO-2, trained on paired tumour images and spatial gene-expression data, to infer gene activity from standard H&E slides taken before treatment in the Phase 1b trial C-800-01. Among 49 patients with microsatellite-stable colorectal cancer and no active liver metastases, the 30% the model ranked highest had a 64% response rate against 9% for the rest (AUROC 0.81, permutation p=0.04); the survival hazard ratio was 0.18 in 47. It beat two open H&E-only pathology models.
Why it matters: It tests whether a model that learned spatial biology can read a drug response off the cheapest slide in cancer care. The answer here is a hindsight analysis of a small, fully enrolled trial: 49 patients in the main group, a model fitted to those patients' own outcomes under cross-validation, written by Noetik with the drug's own developer and not peer-reviewed. Ovarian cancer and sarcoma looked similar but did not reach significance. Only a prospective trial that selects patients this way would show it works.
Radical Numerics, started by the team behind the Evo and Evo 2 DNA models, launches with a $50M seed round led by Emergence Capital to build one model that reads DNA, RNA and proteins together
What happened: Obvious Ventures, Triatomic Capital, Factory and First Spark Ventures joined, and Patrick Collison backed it at pre-seed, per Fortune. Eric Nguyen is CEO, with Michael Poli, Stefano Massaroli and Armin Thomas. The money is for larger models, more researchers and compute; a data center of NVIDIA Blackwell chips is under construction. It has two early partnerships, a cancer-detection company and a US national lab for pathogen detection, and says its revenue mix is still taking shape: API licences, tuned models for pharma, milestone payments.
Why it matters: The money goes to a lab that treats biology as one language across DNA, RNA and proteins, where most funded biology-AI companies work on a single modality. It also comes from the people who built an open DNA model that others used to generate a working virus genome, so the company is pairing design with defence from the start.
Evidence and limits: No valuation is disclosed. The company's own post gives the amount and no investor list; the investors and the use of funds come from Fortune's exclusive and JustAINews, which quotes Emergence's Gordon Ritter. Both partners are unnamed, and the revenue model is, in the CEO's words, not worked out.
Radical Numerics previews Omnii, a genome language model that scores noncoding variants above CADD and Evo 2 on the company's own benchmark runs
What happened: Omnii takes DNA plus conservation tracks, with a 2 million-base context, and is post-trained so it needs no probe on frozen embeddings. On ClinVar noncoding single-letter variants it scores AUROC 0.975 against 0.909 for CADD and 0.911 for Evo 2 40B; on copy-number variants 0.773 against 0.664 for CADD-SV; on TraitGym complex traits AUPRC 0.410 against 0.297 for Borzoi. The post also shows a design method and attention maps; Omnii is in early access.
Why it matters: Base genome models have trailed hand-built scores like CADD on variant interpretation, which is the applied task geneticists care about. This is the first report from Evo's authors that a post-trained genome model passes those scores on noncoding variants, where most disease-linked variation sits.
Evidence and limits: Every number is the company's own: it re-scored the baselines on a protocol borrowed from the EVEE paper, reproduced EVEE's probe itself, and compares only against open-source sequence-to-function models. No outside group has run it, and the Omnii preview posts on biodefence and cancer vaccines were not read.
Shenzhen's Jindu Shengke raises nearly 100M yuan in a Series A for an omics model tied to agents that run lab instruments
What happened: Gaotejia Investment led; no other investor and no valuation are named. Founder and CEO Yongcheng Jin describes two parts: GeneLLM, a multi-omics model pretrained on individuals' raw RNA, protein and metabolite data, which the company puts at 1.5 billion parameters and 3.5 trillion bases; and BioFord Agent, agents that design experiments and schedule sequencers, liquid handlers, PCR and mass-spec instruments from different vendors. The money goes to its physics-AI architecture, the agent platform and overseas markets.
Why it matters: Jindu Shengke sells biology labs the closed loop: the model proposes, instruments test, results come back. Every number is the company's own, and it shows no benchmark for GeneLLM, no paper and no experiment the loop has run. What the round does show is a fund that invests only in healthcare paying for lab agents.
Terray's TerraSynth designs molecules with a route to make them, and in Terray's own projects chemists made half its compounds in 4 weeks instead of 8
What happened: A 2-billion-parameter model built on Terray's COATI proposes each molecule with a route drawn from 260 reliable reactions. On 1,000 ChEMBL molecules it rebuilt 53.8%, against 32.1% for ReaSyn, which takes about 1,000 times longer; PrexSyn, as fast, rebuilt 24.0%. In 5 Terray projects, half its designs were made within 4 weeks, against 8 weeks in 8 earlier projects whose designs had no route; in 4 later projects, with a chemist picking 15 of every 22, 90% were made in 4 weeks.
Why it matters: Making a molecule is the slowest step in drug design, and AI designs often stall there. This is the company's own post, not a paper: the project comparison is before and after a date cutoff, not randomised, over 244 compounds, and the check that designs stay potent used Terray's own model, not an assay.
Recursive reports its first results from an automated AI research system, beating the best public scores on two small-model training benchmarks and a GPU-kernel benchmark
What happened: Recursive Superintelligence's system proposes an idea, implements it, runs the experiment, checks the result and picks the next one. It reports lowering validation loss on Karpathy's five-minute NanoChat task from 0.9372 to 0.9109 bits per byte, beating the community's best; cutting the NanoGPT Speedrun from 79.7 to 77.5 seconds; and raising the mean score on NVIDIA's 235-kernel SOL-ExecBench from 0.699 to 0.754. It is open-sourcing the run outputs.
Why it matters: Recursive raised $650M to build AI that improves AI, and its founders, Richard Socher, Tim Rocktäschel and Jeff Clune among them, say the target is automating discovery in science. These are its first public results: an automated loop that beat the best public scores, one of them on a speedrun humans have tuned for two years, on tasks chosen because each experiment is fast and the score is clear. Science rarely offers that, which is the gap between this and a lab.
Evidence and limits: Every number is Recursive's own. The speedrun time was measured on rented nodes, not the leaderboard's official hardware, and had not been submitted. The company says some candidates gamed the kernel benchmark and were filtered by its own checks. It also says the base models may already know the public techniques the system used.
XtalPi signs an unnamed international drugmaker to an oral small-molecule programme for a GPCR target, with a potential total value over $400M
What happened: In a voluntary HKEX filing, XtalPi says a pilot phase on a metabolic GPCR target led to the full deal: the partner pays an upfront sum and funds XtalPi's early R&D, and preclinical, clinical and commercial milestones plus royalties bring the potential total above $400M. The upfront is not stated. The target has no public high-resolution structure bound to a small molecule, so XtalPi says it simulated the receptor's shapes, screened hundreds of millions of compounds against them, and scored binding with free-energy calculations.
Why it matters: It asks whether simulation can stand in for a small-molecule-bound structure nobody has published, and a drugmaker paid to continue after seeing the pilot. But the partner, the target and the upfront are all undisclosed, the pilot's hit rates are called a breakthrough without a number, and most of the $400M depends on milestones. It shows what one buyer will pay for, not that the method works.
Anthropic tests Claude against ChemDraw and MestReNova on NMR prediction for 20 new compounds
What happened: David Kamber, a chemist at Anthropic, took 20 compounds from ChemRxiv preprints posted after the models' training cutoff. Predicting their 1H shifts, Opus 4.7 averaged 0.079 ppm off, the most accurate of the tools; on 13C it tied MestReNova, 1.37 against 1.48 ppm. Working backwards from formula and 1D spectra, it recovered all eight simpler structures on every run, and the seven harder ones, given the starting material, on two or three runs of three.
Why it matters: Reading spectra is one of the slowest routine steps in synthesis: ChemDraw and MestReNova only work forward, from a structure to its spectrum, and elucidation software needs 2D spectra. A general model with no chemistry fine-tuning that matches that software, and can also propose a structure from the 1D data a chemist would paste into a chat, means a chemist can start to hand NMR checking off and then audit it.
Evidence and limits: A small in-house test: 20 compounds in four scaffold families, 15 elucidation problems, three solvents, and no 2D NMR or stereochemistry. The harder structures needed the starting material as a hint, and the authors call the results indicative rather than precise.
Generalist AI raises $400M led by Radical Ventures for robot foundation models; a Form D later shows $198M more
What happened: Pete Florence and Andy Zeng, formerly of Google DeepMind, and Andrew Barry of Boston Dynamics founded the company in 2024. Radical Ventures leads the $400M; 8VC, Union Square Ventures, Hanabi Capital and Norwest are new, Fei-Fei Li and Naval Ravikant invest as angels, and NVIDIA, Spark Capital and Bezos Expeditions follow on. Total raised passes $500M. A Form D filed on 24 August reports $198.2M more sold from 7 August.
Why it matters: Robots that handle objects reliably are what a self-driving lab needs, and Generalist's bet is that robot skill scales with real physical interaction data, the claim behind its GEN-0 and GEN-1 models. The money goes to that data engine and to training compute, so its next models test whether the scaling holds.
Evidence and limits: The company states no stage or valuation; the $2B is SiliconANGLE's and TechCrunch's figure. The 8VC lead and $3B valuation for the August money come from TechCrunch's unnamed sources, not the filing. GEN-1's reliability figures are the company's own.
Alnylam pays Inceptive $30M upfront to design siRNA drugs with its generative models, in a deal worth up to $2B
What happened: Alnylam, the company that pioneered RNAi drugs, signed a collaboration with Inceptive, co-founded by Transformer co-author Jakob Uszkoreit. The $30M upfront is cash plus a purchase of Inceptive equity; the rest of the up to $2B depends on preclinical, regulatory and sales milestones. Inceptive's models are to model target mRNAs, explore siRNA sequences and chemical modifications, and predict which candidates will do best in preclinical models, for Alnylam to develop.
Why it matters: Alnylam has six approved drugs and says it holds more than 20 years of its own siRNA data, so the deal asks whether a general sequence model adds anything where that much in-house data already exists. Part of the upfront is an equity stake, and the release calls the joint exploratory results exceptional without giving a number: this shows what Alnylam will pay for, not that the models work.
Lila Sciences is in talks to raise about $2 billion at an $8.5 billion pre-money valuation, Bloomberg reports
What happened: People familiar with the matter told Bloomberg the Series B is anchored by CalPERS, California's public pension fund, and NVIDIA's NVentures. Lila and CalPERS declined to comment. As of 2026-10-07 neither Lila nor any investor has announced a close.
Why it matters: If it closes, the price would be several times what investors paid eight months earlier, set before Lila had published a result. A public pension fund anchoring the round would put retirement savings behind the bet that a lab can be a model's data source. It is still talks, from unnamed sources.
Genesis says its Pearl system, given only an unbound protein structure, posed 78% of 802 ligands within its best 25 tries in OpenBind's EV-A71 benchmark, above every co-folding model OpenBind tested
What happened: The Pearl system adds sampling at inference time and physics- and AI-based ranking to Pearl, Genesis's co-folding model. Given the protease sequence, each ligand's 2D structure and an apo structure released after Pearl's training cutoff, it had a pose meeting OpenBind's joint test (within 2 Å, physically valid, LDDT-PLI at least 0.8) among its top 25 for 78% of ligands. Within 1 Å it reached 60%, against 27% for Protenix, the best co-folding model OpenBind tested, and 52% for GNINA docking into the bound structure.
Why it matters: On this target a loop next to the pocket moves about 3.9 Å when a drug binds, and docking into the unbound structure succeeds under 5% of the time. A model that predicts that movement from sequence can be used on targets with no drug-bound crystal structure, which is where co-folding is meant to pay off.
Evidence and limits: This is Genesis's own run on one protease, posted on its blog with no code or model released; the baselines are OpenBind's published numbers. Best-of-25 assumes the right pose can be picked out; with one pose, its joint success within 2 Å was 55%.
Orbital Industries releases OrbMol-v2, an open molecular simulation model it scores ahead of the B3LYP and PBE0 functionals on a public chemistry benchmark
What happened: OrbMol-v2 adds learned atomic charges and an explicit long-range electric term to Orbital's earlier molecular model. On 5,152 reactions from the GSCDB benchmark its error was 1.62 times that of the best hybrid DFT functionals, which the company says beats the widely used B3LYP and PBE0 and is 21% lower than the next machine-learned model, at about a millionth of DFT's cost. Weights and code are open under Apache 2.0. The company says LLM agents wrote the first version of the new module.
Why it matters: Hybrid functionals such as B3LYP are the default way chemists compute molecular energies, and they are slow at the scale of proteins or electrolytes. A machine-learned model that matches them on a broad outside benchmark, with few large misses (one reaction off by more than 100 kcal/mol, against 59 for the earlier version), could become the first calculation a chemist runs. Without long-range charges, earlier models got even a separating ion pair qualitatively wrong.
Evidence and limits: Orbital ran the comparison itself, in a blog post with no paper. It dropped 28 of GSCDB's 139 subsets and every reaction involving a lone atom, which such models cannot compute, and scored the DFT functionals on the same subset. Its error is still larger than the best hybrid functionals', and it learned from calculations at one functional's level, ωB97M-V.
FieldQuantum raises several hundred million yuan in two angel rounds to put quantum computing behind an AI4S platform
What happened: InnoAngel Fund led the angel and angel+ rounds, with Guoqi Investment, Beigong Investment, BV Baidu Ventures and the Shuimu Tsinghua Alumni Fund joining and Lighthouse Capital as adviser. Founder Dingshun Lyu, one of Tsinghua University's first quantum-computing doctoral students and then seven years at Huawei and ByteDance's AI4S Lab, founded the Beijing company in January 2026. The money goes to core research, the product platform, industry deployments and hiring. No amount beyond "several hundred million yuan" was given.
Why it matters: The bet is that AI models for materials and chemistry are limited by the accuracy of the data they train on, and that quantum computing is where more accurate data comes from. It is a bet on a supply chain, not a result: the company is months old and has shown no calculation anyone outside it has checked.
