The Bio AI Race Is Shifting From Open Benchmarks to Closed Engines
In the last couple of weeks, one of the clearest signals in bio AI has been that the center of gravity is moving from public benchmark wins to proprietary interaction engines. Nature reported that Isomorphic Labs has announced a new drug discovery model that researchers describe in almost mythical terms, because it appears to push beyond what AlphaFold style systems enabled, but it is not being released. The headline is not just performance. The headline is that the most valuable layer is now interaction, meaning protein plus ligand plus binding pocket plus affinity, and whoever owns that layer can compress years of iteration into something closer to a software loop. https://www.nature.com/articles/d41586-026-00365-7
Isomorphic’s own description is explicit about the target. They frame their Drug Design Engine, IsoDDE, as outperforming AlphaFold 3 on a protein ligand generalisation benchmark and as predicting binding affinities with accuracy they claim exceeds physics based methods at a fraction of time and cost. Whether you take every claim at face value or not, the strategic message is obvious: the model is being positioned as a productized scientific instrument, not a paper. https://www.isomorphiclabs.com/articles/the-isomorphic-labs-drug-design-engine-unlocks-a-new-frontier
At the same time, pharma is doubling down on the same interaction centric idea through partnerships that look more like platform buys than single program collaborations. Reuters reported that Takeda expanded its AI push with a deal tied to Iambic and its NeuralPLexer model for predicting how small molecules bind proteins, with the promise of compressing preclinical timelines. This is the same playbook, build a binding prediction engine, connect it to automation, and turn discovery into a closed loop where the model proposes, the lab tests, and the model learns. https://www.reuters.com/business/healthcare-pharmaceuticals/takeda-deepens-ai-drug-discovery-push-with-17-billion-iambic-deal-2026-02-09/
What this does to biology as a field is subtle but huge. In the AlphaFold era, the conversation was about democratizing structural biology. In this next phase, the advantage comes from integrating models with experimental throughput, data rights, and manufacturing constraints, which naturally pushes the best systems behind the firewall. Berkeley Lab recently described a version of this future in its work on foundational models for autonomous research, where the goal is prediction plus design plus smart data collection, not just analysis. https://newscenter.lbl.gov/2026/02/02/foundational-ai-models-to-accelerate-biological-discovery/
The practical takeaway is that the competitive boundary is moving. It is no longer enough to publish a model that predicts a structure or a sequence. The durable advantage is an engine that can predict interactions, propose interventions, and learn from real experiments faster than competitors can run their own cycles. That is where biology starts to look less like a set of assays and more like an operating system for discovery.
Sources https://www.nature.com/articles/d41586-026-00365-7
https://www.isomorphiclabs.com/articles/the-isomorphic-labs-drug-design-engine-unlocks-a-new-frontier
https://www.reuters.com/business/healthcare-pharmaceuticals/takeda-deepens-ai-drug-discovery-push-with-17-billion-iambic-deal-2026-02-09/
https://newscenter.lbl.gov/2026/02/02/foundational-ai-models-to-accelerate-biological-discovery/