(PA-229) Machine Learning–Based Unsupervised Molecular Subtyping of Malignant Plasma Single-Cell RNA Data Reveals Tumor-Immune Features Linked to Poor Outcomes in Multiple Myeloma

Introduction: Multiple myeloma (MM) patients exhibit diverse clinical progression and therapeutic response, driven by tumor intrinsic features such as somatic alterations impacting the genetic and epigenetic landscape of a patient as well immune dysfunction in the tumor microenvironment (TME). We hypothesized that identifying molecular subtypes using unsupervised machine learning on gene expression profiles of malignant plasmas cells would uncover clinically actionable subtypes. In the previous pilot study, (Hamidi et al., 2023) we applied this approach to samples from 95 patients. In this study, we expanded our cohort to 521 samples and performed clustering on malignant plasma cells and not total plasma cells to better capture the heterogeneity in the tumors.

Methods: Baseline samples from 521 MM patients from the MMRF CoMMpass study (NCT01454297), with more than four malignant plasma cells were pseudobulked and expression profiles were subjected to unsupervised learning based on consensus non-negative matrix factorization (cNMF) to identify molecular subtypes.

Results: 2,921,589 cells passed through quality control and doublet detection, 13.00% of which were plasma cells (n=379,882). Five distinct molecular subsets were discovered using cNMF. These subtypes were significantly associated with both T2T (p=0.042) and overall survival (OS; P=0.001).  
Subtype pNMF5.1 was associated with the worst clinical outcomes, including the shortest median T2T and OS. A multivariable Cox proportional hazards model incorporating clinical covariates such as age, gender, race, ISS stage and treatment regimen confirmed that the molecular subtypes were independently associated with OS.  
Biologically, plasma cells in worst outcome subset, pNMF5.1, were enriched in inflammatory and stress-related pathways including TNFa signaling, IL2-STAT5 signaling, apoptosis, hypoxia and unfolded protein response. This subtype also exhibited transitional EMT markers, indicating a tumor state marked by inflammation and cellular stress.  
Analyzing the TME scRNA-seq data revealed that the worst patient subset, pNMF5.1, had reduction in several key regulatory and adaptive cell populations including naïve CD4⁺ T, Tregs, naïve B, Memory B, and NK CD56 bright cells. Conversely, there was an expansion of pro-inflammatory and myeloid populations including inflammatory monocytes, granulocytes, CD8+ T effector and NK CD56 dim cells. This immune environment imbalance, characterized by chronic inflammation, loss of regulatory control and a diminished adaptive immune presence, likely contributes to disease progression and poor clinical outcomes.

Conclusions: Applying innovative unsupervised machine learning on the scRNA-seq data of malignant plasma cells improves our ability to more accurately identify molecular subtypes of MM. This approach offers the potential for accurate prediction of therapeutic response as well as capturing biologically and clinically informative features beyond traditional risk factors.