AUTHORS
Fieke W Hoff, MD, PhD, Wei Xiao, PhD2, Ivo Veletic, MD3, Annie Hoai Hoai Nguyen4, Taghi Manshouri5, Bofei Wang, PhD6, Sherry Pierce, BSN, RN7, Jennifer Marvin-Peek, MD8, Mahesh Swaminathan, MD9, Naveen Pemmaraju, MD10, Prithviraj Bose, MD11, Lucia Masarova, MD12, Hussein Ali Abbas, MD, PhD13
AFFILIATIONS
1 University of Texas MD Anderson Cancer Center, Houston, Texas, United States
2 University of Texas MD Anderson Cancer Center, Houston, Texas, United States
3 University of Texas MD Anderson Cancer Center, Houston, Texas, United States
4 University of Texas MD Anderson Cancer Center, Houston, Texas, United States
5 University of Texas MD Anderson Cancer Center, Houston, Texas, United States
6 University of Texas MD Anderson Cancer Center, Houston, Texas, United States
7 University of Texas MD Anderson Cancer Center, Houston, Texas, United States
8 University of Texas MD Anderson Cancer Center, Houston, Texas, United States
9 University of Texas MD Anderson Cancer Center, Houston, Texas, United States
10 University of Texas MD Anderson Cancer Center, Houston, Texas, United States
11 University of Texas MD Anderson Cancer Center, Houston, Texas, United States
12 University of Texas MD Anderson Cancer Center, Houston, Texas, United States
13 University of Texas MD Anderson Cancer Center, Houston, Texas, United States
ABSTRACT
Background: Myelofibrosis (MF) is a heterogenous myeloproliferative neoplasm characterized by progressive bone marrow fibrosis, cytopenia, splenomegaly, and constitutional symptoms. Constitutive JAK-STAT pathway activation driven by JAK2, CALR, or MPL mutations is present in >90% of patients but does not fully explain the heterogeneity in disease phenotype, symptom burden, and treatment response. Chronic inflammation is a hallmark of MF, yet circulating inflammatory programs remain incompletely defined. We employed AI-driven plasma proteomics to define baseline inflammatory states associated with clinical phenotype and longitudinal remodeling during JAK inhibitor therapy.
Methods: Plasma proteins were quantified using a 250-plex NULISA inflammatory panel, an ultrasensitive proximity extension assay enabling attomolar (10-18 M) detection. A total of 165 MF patients were analyzed; 131 received ruxolitinib (RUX)-based therapy, and 34 untreated. Among treated patients, 39 had paired baseline (BSL) and follow-up (FU) samples. 19 healthy donors were also profiled. Comprehensive clinical, molecular, and genomic annotations were integrated. Unsupervised hierarchical clustering of BSL samples was performed and associated with clinical and survival outcomes. Paired BSL-FU differential expression was FDR-corrected; Random Forest/Boruta selection across iterative resampling identified the most stable proteins, with performance assessed by ROC-AUC on held-out data at each iteration.
Results: Of the 165 patients, 96 (58%) had primary MF and 69 (42%) had secondary MF. Driver mutations included JAK2 (73%), CALR (13%) or MPL (6%), and 8 patients were triple negative. Inflammatory proteins detectability was high (98%). Principal component analysis of BSL samples and healthy donors revealed distinct clustering. AI-driven unsupervised machine learning was applied to the high-dimensional proteomic dataset, identifying 5 distinct BSL inflammatory protein signatures (IPSs). IPSs were significantly associated with clinically meaningful features of disease burden, including leukocytosis (≥25K), hemoglobin (anemia), neutrophil count, fibrosis grade, splenomegaly, and DIPSS risk. The most adverse features were enriched in IPS5 and least common in IPS1 suggesting that distinct inflammatory protein states capture biological heterogeneity and correlate with disease severity. Notably, driver mutations, DIPSS-Plus and MF subtype were not associated with the IPSs, suggesting that inflammatory heterogeneity represents an independent layer of disease biology, not captured by established genetic and phenotypic classifications. Furthermore, among RUX-treated patients, IPS1 was associated with a significantly longer median OS than IPS5 (median OS, NR vs 5.4 years, p=0.006), supporting the prognostic relevance of inflammatory state classification.
IPS1 and 5 demonstrated strikingly different expression patterns. IPS1 was uniquely enriched for platelet activation, PDGF signaling, and extracellular matrix interactions, consistent with platelet- and stromal-associated tissue remodeling phenotype, whereas IPS5 was enriched for IFN signaling, innate/adaptive immune pathways, neutrophil degranulation, T-cell receptor signaling, and multiple interleukin pathways, consistent with an activated IFN-driven immune activation phenotype.
We next evaluated paired samples obtained before and after treatment initiation. Iterative Random Forest/Boruta selection (100 random resamples) ranked proteins by selection frequency; the top 20 most frequent formed the FUP-vs-BSL signature, achieving mean test-set AUC=0.83. All proteins were significantly downregulated after RUX treatment; however, the magnitude of suppression was variable, and was less pronounced in CALR-mutant disease, suggesting that residual inflammatory signaling proteins persist in defined patient subsets and may contribute to heterogeneity in treatment response.
Discussion: Leveraging AI-enabled analysis of ultra-sensitive high-dimensional proteomics, we identified inflammatory states that stratify MF patients according to clinical phenotype, disease burden, and survival, independent of driver mutations and DIPPS-Plus risk score. These findings highlight the biological relevance of inflammatory heterogeneity in MF and support proteomic profiling as a potential tool for risk stratification. Ongoing paired and other AI-based applications will be presented at ASH, and external validation studies are ongoing.
DISCLOSURES
Fieke W Hoff, MD, PhD: No relevant conflicts of interest to declare.