An adaptive, continuous-learning framework for clinical decision-making from proteome-wide biofluid data.

Abstract

Mass spectrometry (MS)-based proteomics provides deep molecular insights from patient samples, but clinical use has been limited by missing values, static biomarker panels, and the need for targeted assay development. We present a framework - Adaptive Diagnostic Architecture for Personalized Testing by Mass Spectrometry (ADAPT-MS) - that enables direct diagnostic and prognostic interpretation of discovery-mode proteomics data at the level of individual samples. ADAPT-MS dynamically retrains simple, robust classifiers based on the proteins quantified in each sample, eliminating the need for imputation or fixed panels. Applied to plasma and cerebrospinal fluid datasets across diseases and clinical centers, it achieves high performance and generalizability using robust, transparent and generalizable statistical models. A single proteomic measurement can support multiple diagnostic questions via retrospective cohort matching, with each classification taking only seconds. As population-scale proteomics datasets grow, this approach lays the foundation for scalable, real-time, and personalized diagnostics directly from proteome-wide data. Such a community effort may help to transform discovery proteomics into a routine clinical tool.