IsoDDE

Overview

IsoDDE (the Isomorphic Labs Drug Design Engine) is a proprietary unified AI drug design system from Isomorphic Labs, the DeepMind spin-off founded by Demis Hassabis to apply foundation-model approaches to therapeutic discovery. Announced via technical report in February 2026 and covered widely in the scientific press in March 2026, IsoDDE is a single integrated engine that handles protein-ligand structure prediction, binding affinity estimation, and compound generation, replacing the patchwork of separate models that traditional pharma pipelines maintain.

The technical report claims IsoDDE roughly doubles AlphaFold 3 accuracy on protein-ligand structure prediction benchmarks and approaches gold-standard free energy perturbation (FEP+) methods on binding affinity at orders-of-magnitude lower computational cost. Several leading computational chemists publicly characterized the work as "AlphaFold 4-class" — the first model to substantially surpass AF3 on the drug-relevant subset of structure prediction tasks. The release also reignited debate about open versus proprietary biological foundation models, since IsoDDE is closed-source and accessible only to Isomorphic Labs and its pharma partners (Eli Lilly, Novartis).

Key Features

• Unified prediction + design: One engine spans structure, affinity, and generative design tasks, removing the need to chain together separate co-folding, docking, scoring, and generation models.
• AF3-surpassing structure accuracy: Approximately 2x improvement over AlphaFold 3 on protein-ligand pose prediction benchmarks, with particularly strong gains on cryptic pockets and ligand-induced conformational changes.
• FEP+-class binding affinity: Achieves accuracy comparable to commercial FEP+ workflows at a fraction of the compute, enabling affinity ranking at scales previously infeasible.
• Generative compound design: Native support for de novo small molecule generation conditioned on protein targets, integrated into the same model that scores binding.
• Production validation in Isomorphic's pipeline: Deployed across active drug discovery programs at Isomorphic Labs and used by Eli Lilly and Novartis under their multi-year partnerships.

Technical Details

The IsoDDE technical report describes a unified diffusion-transformer architecture trained on extensive proprietary structure-activity data, including unpublished crystallographic and binding affinity datasets accumulated through Isomorphic's pharma partnerships. Specific architectural details, parameter counts, and training corpus are not publicly disclosed. Benchmarks reported in the document focus on PoseBusters subsets, Astex Diverse, and proprietary internal hit-discovery campaigns.

The engine is integrated with Isomorphic's drug-discovery cloud platform announced in 2025; access is gated to internal teams and partners. No code, weights, or API are publicly available.

Applications

IsoDDE is positioned for end-to-end drug discovery workflows: hit identification, lead optimization, off-target screening, and structure-activity relationship analysis. Its FEP+-class affinity prediction is particularly valuable for late-stage lead optimization, where the cost of FEP simulations becomes a bottleneck. Generative design conditioned on a target structure supports de novo scaffold generation for difficult target classes such as protein-protein interaction interfaces.

Impact

IsoDDE marks the first publicly disclosed AI drug design system to substantially exceed AlphaFold 3 accuracy on drug-relevant tasks, raising the question of how long the open-source ecosystem (RoseTTAFold, Boltz, Chai, Protenix) will lag behind proprietary frontier systems. The release intensified debate within the field about the trade-off between open scientific tools and the commercial incentives that fund the largest training runs. As a closed model accessible only through partnerships, IsoDDE will not directly enable academic research but sets a public benchmark that competing open efforts will target.