IsoDDE is a proprietary drug-design engine from Isomorphic Labs, the Google DeepMind drug-discovery spinoff, announced on February 10, 2026 via a company article and an accompanying technical report. It builds on AlphaFold 3 and is closed-weight, available for internal use and partner deployments rather than as open weights (Source: isomorphiclabs.com).
| Field | Value | |
|---|---|---|
| Developer | [[companies/isomorphic-labs | Isomorphic Labs]] (Google DeepMind spinoff) |
| Announced | February 10, 2026 (company article + technical report) | |
| Lineage | Builds on [[models/alphafold | AlphaFold 3]] (DeepMind, May 2024) |
| Access | Closed; internal use + partner deployments (Eli Lilly, Novartis) | |
| Open weights | No |
Capabilities and benchmarks
Isomorphic describes IsoDDE as a unified computational drug-design system that extends AlphaFold 3 from structure prediction toward downstream drug-discovery tasks. On the "Runs N' Poses" benchmark (Škrinjar et al., 2025), which is designed to test generalization to novel binding pockets and ligands, the company reports that IsoDDE more than doubles AlphaFold 3's accuracy on the hardest, most dissimilar systems. For antibody-antigen interfaces it reports outperforming AlphaFold 3 by 2.3× and the open model Boltz-2 by 19.8× in the high-fidelity regime (DockQ > 0.8) on a low-homology test set, including on the variable CDR-H3 loop. The company also reports that IsoDDE surpasses other deep-learning methods on the FEP+, OpenFE, and CASP16 blind binding-affinity tasks and can match physics-based free-energy-perturbation methods without requiring experimental crystal structures, and that it can perform "blind" pocket identification from sequence alone — recapitulating a recently discovered cryptic, allosteric pocket on cereblon (Dippon et al., 2026) using only the protein sequence as input (Source: isomorphiclabs.com; technical report at storage.googleapis.com).
These figures come from the company's own article and technical report (cited as Zenodo record 10.5281/zenodo.19699685) and from Max Jaderberg's remarks at WIRED Health; independent benchmarking has not yet been reported, and confidence in the claims is rated medium for that reason (Source: zenodo.org) (Source: wired.com).
Architecture and role
IsoDDE sits as the proprietary drug-design layer above AlphaFold 3, which provides the public structural-prediction substrate. IsoDDE converts predicted structures into engineered candidate molecules, and together the two form the technical core of Isomorphic's clinical-trial pipeline in oncology and immunology.
Availability and access
IsoDDE is closed-weight and available only for internal use and partner deployments, including with Eli Lilly and Novartis. This continues the AlphaFold 2 → AlphaFold 3 progression from open to closed access, in which access tightened as the capability moved closer to drug-pipeline value. The closed posture is relevant to open-source-AI debates.
Relationships
- depends-on: AlphaFold 3 — structural-prediction substrate
- related: Isomorphic Labs — developer
- deploys-in: Healthcare (drug discovery)
- instance-of: AI for Science