Google Research presents CoDaS, a multi-agent system that structures work as a research loop: generating hypotheses, running statistical analysis, training models, grounding interpretations in literature and adversarially checking results under human oversight.

Across three cohorts totalling 9,279 participant-observations, the system identified circadian-variability signals associated with depression and a steps-to-resting-heart-rate index associated with insulin resistance. Adding the features to demographic models produced modest predictive gains.

The authors frame the output as candidate prioritisation rather than causal discovery. No identical biomarker replicated across cohorts, and clinicians rated actionable value below methodological validity. Prospective studies are required before health decisions.

Why it matters

The workflow can transfer to nutraceutical scouting: hypotheses, data, literature and criticism become a traceable chain rather than an isolated generative answer.

Evidence limit

A preprint and institutional research report. Associations are hypothesis-generating, predictive gains are modest and neither causality nor clinical utility has been shown.

Editorial note. This content is intended for industry professionals. It does not constitute medical advice, therapeutic guidance or regulatory advice. Read our editorial method.

Read the original source · Google Research ↗