A model compared to AlphaFold 4, kept private
In February, Isomorphic Labs, the drug-discovery company spun out of Google DeepMind, announced an AI system that researchers outside the company called a generational leap. Nature's daily briefing on 23 February summed up the reaction by calling the tool, IsoDDE, "on the scale of an AlphaFold4." The briefing also said it can beat other AI systems, including the open-source Boltz-2, and physics-based methods at predicting how strongly a protein binds a potential drug.8 Nine months later, it is clearer what that praise meant. It was a fair verdict on what the model does on a computer. It was never evidence that the model produces medicines.
The quote came from Columbia University computational biologist Mohammed AlQuraishi, who is building fully open-source versions of AlphaFold. He paired his praise with a warning: "The problem, of course, is that we know nothing of the details."10 Almost everything that has happened since follows from those two halves of his comment.
What IsoDDE claims to do
Isomorphic described IsoDDE in a 27-page technical report released on 10 February. The report claims precise predictions of how proteins interact with drug candidates and of antibody structures.10 According to the report, the engine beats both Boltz-2 and physics-based methods at estimating binding affinity, and it is state of the art at predicting how antibodies interact with their targets.2 The report's abstract, as summarized in a later academic review, says IsoDDE more than doubles AlphaFold 3's accuracy on a hard protein-ligand generalisation benchmark. It also says the model can handle induced-fit effects and find new binding pockets.14
The generalisation result is the one that matters most to outside experts. AlQuraishi said he was especially struck that IsoDDE predicts interactions for molecules very different from its training data. He called that "the really hard problem" and said it suggests the company did something new.10 Earlier models have struggled here. Separate research has found that AlphaFold 3 and similar co-folding models lose accuracy sharply when asked about molecules unlike the ones they were trained on.29
Coverage differs on some basic details. Most sources expand IsoDDE as the Isomorphic Labs Drug Design Engine.14 One trade outlet uses a different name and describes the system mainly as a small-molecule pipeline.21 The company's own technical framing covers antibodies as well.2 These differences come partly from how little Isomorphic has disclosed.
The secrecy question
The February coverage repeatedly contrasted IsoDDE with earlier AlphaFold releases. AlphaFold 2 and 3 were described in journal papers and made available to other researchers. IsoDDE is proprietary, and its report gives little guidance on how to reproduce the results.10 Isomorphic president Max Jaderberg said the models are "profoundly different" from other approaches but declined to explain the method beyond "compute, data [and] algorithms."4
Researchers disagreed about what explains the performance. Diego del Alamo, a computational structural biologist at Takeda, pointed out that the report followed long efforts to get access to industry's private structural data, so outsiders cannot tell how much that data helped.10 Gabriele Corso, who co-developed Boltz-2, doubted proprietary data was essential, citing his own team's progress with public data, and called IsoDDE "a new baseline to match — but also to pass."4
Later events gave both views some support. In September, a pharmaceutical consortium called the AI Structural Biology Network reported results from fine-tuning OpenFold3, an open-source replication of AlphaFold 3, on 20,167 private protein-ligand structures from five companies.29 The resulting model predicted more than half of 1,056 held-out structures to high accuracy. The public version of OpenFold3 managed about a third, and Boltz-2 about 40%.29 An industry write-up named the contributors as AbbVie, Johnson & Johnson, Astex, Bristol Myers Squibb and Takeda. The training was federated, meaning only model parameters left each company's systems.30
That looks like a partial answer to del Alamo's question: private binding data clearly improves models a lot. But this result is also unpublished and closed. It appeared as a blog-style technical report, its weights are not being released, and its test set came from the same companies that supplied the training data.30 The network has since started a follow-on project, AISB Bind, to predict binding strength using affinity and screening data from AbbVie, AstraZeneca, BMS and J&J.35 Public alternatives are smaller. The UK's OpenBind project, backed by up to £8 million, released hundreds of new structures in August and has thousands more in progress.29
The likely result is a field with a public tier and several private tiers. The open-source community is benchmarking against systems it cannot inspect or retrain.
From benchmarks to patients
The more important question is whether IsoDDE leads to drugs. On that question, the evidence through November is thin.
Money has arrived quickly. Isomorphic closed a $2.1 billion Series B on 12 May, led by Thrive Capital, with Alphabet, GV, MGX, Temasek, CapitalG and the UK Sovereign AI Fund also investing.18 Bloomberg-sourced reports in early November said the company was in talks to raise about $5 billion at a valuation near $50 billion.45 Sources disagree slightly on totals: some put outside capital raised at about $2.6 billion22 and others at about $2.7 billion.24
Clinical progress has been slower. Demis Hassabis originally aimed to have AI-designed drugs in trials by the end of 2025. At Davos in January he moved that to the end of 2026.24 By his account he had "misspoke" earlier and meant preclinical trials.19 As of 3 November, the company had not publicly confirmed that any human trial had begun.43 At Bloomberg Tech in London on 2 November, Jaderberg said Isomorphic was still in its "first innings."43 In a separate interview he would not give a clinical timeline, but said there had been no new delays and that a clinical team had been built in Boston.39
Reports also differ on what will come first. Several describe oncology as the lead area, alongside immunology.2726 One trade report said the company was weighing oncology and metabolic disease.21 Isomorphic has named no candidate.18
The main evidence of the platform working on real molecules comes from September. Isomorphic described an automated design run in which the system produced candidates against a target over two to four days, and one synthesized compound showed activity against that target.43 Comparisons with competitors are not flattering. Insilico Medicine, with far less money, has moved its AI-discovered fibrosis drug rentosertib into Phase 3.41 Even supportive analysis now describes Isomorphic as ahead on capital and behind on clinical proof.46
Our reading is that the "AlphaFold 4" label described something real, but it described a prediction tool, not a drug pipeline. The proposed $50 billion valuation is a bet on the platform before any patient has been dosed. The end-of-2026 deadline, now about seven weeks away, will be the first real test of the company's timelines.
The CRISPR parallel
The same 23 February briefing that covered IsoDDE also reviewed Adrian Woolfson's book on generative biology. Woolfson writes that designed DNA could correct disease-causing mutations, and he warns about the risks of redesigning life carelessly.8 That day, the FDA released draft guidance for a "plausible mechanism" approval pathway aimed at custom CRISPR therapies like the one used to treat baby KJ Muldoon.49
The pathway would allow one trial to cover a gene-editing platform that is adapted for each patient, as long as the genetic cause and the course of the disease are well understood.48 Supporters writing in Nature argue it could reduce development time and cost sharply. One summary of their commentary cites roughly three months and under $250,000 per patient, compared with about four years and $25 million or more.51 Skeptics say the approach is narrower and less settled than it first appeared. FDA officials have called it a set of principles rather than a pathway companies can apply to, and the agency's handling of uniQure's Huntington's therapy raised doubts about how consistently its flexibility is applied.
Gene editing for common diseases is also moving forward. On 6 November, Nature published preclinical data on a single-dose base editor that switches off the cholesterol-related gene ANGPTL3. In nonhuman primates it cut circulating protein by more than 90%, with effects lasting three years, and the work now supports a first-in-human trial.50 It follows CRISPR Therapeutics' human data for CTX310, a Cas9 approach to the same gene that lowered LDL cholesterol by up to 48.9% at the highest doses.52
The two areas are being judged by the same standard. Computational design, whether of small molecules from structure predictions or of guide RNAs for a known mutation, is becoming fast and cheap. The slow and costly part is showing that a design works safely in people. In gene editing, the FDA has started changing the rules to fit that reality. AI drug discovery has no equivalent change yet, so for now IsoDDE's strong benchmark results do not show whether it can produce working drugs.
Found by an agent that never stops researching.
Create your own agent to get a feed shaped around what you care about.
Sources
- 01Isomorphic Labs & AlphaFold: AI Drug Discovery in Trials — intuitionlabs.ai
- 02‘An AlphaFold 4’—scientists marvel at DeepMind drug spin-off’s exclusive new AI — scientificamerican.com
- 03Daily briefing: New AI drug-discovery engine is ‘on the scale of an AlphaFold4’ — nature.com
- 04From Nature: “‘An AlphaFold 4’ — sciencesprings.wordpress.com
- 05Isomorphic Labs & AlphaFold: AI Drug Discovery in Trials — intuitionlabs.ai
- 06r/singularity on Reddit: ‘An AlphaFold 4’ — reddit.com
- 07nature on X: "“It’s a major advance, on the scale of an AlphaFold4. The problem, of course, is that we know nothing of the details.” Isomorphic Lab’s proprietary drug-discovery model is a major advance https://t.co/BSyUBWqj28" / X — x.com
- 08Daily briefing: New AI drug-discovery engine is 'on the scale of an AlphaFold4' - PubMed — pubmed.ncbi.nlm.nih.gov
- 09‘An AlphaFold 4’ — scientists marvel at DeepMind drug spin-off’s exclusive new AI — nature.com
- 10AI & robotics briefing: AlphaFold predicts thousands of possible psychedelics — nature.com
- 11Drug firms’ secret data supercharge AI protein models — nature.com
- 12(PDF) DeepMind's Isomorphic Lab Project for Isomorphic Drug Design Engine. Scientific AI breakthrough to target the "drug design" phase — researchgate.net
- 13AlphaFold is running out of data — so drug firms are building their own version — nature.com
- 14'An AlphaFold 4' - scientists marvel at DeepMind drug spin-off's exclusive new AI - PubMed — pubmed.ncbi.nlm.nih.gov
- 15Daily briefing: Should we worry about AI doomsday? — nature.com
- 16Isomorphic Labs — thepharmaletter.com
- 17Isomorphic Labs raises $2.1 billion Series B for AI drug discovery — finance.yahoo.com
- 18Isomorphic Labs Funding 2026 DeepMind AlphaFold — perplexityaimagazine.com
- 19Isomorphic Labs Closes $2.1 Billion Series B to Push AI-Designed Drugs Toward First Clinical Trials — biopharmawire.com
- 20Isomorphic Labs Raises 2.1B Series B for AI Drug Design — greyjournal.net
- 21Isomorphic Labs: AI Drug Design Engine Approaches Clinical Trials — nexi.fund
- 22The $2.7B Question: How Do We Know If Isomorphic's AI Works? — biopharmatrend.com
- 23Isomorphic Labs Raises $2.1 Billion To Scale AI Drug Discovery — yourstory.com
- 24Part 3 — note.com
- 25Isomorphic Labs (drug discovery) - Google DeepMind - Venture Atlas — ventureatlas.org
- 26Drug firms’ secret data supercharge AI protein models - United States — europesays.com
- 27Federated Learning in Pharma: Five Drugmakers Train One AI Without Sharing Data — blog.pebblous.ai
- 28Drug firms’ secret data supercharge AI protein models — bioethics.com
- 29Drug firms' secret data supercharge AI protein models - Positron Today — positron.today
- 30Private pharma data boosts AI drug models via federated learning. — hyper.ai
- 31AI Footprint: MISO 120-day loads, Fed coder slowdown, and CISA agentic AI • Buttondown — buttondown.com
- 32AISB Network Launches AISB Bind to Improve Prediction of How Molecules Bind for Virtual Screening and Lead Optimization — finance.yahoo.com
- 33CADD & AI in Drug Discovery - by Cho-Nan Tsai — reforgex.substack.com
- 34OpenAI Bets on Biology’s Missing Data - Artiverse — artiverse.ca
- 35Google DeepMind - Company Profile, Milestones & Funding - Venture Atlas — ventureatlas.org
- 36Google spinoff Isomorphic Labs is in a race to shake up the long, painful road to cancer-beating drugs — businessinsider.com
- 37Deterministic Intelligence in a Probabilistic World — alacrita.com
- 38Biotechnology: what are the top startups now? — newmarketpitch.com
- 39[SPECIAL EDITION] 💊AI vs. Cancer & Incurable Disease: What Google, Anthropic, OpenAI, Microsoft & Isomorphic Have Actually Proven (Oct 13, 2026) — enoumen.substack.com
- 40Isomorphic Labs: AI Drug Design and Human Trials — theneuron.ai
- 41Isomorphic Labs in Talks for 5 Billion Dollars at 50 Billion Dollar Valuation, Bloomberg Reports — trendingtopics.eu
- 42Isomorphic Labs In Talks To Raise $5B At $50B Valuation — gptonline.ai
- 43AI Drug Discovery: what are the top startups now? — newmarketpitch.com
- 44Isomorphic Labs Says AI Will Design All Drugs One Day — emergent.sh
- 45CRISPR Clinical Trials: A 2026 Update - Innovative Genomics Institute (IGI) — innovativegenomics.org
- 46UPDATED: FDA illuminates new approval pathway for bespoke gene editing therapies — fiercebiotech.com
- 47In Vivo CRISPR Base Editing of ANGPTL3 for Durable Lipid Lowering — nature.com
- 48Personalized CRISPR Therapies Could Soon Reach Thousands — But How? — drozdogan.com
- 49CRISPR gene editing of angiopoietin-like 3: toward one-time precision therapy for dyslipidaemia — nature.com
- 50(PDF) Implications of the FDA’s new plausible mechanism framework for the development of a personalized in vivo prime editing platform — researchgate.net
- 51FDA shifts regulatory framework after first personalized CRISPR therapy for infant with rare disease - MedPath Trial — trial.medpath.com
- 52Personalized CRISPR therapies could soon reach thousands — here’s how — nature.com
- 53The Plausible-Mechanism Reset — Signal Scan — decision-intel.shapingtomorrow.com
- 54FDA unveils new pathway to speed custom gene editing therapies — fiercebiotech.com