A biotech startup says the AI drug-discovery industry’s biggest problem is not compute power or model size, but the data these systems are trained on: static snapshots of cells and animal models that fail to capture how living human tissue actually behaves.
Vivodyne, a University of Pennsylvania spinout that has raised just under $80 million from investors including Khosla Ventures, last week opened what it calls the world’s largest human data center just outside San Francisco. The facility houses hundreds of the company’s HIVE robotic labs, which autonomously grow 20 kinds of human tissue, dose them with compounds, and track biological responses in real time.
The Problem: AI Cures Cancer in Mice
CEO Andrei Georgescu argues that current AI drug-discovery models lack the causal data needed to understand human biology. Existing training datasets mostly come from animal testing or studies of isolated cells and proteins, not living tissue systems.
Absent human testing, what are these models going to do? They are going to cure cancer in mice.
The critique comes at a moment when leading AI figures have tempered expectations. Anthropic CEO Dario Amodei wrote over the weekend that claims AI will cure cancer have become more clich than credible. Google DeepMind CEO Demis Hassabis last year predicted AI could cure all disease within a decade, but Isomorphic Labs, the company founded to commercialize Nobel-prize-winning AlphaFold, has yet to produce its first clinical drug.
The pharmaceutical industry already faces a brutal attrition rate: roughly 90 percent of drugs that appear effective in animal testing fail to win regulatory approval for humans. Georgescu says this failure stems from a fundamental data problem rather than a modeling one.
Causal Data, Not Just More Data
A study published in Nature Methods last month found no clear data scaling laws when training generative AI models on existing cellular data, suggesting that simply feeding more of the same data into larger models will not solve the problem.
Vivodyne’s HIVE labs track hundreds of thousands of ongoing experiments where diseased tissue is exposed to stimuli, generating what Georgescu calls causal biological data: not just what a cell looks like, but how it got there. The company says its liver cells achieve 94 percent predictive accuracy for human toxicity tests, its airway tissue matches real human tissue behavior 96 percent of the time, and its bone marrow tests showed 100 percent concordance across 20 different chemotherapy drugs.
The company claims it is already achieving twice the throughput of all animal trials currently being conducted in the United States.
From Crash Tests to Combination Therapies
Georgescu compares the approach to automotive crash testing: automakers are typically confident their vehicles will pass regulatory requirements before formal testing, but drugmakers rarely have comparable confidence before entering expensive clinical trials. Vivodyne says it is working with multiple major pharmaceutical companies, though it has not disclosed their names publicly.
The longer-term vision involves training AI models on Vivodyne’s causal dataset using reinforcement learning, which Georgescu believes could unlock progress on combination therapies that target multiple disease pathways simultaneously. The search space for multi-drug combinations is so vast that traditional experimental approaches cannot cover it, he argues.
Sources: TechCrunch; Nature Methods; Isomorphic Labs
discussion