IsoDDE
Last reviewed
May 16, 2026
Sources
13 citations
Review status
Source-backed
Revision
v1 ยท 3,995 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 16, 2026
Sources
13 citations
Review status
Source-backed
Revision
v1 ยท 3,995 words
Add missing citations, update stale details, or suggest a clearer explanation.
IsoDDE, short for Isomorphic Labs Drug Design Engine, is a unified computational drug design system developed by Isomorphic Labs, the Alphabet subsidiary spun out of Google DeepMind in 2021. The engine was introduced in a 27 page technical report titled Accurate Predictions of Novel Biomolecular Interactions with the Isomorphic Labs Drug Design Engine, published on February 10, 2026 on the company website and deposited with a Zenodo DOI. IsoDDE extends ideas pioneered by AlphaFold 3 to a single multi task system that performs protein ligand structure prediction, antibody antigen interface modelling, binding affinity estimation, and ligandable pocket identification from amino acid sequence alone.
The model attracted unusual attention for two reasons. First, on a generalisation benchmark that limits each test case to less than 20 percent similarity with anything in the training set, IsoDDE more than doubled the accuracy of AlphaFold 3, lifting success rates from 23.3 percent to roughly 50 percent. Second, IsoDDE's binding affinity predictions matched or exceeded gold standard physics based free energy perturbation methods on the FEP+, OpenFE, and CASP16 benchmarks, while running in seconds rather than the hours required by molecular dynamics. Mohammed AlQuraishi, a computational biologist at Columbia University, described the system as "a major advance, on the scale of an AlphaFold 4," while noting that the underlying methodology was almost entirely undisclosed.
Unlike AlphaFold 2 and AlphaFold 3, which were eventually released to academic users, IsoDDE is fully proprietary. Isomorphic Labs maintains it as the technical foundation for its own internal drug pipeline and for its multi billion dollar research collaborations with Eli Lilly, Novartis, and Johnson and Johnson. The release intensified an ongoing debate in AI drug discovery about whether the most capable structure and affinity models will remain inside commercial laboratories or be reproduced by open source projects such as Boltz, Chai-1, and OpenFold.
Isomorphic Labs was founded in November 2021 by Demis Hassabis, who continues to serve as its chief executive while also leading Google DeepMind. The company was created to apply AlphaFold style methods to the long industrial pipeline of small molecule drug discovery, an area where structure prediction alone is rarely enough. Drug projects depend not only on knowing the shape of a protein and how a candidate molecule binds to it, but also on quantifying how tightly the molecule binds, how selectively it picks out one target from related proteins, and whether new binding pockets might exist that biology has not yet exploited.
AlphaFold 3, released in May 2024 as a joint publication with Google DeepMind, was the first model that could place a protein and a small molecule into the same prediction in one shot. It also extended the original AlphaFold framework to nucleic acids, ions, and chemical modifications, with a diffusion based generative head replacing the Evoformer module of AlphaFold 2. Its PoseBusters score of about 76 percent for protein ligand poses outperformed all previous docking tools. However, AlphaFold 3's accuracy drops sharply for targets and chemistries that differ from those in its training set, a familiar weakness for structure prediction systems that are heavily anchored to the Protein Data Bank.
IsoDDE was designed to close that generalisation gap and to fold several adjacent computational tasks into one engine. The report frames IsoDDE not as a successor model that replaces AlphaFold 3 outright but as a unified system that absorbs the structure prediction capabilities of AlphaFold 3 and surrounds them with task heads for affinity, pocket detection, and antibody modelling, all trained against benchmarks that explicitly target real drug discovery problems.
Isomorphic Labs published the technical report and an accompanying overview article on its website on February 10, 2026. The report is 27 pages long and is credited to the Isomorphic Labs team rather than a single first author. A PDF copy is hosted on Zenodo, and the company also added IsoDDE to the technology page on isomorphiclabs.com, where the engine is described as "a unified drug design engine for a new era of discovery."
Within a day, coverage appeared in Scientific American, BioPharma Trend, Analytics Drift, Clinical Laboratory International, Winbuzzer, and several specialist newsletters. Scientific American led with the AlQuraishi quote and framed the launch as a flashpoint for the open science debate that AlphaFold itself had triggered in 2024, when more than a thousand scientists signed a petition asking for the AlphaFold 3 source code to be released.
Isomorphic Labs did not release model weights, training code, or inference code. The technical report itself provided benchmark numbers and architectural sketches but withheld specific details about training data, loss functions, and the exact relationship between IsoDDE and AlphaFold 3. The company stated that different internal versions of IsoDDE incorporate different data sources, with some variants trained on private structural and assay data contributed by pharmaceutical partners.
The technical report describes IsoDDE as a single engine that produces several different kinds of output, with shared internal representations across tasks. Four principal capabilities are reported in detail.
| Capability | What IsoDDE produces | Why it matters for drug design |
|---|---|---|
| Protein ligand structure prediction | Three dimensional pose of a small molecule bound to a target protein, with predicted local confidence | Replaces traditional molecular docking with a generative model that can capture induced fit and cryptic pocket opening |
| Binding affinity estimation | Numerical estimate of binding free energy or affinity score, with rank ordering across a chemical series | Allows fast hit triage and lead optimisation without running expensive free energy perturbation calculations |
| Antibody antigen interface modelling | Full structure of an antibody bound to an antigen, with explicit modelling of complementarity determining regions including the highly variable CDR H3 loop | Supports antibody optimisation, epitope mapping, and rational design of biologics |
| Ligandable pocket identification | Map of potential drug binding pockets on a target, including cryptic and allosteric sites, from amino acid sequence alone | Surfaces novel chemical handles on "undruggable" targets without requiring known ligands or co crystal structures |
A fifth implicit capability runs through the report: by sharing a backbone representation across these tasks, the engine can be queried iteratively in a drug discovery workflow, for example by predicting a pocket, generating candidate poses for a virtual screening library, ranking the resulting compounds by affinity, and then folding the same pipeline back onto antibody style modalities for targets where small molecules are insufficient.
The technical report describes IsoDDE as a multi model system that extends AlphaFold 3's diffusion based architecture rather than discarding it. The structure prediction core handles the same broad range of biomolecules as AlphaFold 3, including proteins, nucleic acids, small molecules, ions, and chemical modifications, and produces full atom three dimensional coordinates through a generative process. On top of this core, IsoDDE adds task specific heads for affinity prediction, pocket scoring, and antibody specific modelling.
The report does not disclose layer counts, hyperparameters, training set composition, or the precise loss functions used. It states that the model was trained on a combination of public structural data, proprietary internal data, and additional supervision signals from biophysical and assay measurements, but does not quantify the relative contributions. Several commentators noted that this level of disclosure is less than was provided for AlphaFold 3, which itself was criticised for the brevity of its supplementary materials.
The engine emphasises robustness to chemistry and protein space not seen in training. The report repeatedly highlights cases where IsoDDE recovers induced fit conformations or cryptic pocket openings for systems that AlphaFold 3 modelled as rigid, suggesting that the generative diffusion process has been tuned to sample alternative protein conformations more aggressively. Isomorphic Labs has not stated whether new equivariant architectures, longer diffusion trajectories, or different conditioning strategies are responsible for this behaviour.
The overview article and the technical report both stress that IsoDDE is used inside Isomorphic Labs as a screen and design platform, not just a stand alone predictor. In practice this means generating large libraries of candidate molecules, ranking them by predicted pose and affinity, and feeding the survivors into more expensive simulation and synthesis steps. The report includes a case study on cereblon, a well known E3 ligase used in proteolysis targeting chimera (PROTAC) drug design, in which IsoDDE recovered novel pose families that earlier docking tools had missed.
Isomorphic Labs has not published the generative chemistry component as a separate system, but the report describes IsoDDE as one of two pillars of its internal drug design stack, the other being a suite of generative molecular design models.
IsoDDE's release was accompanied by an unusually concentrated set of benchmark comparisons against both AlphaFold 3 and open source competitors. The report focuses on tasks where structure prediction quality directly affects drug design decisions.
The headline benchmark is Runs N' Poses, an internal generalisation benchmark designed to test prediction quality on systems whose proteins and ligands are dissimilar to anything in the training data. The benchmark partitions test cases into bins based on similarity to training data, with the hardest bin containing cases that are less than 20 percent similar by the report's similarity metric. On this hardest bin, IsoDDE achieved a success rate of roughly 50 percent compared with 23.3 percent for AlphaFold 3.
On the independent FoldBench benchmark, IsoDDE reported 75.99 percent accuracy on protein ligand interfaces against 64.90 percent for AlphaFold 3, and 75.58 percent on antibody antigen interfaces against 47.90 percent for AlphaFold 3.
| Benchmark | Task | IsoDDE | AlphaFold 3 | Notes |
|---|---|---|---|---|
| Runs N' Poses, 0-20 percent similarity bin | Protein ligand structure prediction on novel targets | ~50 percent | 23.3 percent | Internal generalisation benchmark from the IsoDDE report |
| FoldBench, protein ligand interfaces | Protein ligand structure prediction | 75.99 percent | 64.90 percent | Independent benchmark cited in the report |
| FoldBench, antibody antigen interfaces | Antibody antigen structure prediction | 75.58 percent | 47.90 percent | Independent benchmark cited in the report |
| Antibody antigen, low similarity, DockQ > 0.8 | High fidelity antibody antigen complex prediction | 39 percent | 17 percent | Compared on 334 low similarity complexes |
| CDR H3 loop, backbone RMSD <= 2 angstrom | Antibody loop modelling | 70 percent | 58 percent | Critical loop for antibody specificity |
| FEP+ 4, Pearson correlation | Small molecule binding affinity | 0.85 | not applicable | Compared against FEP+ at 0.78, a physics based gold standard |
| CASP16 blind binding affinity | Binding affinity, blind community benchmark | 0.75 | not applicable | Next best deep learning system scored 0.65 |
| Pocket identification, AUPRC | Ligandable pocket identification | 0.75 | not applicable | Compared against P2Rank at 0.51 |
The antibody antigen benchmarks are particularly large jumps over both AlphaFold 3 and other open source models. On a set of 334 low similarity antibody antigen complexes, IsoDDE reached 39 percent accuracy at the strict DockQ > 0.8 high fidelity threshold, compared with 17 percent for AlphaFold 3 and 2 percent for Boltz-2. For the critical CDR H3 loop, which dominates antibody specificity and is the hardest loop to predict, IsoDDE reached backbone root mean square deviation of two angstroms or better in 70 percent of cases, compared with 58 percent for AlphaFold 3 and 43 percent for Boltz-2.
These antibody results matter because most existing antibody specific tools, including dedicated antibody loop predictors, struggle to reach high fidelity accuracy on novel antigens. IsoDDE's combination of a general purpose biomolecular core with antibody specific training signals appears to deliver substantial gains on a task that has been resistant to deep learning approaches.
For small molecule affinity, IsoDDE was benchmarked against the FEP+ congeneric series, the OpenFE collaborative free energy benchmark, and the binding affinity prediction task in CASP16, a blind community challenge held in 2024 and 2025. On FEP+ 4 the Pearson correlation between IsoDDE predictions and experimental affinities was 0.85, compared with 0.78 for traditional FEP+ calculations. On CASP16 the IsoDDE score was 0.75 compared with 0.65 for the next best deep learning approach.
The report emphasises that IsoDDE achieves these results without requiring an experimental crystal structure of the target protein bound to a reference ligand, a starting point that traditional FEP methods cannot operate without. This eliminates a major bottleneck in early stage drug discovery, where co crystal structures are often unavailable. IsoDDE also runs in seconds per prediction on standard accelerator hardware, while a typical FEP calculation can take hours per ligand on dedicated molecular dynamics infrastructure.
The pocket detection benchmark tests whether IsoDDE can identify drug binding pockets on a target protein from amino acid sequence alone, without any input ligand or known binding site. The reported AUPRC of 0.75 substantially exceeds the AUPRC of 0.51 reported for P2Rank, a widely used pocket prediction tool. The report also describes blind retrieval of cryptic and allosteric pockets that are absent in canonical apo crystal structures, comparing the result to experimental fragment soaking campaigns that can take weeks or months in a wet laboratory.
The ability to surface pockets without prior chemical matter is significant for so called undruggable targets, where the absence of known binders has historically blocked entry into a drug discovery programme.
IsoDDE was released into a crowded landscape of AlphaFold inspired structure prediction systems, with several open source projects having narrowed the gap with AlphaFold 3 in 2025. Direct comparisons are complicated because IsoDDE is closed source and the reported benchmarks are largely those that Isomorphic Labs chose to publish, but a clear pattern emerges from the available numbers.
| System | Developer | Release | Open or closed | Primary capability | Position relative to IsoDDE |
|---|---|---|---|---|---|
| IsoDDE | Isomorphic Labs | February 2026 | Closed source | Unified drug design engine: structure, affinity, antibody, pocket | Reference system, leads on every reported benchmark |
| AlphaFold 3 | Google DeepMind and Isomorphic Labs | May 2024 | Inference code and weights released for non commercial use, November 2024 | General biomolecular structure prediction | Trailing on every directly compared benchmark; especially weak on novel targets and antibodies |
| Boltz-1 | MIT | November 2024 | Open source under MIT license | Biomolecular structure prediction, AlphaFold 3 style | Comparable to AlphaFold 3 on PoseBusters but not on IsoDDE's generalisation benchmarks |
| Boltz-2 | MIT and Recursion Pharmaceuticals | June 2025 | Open source under MIT license | Structure plus binding affinity, approaches FEP at roughly 1000x speed | Far behind IsoDDE on antibody antigen and generalisation, broadly competitive on small molecule affinity for in distribution cases |
| Chai-1 | Chai Discovery | September 2024 | Weights for non commercial use | Biomolecular structure prediction | Roughly matches AlphaFold 3 on standard benchmarks, no published numbers on IsoDDE style generalisation benchmarks |
| OpenFold | Columbia University and contributors | 2022 onwards, with regular updates | Fully open source | AlphaFold 2 reimplementation and successors | Important for academic reproducibility, lags on AlphaFold 3 era capabilities |
Diego del Alamo of Takeda Pharmaceuticals, quoted by Scientific American, suggested that Isomorphic's collaborations with major pharmaceutical companies likely provide it with private structural and assay data not available to open source projects, which may help explain part of the performance gap on antibody and affinity benchmarks. Gabriele Corso, one of the leads of the Boltz project, told the magazine that public datasets still offer substantial room for improvement and that IsoDDE should be seen as "a new baseline to match, but also to pass."
The other major comparison class for IsoDDE is the family of physics based free energy methods that have been the gold standard in computational drug design for decades. Free energy perturbation, alchemical free energy, and related molecular dynamics methods are accurate but computationally expensive and require careful preparation, including a high quality starting structure and force field parameters for every ligand.
| Method | Typical use | Strengths | Weaknesses | IsoDDE comparison |
|---|---|---|---|---|
| Free energy perturbation (FEP+) | Lead optimisation, congeneric series ranking | Well validated, accurate when carefully run | Requires co crystal structure, force field parameterisation, and many CPU or GPU hours per ligand | IsoDDE matches or exceeds FEP+ accuracy on the FEP+ 4 benchmark without requiring a co crystal structure, and runs in seconds rather than hours |
| OpenFE alchemical free energy | Community open source alternative to FEP+ | Open and transparent, growing benchmark coverage | Same speed and structural data requirements as FEP+ | IsoDDE outperforms on the OpenFE benchmark and avoids the structural data requirement |
| Traditional docking, e.g. Glide or AutoDock | Virtual screening | Fast, scalable to libraries of millions | Limited accuracy, especially for novel chemistry and induced fit | IsoDDE substantially outperforms on novel chemistry while remaining fast enough for screening |
| P2Rank pocket detection | Pocket finding from structure | Simple, widely used | Misses cryptic and allosteric pockets, requires structure as input | IsoDDE works from sequence alone and reports much higher AUPRC |
IsoDDE is not offered as a service to outside researchers. It is the core technical platform that Isomorphic Labs uses internally and that it provides to its pharmaceutical partners under collaboration agreements. The economics of the company depend on these partnerships rather than on direct sales of model access.
In January 2024, Isomorphic Labs announced two major strategic deals. Eli Lilly entered into a research collaboration that includes up to 1.7 billion United States dollars in performance based milestone payments, and Novartis entered into a similar agreement with up to 1.2 billion United States dollars in milestones. Johnson and Johnson followed with a separate collaboration whose financial terms have not been disclosed. According to the DeepCeutix briefing on the IsoDDE launch, the combined potential value of the Lilly and Novartis deals exceeds three billion United States dollars before any royalty payments on commercialised drugs.
The company's own drug pipeline included 17 programmes at the time of the IsoDDE release. Isomorphic Labs has publicly stated that its first AI designed cancer drug candidate is on track to enter a Phase 1 clinical trial by the end of 2026.
Isomorphic Labs raised 600 million United States dollars in its Series A round in March 2025, led by Thrive Capital with participation from existing investor GV and Alphabet, valuing the company at approximately 2.5 billion United States dollars. By February 2026 European Biotechnology Magazine and other outlets reported that the company had raised a further 2.1 billion United States dollars in additional financing. Alphabet remains the largest shareholder.
The choice not to release IsoDDE was deliberate and was framed by Isomorphic Labs as necessary to protect commercially sensitive training data and pipeline programmes. In Scientific American, AlQuraishi observed that the limited disclosure left external researchers "essentially guessing about methodology," while the DeepCeutix briefing noted that the release represents a clear departure from the partial code and weights releases of AlphaFold 2 and AlphaFold 3.
The scientific reception combined strong technical admiration with explicit concern about reproducibility and open science. Several reactions appeared within days of the February 10 release.
Mohammed AlQuraishi, of the AlQuraishi Lab at Columbia University, called IsoDDE "a major advance, on the scale of an AlphaFold 4," and added that the ability to predict drug protein interactions for molecules vastly different from training data "suggests that they must have done something pretty novel." He also said that "the problem, of course, is that we know nothing of the details."
Diego del Alamo, a computational scientist at Takeda Pharmaceuticals, suggested in Scientific American that Isomorphic Labs' private collaborations with large pharmaceutical companies likely provide it with proprietary structural and assay datasets not available to academic competitors, which may contribute to its lead on antibody and affinity benchmarks.
Gabriele Corso, one of the leads on the open source Boltz models, took an optimistic view, suggesting that publicly available data still has substantial untapped potential and that IsoDDE "should be seen as a new baseline to match, but also to pass."
In industry coverage, BioPharma Trend described IsoDDE as the first system that genuinely shifts the centre of computational drug design from physics based methods to deep learning, while Analytics Drift and Clinical Laboratory International focused on the antibody and pocket detection numbers as the most surprising results. Commentary in Nature and other journals raised the broader question of whether the next generation of biological foundation models will be available to academic researchers at all, given that the most capable systems are now being developed inside commercial laboratories with access to private data and large compute budgets.
IsoDDE is best understood as a parallel rather than a strict successor to AlphaFold 3. The two systems share a lineage in Google DeepMind's AlphaFold work and overlapping personnel, since Demis Hassabis leads both DeepMind and Isomorphic Labs, but they target different audiences. AlphaFold 3 was released through DeepMind's research channels with inference code and weights eventually shared for non commercial use. IsoDDE is an Isomorphic Labs product designed for industrial drug discovery and is not being released. Isomorphic Labs has avoided the AlphaFold branding deliberately, both to signal that IsoDDE is a different kind of system and to retain commercial flexibility around licensing and IP.
The "AlphaFold 4" framing in AlQuraishi's quote refers to the magnitude of the technical leap rather than to an actual unreleased DeepMind successor model. There is no public AlphaFold 4 product or paper as of the IsoDDE release. The phrase has been repeated by several outlets as shorthand for the scale of the advance, but it is not Isomorphic Labs' or Google DeepMind's own description of the system.