The industry's expectations in recent years have been high around the topic “AI's ability to accelerate the treatment of cancer”, but there are still limited cases of real clinical and breakthroughs. VIvodyne, a biotechnology start-up company, argues that bottlenecks are not primarily in modelling capacity, but that training data are still far from real human life processes.

According to the company, a number of AI drug research and development systems currently rely on data from animal experiments, single-cell studies or protein levels. These data help model structures and state, but it is difficult to explain why human tissue changes after irritation, medication or disease.

Progress is still limited

In the past few years, AI has been frequently used for drug screening, target detection and molecular design. A number of candidate drugs designed with AI participation have been introduced into human trials, with individual projects advancing to clinical phase III. However, in terms of overall results, there is still no evidence in this area that AI has significantly rewritten the success rate of new drug development.

The article mentions that AlphaFold has made significant progress on protein structural predictions, but has not yet produced a new drug directly. DeepMind's hatching Isomorphic Labs is still waiting for the first trials to advance. Vivodyne judged that existing models were not sufficient to cover the complexity of human biology.

HIVE develops 20 human tissues

Vivodyne was founded in 2021 from the University of Pennsylvania. The company developed HIVE, a modular robotic laboratory system that nurtures 20 human tissues and automatically completes the process of delivery, observation and documentation. The goal is to identify, before entering clinical practice, which candidate drugs are more likely to be effective or have a toxic risk.

  • Hepatotoxic predictions match 94%.
  • Airway behavioral match is 96%.
  • The bone marrow model was 100% consistent in 20 chemotherapy tests.

Last week, Vivodyne launched what it calls the world's largest “human data centre” near the San Francisco Bay region. According to the company, nearly US$ 80 million in financing had been completed, from two rounds of financing, with Khosla Ventures as one of the lead investors.

The company is betting on cause-and-effect biological data.

According to Vivodyne, what is really missing from drug research and development is “causal data”, that is, not only what the cell or tissue is in, but also how it is formed. In this line of thinking, models cannot only learn static state, but also need to understand the relationship between irritation, inflammation or delivery of drugs and outcomes.

According to the company, HIVE is following hundreds of thousands of ongoing experiments to document the evolution of the pathological organization under different incentives. Georgescu believes that such continuous data are better suited to train the next generation of biological models and have a better chance of helping AI understand human reactions rather than just static identification.

In his view, this is particularly important for future combination therapy. As the treatment of diseases increasingly involves multi-road interventions, the number of candidate combinations will rapidly increase and it will be difficult to complete the screening alone by traditional experimental methods. Vivodyne's bet is to build a data base closer to the real human reaction, and then talk about AI's massive breakthrough in the medical field.