Isomorphic's IsoDDE moves AI drug design from prediction to pipeline

Isomorphic Labs, the drug-discovery company spun out of Google DeepMind in 2021, published the technical report for its Drug Design Engine, IsoDDE, on February 10, 2026.^{1 Isomorphic Labs 2026-02-10 Isomorphic published the IsoDDE technical report on February 10, 2026: a unified engine for protein-ligand structure, binding affinity, blind pocket detection, and antibody-antigen modeling that more than doubles AlphaFold 3 accuracy on a hard benchmark and surpasses FEP on affinity. Open source 3 IntuitionLabs 2026-02-15 IsoDDE, from the 2021 DeepMind spinout, advanced Isomorphic's oncology and immunology candidates toward first-in-human trials and remains proprietary. Open source} Unlike AlphaFold 3, which predicts biomolecular structure, IsoDDE unifies protein-ligand structure prediction, binding affinity estimation, blind pocket detection, and antibody-antigen interface modeling in a single system, and it more than doubles AlphaFold 3's accuracy on a hard generalization benchmark while surpassing physics-based free energy perturbation on affinity prediction.^{1 Isomorphic Labs 2026-02-10 Isomorphic published the IsoDDE technical report on February 10, 2026: a unified engine for protein-ligand structure, binding affinity, blind pocket detection, and antibody-antigen modeling that more than doubles AlphaFold 3 accuracy on a hard benchmark and surpasses FEP on affinity. Open source} We assess, with moderate confidence, that the significant shift is architectural rather than a single accuracy jump: integrating the steps a drug program needs into one engine is what moves AI from producing insight to running a pipeline, and the hardest of those steps, binding on low-similarity targets, is where IsoDDE reports the largest gains.

What the numbers show

The reported metrics are specific and consistent across the company report and independent coverage. On protein-ligand structure prediction for novel targets, the hardest similarity bin of 0 to 20 percent, IsoDDE reports 50 percent success versus AlphaFold 3's 23.3 percent.^{2 Clinical Laboratory International 2026-02-11 IsoDDE reports 50% success on novel targets versus AlphaFold 3's 23.3%, 39% high-fidelity antibody-antigen accuracy versus 17% for AF3 and 2% for Boltz-2, and 0.85 mean Pearson affinity correlation versus 0.78 for FEP+. Open source} On antibody-antigen interfaces in the high-fidelity regime, it reports 39 percent accuracy against 17 percent for AlphaFold 3 and 2 percent for the open model Boltz-2, a roughly 2.3-fold gain over AF3.^{2 Clinical Laboratory International 2026-02-11 IsoDDE reports 50% success on novel targets versus AlphaFold 3's 23.3%, 39% high-fidelity antibody-antigen accuracy versus 17% for AF3 and 2% for Boltz-2, and 0.85 mean Pearson affinity correlation versus 0.78 for FEP+. Open source} On binding affinity, it reports a mean Pearson correlation of 0.85 on the FEP+ benchmark versus 0.78 for FEP+ itself, meaning it matches or beats an established physics method without needing a crystallographic starting structure.^{1 Isomorphic Labs 2026-02-10 Isomorphic published the IsoDDE technical report on February 10, 2026: a unified engine for protein-ligand structure, binding affinity, blind pocket detection, and antibody-antigen modeling that more than doubles AlphaFold 3 accuracy on a hard benchmark and surpasses FEP on affinity. Open source 2 Clinical Laboratory International 2026-02-11 IsoDDE reports 50% success on novel targets versus AlphaFold 3's 23.3%, 39% high-fidelity antibody-antigen accuracy versus 17% for AF3 and 2% for Boltz-2, and 0.85 mean Pearson affinity correlation versus 0.78 for FEP+. Open source}

The generalization number is the one that matters most. Predicting how a drug binds a target that looks nothing like anything in the training data is the case where structure predictors usually fail, and it is the case real discovery programs face constantly. Doubling accuracy on low-similarity targets, from 23.3 to 50 percent, is a claim about the part of the problem that has resisted machine learning.^{2 Clinical Laboratory International 2026-02-11 IsoDDE reports 50% success on novel targets versus AlphaFold 3's 23.3%, 39% high-fidelity antibody-antigen accuracy versus 17% for AF3 and 2% for Boltz-2, and 0.85 mean Pearson affinity correlation versus 0.78 for FEP+. Open source} The company also reports strong performance on the CDR-H3 loop, the most variable and hardest-to-predict region of an antibody, which is the bottleneck for de novo antibody design.^{1 Isomorphic Labs 2026-02-10 Isomorphic published the IsoDDE technical report on February 10, 2026: a unified engine for protein-ligand structure, binding affinity, blind pocket detection, and antibody-antigen modeling that more than doubles AlphaFold 3 accuracy on a hard benchmark and surpasses FEP on affinity. Open source}

Why a unified engine, not just a better predictor

A structure predictor answers one question. A drug program asks a chain of them: where does a molecule bind, how tightly, is there a druggable pocket, and can an antibody be designed against the target. IsoDDE folds those into one system that works from amino-acid sequence, so a program does not stitch together separate tools with separate error modes.^{1 Isomorphic Labs 2026-02-10 Isomorphic published the IsoDDE technical report on February 10, 2026: a unified engine for protein-ligand structure, binding affinity, blind pocket detection, and antibody-antigen modeling that more than doubles AlphaFold 3 accuracy on a hard benchmark and surpasses FEP on affinity. Open source} That integration is why Isomorphic frames it as moving beyond AlphaFold rather than as an incremental version bump, and why some outside observers described the leap as effectively an AlphaFold 4.^{3 IntuitionLabs 2026-02-15 IsoDDE, from the 2021 DeepMind spinout, advanced Isomorphic's oncology and immunology candidates toward first-in-human trials and remains proprietary. Open source}

The practical payoff Isomorphic claims is speed and reach. Accurate affinity prediction without a crystal structure means candidates can be triaged computationally before wet-lab work, and de novo antibody design capability compresses screening that otherwise takes months.^{2 Clinical Laboratory International 2026-02-11 IsoDDE reports 50% success on novel targets versus AlphaFold 3's 23.3%, 39% high-fidelity antibody-antigen accuracy versus 17% for AF3 and 2% for Boltz-2, and 0.85 mean Pearson affinity correlation versus 0.78 for FEP+. Open source} The engine is reported to have advanced Isomorphic's own oncology and immunology candidates toward first-in-human trials, which is the test that separates a benchmark from a drug.^{3 IntuitionLabs 2026-02-15 IsoDDE, from the 2021 DeepMind spinout, advanced Isomorphic's oncology and immunology candidates toward first-in-human trials and remains proprietary. Open source}

Second order effects and the ledger

Who gains. Isomorphic gains an internal pipeline advantage and a proof point for its partnership-and-in-house model, since the engine directly fed its clinical candidates.^{3 IntuitionLabs 2026-02-15 IsoDDE, from the 2021 DeepMind spinout, advanced Isomorphic's oncology and immunology candidates toward first-in-human trials and remains proprietary. Open source} Its pharma partners gain access to a design tool that reduces reliance on expensive structural biology. Antibody-focused biotech gains the most specific benefit, because the CDR-H3 and antibody-antigen results target the exact step that gates biologic design.^{1 Isomorphic Labs 2026-02-10 Isomorphic published the IsoDDE technical report on February 10, 2026: a unified engine for protein-ligand structure, binding affinity, blind pocket detection, and antibody-antigen modeling that more than doubles AlphaFold 3 accuracy on a hard benchmark and surpasses FEP on affinity. Open source}

Who loses, or is pressured. Vendors of standalone structure-prediction or affinity tools lose ground to an integrated engine that does several jobs at once.^{2 Clinical Laboratory International 2026-02-11 IsoDDE reports 50% success on novel targets versus AlphaFold 3's 23.3%, 39% high-fidelity antibody-antigen accuracy versus 17% for AF3 and 2% for Boltz-2, and 0.85 mean Pearson affinity correlation versus 0.78 for FEP+. Open source} Open-source efforts like Boltz-2, which trailed badly on the antibody metric, face a widening capability gap against a proprietary system.^{2 Clinical Laboratory International 2026-02-11 IsoDDE reports 50% success on novel targets versus AlphaFold 3's 23.3%, 39% high-fidelity antibody-antigen accuracy versus 17% for AF3 and 2% for Boltz-2, and 0.85 mean Pearson affinity correlation versus 0.78 for FEP+. Open source} And the broader field loses transparency: IsoDDE remains proprietary and not publicly available, so its claims cannot be independently reproduced the way an open model's can, which means the reported numbers rest substantially on the company's own technical report.^{3 IntuitionLabs 2026-02-15 IsoDDE, from the 2021 DeepMind spinout, advanced Isomorphic's oncology and immunology candidates toward first-in-human trials and remains proprietary. Open source}

The counter-case

The strongest reason for caution is that benchmark accuracy is not clinical success, and a closed engine's self-reported metrics are hard to verify. The affinity and structure numbers come from Isomorphic's own report and coverage of it, not from independent replication, and the system is not accessible for outside testing.^{1 Isomorphic Labs 2026-02-10 Isomorphic published the IsoDDE technical report on February 10, 2026: a unified engine for protein-ligand structure, binding affinity, blind pocket detection, and antibody-antigen modeling that more than doubles AlphaFold 3 accuracy on a hard benchmark and surpasses FEP on affinity. Open source 3 IntuitionLabs 2026-02-15 IsoDDE, from the 2021 DeepMind spinout, advanced Isomorphic's oncology and immunology candidates toward first-in-human trials and remains proprietary. Open source} The history of computational drug discovery is full of tools that improved in-silico metrics without shortening the path to an approved drug, because the failures that kill candidates, toxicity, pharmacokinetics, trial-stage efficacy, are not what a binding predictor addresses. For the pipeline thesis to hold, IsoDDE-designed candidates have to clear human trials, and that evidence does not yet exist; the candidates are reported as advancing toward first-in-human, not through it.^{3 IntuitionLabs 2026-02-15 IsoDDE, from the 2021 DeepMind spinout, advanced Isomorphic's oncology and immunology candidates toward first-in-human trials and remains proprietary. Open source} The honest position is that the engine is a real step in the design stage and an unproven one in the outcome that matters.

What to watch

• Independent or peer-reviewed validation. Watch over the next year for third-party or published confirmation of the 50 percent novel-target and 0.85 affinity numbers; without access to the model, external checks are the only way the claims move from company report to established result.^{2 Clinical Laboratory International 2026-02-11 IsoDDE reports 50% success on novel targets versus AlphaFold 3's 23.3%, 39% high-fidelity antibody-antigen accuracy versus 17% for AF3 and 2% for Boltz-2, and 0.85 mean Pearson affinity correlation versus 0.78 for FEP+. Open source}
• Clinical candidates enter trials. If IsoDDE-designed oncology or immunology molecules begin first-in-human studies within twelve to eighteen months, the pipeline claim gains substance; delay would suggest the design-stage gains have not yet propagated to the clinic.^{3 IntuitionLabs 2026-02-15 IsoDDE, from the 2021 DeepMind spinout, advanced Isomorphic's oncology and immunology candidates toward first-in-human trials and remains proprietary. Open source}
• Partnership economics. New or expanded pharma deals referencing the engine would show partners are paying for the integrated capability rather than treating it as a research demo.
• The transparency gap. If Isomorphic keeps IsoDDE fully closed while open models like Boltz-2 catch up on the antibody metrics, the durable question becomes whether a proprietary lead in AI drug design is defensible, or whether the field converges as it did after AlphaFold 2's methods spread.^{2 Clinical Laboratory International 2026-02-11 IsoDDE reports 50% success on novel targets versus AlphaFold 3's 23.3%, 39% high-fidelity antibody-antigen accuracy versus 17% for AF3 and 2% for Boltz-2, and 0.85 mean Pearson affinity correlation versus 0.78 for FEP+. Open source}