Dynamic Prediction of Survival Outcomes in Multiple Myeloma

Background: Multiple myeloma (MM) remains an incurable plasma cell malignancy characterized by marked clinical heterogeneity. Existing prognostic frameworks, including the International Staging System (ISS) and FISH-defined cytogenetic risk, are anchored at diagnosis and do not capture the evolutionary dynamics of disease or treatment response, leaving an unmet need for risk models that retain prognostic validity longitudinally. Methods: Using transcriptomic data from 762 CD138-selected MM plasma cells from newly diagnosed patient samples in the MMRF CoMMpass study (NCT01454297), we computed single-sample pathway activity scores for 469 curated cancer-relevant pathways (MSigDB Hallmark; Reactome) and learned a Bayesian causal network linking pathway activity to survival. The model was validated in five independent diagnostic cohorts (n = 1255) and in two independent treatment and relapsed/refractory cohorts (n = 319). Longitudinal risk tracking was additionally assessed in a 46-patient subset of the discovery cohort with serial pre- and post-treatment sampling. Results: The network identified five pathways associated with survival: unfolded protein response (UPR), FLT3 signaling through SRC family kinases, G2M DNA replication checkpoint, metabolism of selenium compound (SeMet), and nicotinate metabolism. The composite survival score stratified patients into high-risk (n = 76; 10%) and standard-risk groups with markedly divergent survival (median 1170 days vs. not reached; p < 0.0001). The score remained an independent prognostic factor after adjustment for age, sex, ISS stage, and KRAS, TP53, and UBR5 mutational status (HR 4.93; 95% CI 2.96–8.19; p < 0.001), and replicated across all five external diagnostic cohorts. Critically, the model retained prognostic discrimination in previously treated (GSE57317; p < 0.0001) and relapsed/refractory (GSE9782; p < 0.0001) settings, and patients transitioning from standard- to high-risk between serial samples exhibited significantly inferior survival compared to standard-risk patients. Conclusions: This pathway-based Bayesian network provides a reproducible, dynamically applicable risk model for MM that captures information complementary to ISS and FISH-defined cytogenetics. The framework supports longitudinal patient monitoring and may inform trial enrichment strategies and closer surveillance for high-risk subpopulations.