Extended Data Fig. 8: Additional Model Diagnostics and Predictive Modeling of Overall Survival and Progression in MM Using Immune Signatures, Cytogenetics, and Clinical Features.
Panels (a-c) show model AUCs in the discovery cohort across immune compartments: (a) AUCs by immune cell compartments and variable combinations, (b) Boxplots showing the distribution of AUCs for models integrating a single cell type with various covariates, along with the AUC for the top models integrating all clinical covariates and either 7, 11, 34, or all immune populations, and (c) Boxplots showing the AUCs for models derived from individual immune populations, grouped by cellular compartment. In the box plots, bounds of the box represent the 25th and 75th percentile, with the center displaying the median. Whiskers extend to 1.5*IQR beyond the bounds of the box. Whiskers extend to 1.5*IQR beyond the bounds of the box. (d-i) Receiver operating characteristic (ROC) and Kaplan-Meier (KM) analysis for overall survival (OS) prediction in the discovery cohort. (d) ROC curves for models with single immune subclusters (SubC), clinical covariates (CoV), cytogenetics, and combinations. Covariates include age, batch, site, ISS stage and cytogenetics. KM curves depict predicted OS based on (e) clinical covariates (f) cytogenetics + clinical covariates, (g) Immune Atlas Signature + clinical covariates, and (h) the top 20 predictive immune subclusters + clinical covariates. (i) The importance of immune subclusters for predicting the OS colored by favorable (blue) or poor (red) OS association. (j-o) ASCT’s contribution to survival prediction in discovery and validation cohorts. KM curves showing ASCT association with OS in discovery (j) and validation (k) cohorts. Predictive models including ASCT, cytogenetics, clinical variables, and top immune features in discovery (l) and validation (m). Equivalent models excluding ASCT as a variable, still stratifying high- versus low-risk patients in discovery (n) and validation (o) cohorts. (p-t) OS prediction in the validation cohort. (p) A forest plot based on a multivariate CoxPH illustrating the bias of ASCT and ISS for OS. Two-sided p-values from the CoxPH model are displayed. (q) Box plot of bootstrapped AUCs for models using various immune compartments (npatient=71), with the integrative model as a superior option. However, the AUC for OS remained below 0.75 in general. (r) Box plot of bootstrapping applied to an integrative model of feature selection based on AUC to identify the optimal model for OS prediction using immune populations, clinical information, cytogenetics, and ASCT. In the box plots, bounds of the box represent the 25th and 75th percentile, with the center displaying the median. Whiskers extend to 1.5*IQR beyond the bounds of the box. (s) KM curves show the effect of ASCT on the prediction of OS. (t) Integrative model using the top 20 immune signatures combined with clinical and cytogenetic information in the prediction of OS.
