Isomorphic Labs Unveils IsoDDE: AlphaFold 4-Level Breakthrough, Closed-Source

A paradigm shift in AI-driven drug discovery — unprecedented performance, unprecedented opacity.

🚀 Technical Leap: Beyond AlphaFold 3

Isomorphic Labs — the DeepMind spin-off co-founded by Demis Hassabis and led by him as CEO — has released IsoDDE, a next-generation AI drug design engine hailed by Nature and computational biologists as the de facto “AlphaFold 4“.

Unlike its predecessors, IsoDDE is not an incremental upgrade — it’s a unified architecture integrating:

Protein structure prediction
Binding affinity estimation (ΔG)
Hidden cryptic binding site discovery
Antibody–antigen interaction modeling

⚡ Benchmark Dominance

In the rigorous Runs N’ Poses benchmark — designed to test generalization on unseen protein folds — IsoDDE outperforms AlphaFold 3 (AF3) by over 2× when sequence similarity to training data drops below 20% (the hardest regime).

Metric	IsoDDE	AlphaFold 3	Boltz-2
Antibody Target Recognition (High-Accuracy)	✅ 2.3× AF3	—	✅ ~20×
Cryptic Site Detection (Cereblon)	✅ Both sites (incl. 15-yr-hidden)	❌ Missed second site	—
Binding Affinity Prediction	✅ Surpasses FEP simulations	❌ Lags behind physics-based methods	❌ Sub-FEP accuracy

IsoDDE vs AlphaFold 3 on cereblon binding site prediction
IsoDDE identifies both known and cryptic binding pockets in cereblon — including one hidden for 15 years. AlphaFold 3 fails on the second.

⏱️ Speed & Accessibility

Discovers cryptic binding sites in seconds, versus weeks/months of crystallographic soaking experiments.
Requires only amino acid sequence input — no experimental structural data needed.
Outperforms Free Energy Perturbation (FEP), the gold-standard physical simulation method — without any lab-derived starting structures.

Performance comparison on Runs N' Poses benchmark

🔒 The Closed-Source Turn: A Strategic Pivot

Despite its scientific magnitude, IsoDDE breaks from AlphaFold’s open ethos:

❌ No source code release
❌ No peer-reviewed publication (only a 27-page technical report)
❌ No model architecture or training methodology disclosed
❌ No public API or academic access pathway

“We do not intend to disclose the ‘secret sauce.'”
— Max Jaderberg, President of Isomorphic Labs (Nature, Feb 2026)

Isomorphic Labs technical report cover

This marks a decisive departure from AlphaFold’s legacy: over 3 million researchers globally used the open models to accelerate discoveries — from malaria vaccine design to neurodegenerative disease targets.

🧪 Why Does Closure Matter? Data, Not Just Code

Critics highlight a deeper concern: IsoDDE’s edge may stem less from algorithmic novelty and more from exclusive private datasets — proprietary protein–ligand complexes obtained via partnerships with Eli Lilly ($1.7B) and Novartis ($11.5B), and internal pipelines across 17 drug development programs.

As Diego del Alamo (Computational Structural Biologist, Takeda) notes:

“The advantage isn’t necessarily architectural — it’s data moat depth. And that moat isn’t replicable by academia.”

Isomorphic Labs partnership announcement

🌐 Open Ecosystem Response: Acceleration, Not Surrender

The closed release has galvanized open-source innovation:

Boltz-2 (nonprofit Boltz): Claims “full transparency + competitive accuracy” — trained exclusively on public data (PDBbind, BindingDB). Founder Gabriele Corso asserts: “The ceiling is higher than we thought — and it’s reachable without private vaults.”
Deep Origin’s DODock: Announced same-week performance parity on Runs N’ Poses using a novel diffusion-based docking framework — entirely open methodology.
Community momentum: AlphaFold 3-inspired models like Chai-1, Protenix, and RoseTTAFold All-Atom now match or exceed AF3 on key subtasks — proving rapid open iteration remains viable.

DODock vs IsoDDE performance claim

🚪 The Bigger Question: Who Controls Scientific AI?

AlphaFold symbolized AI as global public infrastructure. IsoDDE signals a pivot toward IP-locked commercial infrastructure.

This raises foundational questions:

Can science advance equitably when the most powerful tools are gated by licensing and capital?
Does reproducibility survive when methods are undisclosed and data is siloed?
Is the Nobel-caliber ideal — “knowledge for all” — giving way to “advantage for those who pay”?

As Mohammed AlQuraishi (Columbia University) puts it:

“We see the output. We admire it. But we’re left guessing how — and that’s not science. That’s spectacle.”

AlphaFold 2 open release milestone

🔗 Key Resources

Article originally published by Xin Zhī Yuán (New Intelligence Era). Translated and reformatted for global scientific audience.