The most visible impact of AI seems to be in coding, but it could also be poised to disrupt other hard sciences.
Isomorphic Labs, a drug discovery subsidiary of Google’s parent company Alphabet, announced today a major advancement in computational drug design that more than doubles the accuracy of previous AI models in predicting how molecules interact with proteins—a fundamental challenge in developing new medicines.
The company unveiled its Isomorphic Labs Drug Design Engine (IsoDDE), a unified computational system that represents what researchers describe as a significant leap beyond AlphaFold 3, the groundbreaking protein structure prediction model released in 2024. While AlphaFold 3 revolutionized scientists’ ability to predict the 3D shapes of proteins, IsoDDE goes further by addressing practical challenges required to actually design drugs on computers.
Understanding the Breakthrough
To appreciate what IsoDDE accomplishes, it helps to understand drug discovery’s core challenge: finding molecules that fit precisely into specific pockets on disease-causing proteins, like finding the right key for a lock. Traditional methods involve years of laboratory experiments, testing thousands of candidate molecules to find ones that bind effectively.
AI models like AlphaFold 3 made headlines by predicting protein structures with remarkable accuracy, but they struggled when faced with truly novel biological systems—scenarios that looked different from their training data. This is precisely where real drug discovery happens, since researchers are often pursuing unprecedented targets or mechanisms.
IsoDDE demonstrates dramatic improvements in three critical areas:
Structure Prediction for Novel Systems: On the “Runs N’ Poses” benchmark, specifically designed to test AI models on unfamiliar protein-drug combinations, IsoDDE achieved 50% accuracy on the most difficult cases—more than doubling AlphaFold 3’s 23.3% success rate. This means the system can reliably predict molecular structures even when dealing with proteins and drug candidates it has never encountered before.
The model can also predict complex biological phenomena like “induced fits,” where proteins reshape themselves to accommodate binding molecules, and “cryptic pockets”—hidden binding sites that only reveal themselves when the right molecule approaches. These are sophisticated real-world scenarios that previous models frequently missed.
Binding Affinity Prediction: Knowing the structure is only half the battle; drug developers need to know how strongly a candidate molecule will bind to its target. IsoDDE achieves a Pearson correlation of 0.85 on binding affinity predictions, surpassing traditional physics-based methods that require experimental crystal structures and take significantly longer to compute. This allows researchers to rapidly rank thousands of potential drug candidates.
Pocket Identification: Perhaps most intriguingly, IsoDDE can identify potential drug binding sites on proteins using only the amino acid sequence as input—no prior knowledge of where drugs might attach required. The system demonstrated this by rediscovering a novel binding pocket on cereblon, a protein involved in tagging damaged proteins for destruction. For 15 years, scientists believed there was only one way to drug cereblon, until a recent study found a previously hidden pocket. IsoDDE identified both the known and novel sites using sequence data alone.
From Antibodies to Aspirin
The system’s capabilities extend beyond small molecule drugs like aspirin. For complex biologics like antibodies—large protein-based therapeutics such as insulin—IsoDDE outperforms AlphaFold 3 by 2.3 times and competing model Boltz-2 by nearly 20 times in predicting high-quality antibody-antigen interactions.
Particularly notable is IsoDDE’s performance on the CDR-H3 loop, the most variable and challenging region of antibodies to predict accurately. This advancement could unlock new possibilities for designing antibodies from scratch rather than discovering them through traditional laboratory methods.
Implications for Drug Discovery
The announcement arrives at a pivotal moment for AI-driven drug discovery. While AlphaFold 3 accelerated basic research—with over 3 million researchers in more than 190 countries using the freely available AlphaFold Protein Database—translating structural predictions into actual drug development programs has remained challenging.
IsoDDE’s combination of improved accuracy, speed, and breadth suggests a potential shift toward genuinely computational drug design. Where physics-based simulation methods might take hours or days and require experimental starting structures, IsoDDE can generate predictions in seconds without such constraints.
Isomorphic Labs reports that its drug design teams are already deploying these capabilities across their programs to understand novel structures, identify previously uncharacterized binding pockets, and design new chemical compounds.
The Bigger Picture
The development represents more than incremental improvement. By achieving what Isomorphic Labs calls “predictive fidelity” across diverse biochemical properties while maintaining the ability to generalize to unexplored biological systems, IsoDDE addresses fundamental limitations that have constrained computational drug design. Google DeepMind CEO Demis Hassabis is bullish on the field in general, having said that humanity has a shot at curing all diseases through AI, and Isomorphic Labs has prevoiusly said it’s on the cusp of starting human trials for AI-created medicines.
Whether all this translates to faster, cheaper drug development or enables entirely new classes of therapeutics remains to be seen. Drug discovery involves many steps beyond initial molecular design, including extensive clinical testing. But for the computational chemistry component, at least, the tools appear to be reaching a new level of sophistication.
As AI systems demonstrate increasing capabilities in scientific domains—from protein folding to materials science—IsoDDE suggests that some of AI’s most transformative impacts may ultimately emerge not in consumer applications or software development, but in accelerating humanity’s understanding and manipulation of the molecular world.