Prompt-as-a-Protocol with Agentic Research Automation: A Validated Methodology for AI-Native Biomedical Data Science Producing Peer-Reviewed Findings Across Genomics and Biosensor Domains

Recent advances in large language model (LLM)-driven agentic systems have demonstrated capabilities for automating biomedical data analysis workflows. However, existing systems are predominantly validated on synthetic benchmarks or retrospective tasks, without demonstrated production of novel, peer-reviewed biological findings. Here we describe Prompt-as-a-Protocol (PaaP), a methodology that converts natural-language research questions into structured, executable analytical protocols, coupled with an agentic Generative Research Automation (GRA) layer that executes these protocols with integrated quality assurance. We validate PaaP+GRA through two prospective case studies in distinct biomedical domains. In single-cell transcriptomics, the pipeline identified KIF3A as the sole genome-wide significant correlate of ciliation status in human striatal parvalbumin interneurons (adjusted p = 1.37e^-14), while autonomously detecting and correcting a sex composition confound (p = 8.1e^-44) that would have produced spurious results. In biosensor analytics, the pipeline achieved ECG-derived HRV stress-state classification with ROC AUC of 0.946 (95% CI: 0.871-0.998) under leave-one-subject-out cross-validation, while identifying and correcting subject leakage in the original train-test split that had inflated performance estimates. Both studies have been submitted for peer review. In both case studies, the researcher's original prompt did not anticipate the critical methodological threat; the GRA quality assurance layer identified and corrected it autonomously, producing stronger results than the researcher would have achieved alone. To our knowledge, PaaP+GRA is the first prompt-to-protocol methodology validated by prospective production of novel biological findings across multiple domains, with autonomous detection of both statistical confounds and methodological flaws, rather than benchmark performance alone.