A Multi-Agent Automated Platform for Scientific Discovery in Systems Biology

The automation of scientific research has the potential to transform science. Large Language Models (LLMs) are revolutionizing AI, achieving impressive results on tasks that previously required human intelligence. Here we show that LLMs can be provided with a logical scaffold for scientific reasoning through explicit mathematical models and logical hypothesis generation. This reduces incoherence and enhances output reliability. We demonstrate the use of LLMs to automate experimental design, employing relational learning–derived hypotheses, and physical laboratory constraints. This methodology is integrated with an automated laboratory cell and metabolomics platform, presenting a flexible and efficient approach to automated scientific discovery in a user-defined hypothesis space.

Hypotheses, experimental designs, and empirical data are all recorded in a graph database using controlled vocabularies. Existing ontologies are extended, and a novel representation of scientific hypotheses is presented using description logics.

We evaluate our approach on multiple interaction experiments in the yeast Saccharomyces cerevisiae, revealing, among other findings, synergistic growth inhibition when glutamate is added to spermine-treated cells. Additionally, we offer a proof-of-concept demonstrating how metabolomics data can automatically refine generated hypotheses, revealing a previously unknown interaction.