Isomorphic Labs, the Google DeepMind spinoff founded by Nobel laureate Demis Hassabis, publicly revealed its Isomorphic Drug Design Engine (IsoDDE) on February 10. The system represents a significant step beyond AlphaFold 3, the protein structure prediction tool that helped earn Hassabis and colleague John Jumper their 2024 Nobel Prize in Chemistry.
Where AlphaFold 3 predicts how proteins fold, IsoDDE attempts something more ambitious: running entire drug discovery workflows computationally. The system unifies protein structure prediction, ligand binding, binding affinity estimation, antibody interactions, and pocket discovery into a single engine.
The Numbers
On a challenging protein-ligand generalization benchmark (Runs N’ Poses, specifically the 0-20% similarity bin for out-of-distribution data), IsoDDE more than doubles AlphaFold 3’s accuracy. For antibody-antigen prediction, it outperforms AlphaFold 3 by 2.3x and the open-source Boltz-2 by 19.8x on high-fidelity predictions.
The binding affinity predictions surpass all existing deep-learning methods “by a considerable margin” on established benchmarks including FEP+, OpenFE, and CASP16, matching the accuracy of physics-based free energy perturbation methods at a fraction of the computational cost and time.
Perhaps most notable is the blind pocket identification capability. IsoDDE can identify druggable binding sites on proteins using nothing but the amino acid sequence as input. In tests, it found both known and novel cryptic binding sites on cereblon - a key drug target - with accuracy that Isomorphic says approaches experimental techniques like fragment soaking. Finding these hidden pockets is one of the hardest problems in drug design, since many proteins only reveal their binding sites when they change shape during molecular interaction.
What It Actually Means for Drug Development
Hassabis has framed the potential in characteristically bold terms, telling 60 Minutes last August that AI could “reduce [drug discovery timelines] from years to maybe months or maybe even weeks.”
The reality is more measured. IsoDDE compresses the early discovery phase - the part where researchers identify promising molecules and predict how they’ll interact with targets. This stage typically consumes years of lab time and enormous amounts of capital. If the system works as claimed, it could meaningfully accelerate the pipeline before a drug candidate ever reaches clinical trials.
But it doesn’t eliminate clinical trials, safety studies, or the regulatory gauntlet. No AI-designed drug has received FDA approval yet. Isomorphic says it expects its first AI-designed drugs to enter clinical trials by the end of 2026, primarily targeting oncology, though even that timeline has reportedly shifted from earlier projections.
The Competitive Picture
Isomorphic isn’t operating in a vacuum. The AI drug discovery space has attracted enormous capital:
Eli Lilly and NVIDIA announced a $1 billion co-innovation lab in January focused on AI-driven drug discovery. Iambic Therapeutics signed a multi-year collaboration with Takeda worth $1.7 billion. Researchers in China released DrugCLIP with a 10,000-protein database for open scientific use.
What sets IsoDDE apart - if the benchmarks hold - is the breadth of the unified system. Rather than stitching together separate AI models for different tasks, it handles the entire computational drug design pipeline in one engine. That integration matters because errors compound when you chain multiple models together.
The Caveats
Benchmarks aren’t drugs. The gap between impressive computational predictions and medicines that work in human bodies remains vast. Physics-based drug design methods have been promising similar revolutions for decades without fully delivering.
Isomorphic’s benchmarks are self-reported. Independent verification on real-world drug programs, not curated academic datasets, will be the true test. The company has been using IsoDDE internally for some time before this public announcement, which could mean the results have been validated against actual drug programs - or it could mean the public benchmarks were selected to showcase strengths.
The oncology focus for initial clinical candidates is telling. Cancer drug development has relatively clear biological targets and well-established trial frameworks, making it a logical proving ground. The harder test will be whether AI-designed drugs can tackle diseases with less understood biology.
The Bigger Picture
We’re watching the AI drug discovery industry approach a moment of truth. After years of fundraising, partnerships, and benchmarks, the first wave of AI-designed molecules is heading toward human trials. Isomorphic, backed by Google’s deep pockets and the AlphaFold team’s track record, is arguably the best-positioned player in the field.
If their candidates survive clinical trials, it could fundamentally change how drugs get made. If they don’t, it won’t invalidate the technology - drug development has always had high failure rates - but it will force a more honest conversation about timelines and expectations.
Either way, IsoDDE represents a genuine technical milestone: the first publicly disclosed system that unifies the major computational steps of drug design into a single, high-performing engine. What happens next depends entirely on whether computational elegance translates to biological reality.