IsoDDE vs AlphaFold 3: Faster Drug Design, No Approval - philippdubach.comSkip to main contentphilippdubach<br>Quantitative finance, AI, and the economics underneath.
IsoDDE hits 50% on the hardest protein-ligand prediction benchmark versus 23% for AlphaFold 3 and beats the physics-based gold standard FEP+ on binding affinity with a 0.85 Pearson correlation<br>AI-discovered drugs show 80-90% Phase I success rates versus a 40-65% historical average, but Phase II efficacy rates remain roughly 40% for both AI and traditional drugs<br>Isomorphic’s pharma deals total over $4 billion in headline value but only $82.5 million in upfront cash, a 50:1 ratio that reflects how much pharma is betting on contingent outcomes<br>No AI-discovered drug has received FDA approval as of February 2026, and Isomorphic targets its first clinical candidates for late 2026<br>No AI-discovered drug has ever received FDA approval. That sentence should sit uncomfortably next to every headline about Alphabet’s drug discovery spinoff.
On February 10, Isomorphic Labs, the Google DeepMind spinoff focused on computational drug design, released IsoDDE: its Drug Design Engine. This isn’t a model or an AlphaFold upgrade. IsoDDE is a unified in silico drug discovery system that runs protein structure prediction, ligand binding, affinity estimation, and pocket identification in concert, generating in seconds what used to take days of physics-based simulation. On the hardest molecular prediction tasks, the “Runs N’ Poses” benchmark designed to test generalization to unfamiliar proteins, IsoDDE hits a 50% success rate. AlphaFold 3 manages roughly 23% . On antibody-antigen modeling, IsoDDE beats AlphaFold 3 by 2.3× and the open-source Boltz-2 by 19.8×. On binding affinity prediction, it achieves a Pearson correlation of 0.85, beating the physics-based gold standard FEP+ at 0.78.<br>I would assume that these are large enough improvements that the computational bottleneck in drug design may no longer be the binding question.<br>What pharma believes<br>Isomorphic has signed partnerships with Eli Lilly, Novartis, and Johnson & Johnson worth a combined $4 billion+ in potential value. But look at the structure. Lilly paid $45 million upfront against $1.7 billion in milestones. Novartis paid $37.5 million upfront against $1.2 billion. That’s a 50:1 ratio between what pharma promises in biobucks and what it actually wires.<br>This ratio is standard across AI drug discovery deals in 2025. Pharma is enthusiastic enough to sign but cautious enough to make nearly all the economics contingent on clinical results that don’t exist yet. The upfront payments fund research. The milestone payments are structured so that pharma loses almost nothing if the drugs fail. The royalties only matter if a drug reaches blockbuster status, which for an AI-designed molecule has never happened.<br>Novartis expanded its partnership in February 2025, doubling the number of programs to six, targeting what Novartis described as “particularly challenging” and previously undruggable targets, on the same financial terms. That’s a positive signal: it means internal results impressed Novartis scientists enough to commit more targets. The J&J deal, announced January 2026, goes further, covering small molecules, antibodies, peptides, and molecular glues. But “expanded partnerships” and “approved drugs” remain separated by the most unforgiving filter in business: human biology.<br>Phase II wall<br>Most commentary on AI drug discovery stops too early. Jayatunga et al. (2024), in the first systematic analysis of AI-discovered drugs in clinical trials, showed AI-discovered molecules achieving 80-90% success rates in Phase I trials, well above the historical 40-65% average. AI is good at designing molecules that are safe and have decent pharmacokinetic properties: they get absorbed, distributed, metabolized, and excreted the way you’d want. Phase I is mostly about safety. AI passes it.<br>But Phase II is about efficacy. Does the drug actually treat the disease? And here the numbers are sobering: AI-discovered drugs show roughly 40% Phase II success rates, which is about the same as traditionally discovered drugs. AI has not yet demonstrated it can predict whether a molecule will work in a patient, only that it can predict whether a molecule will be tolerable in a patient.<br>If both trends hold, end-to-end success rates could rise from the historical 5-10% to something like 9-18%. That would roughly double R&D productivity, which in a trillion-dollar industry is worth an enormous amount. McKinsey estimates generative AI could generate $60-110 billion annually in economic value for pharma and medical products. But it’s a far cry from the narrative that generative AI will “solve” drug discovery. It would make drug development somewhat cheaper and faster. An improvement, not a revolution.<br>The counterargument, and it’s a...