CoDaS Prioritizes 66 Biomarker Candidates From Wearables

Google Research introduced the Biomarker Discovery Framework on August 21, 2026, a multi-agent pipeline for prioritizing digital biomarker candidates from continuous wearable data. The system, called CoDaS, or AI Co-Data-Scientist, targets a bottleneck in health research: converting high-dimensional sensor streams into statistically defensible signals that clinicians can evaluate.

Wearables produce streams from PPG heart-rate dynamics, electrodermal activity, sleep, skin temperature, and movement sensors. The volume is manageable at collection time, while feature discovery remains difficult because physiological data contains temporal dependencies, missingness, confounding, and opportunities for target leakage.

A research loop with separate reasoning and computation

CoDaS coordinates six phases through an Orchestrator agent. Scout agents inspect schemas, missingness, temporal structure, and clinical endpoints. Literature and Hypotheses agents ground candidate features in biomedical research before Statistical and ML agents run the analysis.

The architecture draws a strict boundary between generative reasoning and deterministic execution. LLMs propose hypotheses, search literature, and synthesize findings. Code performs numerical calculations, statistical adjustments, feature extraction, and model training. This separation gives each result an auditable path from raw sensor data to a candidate biomarker.

The discovery loop includes multiple-testing corrections such as FDR and Bonferroni. Critic and Defender agents then apply an 11-test adversarial validation battery covering target leakage, overfitting, confounding sensitivity, construct overlap, and tautological feature definitions. A feature derived directly from squared glucose, for example, would be challenged when used to predict insulin resistance.

This design reflects a broader pattern in multi-agent systems: specialized agents divide work, while a coordinating layer enforces workflow and review boundaries. In scientific applications, specialization matters because a fluent explanation cannot substitute for a reproducible statistical test.

66 candidates across mental and metabolic health

The framework ran across three real-world cohorts totaling 9,279 participant-observations and produced 66 health signal candidates, including 41 mental health candidates and 25 metabolic candidates.

Domain	Candidate signal	Association
Depression, DWB	Sleep-duration variability and PHQ-8 severity	ρ = 0.252, p < 0.001
Depression, GLOBEM	Sleep-onset variability and PHQ-4	ρ = 0.126, p < 0.001
Metabolic health	Wearable cardiovascular fitness index and insulin resistance	ρ = -0.374, p < 0.001

The depression results converged on circadian instability, with variability in sleep duration or sleep onset tracking symptom severity across two cohorts. The metabolic result combined step count and resting heart rate into a wearable cardiovascular fitness index that moved inversely with insulin resistance.

Adding the AI-selected wearable features to baseline demographic models improved predictive performance. Depression severity gained ΔR² = 0.040, while insulin resistance gained ΔR² = 0.021. These are incremental improvements, which is the relevant standard for a biomarker pipeline: the wearable signal must add information beyond variables already available to a model.

Expert review favored validation-heavy outputs

A blinded evaluation with 15 medical and clinical experts compared CoDaS-generated manuscripts and candidate profiles with outputs from baseline science agents, including AI Co-Scientist and Biomni. Experts retained an average of 56.9% of the CoDaS content, compared with 18.8% to 30.4% for the baseline systems.

In a separate review involving 12 clinicians, validity judgments aligned with CoDaS confidence tiers at ρ = 0.67, p = 0.005. The result connects the system’s internal prioritization with expert assessment, while clinical utility still depends on human judgment in the final phase.

For developers building research agents, the implementation lesson is concrete. Give language models responsibility for search, hypothesis formation, and explanation, then route every statistical claim through fixed tools, leakage controls, correction procedures, and adversarial tests. Evaluating AI agents requires measuring the full workflow, including failure detection and review quality, rather than judging generated prose alone.

If you build a wearable-health pipeline, treat CoDaS’s feature candidates as hypotheses for external validation, not diagnostic conclusions. Preserve cohort boundaries, run temporal and confounding checks before model comparison, and make clinician review part of the production workflow.