Researchers at Google DeepMind have developed a new AI system that accelerates hypothesis generation, experimental testing and result analysis, according to a paper published in the journal Nature Communications.
The platform consists of several neural networks, each performing a typical role of a research scientist. They analyse scientific literature, generate hypotheses, and assess their potential and possible connections to other fields of science.
The AI system is not autonomous, however. It requires a scientist to correct and guide it.
The researchers tested the platform on three molecular biology tasks: finding new uses for existing drugs, identifying new protein targets for treating pulmonary fibrosis and related diseases, and uncovering mechanisms of antibiotic resistance in pathogenic microbes.
In all three cases, the AI system generated interesting hypotheses, some of which were confirmed in practice. For example, drug combinations proposed by the neural networks suppressed the growth of several acute myeloid leukaemia cell lines.
A similar AI system has already been launched by US company OpenAI – called GPT-Rosalind. It is being used for research by Amgen, Moderna and Thermo Fisher Scientific, among others.