Skip to content

Autonomous Agent-Orchestrated Digital Twins (AADT): Leveraging the OpenClaw Framework for State Synchronization in Rare Genetic Disorders

🕒 Published (v1): 2026-03-28 03:18 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

AADT is an agent-orchestrated Medical Digital Twin (MDT) framework for rare genetic disorders that uses OpenClaw's scheduled "heartbeat" mechanism and modular Agent Skills to continuously synchronize a patient's virtual state with evolving phenotype reports and external genomic database reclassifications. Rather than building a new predictive model, it provides orchestration infrastructure that bridges the synchronization gap between static MDT snapshots and continuously changing clinical reality. Two illustrative case studies demonstrate autonomous variant reinterpretation and longitudinal phenotype tracking in rare undiagnosed diseases.

Problem

Most MDT implementations are static or only passively updated, creating a synchronization gap that is especially severe in rare genetic disorders where (1) patient phenotypes evolve slowly and non-specifically before crystallizing into diagnosable patterns, (2) variant classifications in databases like ClinVar are revised frequently, and (3) reinterpretation workflows are manual and clinician-initiated, introducing latency between new evidence and clinical action.

Method

The framework layers a constrained local AI agent (OpenClaw v2026.3) as an orchestration middleware over deterministic bioinformatics tooling:

  • Heartbeat mechanism: OpenClaw runs a scheduled polling cycle (configurable cadence) that autonomously monitors local file directories and remote databases without requiring human prompting.
  • AgentSkill system: Domain-specific automation routines are packaged as installable skills via ClawHub. A single skill, PhenoSkill, implements two functions:
  • Phenotype monitoring: On caregiver free-text input, invokes PhenoSnap (a deterministic Python CLI using spaCy's PhraseMatcher) to extract HPO terms with negation detection, redact PII, and append timestamped PhenoPacket-compliant JSON records to the patient's digital twin.
  • ClinVar Monthly Scan: A time-triggered heartbeat task that converts patient variants to SPDI notation and queries NCBI E-utilities (esearch + esummary) for reclassifications from VUS to Pathogenic/Likely Pathogenic at \(\geq\) two-star review confidence.
  • Event-sourced state: The digital twin state is a single JSON file assembled from all patient data files; each update is an immutable, timestamped event preserving full provenance. The LLM handles only orchestration (event detection, natural language understanding, workflow routing); all clinical data transformation is performed by PhenoSnap, making results reproducible regardless of which LLM backs OpenClaw.

Key Contributions

  • AADT framework architecture: event-driven, local-first, constrained-agency MDT orchestration via OpenClaw.
  • PhenoSkill: an OpenClaw Agent Skill (ClawHub-installable) combining phenotype extraction and variant monitoring in a single heartbeat-driven routine.
  • PhenoSnap: a deterministic, version-controlled phenotype extraction and ClinVar variant lookup toolkit that separates AI inference from structured clinical data production.
  • Event-sourced Rare Disease MDT (RDMDT) state representation integrating VCF genotype, longitudinal HPO phenotype, and auxiliary clinical data (EEG, imaging).
  • Two illustrative case studies demonstrating autonomous diagnostic re-ranking (phenotype crystallization) and automated VUS-to-Likely-Pathogenic reclassification detection.

Results

  • Case Study 1 (Angelman syndrome phenotype tracking): Correct diagnosis rose from rank 327th at birth to rank 2nd at age 3 years and 1st at age 4 years as caregiver-submitted HPO terms accumulated. Interval from phenotype emergence to diagnostic re-prediction reduced from a typical multi-week clinical review cycle to under 15 minutes.
  • Case Study 2 (EYS variant reinterpretation): After 8 months of monthly silent scans with no reclassification event, the heartbeat detected NM_001142800.2(EYS):c.2137+1G>A upgraded from VUS to Likely Pathogenic (new splicing assay data, concordant multi-lab submissions), automatically surfacing a structured alert with evidentiary summary to the clinical team.
  • Both results are from illustrative case studies; no prospective quantitative benchmarks or controlled comparisons against baselines are reported.

Limitations

  • Case studies are illustrative, not prospective; the paper makes no claim about real-world impact on patient outcomes or clinical efficiency.
  • Requires interoperable data sources, local computing infrastructure, and institutional governance mechanisms that may not be uniformly available.
  • Lacks formal mechanisms for federated or multi-institutional deployment, limiting rare-disease cohort aggregation.
  • No prospective clinical validation; external validity across diverse care settings is unestablished.
  • Vulnerability surface of extensible agent ecosystems (prompt injection, permission misconfiguration, PHI exfiltration) is acknowledged but mitigated only architecturally, not formally evaluated.

Relevance to Harnesses / Meta-Harnesses

AADT is a concrete deployed instance of a meta-harness pattern: a scheduled, event-driven controller (OpenClaw's heartbeat) that dispatches typed sub-tasks (Agent Skills) while constraining LLM involvement to orchestration and routing, delegating all deterministic data work to versioned CLI tools. The ClawHub skill registry is architecturally equivalent to a plugin registry in modern harness systems, enabling composable, installable capabilities without monolithic agent re-training. The strict separation between LLM-mediated inference and deterministic tool execution directly instantiates the harness principle of isolating non-reproducible AI calls from auditable, reproducible pipeline steps — a design pattern increasingly advocated for in agentic AI safety literature. For researchers tracking meta-harnesses, AADT provides a domain-grounded case study of how heartbeat scheduling, modular skill dispatch, event-sourced state, and human-in-the-loop escalation compose into a production-viable orchestration layer.