AUTHORS
David G Coffey, MD, Gargi Damle2, Deniz Demircioglu, PhD3, Geoffrey Kelly4, Travis Dawson5, Manishkumar Patel6, Darwin D'Souza7, Dan Hasson, PhD8, Michael Mason, PhD9, Alexandre Alloy10, Shameek Biswas11, Ola Landgren, MD, PhD12, Seunghee Kim-Schulze, PhD13, Sacha Gnjatic, PhD14, Samir Parekh, MD15, Sundar Jagannath, M.D.16, Hearn Cho, MD, PhD17, Selma Bekri18
AFFILIATIONS
1 University of Miami, Miami, Florida, United States
2 Tisch Cancer Institute, New York, New York, United States
3 Tisch Cancer Institute, New York, New York, United States
4 Icahn School of Medicine at Mount Sinai, New York, New York, United States
5 Icahn School of Medicine at Mount Sinai, New York, New York, United States
6 Icahn School of Medicine at Mount Sinai, New York, New York, United States
7 Icahn School of Medicine at Mount Sinai, New York, New York, United States
8 Tisch Cancer Institute, New York, New York, United States
9 Bristol Myers-Squibb, Seattle, United States
10 Bristol Myers-Squibb, Seattle, United States
11 Bristol Myers-Squibb, Seattle, United States
12 University of Miami, Miami, Florida, United States
13 Icahn School of Medicine at Mount Sinai, New York, New York, United States
14 Icahn School of Medicine at Mount Sinai, New York, New York, United States
15 Tisch Cancer Institute, New York, New York, United States
16 Icahn School of Medicine at Mount Sinai, New York, New York, United States
17 Tisch Cancer Institute, New York, New York, United States
18 Icahn School of Medicine at Mount Sinai, New York, New York, United States
ABSTRACT
Background:
Tumor progression in multiple myeloma (MM) is shaped by communication between malignant plasma cells and neighboring immune and stromal cells within the bone marrow tumor microenvironment (TME). Existing computational approaches primarily infer these interactions from ligand and receptor expression and typically do not incorporate tumor-specific genomic alterations that may modify signaling or its downstream consequences. We developed a novel tumor-aware framework that integrates transcriptomic, functional, and genomic features to estimate patient-specific cell-cell communication in MM.
Methods:
We applied this approach to 224 patients with newly diagnosed MM from the MMRF CoMMpass study. High-confidence secreted or membrane-associated ligand-receptor pairs were identified and integrated with immune-cell pseudobulk single-cell RNA sequencing, tumor bulk RNA sequencing, and somatic copy-number variants (CNVs) and single-nucleotide variants derived from whole-exome sequencing. For each candidate interaction, we combined ligand-receptor expression with receiver-cell functional gene programs, pathway activity, and genomic weights reflecting tumor-specific alterations. Each patient was represented as a weighted, directed graph in which nodes corresponded to cell populations and edges represented inferred communication strength. A graph neural network learned latent patient representations, which were grouped using Louvain community detection. Generalizability was evaluated in an external Mount Sinai cohort of 43 patients with newly diagnosed or relapsed MM.
Results:
The model preferentially assigned high scores to established MM-associated ligand-receptor interactions, including MIF-CD74/CXCR4, TGFB1-TGFBR2, TNFSF13/TNFSF13B-TNFRSF17, and IL6-IL6R signaling, whereas interactions specific to solid-tumor malignancies were not similarly enriched.
The graph neural network identified six communication subgroups associated with both progression-free survival (log-rank P = 1.28 x 10-4) and overall survival (P = 0.0446). Patients assigned to the same group had more similar communication profiles than patients assigned to different groups, supporting the robustness of the classification. The communication groups exhibited distinct network topologies, ligand-receptor interactions, functional programs, and pathway activities, although fibroblasts, Tem/Effector helper T cells, and non-classical monocytes were recurrent dominant signaling hubs.
Communication groups captured TME biology distinct from established tumor-intrinsic and clinical classifications. Removing SNV and CNV features preserved the association with PFS and OS, whereas removing downstream pathway and functional gene-set features eliminated statistical significance, supporting a primary contribution from functional signaling consequences rather than genomic features alone. External-cohort patients mapped to five of the six discovery groups. Group-specific hazard ratios were highly concordant between discovery and validation cohorts (Spearman R = 0.80 for PFS, 0.70 for OS) with no statistically significant differences across communication groups (P > 0.30).
In 29 patients with serial samples, within-patient communication profiles were more similar than interpatient profiles, and 59% retained their baseline communication group. Despite this stability, relapse was accompanied by reductions in interactions involving cytotoxic T cells, memory B cells, erythroid and myeloid progenitors, and tumor cells. TME rewiring was greater among patients whose tumor RNA subtype changed, demonstrating both stable patient-specific organization and dynamic remodeling during disease evolution.
Conclusions:
Integrating tumor genomics with the downstream consequences of ligand-receptor signaling identified six biologically distinct communication groups that captured TME biology not reflected by established MM classifications and were associated with PFS and OS. These groups were predominantly stable within patients, yet relapse was accompanied by coordinated loss of immune- and progenitor-cell communication, revealing dynamic TME remodeling during disease evolution. Together, these findings establish intercellular communication as a clinically relevant dimension of MM biology and support its development as a prognostic and longitudinal biomarker, while identifying network hubs and interactions for therapeutic investigation.
DISCLOSURES
David G Coffey, MD: No relevant conflicts of interest to declare.