Comment from Vyona Health

Vyona Health submits the following comments on the above-referenced draft guidance. Vyona Health is a health technology company developing causal AI infrastructure for longitudinal patient data, with an active Stanford Research to the People Program grant on multi-omic causal modeling in pulmonary fibrosis. Our comments address a foundational evidentiary challenge the Plausible Mechanism Framework raises but does not yet resolve: the inferential standards required to distinguish genuine treatment effects from natural variability in small, heterogeneous patient populations. I. The Guidance Requires Causal Distinction But Provides No Methodological Framework for Achieving It The draft guidance requires sponsors to demonstrate "that the improvement in outcome cannot reasonably be attributed to alternative treatments or natural variability in the disease phenotype" (lines 119-121) and that natural history data be "adequate to allow for the treatment effect to be reasonably distinguished from natural variability" (lines 470-471). These are causal claims that require explicit confounding structure modeling, not just data collection. Without methodological guidance, sponsors may rely on unarticulated causal reasoning that produces inconsistent evidentiary standards across submissions. We recommend FDA develop supplementary guidance on minimum causal inference standards for natural history external controls, including explicit confounding structure representation, pre-specified adjustment approaches, and sensitivity analyses for unmeasured confounding. II. The Framework's Biomarker Standards Implicitly Require Causal Pathway Modeling The guidance states that determining whether a biomarker predicts clinical benefit "depends on an understanding of the role of the biomarker in the causal pathway of the disease" (lines 578-580), but does not specify what constitutes adequate causal pathway characterization for novel diseases without established surrogate endpoints. We recommend the guidance specify what evidence FDA considers necessary to establish a biomarker's causal role for surrogate endpoint qualification. Multi-omic causal modeling — integrating genomic, proteomic, and clinical data to reconstruct causal pathway structure — should be explicitly recognized as a valid evidence source for this purpose. III. The Data Sharing Recommendation Should Address Federated Evidence Generation The guidance encourages data sharing (lines 734-743) but limits this to consent language permitting future data use. This does not address the practical barriers to direct data sharing: HIPAA requirements, institutional data governance, and proprietary clinical data. Federated causal modeling — pooling model parameters and causal effect estimates across institutions without sharing patient records — offers a mature path to achieving the compounding-evidence objective under real-world constraints. FDA should explicitly recognize federated evidence generation as valid for constructing natural history external controls and shared biomarker reference data, and provide clarity on how such analyses should be documented in submissions. Vyona Health would welcome the opportunity to engage further with FDA on these topics. Respectfully submitted, Julie Vaughn Co-Founder and CEO, Vyona Health julie@vyonahealth.com https://www.vyonahealth.com