Cross-Platform Proteomics and Machine Learning Nominate Plasma Biomarkers of Stroke Diagnosis

10 Mar 20264 min read0 viewsJournal Feed

GIST

This study investigates plasma biomarkers to aid stroke diagnosis using cross-platform proteomics and machine learning. In a case–control design, adults with suspected stroke were enrolled in emergency departments before treatment, with discovery (n=100) spanning acute ischemic stroke, intracerebral hemorrhage, transient ischemic attack, and stroke mimics, and external validation (n=80) with equal representation across groups.

SomaScan proteomics quantified 7307 proteins, yielding 61 proteins differentiating stroke subtypes. Protein panels were derived via LASSO logistic regression, with internal validation by repeated nested cross-validation (rCV) and targeted MS, and external validation by data-independent acquisition MS in an independent Yale cohort.

Reported performance shows subtype-specific classifiers: 7 for acute ischemic stroke (rCV-AUC 0.82; 95% CI 0.78–0.86), 6 for intracerebral hemorrhage (AUC 0.70; 95% CI 0.64–0.76), 8 for transient ischemic attack (AUC 0.78; 95% CI 0.73–0.84), and 7 for stroke mimics (AUC 0.81; 95% CI 0.77–0.86). Internal validation identified 11 proteins; external validation identified 32 proteins, with VTN, PLG, and S100A9 highlighted among top classifiers.

Uncertainty remains regarding generalizability beyond the study cohorts, and larger multicenter validation is suggested.

Clinical Editorial

Overview and purpose

The study investigates plasma-based biomarkers to aid stroke subtyping in emergency settings, employing cross-platform proteomics and machine learning to distinguish major diagnostic groups.

Design and populations

Case–control framework leveraging two biorepositories.
Participants were adults with suspected stroke presenting to the emergency department before treatment.
Discovery cohort: 100 subjects divided into four groups (acute ischemic stroke, intracerebral hemorrhage, transient ischemic attack, stroke mimics) with 40, 20, 20, and 20 individuals respectively.
External validation cohort: 80 subjects, with 20 per diagnostic group.

Analytical platforms and discovery approach

Proteomic discovery employed SomaScan to quantify plasma proteins, yielding a broad proteome readout.
Differential enrichment analyses compared across the four diagnostic categories, with adjustments for clinical covariates.
Nomination of protein panels occurred via pairwise and multigroup comparisons.
Model construction used least absolute shrinkage and selection operator (LASSO) logistic regression to derive classifier panels.

Key findings and diagnostic signatures

SomaScan quantified 7307 proteins; 61 showed differential enrichment across stroke subtypes.
Acute ischemic stroke classifiers: seven protein panel with internal cross-validation AUC 0.82 (95% CI 0.78–0.86).
Intracerebral hemorrhage classifiers: six-protein panel with internal AUC 0.70 (95% CI 0.64–0.76).
Transient ischemic attack classifiers: eight-protein panel with internal AUC 0.78 (95% CI 0.73–0.84).
Stroke mimics classifiers: seven-protein panel with internal AUC 0.81 (95% CI 0.77–0.86).
Targeted proteomics provided internal validation for 11 proteins; data-independent acquisition (DIA) MS external validation identified 32 proteins, with notable top classifiers including VTN (vitronectin), PLG (plasminogen), and S100A9 for stroke mimics, TIA, and intracerebral hemorrhage respectively.

Validation and corroboration

Internal validation utilized repeated nested cross-validation (rCV) to assess model robustness.
External validation employed DIA-MS in an independent Yale cohort, reinforcing a subset of protein classifiers.

Context and limitations

The investigation demonstrates the feasibility of plasma proteomics to nominate biomarkers discriminating stroke subtypes.
The authors call for larger, multicenter validation to establish biomarker-guided diagnostic utility in acute care.