Mechanism-Aware ML Identifies a Protease-Chemokine-Galectin (PCG) Axis that Links Plasma Proteomics to Single-Cell Signaling and Enables Compact Severity Classification in COVID-19

Document Type

Poster Presentation

Publication Date

3-8-2026

Abstract

Background: Machine learning (ML) can accurately classify infectious disease severity, such as with COVID-19, yet many such approaches remain black boxes, providing little mechanistic insight into biochemical drivers of disease. To advance interpretability, we sought to re-analyze COVID-19 plasma omic and CyTOF datasets with mechanism-aware machine learning. Our purpose was to identify a compact, low-plex protein panel that captures the biochemical basis of immune dysregulation, with the potential to be extended to other respiratory viral diseases.   Methods: We re-analyzed the ImmProt (https://www.immport.org) SDY2129 cohort, comprising Olink NPX proteomics (~1,400 proteins) and CyTOF single-cell signaling (~3,500 frequency features) from patients with mild, moderate, or severe COVID-19. Data were cleaned, scaled, and merged at the patient level, with imputation for missing values and removal of near-zero variances or high collinear features (|r| > 0.95), yielding 150 ranked variables. We trained Random Forest (RF) models with stratified splits and nested cross-validation, and interpretability was achieved with Shapley Additive Explanations (SHAP). SHAP-derived contributions were grouped into modules. Connections between proteomic and CyTOF phopho-states were visualized through network schematics, and performance assessed relative to feature set size. Results: The RF models revealed that PCG proteins and their signaling correlates contributed significantly to severity classification. High-ranking features LGALS1, CCL7, PLAUR, FURIN< and TNFRSF10A together with stimulus-responses pSTAT, pERK, and pNF-kB readouts—these features aligned with reduced dendritic-cell antiviral signaling, indicating systemic dampening of interferon pathways. THE PCG axis explained a substantial share of the model variables, demonstrating direct biochemical coupling between extracellular proteolysis, chemotactic cell recruitment, and apoptotic or glycan-mediated immune regulation. Notable, a PCG-only proteomic subset achieved classification metrics comparable to the full multi-omic model (accuracy = 0.83; AUC = 0.94).   Conclusions: This ImmPort re-analysis with Mechanism-aware ML reframes multi-omic COVID-19 proteomic and CyToF data into a Protease-Chemokine-Galectin signaling axis that mechanistically links proteolytic activation (FURIN, PLAUR), chemokine trafficking (CCL7), and immunomodulatory lectins (LGALS1) to intracellular phosphor-networks controlling inflammatory reactivity. This biochemically interpretable framework yields a compact, explainable market set suitable for translation and extension to environmental-respiratory viral immunology beyond COVID-19. Future work will validate the PCG axis in other datasets and test its generalizability environmental-respiratory viral immunology studies.

Host

Gaylord National Harbor Resort and Convention Center

Conference/Symposium

American Society for Biochemistry and Molecular Biology (ASBMB) annual meeting

City/State

Oxon Hill, MD

Department

College of Arts and Sciences

Comments

Funding Information: We gratefully acknowledge ImmPort (https://www.immport.org), funded by NIAID/NIH, for providing the immunological data used. Data supporting this publication are available under study accession SDY2129. We thank the original investigators for generating these valuable resources.

----------------------------------------------------------------------

Félix Rivera-Mariani, associate professor in the College of Arts and Sciences, presented research at the 2026 American Society for Biochemistry and Molecular Biology (ASBMB) annual meeting, highlighting a mechanism-aware machine learning approach to identify immune signaling drivers of COVID-19 severity. His study re-analyzed large multi-omic datasets and identified a Protease–Chemokine–Galectin signaling axis linking plasma proteomics to intracellular immune signaling, enabling accurate, interpretable classification of disease severity with a compact biomarker panel. 

Rivera-Mariani conducted the research in collaboration with mentees, including Lynn University students and alumni, as well as a trainee from his alma mater, the Department of Microbiology and Immunology at the University of Puerto Rico School of Medicine (Medical Sciences Campus). This collaborative mentorship effort highlights the integration of computational biology, immunology and translational research to advance explainable artificial intelligence approaches for infectious disease and respiratory immunology research.

This document is currently not available here.


Share

COinS