Google releases AI drug design engine IsoDDE: hailed as "AlphaFold 4," outperforming the previous generation, but no longer open source

Google’s subsidiary, Isomorphic Labs, led by DeepMind CEO Demis Hassabis, has released a new generation AI drug design engine called IsoDDE, which Nature calls “AlphaFold 4.”

It completely outperforms its predecessor but is entirely closed source. The golden age of AI benefiting science may be closing its doors.

In 2024, Demis Hassabis will stand on the Nobel stage because of AlphaFold.

This AI model, capable of predicting the three-dimensional structure of proteins, is used by over 3 million researchers in more than 190 countries, making it a benchmark example of AI benefiting all humanity.

The Nobel Committee’s recognition is less about an algorithm and more about a spirit — providing the most powerful scientific tools free of charge to every researcher.

Sixteen months later, AlphaFold’s successor has arrived.

On February 10, Hassabis’s AI pharmaceutical company, Isomorphic Labs, released a 27-page technical report showcasing a drug design engine called IsoDDE, which fully surpasses AlphaFold 3 and is described by Columbia University computational biologist Mohammed AlQuraishi as “a major leap akin to AlphaFold 4.”

But this time, the code will not be open, the paper will not be published, and the methods will not be shared.

Max Jaderberg, President of Isomorphic Labs, told Nature straightforwardly: “We do not intend to disclose the ‘secret recipe.’”

The open-source legend of AlphaFold may very well end with the third generation.

The capabilities are truly astonishing

First, let’s clarify what IsoDDE has achieved, which helps explain why the controversy is so intense.

To use a rough analogy: imagine proteins as locks and drug molecules as keys. AlphaFold’s job is to help you see what the lock looks like.

But just seeing the lock isn’t enough — you need to know if the key can turn, how tightly it fits, and whether there are other hidden keyholes you haven’t noticed.

IsoDDE aims to answer these more difficult questions.

It is a unified engine that integrates structure prediction, binding affinity calculation, hidden binding site discovery, and more.

Numbers speak volumes.

In a test specifically designed to evaluate AI’s ability to handle “never seen before” new protein structures (the Runs N’ Poses benchmark), when the similarity between test samples and training data drops to 0-20% (the most challenging scenario), IsoDDE’s success rate is more than double that of AlphaFold 3 (AF3).

Out of 60 most difficult cases, 17 were completely failed by AlphaFold 3 but correctly predicted by IsoDDE.

AlphaFold 3 fails in this example, IsoDDE succeeds

In predicting how antibodies recognize their targets, IsoDDE’s high-precision prediction success rate is 2.3 times that of AlphaFold 3, and nearly 20 times that of another mainstream open-source model, Boltz-2.

What surprised colleagues most was the combined affinity prediction — assessing how tightly a drug molecule binds to its target.

Traditionally, this task relies on a physics simulation method called FEP, which is computationally expensive and requires crystal structures as starting points.

In multiple public tests, IsoDDE not only outperformed all AI methods but also surpassed FEP, all without needing any experimental data as a starting point.

The technical report also features a particularly impressive case.

A protein called cereblon, which scientists believed for 15 years had only one drug-binding site, was recently found to have a second hidden site through experimental research.

IsoDDE, simply inputting the amino acid sequence of this protein, identified both sites — including the one hidden for 15 years.

Laboratories would need costly crystal soaking experiments and extensive time to do the same, but IsoDDE does it in seconds.

AlQuraishi said he was most shocked by IsoDDE’s generalization ability on completely unfamiliar molecular systems, “which indicates they must have done something very innovative.”

Closed source: the real concern

If IsoDDE were just a typical commercial software, being closed source would be understandable and unremarkable.

The problem is, its predecessor AlphaFold represented a very different set of values.

AlphaFold 2 was open-sourced in 2021, with the accompanying paper published in Nature, and its predictions freely available worldwide.

This was more than just a technical achievement — it demonstrated a possibility: cutting-edge AI research funded by tech giants can truly become a public good for all humanity.

Over 3 million scientists used it for their research, accelerating countless projects and transforming the entire biological research landscape.

AlphaFold 3, released in 2024, also published a paper. Although there was controversy over the open-sourcing speed, it was ultimately made available to the academic community.

IsoDDE breaks this tradition.

The 27-page technical report contains almost no details about the model architecture or training methods.

Nature’s headline was blunt: “Scientists can only guess how similar results were achieved.”

Jaderberg’s words to Nature are thought-provoking. He hopes the report will “inspire” other teams.

But AlQuraishi’s reaction perhaps better reflects the academic community’s true feelings: “The problem is, we know nothing about the details.”

Some believe it is reasonable for Isomorphic Labs, as a commercial company, to protect its core technology. That’s certainly true.

But the question worth asking is: as AI capabilities in science grow stronger and more concentrated in a few companies, who will decide the openness of these capabilities?

Isomorphic Labs has already raised $600 million, signed potential collaborations worth nearly $3 billion with Eli Lilly and Novartis, and operates 17 drug pipelines internally.

Hassabis stated in Davos this January that the first AI-designed drugs are expected to enter clinical trials by the end of 2026.

The company is transforming from a research institution into a commercial machine.

Diego del Alamo, a computational structural biologist at Takeda Pharmaceuticals, pointed out another subtlety: Isomorphic Labs previously invested heavily in collaborations with pharmaceutical companies, possibly gaining access to large amounts of proprietary experimental data.

How much these additional data contribute to IsoDDE’s performance remains unknown.

If the core advantage relies on data barriers rather than algorithmic innovation, then the so-called “incentive” is more of a gesture.

Open-source community is not giving up

Closed source has sparked anxiety but also fueled competition.

Gabriele Corso, co-developer of Boltz-2 and founder of nonprofit Boltz, made his stance clear: he doesn’t believe proprietary data is the key factor, as there is still much room for improvement in publicly available data.

IsoDDE has set a new performance baseline, “it needs to be caught up with and can be surpassed.”

Another company, Deep Origin, has been more high-profile, announcing the day after IsoDDE’s release that its DODock engine had already achieved comparable performance on the same benchmark in August 2025 — using a completely different technical approach.

Over the past two years, the open-source community has been active. After AlphaFold 3, several teams have developed open-source models that approach or even surpass its performance, including Boltz-1/2, Chai-1, Protenix, and others.

The AI drug discovery field is replaying the script from large language models: a company unveils impressive closed-source results, prompting the open-source community to rapidly follow, narrowing the gap from a generational distance to a raceable level.

But there is a key difference.

Training data for language models: internet text, an almost unlimited public resource.

In contrast, training data for AI drug discovery, especially high-quality protein-drug experimental data, is partly held by pharmaceutical companies.

If the moat of closed-source models is built on proprietary data, the difficulty for open-source models to catch up is much greater.

The door is closing

The impact of this may go beyond drug development itself.

In recent years, “AI open source drives scientific progress” has been a widely accepted narrative. AlphaFold was the strongest evidence for this.

Whenever doubts arose about whether tech giants’ AI research truly benefits everyone, AlphaFold was the best answer — look, 3 million scientists worldwide are using it for free.

Now, as AlphaFold’s direct descendants choose to go closed source, this narrative is being torn apart.

It hints at a possible future:

AI as the most powerful tool in basic science gradually shifting from a public good to a commercial asset;

Breakthrough results being released as technical reports rather than peer-reviewed papers;

The academic community can see the results but never the methods.

Hassabis once said that applying AI to science is a richer endeavor than language models. That’s true. But this richness depends on openness.

When the most advanced scientific AI is only accessible to paying customers, the vast majority of the scientific community can only watch from outside the fence.

The Nobel medal on AlphaFold bears the ideal of sharing knowledge with everyone. The technical report of IsoDDE describes a more powerful future.

The distance between the two reflects the choices this era is making.