The Biomarker Discovery Framework is a multi-agent system designed to prioritize candidate biomarkers derived from wearable-device time series. Developed by researchers at Google Research (authors include Yubin Kim, Hamid Palangi, and Daniel McDuff), the framework was applied across three cohorts totaling 9,279 participant-observations in mental-health and metabolic domains.
Why this matters
Wearable devices now collect population-scale continuous physiological signals such as heart-rate dynamics and sleep behavior. The central challenge has shifted from data collection to converting these noisy streams into reliable, clinically meaningful biomarkers. Existing language-model-based agent systems often optimize predictive performance at the expense of statistical validity, producing spurious correlations, leakage, and fragile features.
Framework design
The Biomarker Discovery Framework structures candidate prioritization as an iterative research loop with human supervision, combining deterministic numerical computation (statistics, model training) and generative reasoning (hypothesis formation, literature grounding). An Orchestrator agent translates natural-language research directives into execution plans and coordinates specialized agents through a six-phase process. Shared memory, a structured fact sheet, and common tools maintain traceability across the workflow.
The six phases are:
- Data understanding: Scout agents map schema, missingness, temporal structure, and clinical endpoints; leakage controls separate target labels from feature construction.
- Candidate hypotheses grounding: Literature and Hypotheses agents retrieve and verify prior evidence, proposing physiologically plausible features and composite measures.
- Iterative discovery loop: Statistical and ML agents run deterministic code to construct features, estimate associations, adjust for multiple testing, and evaluate predictive signals. A Critic agent flags weak assumptions and gaps.
- Adversarial validation: Critic and Defender agents stress-test candidates for target leakage, overfitting, confounding sensitivity, construct overlap, instability, and physiological implausibility. An 11-check internal battery assigns reporting labels (screened, conditional, exploratory, rejected, unstable).
- Deep research & assessment: Mechanism, Novelty, and Strategy agents evaluate biological plausibility, prior literature, and translational relevance without treating associations as causal evidence.
- Report writing & assembly: Report agents verify numerical claims against the fact sheet and compile analyses, figures, literature, and limitations into a draft for expert review.
Example: prioritizing depression-related candidates
When asked to prioritize wearable-derived candidates associated with depression severity, the framework profiled the DWB dataset, proposed sleep-timing variability features, and estimated an association between sleep-duration variability and PHQ-8 severity (Spearman ρ = 0.252). The workflow then checked stability, leakage, subgroup consistency, and alternative explanations before presenting the result as a literature-grounded circadian-instability hypothesis for human review.
Results across three cohorts (N = 9,279)
The framework was run independently across three large-scale cohorts—DWB and GLOBEM (mental health) and WEAR-ME (metabolic disease)—totaling 9,279 participant-observations. Autonomously, the pipeline identified 41 candidate digital biomarkers for mental-health outcomes and 25 for metabolic outcomes.
Selected findings:
- DWB (depression): sleep-duration variability associated with PHQ-8 severity (ρ = 0.252, p < 0.001).
- GLOBEM (depression): sleep-onset variability emerged as an exploratory, low-signal association with PHQ-4 (ρ = 0.126, p < 0.001; cross-validated AUC = 0.535). Because cohorts, endpoints, and feature definitions differ and no identical candidate was replicated, this pattern suggests construct-level convergence rather than direct replication.
The system constructed novel composite features rather than only selecting existing variables. In the mental-health domain it highlighted sleep-duration and sleep-onset variability as top correlates of depression severity; in the metabolic domain it derived a cardiovascular fitness index (steps divided by resting heart rate) as a non-invasive correlate of insulin resistance, linking this index to prior literature on glucose regulation and cardiometabolic fitness.
Impact on predictive performance: When combined with demographic variables, features produced by the Biomarker Discovery Framework improved downstream prediction: ΔR² = 0.040 for depression and ΔR² = 0.021 for insulin resistance.
Expert evaluation
Fifteen domain experts in medicine, biomedical data science, machine learning, bioinformatics, and digital health performed blinded reviews of reports produced by the Biomarker Discovery Framework and three contemporary AI research systems (Google DeepMind’s AI co-scientist, Biomni, and Google ADK’s Data Science Agent). The systems were scored together in 21 sessions, and the Biomarker Discovery Framework was also evaluated separately in 13 sessions using the same instrument.
Key outcomes from the blinded evaluation:
- The Biomarker Discovery Framework achieved the highest mean scores across all seven quality dimensions.
- Under the study’s simulated editorial rubric it was the only system to receive any “Accept” or “Minor Revision” recommendations: 2 Accept, 8 Minor Revision, 8 Major Revision, and 3 Reject.
- Reviewers estimated they would retain on average 56.9% of Biomarker Discovery Framework–generated manuscript content, versus 18.8%–30.4% for baseline systems.
- In four-system ranking sessions, the Biomarker Discovery Framework finished first in 9 of 13 sessions.
Conclusion
As wearable health data continue to scale, the bottleneck in digital medicine lies in principled, rigorous hypothesis generation and validation rather than data collection. Increasing model capacity alone does not ensure scientific rigor. The Biomarker Discovery Framework demonstrates that structuring workflows to separate deterministic computation from generative reasoning, enforcing adversarial debate among agents, and preserving human-in-the-loop review can enable AI systems to safely accelerate the hypothesis-to-validation cycle in clinical research.
Acknowledgements
This work was written by Yubin Kim, Hamid Palangi, and Daniel McDuff from Google Research. The effort was led by MIT PhD student Yubin Kim during a Google internship advised by Daniel McDuff and Hamid Palangi. The authors thank co-authors and collaborators from Google Research, Google DeepMind, and academia for their contributions.
