Self-driven Biological Discovery through Automated Hypothesis Generation and Experimental Validation

Modern scientific challenges demand faster, more reliable research workflows. A primary goal for applying artificial intelligence in science is the automation of research, offering faster discoveries and new insights. Large Language Models (LLMs) are revolutionizing AI, achieving impressive results on tasks previously requiring human intelligence. By using explicit mathematical models and logic to generate hypotheses, LLMs can be provided a logical scaffold to reason around, reducing hallucinations and enhancing the reliability of its outputs.

Here we demonstrate a method that uses LLMs to automate parts of experimental design, using relational learningderived hypotheses and physical laboratory constraints. We integrate this methodology with an automated laboratory cell and metabolomics platform, presenting a flexible and efficient approach to automated scientific discovery in a user-defined hypothesis space. We evaluate the methodology on multiple interaction experiments in Saccharomyces cerevisiae , revealing, among other findings, antagonistic effects between glutamate and spermine.

We record the hypotheses, experimental designs, and empirical data in a graph database, using controlled vocabularies. We extend existing ontologies and present a novel way of representing scientific hypotheses using description logics.

We also offer a proof-of-concept demonstrating how metabolomics data can be used to further refine generated hypotheses.