AlphaFold 4 Goes Closed‑Source: IsoDDE Beats AlphaFold 3 in Drug Design

Isomorphic Labs, the AI drug‑design arm of DeepMind led by CEO Demis Hassabis, released a 27‑page technical report describing IsoDDE, a next‑generation protein‑structure and drug‑design engine that the authors refer to as “AlphaFold 4.” The report claims IsoDDE’s performance “comprehensively crushes” the previous AlphaFold 3 model.

IsoDDE integrates structure prediction, binding‑affinity calculation, and hidden‑site discovery into a single engine. In the “Runs N’ Poses” benchmark, which tests AI on completely unseen protein structures (similarity to training data 0‑20%), IsoDDE’s success rate is more than twice that of AlphaFold 3. Among the 60 hardest cases, IsoDDE solved 17 where AlphaFold 3 failed entirely.

For antibody‑target recognition, IsoDDE achieves a 2.3× higher high‑precision success rate than AlphaFold 3 and nearly 20× higher than the open‑source model Boltz‑2. The report highlights a striking case: the protein cereblon, long thought to have a single drug‑binding pocket, was experimentally shown to have a second hidden pocket; IsoDDE identified both pockets from the amino‑acid sequence alone.

Speed is another advantage: where traditional labs need costly crystal‑soaking experiments and weeks of work, IsoDDE produces predictions in seconds. Max Jaderberg, president of Isomorphic Labs, told Nature that the “recipe” will not be released, and computational biologist Mohammed AlQuraishi called the generalization ability “astonishing.”

While IsoDDE’s results are impressive, the report provides almost no details about model architecture or training methods, and the code and paper are withheld. This contrasts sharply with AlphaFold 2, which was open‑sourced in 2021 and made its predictions freely available to over three million researchers worldwide, a fact highlighted by the Nobel Committee as a paradigm of open science.

Critics argue that the lack of transparency creates a “closed‑door” future for AI‑driven science. If IsoDDE’s superiority stems from proprietary data—potentially sourced from collaborations with pharma giants such as Roche and Eli Lilly—then the community cannot assess the true source of its advantage. AlQuraishi warned that “we know nothing about the details,” underscoring academic concerns.

Nevertheless, the closed‑source release has ignited competition. Open‑source projects like Boltz‑1/2, Chai‑1, and Protenix have already narrowed the gap with AlphaFold 3, and other companies (e.g., Deep Origin’s DODock) claim comparable performance on the same benchmarks using different approaches.

The broader implication is that the era of AI as a public scientific good may be ending. As AI models for drug discovery become commercial assets controlled by a few companies, the open‑science narrative that once celebrated free, global access to powerful tools like AlphaFold is being challenged. The article concludes that the distance between an open, community‑driven future and a closed, proprietary one is the pivotal choice facing the scientific community today.