Isomorphic Labs unveils IsoDDE, a unified AI drug design engine surpassing AlphaFold 3 in biomolecular prediction
Isomorphic Labs, the London-based AI drug discovery company spun out from Google DeepMind, has published a detailed technical report describing its proprietary Isomorphic Labs Drug Design Engine (IsoDDE). The system demonstrates substantial improvements over AlphaFold 3 across four core computational drug design tasks: protein-ligand structure prediction, antibody–antigen interface modelling, binding affinity estimation, and ligandable pocket identification.
London-based Isomorphic Labs has published a technical report detailing the capabilities of its Isomorphic Labs Drug Design Engine (IsoDDE), a proprietary computational system for predicting biomolecular interactions that the company says represents the first substantive step-change in structure prediction accuracy since the release of AlphaFold 3 (AF3) in 2024. The report, authored by the Isomorphic Labs Team and released on 10 February 2026, describes four principal capabilities: protein-ligand structure prediction, antibody–antigen interface modelling, binding affinity estimation, and ligandable pocket identification.
Protein-ligand structure prediction: doubling accuracy on novel targets
The key limitation IsoDDE is designed to address is the failure of existing models to generalise to biomolecular structures that lie far outside their training data. Using the Runs N’ Poses benchmark — specifically constructed to assess performance on novel protein pockets and ligands — IsoDDE achieved a 50% success rate in the hardest similarity bin (0–20%), compared with 23.3% for AF3, a gain that is statistically significant across multiple dataset configurations (Table 4 of the technical report). The model also succeeds on mechanistically complex cases such as induced-fit conformational changes and cryptic pocket opening, events that AF3 fails to capture. On the independent FoldBench dataset, IsoDDE achieved 75.58% accuracy on antibody–antigen interfaces and 75.99% on protein–ligand interfaces, compared with 47.90% and 64.90% respectively for AF3.
Antibody–antigen prediction: implications for biologics development
Antibody-based therapies constitute a large and growing segment of the pharmaceutical market, yet computational modelling of antibody–antigen interfaces has remained technically challenging, with AF3 failure rates exceeding 50% on novel interfaces. IsoDDE addresses this directly. On a held-out test set of 334 novel antibody–antigen complexes, IsoDDE achieved 39% accuracy in the high-fidelity regime (DockQ > 0.8), compared with 17% for AF3 and 2% for Boltz-2, representing a 2.3-fold and 19.8-fold improvement respectively. The CDR-H3 loop — the most variable and immunologically critical region of the antibody — was modelled with backbone RMSD ≤ 2 Å in 70% of cases, versus 58% for AF3 and 43% for Boltz-2. Performance improved further with increased inference-time compute: at 1,000 model seeds, IsoDDE reached 59% high-accuracy DockQ performance and 84% CDR-H3 accuracy. Mohammed AlQuraishi, a computational biologist at Columbia University, described the system to Nature as a major advance, comparable in scale to a putative AlphaFold 4.
Binding affinity prediction: exceeding physics-based methods
Accurate binding affinity estimation is a prerequisite for meaningful structure-activity relationship modelling and compound prioritisation. Traditional free-energy perturbation (FEP) methods, whilst considered the gold standard, are computationally expensive and require carefully prepared crystal structure inputs. IsoDDE achieved a mean Pearson correlation coefficient (r) of 0.85 on the FEP+ 4 benchmark, compared with 0.78 for FEP+ itself, and 0.73 on the OpenFE benchmark versus 0.72 for FEP+. On the CASP16 blind prediction task, IsoDDE scored 0.75 versus 0.65 for the next best approach. Crucially, this performance was achieved without requiring experimental crystal structures as starting inputs. Performance remained consistent across low, medium, and high chemical similarity bins, indicating genuine generalisation rather than interpolation within familiar chemical series.
Pocket identification: expanding the ligandable proteome
IsoDDE incorporates a blind pocket identification capability that can rank ligand-binding residues on a target protein using only its amino acid sequence as input. On a temporally held-out benchmark, it achieved an area under the precision-recall curve (AUPRC) of 0.75, versus 0.51 for P2Rank, a widely used open-source comparator. Importantly, IsoDDE maintained meaningful performance on cryptic sites — pockets absent in the apo protein state and only accessible upon ligand binding — achieving AUPRC of 0.17 versus 0.10 for P2Rank on this subset. In a retrospective analysis, IsoDDE correctly identified both the classical thalidomide-binding pocket and a recently discovered allosteric cryptic site on Cereblon (CRBN) from sequence alone, subsequently cofolding the respective ligands (lenalidomide and SB-405483) with RMSD values of 0.12 Å and 0.33 Å respectively.
Access and openness
Unlike its predecessors, IsoDDE is proprietary and not publicly accessible. The technical report provides benchmark results but discloses little methodological detail regarding architecture or training procedures, a point noted critically by the open-source modelling community. Isomorphic Labs states that its internal drug-discovery programmes are actively using IsoDDE across multiple therapeutic projects.
References
- Isomorphic Labs Team. Accurate predictions of novel biomolecular interactions with IsoDDE. Technical Report, 10 February 2026. DOI: 10.5281/zenodo.18606681. https://doi.org/10.5281/zenodo.18606681
- Isomorphic Labs. The Isomorphic Labs Drug Design Engine unlocks a new frontier beyond AlphaFold. 10 February 2026. https://www.isomorphiclabs.com/articles/the-isomorphic-labs-drug-design-engine-unlocks-a-new-frontier
- Callaway E. ‘An AlphaFold 4’ – scientists marvel at DeepMind drug spin-off’s exclusive new AI. Nature. 19 February 2026. DOI: 10.1038/d41586-026-00365-7 https://doi.org/10.1038/d41586-026-00365-7
