A proteomics-driven machine learning tool for distinguishing ET from pre-PMF

Abstract Background: Essential thrombocythemia (ET) and prefibrotic primary myelofibrosis (pre-PMF) are clinically overlapping but biologically distinct subtypes of myeloproliferative neoplasms (MPNs), making differential diagnosis both critical and challenging. Current diagnostic approaches rely heavily on histopathologic interpretation, which is subject to variability. Methods: We retrospectively analyzed 525 patients, including 434 with ET and 91 with pre-PMF. Clinical predictors were assessed by logistic regression in 440 patients with complete records. A subset of 85 patients underwent data-independent acquisition (DIA)–based proteomic profiling using FFPE bone marrow samples. Diagnostic models were developed based on clinical data, proteomic features, or a combination of both. Model performance was evaluated through nested cross-validation. A 9-protein panel was identified using random forest followed by support vector machine–recursive feature elimination (SVM-RFE). Results: The proteomic model significantly outperformed the clinical model (AUC = 0.849 vs. 0.499), and its performance was comparable to the combined model (AUC = 0.845). The 9-protein panel achieved robust discrimination (AUC = 0.895), and retained high accuracy in JAK2V617F⁺ (AUC = 0.971) and CALR⁺ (AUC = 0.768) molecular subsets. Key discriminatory proteins included ARHGEF19, CAST, and SFTPA2. Conclusion: Proteomic profiling of FFPE bone marrow specimens provides a reproducible, molecular-based approach for distinguishing ET from pre-PMF, outperforming conventional clinical parameters. This approach shows promise for early and accurate diagnosis across molecular subtypes of MPN.