Машинное обучение - Deriveed Immune Gene Signing Previsions Прогнозы прогноза и терапическая вульгарность в Множестве миеломы.

Abbreviations

confidence intervals

Gene Expression Omnibus

Gene Ontology

immune‐related gene signature

multiple myeloma

overall survival

single‐cell RNA sequencing

single‐sample Gene Set Enrichment Analysis

weighted gene co‐expression network analysis

Introduction

Multiple myeloma (MM), characterized by the clonal proliferation of malignant plasma cells in the bone marrow, represents the second most common hematologic malignancy worldwide [ 1 , 2 ]. Therapeutic strategies for MM have advanced substantially with the introduction of proteasome inhibitors, immunomodulatory agents, monoclonal antibodies, and chimeric antigen receptor T‐cell (CAR‐T) therapies, leading to improved outcomes for many patients [ 3 ]. However, MM remains incurable due to its biological heterogeneity, frequent relapse, and the emergence of treatment resistance [ 4 ]. Clinical risk stratification is commonly supported by the International Staging System (ISS) [ 5 ] and the revised ISS (R‐ISS) [ 6 ]. Even so, these systems have a limited ability to support individualized treatment decisions. This underscores the need for robust prognostic signatures that can refine risk assessment and support personalized treatment strategies [ 7 ].

Dysregulation of the bone marrow immune microenvironment is a central feature of MM pathogenesis [ 8 ]. Myeloma cells actively remodel immune surveillance through impaired T‐cell activation, expansion of immunosuppressive cell populations, altered cytokine signaling, and disruption of antigen processing and presentation [ 9 ]. These immune alterations not only facilitate tumor immune evasion but are also closely associated with therapeutic resistance, disease progression, and adverse clinical outcomes [ 10 ]. With the rapid expansion of immunotherapeutic strategies, including anti‐CD38 monoclonal antibodies, BCMA‐targeted agents, and CAR‐T cell therapies, the critical role of immune regulation in MM has become increasingly apparent [ 11 , 12 ]. In this context, prognostic models derived from immune‐related gene sets may capture biologically meaningful variation in immune activity, such as inflammation, immune checkpoint regulation, cytotoxic activity, and antigen presentation, thereby providing a comprehensive molecular representation of the immune landscape in MM [ 13 ]. In contrast to models derived from nonspecific or purely statistical gene selection, immune‐based models offer enhanced biological interpretability by directly reflecting tumor‐immune interactions [ 14 ]. This biological relevance enhances their robustness, reproducibility, and clinical translatability [ 15 ]. Moreover, dynamic alterations in immune gene expression often reflect critical pathological processes, rendering immune‐informed signatures particularly informative for predicting disease progression and therapeutic benefit [ 16 ].

In this study, we integrated transcriptomic data from the Gene Expression Omnibus (GEO) and the MMRF‐CoMMpass cohort with curated immune‐related gene sets to develop an immune‐based prognostic model for MM. We aimed to capture immune‐associated transcriptional heterogeneity and determine whether this information could provide reproducible prognostic stratification across multiple patient cohorts. To further define the biological context of the signature, we examined its relationships with somatic mutational characteristics and transcriptome‐derived drug‐sensitivity profiles. In parallel, single‐cell RNA sequencing data spanning healthy donors, MGUS, SMM, and MM were used to map the cellular distribution of the gene module comprising the prognostic model across the bone marrow ecosystem and to evaluate its association with malignant plasma cell status and differentiation potential. Together, these analyses establish a framework linking immune‐related transcriptional variation with prognosis, cellular context, somatic mutational features, and candidate therapeutic vulnerabilities in MM.

Material and Methods

Data Acquisition and Preprocessing

As described in Table 1 , gene expression profiles from six independent MM cohorts, comprising a total of 2382 individuals, were systematically collected and analyzed. These datasets included one bulk RNA‐seq dataset, four microarray datasets, and one single‐cell RNA‐seq (scRNA‐seq) dataset. The MMRF‐CoMMpass cohort, comprising RNA‐seq data, matched clinical information and somatic mutation data from 764 patients with MM ( https://portal.gdc.cancer.gov/projects/MMRF‐COMMPASS ) [ 17 ], was used as the primary training cohort for signature development. Four independent microarray datasets, including GSE2658 , GSE24080 , GSE136337 , and GSE57317 , were retrieved from the Gene Expression Omnibus datasets (GEO, https://www.ncbi.nlm.nih.gov/geo ) [ 18 ]. Among these, GSE2658 and GSE24080 were utilized as internal validation cohorts, whereas GSE136337 and GSE57317 served as independent external validation cohorts to comprehensively evaluate the robustness and generalizability of the proposed gene signature. Only cohorts with available overall survival (OS) information were included in the construction and validation of the prognostic model. For each dataset, gene expression profiles were annotated using platform‐specific annotation files. To ensure comparability with the microarray‐based data, the R package TDM (v0.3) was employed to transform RNA‐seq counts into microarray‐like distribution [ 19 ]. The scRNA‐seq dataset GSE271107 , comprising samples from six patients with monoclonal gammopathy of undetermined significance (MGUS), four patients with smoldering MM (SMM), four patients with newly diagnosed MM, and five healthy donors (HD), was further incorporated to characterize transcriptional alterations associated with MM disease progression at single‐cell resolution.

Cells Infiltration Estimation

The relative infiltration levels of immune cell populations in the MMRF cohort were quantified using single‐sample Gene Set Enrichment Analysis (ssGSEA), implemented via the GSVA R package (v1.50.5) [ 20 ]. This approach calculated enrichment scores for 28 distinct immune cell types based on curated marker gene signatures [ 21 ]. To strengthen the robustness and reliability of the ssGSEA‐derived estimates, we performed cross‐validation using four complementary algorithms: CIBERSORT [ 22 ], xCell [ 23 ], EPIC [ 24 ], and ESTIMATE [ 25 ]. These algorithms employ distinct methodological principles, spanning reference‐based deconvolution, cell‐type enrichment scoring, and stromal/immune content inference, thereby reducing algorithm‐specific bias and ensuring a reliable characterization of the tumor immune microenvironment.

Consensus Clustering

Based on the infiltration profiles of immune cell populations, a resampling‐based consensus clustering method was applied to identify robust immune subtypes using the ConsensusClusterPlus package (v1.70.0) [ 26 ]. This procedure ensures the stability of classification through repeated subsampling of both items and features. Subsequently, the consensus matrix heatmap, cumulative distribution function (CDF) curves, and proportion of ambiguous clustering (PAC) scores were synthetically used to determine the optimal number of clusters (k). To further validate the optimal k, independent recommendations from the NbClust package (v3.0.1) [ 27 ] were integrated. Finally, principal component analysis (PCA) was performed to visualize the spatial distribution of samples, confirming the distinct separability of the identified clusters.

Weighted Gene Co‐Expression Network Analysis (WGCNA)

WGCNA was performed on the gene expression data from the MMRF cohort using the WGCNA package (v1.73) [ 28 ] to construct a scale‐free co‐expression network. An optimal soft threshold (β) was selected to satisfy the scale‐free topological criterion, which was then used to construct the weighted adjacency matrix. Subsequently, this matrix was transformed into the topological overlap matrix (TOM), and the corresponding dissimilarity matrix (1‐TOM) was generated. Gene modules were identified using the dynamic tree cutting algorithm. To identify modules significantly associated with the immune clusters, we calculated the Pearson's correlation coefficient between the module eigengene of each module and the immune cluster variable. The module displaying the strongest significant positive correlation was selected for further analysis. Within this key module, genes were defined as immune‐related characteristic genes if they exhibited both high module membership and significant gene significance with respect to the immune cluster trait.

Signature Generated From Machine Learning‐Based Integrative Approaches

To develop a robust and high‐precision consensus immune‐related gene signature (IRGS) for MM, we integrated 10 machine learning algorithms and 101 algorithm combinations [ 29 ]. The integrative algorithms included random survival forests (RSF), elastic net (Enet), lasso regression, ridge regression, stepwise Cox regression, CoxBoost, Cox partial least squares regression (plsRcox), supervised principal components (SuperPC), generalized boosted regression modeling (GBM), and survival support vector machines (survival‐SVM). The signature generation workflow proceeded as follows: (a) The prognostic immune‐related genes were first identified using univariate Cox regression in the MMRF cohort. (b) These candidate genes were then used to construct predictive models under a leave‐one‐out cross‐validation (LOOCV) framework across all 101 algorithmic combinations. (c) Each model was subsequently evaluated in other model‐selection datasets (e.g., GSE2658 and GSE24080 ). (d) The Harrell's concordance index (C‐index) was computed for every model across all cohorts. The model achieving the highest average C‐index was considered optimal. For survival stratification, the optimal cut‐off for the continuous IRGS score was determined separately in each cohort using the “surv_cutpoint” function in the “survminer” R package, and patients were subsequently classified into high‐ and low‐IRGS groups according to the corresponding threshold. In addition, the trained model objects, complete executable scripts, and detailed prediction workflow are publicly available at ( https://github.com/wangkai130818/IRGS ).

Functional Enrichment Analysis

The Gene Ontology (GO) enrichment analyses were conducted using the R package clusterProfiler (v4.6.0) [ 30 ]. Significantly enriched pathways with adjusted p value < 0.05 were further visualized with the R package ggplot2 (v4.0.0).

Somatic Mutational Profile Analysis

Fisher's exact test was applied to identify genes that had significantly different mutation frequencies between the two MM risk groups. The top 30 most frequently mutated nonsynonymous genes were subsequently visualized using an OncoPrint [ 31 ] to summarize the mutational landscape. Tumor mutation burden (TMB) was calculated from the single nucleotide variation (SNV) data of the MMRF cohort and defined as the total number of non‐synonymous mutations per megabase (mutations/Mb). The mutation categories included missense, nonsense, frameshift, in‐frame and splice‐site variants, as well as intron, 5′UTR, 3′UTR and multi‐hit variants.

Single‐Cell RNA Sequencing Analysis

The single‐cell RNA‐sequencing dataset GSE271107 was analyzed on a per‐sample basis using Seurat (v5.4.0) [ 32 ]. Cells were retained according to the following quality‐control criteria: 200–7500 detected genes, 200–60,000 unique molecular identifiers (UMIs), and no more than 25% mitochondrial transcripts. Potential doublets were identified and removed using DoubletFinder (v2.0.6) [ 33 ]. After quality control, all samples were merged into a unified Seurat object for downstream analysis. Gene expression matrices were processed using the SCTransform workflow, which performs normalization, variance stabilization, and highly variable feature selection within a unified framework [ 34 ]. Principal component analysis (PCA) was then performed on the SCT‐normalized expression matrix, followed by batch‐effect correction using Harmony (v1.2.4) [ 35 ]. Graph‐based clustering was conducted on the Harmony‐corrected embeddings by constructing a shared nearest‐neighbor graph and applying Louvain community detection. Cells were visualized in two‐dimensional space using t‐distributed stochastic neighbor embedding (t‐SNE) [ 36 ]. Cluster‐specific marker genes were identified using the Seurat FindAllMarkers function with the Wilcoxon rank‐sum test. Cell identities were assigned based on canonical lineage markers and cluster‐enriched differentially expressed genes. To distinguish malignant from normal plasma cells, large‐scale copy‐number variation profiles were inferred using inferCNV (v1.23.0), with immune cells used as the reference population. Plasma cells with evident chromosomal copy‐number alterations were defined as malignant plasma cells. In addition, CytoTRACE (v0.3.3) [ 37 ] was applied to estimate cellular stemness scores.

To characterize the cellular distribution of the genes comprising the bulk‐derived IRGS, the 12 component genes were evaluated as an unweighted expression module at single‐cell resolution. Module activity was primarily quantified using UCell (v2.10.1) based on the log‐normalized, non‐imputed RNA expression matrix, and the resulting measure was termed the 12‐gene IRGS module activity. As sensitivity analyses, module activity was independently recalculated using AUCell (v1.28.0) and Seurat AddModuleScore (Seurat v5.4.0). Concordance among the three scoring approaches was evaluated using Pearson's correlation coefficient. To account for the non‐independence of multiple cells derived from the same individual, differences across disease stages, major cell populations, and malignant versus non‐malignant plasma cells were evaluated using linear mixed‐effects models with patient identity included as a random intercept.

Drug‐Sensitivity Prediction

Drug‐sensitivity prediction was performed using the oncoPredict package (v1.2) to explore potential differences in transcriptome‐derived drug‐response profiles between the IRGS‐defined groups in the MMRF cohort. This method estimates drug sensitivity by mapping patient transcriptomic profiles to pharmacogenomic prediction models trained on cancer cell‐line expression data and experimentally measured half‐maximal inhibitory concentration (IC50) values from the Genomics of Drug Sensitivity in Cancer (GDSC) database. Predicted IC50 values were generated for each patient, with lower values indicating greater computationally inferred sensitivity to a given compound.

Statistical Analysis

Overall survival (OS) was defined as the time interval from study entry to death from any cause or the date of last follow‐up (censored). Survival analyses were performed using the survival (v3.8‐3) and survminer (v0.5.1) packages. The Kaplan–Meier (K–M) method was utilized to estimate survival probabilities, and differences between groups were evaluated with the log‐rank test [ 38 ]. The Cox proportional hazards regression model [ 39 ] was employed to calculate hazard ratios (HRs) and 95% confidence intervals (CIs), with adjustment for relevant clinical covariates (e.g., age, gender, ISS). Optimal cut‐off points for risk stratification were determined using the maximally selected rank statistics implemented in the survminer package. All statistical tests were two‐sided, and p value < 0.05 was considered statistically significant. All data processing, statistical analysis, and visualization were conducted in R software (v4.4.3, https://www.r‐project.org/ ).

Results

Identification and Characterization of Immune Infiltration‐Based Consensus Clusters

We first quantified the infiltration levels of 28 immune cell types across the MMRF cohort using the ssGSEA algorithm. To delineate distinct immune infiltration patterns, consensus clustering was performed with the number of clusters ( k ) ranging from 2 to 6. Evaluation of the cumulative distribution function (CDF) curves of the consensus matrices, together with the proportion of ambiguous clustering (PAC) statistic, indicated that k = 2 yielded the most stable clustering solution (Figures 1A,B and S1A ). The NbClust testing further validated k = 2 as the optimal cluster number (Figure S1B ), and Principal Component Analysis (PCA) demonstrated clear segregation between the two clusters (Figure S1C ).

We next characterized the immune landscapes of the two subgroups. As illustrated by the boxplots and heatmaps (Figure 1C,D ), Cluster 1 (C1) exhibited markedly higher infiltration levels for the majority of immune cell types, whereas Cluster 2 (C2) showed generally lower immune activity. Consistently, the overall ImmuneScore was significantly elevated in C1 compared to C2 ( p < 0.001, Figure 1E ). Thus, we defined C1 as the “immune‐hot” subtype (characterized by high immune infiltration) and C2 as the “immune‐cold” subtype. To ensure that the two consensus clusters were not biased by the analytical algorithm, other immune algorithms, including CIBERSORT, EPIC, and xCell, were used to verify the stability and robustness of the ssGSEA results (Figure S1D ).

Identification of Functional Gene Modules Derived From Immune Infiltration Patterns

The weighted gene co‐expression network analysis (WGCNA) was employed to identify gene modules associated with the established immune infiltration patterns. To ensure the robustness of the co‐expression network, hierarchical clustering analysis was initially performed, leading to the exclusion of two outlier samples (Figure S2A ). The soft threshold β was subsequently determined to be 5, as this value optimized scale independence and mean connectivity, providing a suitable statistical power for network construction (Figure S2B ). Using this parameter, WGCNA successfully identified five distinct gene modules (MEs), represented by different colors in the cluster dendrogram (Figure S2C ). The eigengene, which represents the first principal component of gene expression within a module, was considered as the representative of the module. The adjacency heatmap of the module eigengenes confirmed the independence of the identified modules (Figure 2A,B ).

To establish the biological relevance of these modules, we assessed their correlations with clinical and phenotypic traits, including immune clusters, age, gender, and ISS stage (Figure 2C ). The highest correlation in the module‐trait relationship was observed between the MEblue module and immune clusters ( R = −0.71, p < 0.001). In the MEblue module, the correlation coefficient between gene significance (GS) and module membership (MM) reached 0.94 (Figure 2D , p < 0.001), which suggested that the quality of gene module construction was superior. The 164 genes comprising the MEblue module were thus designated as hub immune‐related genes for further investigation.

To further elucidate the potential function of the immune‐related MEblue module, we performed GO enrichment analysis (Figure S2D ). The results indicated that the genes within the MEblue module were significantly enriched in key immune‐related pathways, such as “leukocyte migration”, “leukocyte chemotaxis”, “regulation of inflammatory” and “positive regulation of response to external stimulus”.

Construction and Definition of the Immune‐Related Gene Signature

Univariate Cox analysis was initially applied to screen 164 immune‐related genes, narrowing the pool to 47 prognostic candidates (Figure S3 ). These genes were subsequently incorporated into a comprehensive machine learning integrative framework to construct a robust consensus immune‐related gene signature (IRGS). Within the MMRF training cohort, we systematically evaluated 101 predictive model combinations using a leave‐one‐out cross‐validation (LOOCV) framework. By benchmarking the concordance index (C‐index) across multiple independent validation datasets, we compared the predictive performance of these algorithms (Figure 3A , top 50). Notably, the integration of Random Survival Forests (RSF) and survival‐SVM emerged as the optimal model, achieving the highest average C‐index (0.605) across all validation cohorts. This ensemble model ultimately identified a final panel of 12 genes, including SLA, SIGLEC5, ADAM8, MMD, CA2, FN1, UPP1, CLEC7A, EMP1, SELE, TLR8, and BTNL8.

Subsequently, an IRGS‐based score was calculated for each patient based on the expression profiles of 12 genes. All patients were stratified into high‐ and low‐score groups according to the optimal cut‐off value determined by the survminer package. As illustrated in Figure 3B–D , Kaplan–Meier survival analysis demonstrated that patients in the high‐score group exhibited significantly prolonged overall survival (OS) compared to the low‐score group in both the MMRF training dataset and two internal model‐selection datasets (all p < 0.05). Meta‐analysis of univariate Cox regression further characterized the IRGS score as a protective factor (Figure 3E ). Moreover, multivariate Cox regression demonstrated that IRGS remained statistically significant (HR = 0.76, 95% CI: 0.66–0.87, p < 0.001) after adjustment for available clinical covariates, such as age, gender, and ISS stage, suggesting that the IRGS serves as a robust and independent prognostic factor for OS (Figure 3F ).

To further characterize the relationship between the IRGS and the immune microenvironment, we examined its association with the immune infiltration in the MMRF cohort. The high‐IRGS group exhibited lower Immune Scores than the low‐IRGS group (Figure S4A , p = 7.43 × 10 −23 ). Pearson's correlation analysis revealed a moderate inverse correlation between the continuous IRGS score and the Immune Score (Figure S4B , R = −0.52, p = 3.61 × 10 −54 ). Consistent with this finding, the IRGS score was negatively correlated with most of the 28 immune‐cell infiltration scores (Figure 3G ).

Validation and Evaluation of the IRGS Model

To substantiate the robust prognostic value of the IRGS, we applied IRGS to other independent validation datasets. In the GSE136337 cohort with 426 MM patients, individuals in the high‐score group exhibited significantly better OS compared to those in the low‐score group (Figure 4A , log‐rank p < 0.001). Multivariate analysis further confirmed that the IRGS remained an independent prognostic factor even after adjusting for available clinical and pathologic covariates (Figure 4B , HR = 0.82, 95% CI: 0.71–0.95, p < 0.001).

We further expanded the evaluation to alternative survival endpoints. The IRGS successfully discriminated outcomes regarding progression‐free survival (PFS) in the GSE136337 cohort and event‐free survival (EFS) in the GSE24080 cohort (Figure 4C,D ). Notably, the prognostic performance of the IRGS was further evaluated in a cohort of patients with previously treated MM ( GSE57317 ). The IRGS consistently stratified patients into subgroups with significantly different overall survival outcomes (Figure 4E ).

Comparison With Established Prognostic Signatures

To further evaluate the prognostic performance of the IRGS, we compared it with previously reported signatures, including IFM15 [ 40 ], MRCIX6 [ 41 ], and a mitophagy‐related signature [ 42 ], across the MMRF, GSE2658 , GSE24080 , and GSE136337 cohorts. The IRGS showed prognostic discrimination comparable to that of IFM15 and generally higher C‐index values than MRCIX6 and the mitophagy‐related signature across the independent validation cohorts (Figure 4C and Table S1 ). Kaplan–Meier analyses further demonstrated that the IRGS consistently stratified patients into groups with significantly different overall survival in all four cohorts. By contrast, the prognostic stratification achieved by the comparator signatures was less consistent (Figure S5 ). Notably, IFM15 and MRCIX6 did not significantly distinguish survival outcomes in GSE136337 . These results support the consistent cross‐dataset prognostic performance of the IRGS.

Somatic Mutational Characteristics of the IRGS‐Defined Prognostic Groups

Among the 764 patients in the MMRF cohort with RNA‐seq profiles, 725 had matched somatic mutation profiles, of whom 508 and 217 were classified into the high‐ and low‐IRGS groups, respectively. Gene‐level comparison of mutation frequencies identified 323 genes with nominally different mutation frequencies between the two groups at p < 0.05, of which 303 were more frequently mutated in the low‐IRGS group (Table S2 ).

Separate oncoplots were generated to display the 30 most frequently mutated genes within each IRGS group (Figure 5A , top 30 genes). MUC16 was the most frequently mutated gene in the high‐IRGS group, whereas DIS3 showed the highest mutation frequency in the low‐IRGS group. The low‐IRGS group also exhibited a modestly higher tumor mutation burden (TMB) than the high‐IRGS group (Figure 5B , p = 0.033). Consistent with this group‐level difference, the IRGS score was weakly inversely correlated with TMB (Figure 5C , R = −0.08, p = 0.03).

Single‐Cell Characterization of the Cellular Distribution of the IRGS Gene Module

To characterize the cellular distribution of the genes comprising the bulk‐derived IRGS, we analyzed scRNA‐seq data from HD ( n = 5) and patients with MGUS ( n = 6), SMM ( n = 4), and MM ( n = 4). After stringent quality control and doublet removal, 119,099 cells were retained. Unsupervised clustering identified eight major bone marrow cell populations, including plasma cells (PlasmaCell, n = 9090), B cells (Bcell, n = 9269), T and natural killer cells (TNK, n = 43,165), mononuclear phagocytes (MPs, n = 23,714), hematopoietic stem cells (HSC, n = 1369), granulocyte‐monocyte progenitors (GMP, n = 2064), erythroid cells (ErythroidCell, n = 28,363), and megakaryocytes ( n = 2065) (Figure 6A,B ).

Coordinated expression of the 12 IRGS component genes was quantified using UCell and is hereafter referred to as the 12‐gene IRGS module activity. UCell‐derived module activity was highly correlated with values independently calculated using AUCell and AddModuleScore, and the three approaches yielded comparable cell‐type distributions and similar descriptive cross‐stage patterns (Figure S6 ). Although the 12‐gene IRGS module activity showed a downward pattern across disease stages (Figure 6C ), the estimated linear trend remained negative but did not reach statistical significance after accounting for the clustering of cells within individuals using a linear mixed‐effects model ( β = −0.0057, p = 0.128). This cross‐stage pattern was therefore considered descriptive. Module activity showed a more pronounced cell type dependent distribution (Figure 6D ). Plasma cells exhibited the lowest activity, whereas MPs and erythroid cells showed higher activity after accounting for patient‐level clustering, indicating that expression of the IRGS component genes is not confined to the plasma cell compartment. Plasma cells were further classified according to inferCNV‐derived copy‐number profiles (Figure 6E ). Cells classified as malignant showed lower module activity than those classified as non‐malignant after accounting for patient‐level clustering ( β = −2.68 × 10 −3 , p = 2.11 × 10 −10 , Figure 6F ).

Finally, the 12‐gene module activity showed a statistically significant but weak inverse correlation with cellular stemness scores (R = −0.13, p < 0.001, Figure 6G ). Given the modest effect size, this association was considered exploratory.

Predicted Drug‐Sensitivity Patterns Across the IRGS‐Defined Groups

Transcriptome‐based drug‐sensitivity prediction was performed to explore differences in predicted IC50 profiles between the high‐ and low‐IRGS groups (Figure 7A ). The two groups exhibited distinct predicted drug‐response patterns. The high‐IRGS group showed lower predicted IC50 values for 39 compounds, including ABT737, Nutlin‐3a, and oxaliplatin. In contrast, the low‐IRGS group showed lower predicted IC50 values for 42 compounds, including cyclophosphamide, JQ1, AZD7762, MK‐8776, and AZD6738.

Correlation analyses were consistent with these group‐level differences. The IRGS score was weakly inversely correlated with the predicted IC50 value for ABT737 (Figure 7B , R = −0.21, p < 0.001) and showed modest positive correlations with the predicted IC50 values for cyclophosphamide (Figure 7C , R = 0.26, p < 0.001) and JQ1 (Figure 7D , R = 0.29, p < 0.001). These findings identify IRGS‐associated differences in transcriptome‐derived drug‐response predictions and may help prioritize candidate agents for further experimental evaluation in different prognostic subgroups.

Discussion

Multiple myeloma is a biologically heterogeneous plasma cell malignancy shaped by both tumor‐intrinsic alterations and the bone marrow microenvironment. In this study, we developed a 12‐gene immune‐related gene signature using an integrative machine‐learning framework. The IRGS consistently stratified overall survival across multiple cohorts and remained independently associated with outcome after adjustment for ISS stage, age, and gender. Direct comparison with established transcriptomic signatures showed that its prognostic discrimination was comparable to that of IFM15 and generally higher than that of MRCIX6 and the mitophagy‐related signature across the evaluated validation datasets. Nevertheless, the moderate C‐index indicates that the IRGS captures only part of the substantial prognostic heterogeneity of MM. Accordingly, the IRGS should be regarded as a complementary molecular biomarker rather than a standalone prognostic tool or a replacement for established clinical staging systems.

The inverse association between the IRGS score and the global Immune Score, together with its negative correlations with most immune‐cell infiltration estimates, supports the immune‐related basis of the signature. Single‐cell analysis further revealed substantial variation in module activity across bone marrow cell populations, with comparatively low activity in plasma cells and higher activity in several non‐plasma compartments. These findings argue against an exclusively malignant plasma cell intrinsic origin and suggest that the bulk IRGS signal reflects a multicellular transcriptional context shaped by both cellular composition and cell‐type‐specific expression. Importantly, the single‐cell module activity was used to characterize the cellular distribution of the component genes and should not be interpreted as a cell‐level counterpart of the nonlinear bulk prognostic model. Moreover, the cross‐stage pattern did not remain statistically significant after accounting for patient‐level clustering, and the weak association with CytoTRACE‐derived differentiation potential does not support a substantial relationship with stem‐like cellular features. Accordingly, the single‐cell findings should be interpreted primarily as evidence of cellular localization rather than mechanistic validation of the bulk IRGS.

The IRGS‐defined groups also differed in their somatic mutational profiles. The relative enrichment of DIS3 mutations in the low‐IRGS group is biologically noteworthy, as DIS3 encodes a catalytic ribonuclease of the RNA exosome and is recurrently altered in MM. Beyond disrupting RNA processing and transcriptome homeostasis, recent mechanistic evidence indicates that the clinically relevant DIS3 G766R variant enhances aberrant activation‐induced cytidine deaminase activity during B‐cell activation, thereby promoting DNA double‐strand breaks and IGH‐associated chromosomal translocations that promote plasma cell transformation [ 43 , 44 ]. These findings provide a plausible biological context linking DIS3 alterations to adverse myeloma features. In contrast, MUC16 mutations were more frequent in the high‐IRGS group. Although MUC16 is best known as the gene encoding CA125 in ovarian cancer, mutations in this gene have been associated with increased neoantigen load and improved immune recognition in several tumor types [ 45 ]. In MM, MUC16 mutations may similarly mark a more immunogenic tumor context, potentially contributing to the immune‐inflamed features observed in the high‐IRGS group.

The predicted drug‐response analysis suggested that the IRGS‐defined groups may differ in pathway‐level therapeutic dependencies. The high‐IRGS group showed lower predicted IC50 values for several agents associated with apoptosis induction and DNA‐damage responses, whereas the low‐IRGS group showed predicted sensitivity to compounds targeting BET‐mediated transcription, checkpoint signaling, and replication‐stress pathways. These patterns may help prioritize agents and generate subgroup‐specific hypotheses for further experimental investigation. However, the correlations between the IRGS score and predicted IC50 values were weak to modest, and the estimates were derived from pharmacogenomic models trained on cancer cell lines. Such predictions do not account for patient‐specific pharmacokinetics, prior treatment exposure, combination regimens, the bone marrow microenvironment, or in vivo mechanisms of drug resistance. Accordingly, these findings should be regarded as hypothesis‐generating and require validation in MM cell lines, primary patient samples, ex vivo drug‐response assays, and clinically annotated cohorts before they can inform individualized treatment selection.

Several limitations should be acknowledged. First, the IRGS was developed from bulk transcriptomic profiles generated from enriched bone marrow samples. Variation in sample purity and the presence of non‐plasma populations may therefore influence the measured expression of immune‐related genes. Second, the single‐cell dataset was cross‐sectional and included relatively few independent individuals, particularly in the SMM and MM groups, limiting the precision of cross‐stage and malignant‐versus‐non‐malignant comparisons. Third, several statistically significant associations had small effect sizes and should be regarded as exploratory rather than evidence of strong biological relationships. Finally, the study was retrospective and predominantly computational, and lacked direct functional validation and prospective clinical assessment.

In conclusion, the IRGS captures an immune‐associated prognostic dimension of MM. Its associations with distinct somatic mutational patterns and predicted drug‐response profiles further provide biologically and therapeutically relevant hypotheses for patient stratification. These findings establish the IRGS as a potentially informative framework for integrating immune context with prognosis in MM and provide a foundation for future clinical‐molecular modeling and functional validation. A graphical overview of the present study is presented in Figure 8 .

Author Contributions

Kai Wang: methodology, software, data curation, conceptualization, investigation, validation, formal analysis, resources, writing – review and editing, writing – original draft, funding acquisition, visualization, project administration, supervision. Chenfei Zhao: conceptualization, methodology, data curation, resources, writing – original draft, investigation, validation, writing – review and editing. Jingru Shi: data curation, resources, investigation, writing – review and editing. Shiwei Liu: investigation, data curation, resources, writing – review and editing. Yi Chen: investigation, data curation, resources, writing – review and editing. Jujuan Wang: investigation, data curation, resources, writing – review and editing. Zhengxu Sun: investigation, data curation, resources, writing – review and editing. Sanmei Wang: investigation, data curation, resources, writing – review and editing. Lei Fan: writing – review and editing, funding acquisition, supervision, resources, project administration. Jin Fan: supervision, writing – review and editing, resources, project administration. Xiaoyan Qu: funding acquisition, supervision, writing – review and editing, resources, project administration.

Funding

This study was supported by the National Natural Science of China (82470186 and 82570256), the Natural Science Foundation of Jiangsu Province (BK20232039), and the postdoctoral research startup funds (2025BSH014). Nanjing Clinical and Basic Research Project (202511074), Beijing Xisike Clinical Oncology Research Foundation (Y‐2024AZ(BTK)MS‐0047 and Y‐SYBLD2022RWR‐0013). The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Disclosure

None of the authors of this manuscript are a current Editor or Editorial Board Member of Cancer Science.

Ethics Statement

The authors have nothing to report.

Consent

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Validation of clustering stability. (A) The proportion of ambiguous clustering (PAC) scores for cluster numbers k = 2 to 6 were displayed, with the lowest value at k = 2 indicating optimal clustering stability. (B) The bar chart displayed the frequency of recommended cluster numbers according to the NbClust package. (C) Principal component analysis (PCA) plot visualized the spatial distribution of samples. (D) Three other algorithms including CIBERSORT, EPIC, and xCell, further verified the stability and robustness of the ssGSEA results.

Figure S2: WGCNA network construction and functional characterization of the hub module. (A) Hierarchical clustering analysis was performed to evaluate sample consistency and identify potential outliers within the cohort. (B) Selection of best soft threshold power. (C) The clustering tree of 5 differentially colored modules. (D) Top 20 Gene Ontology (GO) enrichment categories for biological process in the blue module genes.

Figure S3: Univariate Cox analysis identified 47 prognostic genes in the MMRF cohort. Data were presented as hazard ratio (HR) ± 95% confidence interval (CI).

Figure S4: Association between the IRGS and immune infiltration in the MMRF cohort. (A) Distribution of Immune Scores between the low‐ and high‐IRGS groups. Each point represents one patient. Violin plots show the distribution of Immune Scores, and red points indicate the group means. (B) Pearson correlation between the IRGS score and the Immune Score. Each point represents one patient; the solid line indicates the fitted linear regression, and the shaded band represents the 95% confidence interval.

Figure S5: Prognostic stratification by previously published gene‐expression signatures across MM cohorts. Kaplan–Meier curves compare overall survival between high‐ and low‐risk groups defined by IFM15, MRCIX6, and the mitophagy‐related signature. Each row represents one prognostic signature, and each column represents one dataset: MMRF, GSE2658 , GSE24080 , and GSE136337 . Survival differences were assessed using the log‐rank test.

Figure S6: Sensitivity analysis of the single‐cell IRGS module activity across alternative scoring methods. Correlation of UCell‐derived module activity with AUCell‐derived scores (A) and AddModuleScore‐derived score (D). Distribution of AUCell‐derived (B) and AddModuleScore‐derived (E) module activity across HD, MGUS, SMM, and MM samples. (C, F) Distribution of the corresponding module activity across major bone marrow cell populations. Boxes indicate the interquartile range, horizontal lines indicate the median, and whiskers extend to 1.5 times the interquartile range. Statistical significance is indicated as ** p < 0.01, *** p < 0.001, and **** p < 0.0001.

Table S1: Comparison of the prognostic discrimination of the IRGS and previously published gene‐expression signatures across multiple myeloma cohorts.

Table S2: Genes with different mutation frequencies between the low‐ and high‐IRGS groups.

Acknowledgments

The authors have nothing to report.

Contributor Information

Lei Fan, Email: fanlei@jsph.org.cn.

Jin Fan, Email: fanjin@njmu.edu.cn.

Xiaoyan Qu, Email: quxiaoyan@jsph.org.cn.

Data Availability Statement

Data sharing not applicable to this article as no datasets were generated or analysed during the current study.

References

1. Siegel R. L., Giaquinto A. N., and Jemal A., “Cancer Statistics, 2024,” CA: A Cancer Journal for Clinicians 74, no. 1 (2024): 12–49.

2. Rajkumar S. V., “Multiple Myeloma: 2022 Update on Diagnosis, Risk Stratification, and Management,” American Journal of Hematology 97, no. 8 (2022): 1086–1107.

3. Munshi N. C., L. D. Anderson, Jr. , Shah N., et al., “Idecabtagene Vicleucel in Relapsed and Refractory Multiple Myeloma,” New England Journal of Medicine 384, no. 8 (2021): 705–716.

4. Manier S., Salem K. Z., Park J., Landau D. A., Getz G., and Ghobrial I. M., “Genomic Complexity of Multiple Myeloma and Its Clinical Implications,” Nature Reviews. Clinical Oncology 14, no. 2 (2017): 100–113.

5. Greipp P. R., San Miguel J., Durie B. G., et al., “International Staging System for Multiple Myeloma,” Journal of Clinical Oncology 23, no. 15 (2005): 3412–3420.

6. Palumbo A., Avet‐Loiseau H., Oliva S., et al., “Revised International Staging System for Multiple Myeloma: A Report From International Myeloma Working Group,” Journal of Clinical Oncology 33, no. 26 (2015): 2863–2869.

7. Sonneveld P., “Management of Multiple Myeloma in the Relapsed/Refractory Patient,” Hematology. American Society of Hematology. Education Program 2017, no. 1 (2017): 508–517.

8. Bianchi G. and Munshi N. C., “Pathogenesis Beyond the Cancer Clone(s) in Multiple Myeloma,” Blood 125, no. 20 (2015): 3049–3058.

9. Guillerey C., Nakamura K., Vuckovic S., Hill G. R., and Smyth M. J., “Immune Responses in Multiple Myeloma: Role of the Natural Immune Surveillance and Potential of Immunotherapies,” Cellular and Molecular Life Sciences 73, no. 8 (2016): 1569–1589.

10. Tufail M., Jiang C. H., and Li N., “Immune Evasion in Cancer: Mechanisms and Cutting‐Edge Therapeutic Approaches,” Signal Transduction and Targeted Therapy 10, no. 1 (2025): 227.

11. D'Agostino M., Innorcia S., Boccadoro M., and Bringhen S., “Monoclonal Antibodies to Treat Multiple Myeloma: A Dream Come True,” International Journal of Molecular Sciences 21, no. 21 (2020): 8192.

12. Raje N., Berdeja J., Lin Y., et al., “Anti‐BCMA CAR T‐Cell Therapy bb2121 in Relapsed or Refractory Multiple Myeloma,” New England Journal of Medicine 380, no. 18 (2019): 1726–1737.

13. Thorsson V., Gibbs D. L., Brown S. D., et al., “The Immune Landscape of Cancer,” Immunity 48, no. 4 (2018): 812–830.e14.

14. Finotello F. and Trajanoski Z., “Quantifying Tumor‐Infiltrating Immune Cells From Transcriptomics Data,” Cancer Immunology, Immunotherapy 67, no. 7 (2018): 1031–1040.

15. Wang K., Xiang J., Zhou J., et al., “Development and Validation of a Transcription Factor Regulatory Network‐Based Signature for Individualized Prognostic Risk in Lung Adenocarcinoma,” International Journal of Cancer 156, no. 12 (2025): 2440–2451.

16. Galon J. and Bruni D., “Approaches to Treat Immune Hot, Altered and Cold Tumours With Combination Immunotherapies,” Nature Reviews. Drug Discovery 18, no. 3 (2019): 197–218.

17. Skerget S., Penaherrera D., Chari A., et al., “Comprehensive Molecular Profiling of Multiple Myeloma Identifies Refined Copy Number and Expression Subtypes,” Nature Genetics 56, no. 9 (2024): 1878–1889.

18. Barrett T., Wilhite S. E., Ledoux P., et al., “NCBI GEO: Archive for Functional Genomics Data Sets—Update,” Nucleic Acids Research 41, no. Database issue (2013): D991–D995.

19. Thompson J. A., Tan J., and Greene C. S., “Cross‐Platform Normalization of Microarray and RNA‐Seq Data for Machine Learning Applications,” PeerJ 4 (2016): e1621.

20. Hanzelmann S., Castelo R., and Guinney J., “GSVA: Gene Set Variation Analysis for Microarray and RNA‐Seq Data,” BMC Bioinformatics 14 (2013): 7.

21. Charoentong P., Finotello F., Angelova M., et al., “Pan‐Cancer Immunogenomic Analyses Reveal Genotype‐Immunophenotype Relationships and Predictors of Response to Checkpoint Blockade,” Cell Reports 18, no. 1 (2017): 248–262.

22. Newman A. M., Liu C. L., Green M. R., et al., “Robust Enumeration of Cell Subsets From Tissue Expression Profiles,” Nature Methods 12, no. 5 (2015): 453–457.

23. Aran D., Hu Z., and Butte A. J., “xCell: Digitally Portraying the Tissue Cellular Heterogeneity Landscape,” Genome Biology 18, no. 1 (2017): 220.

24. Racle J., de Jonge K., Baumgaertner P., Speiser D. E., and Gfeller D., “Simultaneous Enumeration of Cancer and Immune Cell Types From Bulk Tumor Gene Expression Data,” eLife 6 (2017): e26476.

25. Yoshihara K., Shahmoradgoli M., Martinez E., et al., “Inferring Tumour Purity and Stromal and Immune Cell Admixture From Expression Data,” Nature Communications 4 (2013): 2612.

26. Wilkerson M. D. and Hayes D. N., “ConsensusClusterPlus: A Class Discovery Tool With Confidence Assessments and Item Tracking,” Bioinformatics 26, no. 12 (2010): 1572–1573.

27. Charrad M., Ghazzali N., Boiteau V., and Niknafs A., “NbClust: AnRPackage for Determining the Relevant Number of Clusters in a Data Set,” Journal of Statistical Software 61, no. 6 (2014): 1–36.

28. Langfelder P. and Horvath S., “WGCNA: An R Package for Weighted Correlation Network Analysis,” BMC Bioinformatics 9 (2008): 559.

29. Liu H., Zhang W., Zhang Y., et al., “Mime: A Flexible Machine‐Learning Framework to Construct and Visualize Models for Clinical Characteristics Prediction and Feature Selection,” Computational and Structural Biotechnology Journal 23 (2024): 2798–2810.

30. Wu T., Hu E., Xu S., et al., “clusterProfiler 4.0: A Universal Enrichment Tool for Interpreting Omics Data,” Innovation (Cambridge) 2, no. 3 (2021): 100141.

31. Gu Z., Eils R., and Schlesner M., “Complex Heatmaps Reveal Patterns and Correlations in Multidimensional Genomic Data,” Bioinformatics 32, no. 18 (2016): 2847–2849.

32. Hao Y., Stuart T., Kowalski M. H., et al., “Dictionary Learning for Integrative, Multimodal and Scalable Single‐Cell Analysis,” Nature Biotechnology 42, no. 2 (2024): 293–304.

33. McGinnis C. S., Murrow L. M., and Gartner Z. J., “DoubletFinder: Doublet Detection in Single‐Cell RNA Sequencing Data Using Artificial Nearest Neighbors,” Cell Systems 8, no. 4 (2019): 329–337.e4.

34. Choudhary S. and Satija R., “Comparison and Evaluation of Statistical Error Models for scRNA‐Seq,” Genome Biology 23, no. 1 (2022): 27.

35. Korsunsky I., Millard N., Fan J., et al., “Fast, Sensitive and Accurate Integration of Single‐Cell Data With Harmony,” Nature Methods 16, no. 12 (2019): 1289–1296.

36. Van der Maaten L. and Hinton G., “Visualizing Data Using t‐SNE,” Journal of Machine Learning Research 9, no. 11 (2008): 2579–2605.

37. Gulati G. S., Sikandar S. S., Wesche D. J., et al., “Single‐Cell Transcriptional Diversity Is a Hallmark of Developmental Potential,” Science 367, no. 6476 (2020): 405–411.

38. Bland J. M. and Altman D. G., “The Logrank Test,” BMJ 328, no. 7447 (2004): 1073.

39. Harrell F. E., Lee K. L., and Mark D. B., “Multivariable Prognostic Models: Issues in Developing Models, Evaluating Assumptions and Adequacy, and Measuring and Reducing Errors,” Statistics in Medicine 15, no. 4 (1996): 361–387.

40. Decaux O., Lode L., Magrangeas F., et al., “Prediction of Survival in Multiple Myeloma Based on Gene Expression Profiles Reveals Cell Cycle and Chromosomal Instability Signatures in High‐Risk Patients and Hyperdiploid Signatures in Low‐Risk Patients: A Study of the Intergroupe Francophone du Myelome,” Journal of Clinical Oncology 26, no. 29 (2008): 4798–4805.

41. Dickens N. J., Walker B. A., Leone P. E., et al., “Homozygous Deletion Mapping in Myeloma Samples Identifies Genes and an Expression Signature Relevant to Pathogenesis and Outcome,” Clinical Cancer Research 16, no. 6 (2010): 1856–1864.

42. Min R., Hu Z., and Zhou Y., “Identifying the Prognostic Significance of Mitophagy‐Associated Genes in Multiple Myeloma: A Novel Risk Model Construction,” Clinical and Experimental Medicine 24, no. 1 (2024): 249.

43. Chapman M. A., Lawrence M. S., Keats J. J., et al., “Initial Genome Sequencing and Analysis of Multiple Myeloma,” Nature 471, no. 7339 (2011): 467–472.

44. Kilchert C., Wittmann S., and Vasiljeva L., “The Regulation and Functions of the Nuclear RNA Exosome Complex,” Nature Reviews. Molecular Cell Biology 17, no. 4 (2016): 227–239.

45. Zhang L., Han X., and Shi Y., “Association of MUC16 Mutation With Response to Immune Checkpoint Inhibitors in Solid Tumors,” JAMA Network Open 3, no. 8 (2020): e2013201.

Associated Data

Supplementary Materials

Figure S1: Validation of clustering stability. (A) The proportion of ambiguous clustering (PAC) scores for cluster numbers k = 2 to 6 were displayed, with the lowest value at k = 2 indicating optimal clustering stability. (B) The bar chart displayed the frequency of recommended cluster numbers according to the NbClust package. (C) Principal component analysis (PCA) plot visualized the spatial distribution of samples. (D) Three other algorithms including CIBERSORT, EPIC, and xCell, further verified the stability and robustness of the ssGSEA results.

Figure S2: WGCNA network construction and functional characterization of the hub module. (A) Hierarchical clustering analysis was performed to evaluate sample consistency and identify potential outliers within the cohort. (B) Selection of best soft threshold power. (C) The clustering tree of 5 differentially colored modules. (D) Top 20 Gene Ontology (GO) enrichment categories for biological process in the blue module genes.

Figure S3: Univariate Cox analysis identified 47 prognostic genes in the MMRF cohort. Data were presented as hazard ratio (HR) ± 95% confidence interval (CI).

Figure S4: Association between the IRGS and immune infiltration in the MMRF cohort. (A) Distribution of Immune Scores between the low‐ and high‐IRGS groups. Each point represents one patient. Violin plots show the distribution of Immune Scores, and red points indicate the group means. (B) Pearson correlation between the IRGS score and the Immune Score. Each point represents one patient; the solid line indicates the fitted linear regression, and the shaded band represents the 95% confidence interval.

Figure S5: Prognostic stratification by previously published gene‐expression signatures across MM cohorts. Kaplan–Meier curves compare overall survival between high‐ and low‐risk groups defined by IFM15, MRCIX6, and the mitophagy‐related signature. Each row represents one prognostic signature, and each column represents one dataset: MMRF, GSE2658 , GSE24080 , and GSE136337 . Survival differences were assessed using the log‐rank test.

Figure S6: Sensitivity analysis of the single‐cell IRGS module activity across alternative scoring methods. Correlation of UCell‐derived module activity with AUCell‐derived scores (A) and AddModuleScore‐derived score (D). Distribution of AUCell‐derived (B) and AddModuleScore‐derived (E) module activity across HD, MGUS, SMM, and MM samples. (C, F) Distribution of the corresponding module activity across major bone marrow cell populations. Boxes indicate the interquartile range, horizontal lines indicate the median, and whiskers extend to 1.5 times the interquartile range. Statistical significance is indicated as ** p < 0.01, *** p < 0.001, and **** p < 0.0001.

Table S1: Comparison of the prognostic discrimination of the IRGS and previously published gene‐expression signatures across multiple myeloma cohorts.

Table S2: Genes with different mutation frequencies between the low‐ and high‐IRGS groups.

Data Availability Statement

Data sharing not applicable to this article as no datasets were generated or analysed during the current study.