Unsupervised proteomic and metabolomic longitudinal serum profiling of anti-CD19 CAR-T cell therapy recipients reveals clinically distinct groups

Abstract Introduction: Despite the clinical success of Chimeric Antigen Receptor (CAR)-T cell therapy, it still faces multiple challenges, such as the development of potentially life-threatening toxicities and limited survival efficacy in some patients. Recently, researchers have increasingly applied proteomic and metabolomic analyses to investigate therapy responses and adverse events, typically employing supervised methods to compare predefined groups. We hypothesize that distinct clinical manifestations following anti-CD19 CAR-T cell therapy are driven by underlying proteomic and metabolomic patterns, which can be uncovered through an unsupervised, data-driven approach. To test this, we conducted preliminary unsupervised serum profiling to identify intrinsic molecular clusters without predefined group bias. Methods: We longitudinally collected serum samples from 17 patients (pts) treated with anti-CD19 cells utilizing 4-1BB costimulatory domain (16 tisa-cel, 1 point-of-care). Fourteen pts were treated for large B-cell lymphoma and 3 for acute lymphoblastic leukemia (median age = 45, median prior therapy lines = 3). Timepoints of interest included day -7 (before lymphodepletion), day 0 (before infusion), and day 7 after the infusion. The clinical outcomes of interest included 1-month complete response (CR) and cytokine release syndrome (CRS) status (ASTCT criteria). We analyzed the samples with mass spectrometry (MS) (shotgun proteomics – ESI-Q-TOF tandem MS with liquid chromatography; shotgun metabolomics – ESI-FT-ICR, direct infusion, Pos/Neg modes). We performed unsupervised analysis via UMAP dimension reduction combined with DBSCAN clustering. We included proteins and metabolites present in ≥80 % of samples per timepoint (avg. 51 proteins, 144 metabolites). We assessed cluster-clinical outcome associations, evaluated the proteins and metabolites by Spearman correlation with UMAP coordinates, and mapped molecular pathways. Results: Before lymphodepletion, unsupervised proteomics identified three distinct clusters significantly associated with the key outcomes. Cluster one consisted of 5 pts, all of whom subsequently achieved CR, while developing no or mild CRS (grade 1 in two pts). Cluster two comprised 6 pts, all CRS-positive (4x g3, 1x g2, 1x g1), and four progressing after CAR-T cells. All but one patient in the third cluster responded to the therapy without severe CRS. Proteins significantly correlating with the UMAP coordinates included, among others, complement factor I, fibrinogen, immunoglobulins, and alpha glycoproteins (r: 0.31-0.56, p<0.05). Metabolomics revealed two distinct clusters with no clinical association. At day 0, unsupervised proteomics revealed four discrete clusters. The first cluster gathered 5 responders without significant CRS and corresponded to cluster one from day minus 7. Cluster two included exclusively 3 non-responders with severe CRS. Clusters three and four combined mixed clinical phenotypes of the remaining pts. Fisher's test showed a significant cluster association with the primary outcomes. Correlation analysis indicated complement factor B, haptoglobin, fibronectin, and plasminogen as the best matches to the UMAP coordinates (r: 0.62-0.75, p<0.05). Again, metabolomics revealed two distinct clusters without clinical association. At day 7, both proteomics and metabolomics identified two separate clusters with overlapping clinical presentations. In both cases, cluster one comprised responders with mild or no CRS, while cluster two combined non-responders with severe CRS and the remaining pts. Albumin, serotransferrin, immunoglobulin kappa, inter-alpha-trypsin inhibitor, and a complex set of amino acid, lipid, and carbohydrate metabolites correlated significantly with UMAP coordinates (r: 0.59-0.97, p<0.05). No association between cluster assignment and gender, diagnosis, or bridging status was detected at any timepoint. Conclusions: The unsupervised profiling enabled the identification of clinically meaningful clusters based on omics data. At each timepoint, we observed at least one favorable cluster associated with treatment response and mild or no CRS. Proteomics proved more informative than metabolomics, capturing both innate and adaptive immune signatures. Metabolomic clustering lacked clinical relevance except for day −7, likely due to the functional non-specificity of many metabolites. Funding: Supported by the National Science Centre, Poland - grant 2024/53/N/NZ5/00933 to J. T.