Immune profiling to identify a functionally high-risk smoldering multiple myeloma patient population.

Abstract

7534 Background: Smoldering multiple myeloma (SMM), a precursor to active multiple myeloma (MM), is characterized by a high plasma cell burden but no evidence of the end-organ damage that defines MM. The Mayo 2/20/20 model, and others, relies on tumor-burden estimates to assign risk scores. We hypothesized that algorithm-assisted and explainable artificial intelligence (xAI) tools could ingest peripheral blood (PB) T cell profiles to identify immune signatures predictive of progression to active MM. Methods: We analyzed a cohort of Mayo risk-matched SMM patients with and without early progression to active MM. This included 9 patients with early progression from SMM to MM (median PFS 2.1 years) and a 9 patient “mayo risk matched” cohort without clinical progression (median follow-up 8.6 years). Using high-dimensional spectral cytometry with a 37-color T cell focused panel, we captured 1.4 million PB T cells from 18 SMM patients banked at the time of SMM diagnosis. Algorithm-assisted analysis was performed using dimensionality reduction analysis with UMAP and cell clustering using PhenoGraph. xAI analysis was performed by training a random forest (RF) classifier to predict clinical outcomes using single-cell data followed by feature importance analysis using Shapley Additive Explanations (SHAP) scores. Results: Analysis identified 21 unique T cell characteristic clusters across all patients. Among these, SMM patients with early progression had enrichment for CD8 + CD45RA + CD62L - CCR7 - T effector cells re-expressing CD45RA when compared to non-progressing SMM patients (4.3-fold increase, p = 0.018). This cluster had the highest mean expression level of CD57 and TOX among all algorithm-defined clusters, demonstrating similar phenotypic characteristics to terminally exhausted effector T cells. The RF model to predict progression had an overall accuracy of 75% (stratified five-fold cross validated, repeated ten times). A UMAP analysis of c misclassified cells did not reveal any obvious patterns. SHAP analysis identified high expression of Granzyme B, CD272, Granzyme K, and CD45RA as the four most influential features for predicting progression. Conclusions: Our results show that patient-specific immune phenotypes could offer a method of prognosticating SMM outcomes separate from traditional tumor burden quantification. Both the clustering-based and feature importance analyses demonstrated that a more differentiated T cell phenotype is associated with early progression in SMM. Recent reports have found more differentiated T cell biology in MM patients compared to SMM patients. Our results support the hypothesis that SMM patients displaying an “MM-like” T cell phenotype are at increased risk of early progression to active MM. These results support work to identify a clinically usable patient-specific immune signature to identify SMM patients at increased risk of progression to overt MM.