Vivodyne argues that the current AI‑driven drug discovery pipeline is limited by a lack of causal, human‑relevant biological data, relying instead on animal models or isolated cell assays. To address this gap, the company has built HIVE, a modular robotic laboratory capable of culturing twenty distinct human tissue types, automatically administering compounds, and continuously monitoring phenotypic responses. According to CEO Andrei Georgescu, without such data AI models are effectively only learning to “cure cancer in mice,” a point echoed by Anthropic’s Dario Amodei, who warned that cancer‑curing claims have become cliché.
The HIVE system generates the kind of mechanistic, dose‑response information that existing foundation models for biology have not been trained on. Georgescu calls for a “sanity check,” noting that the pharmaceutical industry already sees about ninety percent of compounds that succeed in animal studies fail in human trials. While AlphaFold demonstrated a major advance in protein structure prediction, it has not yet yielded a new drug, and Isomorphic Labs—founded to extend that work—expects its first clinical trials by the end of 2024, later than the original 2025 target. Only a handful of AI‑designed molecules have reached human testing, with one advancing to Phase III.
The startup’s approach shifts the data generation bottleneck from costly, low‑throughput animal experiments to an automated, tissue‑centric platform, aiming to supply the high‑fidelity, multi‑scale biological signals that modern AI architectures require for reliable prediction of drug efficacy and safety.
Why this matters
The HIVE platform directly tackles the data scarcity that limits current AI models for drug discovery by providing scalable, human‑tissue‑derived causal datasets. If successful, this could reduce the high attrition rate observed when moving from animal to human studies—a noted 90 % failure—by giving algorithms richer signals on efficacy and toxicity. However, the approach remains experimental; translating automated tissue readouts into reliable predictive features for foundation models has yet to be demonstrated at scale.