Isomorphic Labs Unveils IsoDDE, an 'AlphaFold 4' That Scientists Cannot See Inside

Isomorphic Labs, the pharmaceutical AI spin-off of Google DeepMind, has released a 27-page technical report introducing its Isomorphic Drug Design Engine - IsoDDE - a unified computational system that researchers are already calling "an AlphaFold 4." The catch: unlike the Nobel Prize-winning AlphaFold models that came before it, IsoDDE is fully proprietary. The scientific community can see the results but not the recipe.

The Numbers That Have Scientists Talking

Key Performance Claims

Benchmark	IsoDDE vs AlphaFold 3	IsoDDE vs Boltz-2
Protein-ligand structure prediction	2x more accurate	-
Antibody-antigen prediction (DockQ >0.8)	2.3x better	19.8x better
CDR-H3 loop prediction	1.2x better	1.6x better
Binding affinity prediction	Exceeds physics-based FEP methods	Exceeds state-of-the-art deep learning

IsoDDE is not a single neural network. It is an integrated engine that handles protein structure prediction, ligand binding estimation, affinity prediction, antibody interactions, and pocket discovery - all in a unified system. According to the report published on February 10, the engine more than doubles AlphaFold 3's accuracy on a challenging protein-ligand generalization benchmark and outperforms gold-standard physics-based free energy perturbation (FEP) methods at a fraction of their time and cost.

Perhaps most impressive: IsoDDE can identify previously unknown binding pockets on target proteins using only the amino acid sequence as input - no prior experimental data required. The company says this capability approaches the accuracy of lab experiments, a claim that, if true, could fundamentally change how early-stage drug discovery works.

Antibody Design Gets a Massive Upgrade

The antibody results are particularly striking. In the high-fidelity prediction regime (DockQ scores above 0.8), IsoDDE outperforms the open-source Boltz-2 model by nearly 20x. It also shows what Isomorphic calls "remarkable performance" on the CDR-H3 loop - the most variable and hardest-to-predict region of an antibody. Given that antibody-based therapies generate tens of billions in annual pharmaceutical sales, this is not an academic exercise.

From Prediction to Design

Unlike AlphaFold 3, which primarily predicted structures, IsoDDE bridges the gap between structural prediction and practical drug design workflows. Internal teams are using it to explore unseen protein structures, search for and prioritize binding pockets on poorly characterized targets, and generate and rank new molecular candidates. The shift from "tell me what this looks like" to "help me design a drug" is the whole point.

What It Does Not Tell You

Here is where the story gets complicated. AlphaFold and AlphaFold 2 were published in Nature with full methodological disclosure. AlphaFold 2 won Demis Hassabis a share of the 2024 Nobel Prize in Chemistry, and its predictions were made freely available in a public database covering over 200 million protein structures. That openness accelerated research worldwide and became a poster child for how open science and AI can work together.

IsoDDE takes the opposite approach. The technical report, while showcasing impressive results, offers scant insight into the novel mechanisms that make the system work.

"It's a major advance, on the scale of an AlphaFold 4. The problem, of course, is that we know nothing of the details," said Mohammed AlQuraishi, a computational biologist at Columbia University who develops open-source AlphaFold variants.

AlQuraishi was particularly impressed by IsoDDE's ability to generalize beyond its training data, suggesting "they must've done something pretty novel." But the proprietary wall means outside researchers cannot verify claims, reproduce results, or build on the work.

Max Jaderberg, Isomorphic Labs' president, confirmed the company intends to keep its "secret sauce" proprietary, attributing advances to "a combination of compute, data, and algorithms" without revealing specifics. Different versions of IsoDDE have been tailored for specific pharmaceutical partners, incorporating a mix of publicly available datasets, synthetic training data generated by the AI itself, and licensed proprietary data from partner companies.

The Data Question

Diego del Alamo, a structural biologist at Takeda Pharmaceuticals, noted Isomorphic's "extensive efforts to partner with industry and potentially access their private structural data." This raises a genuine question: how much of IsoDDE's advantage comes from architectural innovation versus privileged access to datasets that no academic lab could match?

Gabriele Corso, co-developer of the open-source Boltz-2 model at MIT, struck a more optimistic note. He believes improvements using publicly available data suggest the gap can be closed: "a new baseline to match - but also to pass."

Billion-Dollar Partnerships and Delayed Trials

Isomorphic Labs has secured drug development deals with Johnson & Johnson, Eli Lilly, and Novartis - partnerships potentially worth billions in milestone payments. The company manages 17 active drug development programs spanning oncology, immunology, and cardiovascular disease.

Partner	Deal Focus	Estimated Value
Eli Lilly	Drug discovery collaboration	~$1.7B in potential milestones
Novartis	Drug discovery collaboration	~$1.2B in potential milestones
Johnson & Johnson	Drug development	Undisclosed

But the clinical timeline tells a more sober story. At the World Economic Forum in January, Hassabis said the company's first AI-designed drug would enter clinical trials by the end of 2026 - a delay from his previous target of late 2025. Drug discovery may be getting faster, but regulatory approval and clinical safety studies remain stubbornly analog processes that AI benchmarks alone cannot shortcut.

"We can maybe reduce that down from years to maybe months or maybe even weeks," Hassabis said in an interview about the typical decade-long drug development timeline.

IsoDDE does not eliminate clinical trials, safety studies, or regulatory review. It compresses the candidate discovery phase - historically the slowest and most expensive segment - but the hard part of proving a drug is safe in humans remains unchanged.

The Open-Source Counter-Attack

The scientific community is not standing still. Boltz-2, the open-source model that IsoDDE outperforms, continues active development. Teams at MIT, Columbia, and other institutions are pushing to democratize the same capabilities Isomorphic is locking behind commercial walls.

The tension here mirrors what we see across the broader AI landscape. Frontier models from major labs push capabilities forward but remain proprietary. Open-source alternatives eventually catch up. The question is whether the gap in drug discovery AI will narrow as quickly as it has in language models - or whether Isomorphic's data moat will prove more durable.

IsoDDE is a genuine technical achievement. The performance numbers are impressive, the scope of the system is ambitious, and the pharmaceutical partnerships suggest serious commercial validation. But calling it "AlphaFold 4" is both a compliment and an indictment. AlphaFold became transformative precisely because it was open. IsoDDE may be more powerful, but the world cannot verify that claim, cannot build on the work, and cannot hold the results to the same standard. In drug discovery, where lives hang in the balance, that opacity deserves more scrutiny than marvel.

Sources: