Automated Biomarker Discovery
The Biomarker Discovery Framework is a multi-agent system designed to identify promising biomarker candidates from continuous physiological data collected by wearable devices. The system’s goal is to transform this data into clinically meaningful biomarkers, addressing the bottleneck of data collection while maintaining statistical validity. It operates through a structured, adversarial pipeline with human oversight, combining deterministic computation with generative reasoning. The framework’s architecture consists of specialized agents – Scout, Critic, Defender, Mechanism, and Report – that collaborate to analyze data and propose biomarker candidates.
At its core, the system utilizes a six-phase process mirroring the human biomarker prioritization lifecycle. The Orchestrator agent decomposes natural-language research directives into execution plans, guiding the specialized agents. Data understanding involves mapping the data schema and identifying potential issues like missingness and temporal structure. Candidate hypotheses grounding uses literature and prior evidence to propose features and composite measures. The iterative discovery loop employs statistical and machine learning agents to construct features and estimate associations, while the Critic agent identifies weaknesses and gaps. Finally, the adversarial validation stage rigorously tests candidates for leakage, overfitting, and instability, followed by deep research and assessment to evaluate biological plausibility.
During a trial run, the framework analyzed the DWB dataset in response to a request to prioritize wearable candidates associated with depression severity. It identified sleep-timing variability features and estimated an association between sleep-duration variability and PHQ-8 severity (ρ = 0.252). The system then checked stability, leakage, subgroup consistency, and alternative explanations before framing the result as a literature-grounded circadian-instability hypothesis for expert review.
Across three cohorts (N = 9,279 participant-observations) spanning mental health (DWB and GLOBEM) and metabolic disease (WEAR-ME) domains, the framework autonomously identified 41 candidate digital biomarkers for mental health and 25 for metabolic outcomes. The system’s ability to construct novel composite features, such as sleep duration variability and sleep onset variability as correlates of depression, highlights its potential for generating new insights. The framework enforces safety mechanisms to ensure statistical validity, separating feature construction from target signals and requiring candidates to pass an 11-test adversarial filtering stage.
