Deep Learning Autoencoder Integrates Multi-Omics Data to Profile Multiple Myeloma Risk

Researchers have developed an artificial intelligence framework called MyeVAE that integrates clinical variables and multi-omics data into a shared representation to profile survival risk in newly diagnosed multiple myeloma. In validation tests utilizing patient records from the Multiple Myeloma Research Foundation CoMMpass study, the multi-modal variational autoencoder achieved a concordance index of 0.70 for overall survival when trained on RNA sequencing gene expression alongside clinical markers. The computational investigation, published in BMC Artificial Intelligence, evaluated whether pairing high-dimensional molecular profiles with deep generative neural networks could enhance individualized prognostic stratification across heterogeneous disease courses. The research team evaluated six biological data modalities and four survival prediction algorithms, determining that transcriptomic sequencing data provided the highest predictive contribution among all tested molecular layers. Beyond risk quantification, the model captured documented genomic alterations and identified specific molecular targets, establishing how deep learning frameworks can extract prognostic signals from complex cancer datasets.

Multiple myeloma remains a clinically and genetically heterogeneous hematological malignancy marked by extreme variations in patient survival, ranging from long-term disease remission exceeding a decade to rapid progression within months of diagnosis. In clinical practice, risk-adapted management schemes are implemented to tailor treatment intensity, identify candidates for autologous stem cell transplantation, and prevent excessive chemotoxicity in lower-risk individuals. Standard clinical staging approaches, primarily the Revised International Staging System and its second revision, rely heavily on serum albumin, beta-2-microglobulin, and lactate dehydrogenase, alongside targeted cytogenetic markers including del(17p), t(4;14), and t(14;16). While these markers reflect renal damage, tumor burden, and specific chromosomal disruptions, they exclude wide swaths of molecular biology that influence disease trajectories. Previous studies have demonstrated that mutational processes such as APOBEC cytidine deaminase activity, structural rearrangements, and gene expression signatures carry substantial prognostic relevance. However, effectively unifying these disparate biological measurements into a single coherent prognostic model has presented significant methodological hurdles.

To overcome analytical challenges posed by high-dimensional molecular profiles, the study authors turned to deep generative modeling rather than classical linear regression. Traditional linear survival models, particularly Cox proportional hazards formulations, frequently experience mathematical breakdown when encountering wide datasets where feature counts vastly exceed sample sizes and collinearity is widespread. Artificial neural networks inherently support non-linear relationships and high-dimensional parameter spaces, making them suitable candidates for multimodal survival analysis. While previous architectures such as AUTOSurv demonstrated the efficacy of variational autoencoders for transcriptomic data across multiple primary cancers, they lacked full integration with diverse genomic inputs. The newly introduced MyeVAE architecture extends the traditional variational autoencoder formulation by mapping up to six distinct data modalities simultaneously into an integrated, lower-dimensional latent representation. By combining unsupervised data reconstruction with a dedicated survival prediction sub-network, the framework directly steers its internal data representations along trajectories that correspond to actual clinical outcomes.

Structurally, the MyeVAE model incorporates an encoder-decoder architecture coupled to a secondary task network designated as the survival network. The variational autoencoder maps heterogeneous biological data into a continuous latent space regularized through Kullback-Leibler divergence relative to a standard multivariate normal distribution. Concurrently, the survival sub-task network takes the omics-derived vector embeddings along with three clinical demographic variables—patient age, sex, and International Staging System stage—to compute a scalar risk score bounded between negative one and positive one. To optimize this dual architecture, the investigators applied a semi-supervised objective function combining mean squared error reconstruction loss with a generalized partial log-likelihood function derived from Cox proportional hazards formulations. Adaptive Moment Estimation served as the optimization algorithm, governed by early stopping criteria over training survival loss. This design enables MyeVAE to orient its latent representation according to survival hazard while simultaneously modeling complex non-linear interactions across diverse molecular profiles.

Model development and internal evaluation relied on interim analysis 21 of the Multiple Myeloma Research Foundation CoMMpass study, which cataloged clinical and genomic measurements across 1143 newly diagnosed patients. An audit of data availability revealed that while clinical attributes were universally recorded, individual molecular assays varied in completeness. Whole exome sequencing mutational signatures were present for 947 patients, whole genome copy number profiles covered 877 patients, while RNA sequencing and canonical immunoglobulin heavy chain translocation calls were available in 767 and 783 cases respectively. Complete records spanning all modalities existed for 628 patients, or approximately 55 percent of the cohort. To prevent train-test data leakage, the authors constructed a scikit-learn preprocessing pipeline. Elastic-net regularized Cox filters selected informative features using hazard coefficient cutoffs of plus or minus 0.05 for gene expression and gene-level copy number. Missing values were subsequently imputed using iterative round-robin random forest estimators fitted strictly on training subsets within cross-validation.

Internal cross-validation on the CoMMpass dataset demonstrated that clinical baseline features alone provide substantial predictive capability, establishing a benchmark against which molecular models were evaluated. A baseline Cox proportional hazards model trained strictly on age, sex, and staging achieved concordance indices of 0.681 for overall survival and 0.634 for progression-free survival, surpassing legacy microarray gene expression profiling signatures like UAMS70 and EMC92. Surprisingly, incorporating RNA sequencing gene expression into standard machine learning algorithms often reduced predictive accuracy for progression-free survival and degraded overall survival accuracy in two out of four models. MyeVAE proved to be the notable exception, lifting the overall survival concordance index from 0.681 to 0.702 with non-overlapping confidence intervals. In contrast, progression-free survival prediction remained difficult across all tested configurations.

To evaluate generalization beyond the primary training cohort, the investigators validated the models across four independent historical clinical datasets profiled on microarray platforms: UAMS-GSE24080, HOVON-65/GMMG-HD4, E-MTAB-4032, and the APEX Phase 3 trial. Because large external cohorts with survival annotations and RNA sequencing were unavailable, the models had to bridge platform-specific measurement differences. Gene expression distributions were log-transformed and standardized to unit variance to harmonize RNA sequencing and microarray values. When compared against established gene expression profiling indices—including UAMS70, EMC92, IFM15, and MRC-IX-6—MyeVAE demonstrated superior cross-platform generalization. Across eight distinct dataset and clinical endpoint combinations, MyeVAE attained the highest concordance index in six settings among models utilizing clinical and expression data. In particular, MyeVAE achieved top performance for overall survival in the HOVON-65 cohort and for progression-free survival in E-MTAB-4032, where several older gene signatures experienced reversed hazard associations yielding concordance indices beneath 0.50.

A primary objective of the investigation was evaluating whether stacking diverse omics modalities alongside clinical parameters and gene expression progressively enhances survival predictions. Contrary to expectations, aggregating additional molecular layers—such as chromosomal band copy numbers, whole-exome mutational signatures, and translocation partners—did not produce widespread, substantial accuracy gains. Concordance indices for MyeVAE and random survival forests plateaued near 0.70, while elastic net Cox and DeepSurv showed only modest numerical gains of one to two percentage points on overall survival. An ablation analysis that stripped gene expression while retaining all remaining modalities caused predictive drops in MyeVAE, DeepSurv, and regularized Cox regression. However, random survival forests trained entirely without gene expression maintained strong performance, achieving a concordance index of 0.638 on progression-free survival and 0.701 on overall survival. The researchers suggested that decision-tree ensembles efficiently parse the dense categorical structures that dominate non-expression modalities when transcriptomic profiling is clinically unavailable.

To ensure that MyeVAE learned biologically grounded patterns rather than relying exclusively on clinical features, the researchers inspected the model latent space using Uniform Manifold Approximation and Projection alongside DeepSHAP attribution algorithms. Two-dimensional projections of 128-dimensional latent embeddings showed smooth gradations of survival risk scores across coordinate axes, confirming that patient clustering corresponded to clinical outcomes. In contrast, models forced into two-dimensional latent spaces showed degraded concordance indices and collapsed feature correlations, demonstrating that higher embedding dimensions are vital for multimodal autoencoders. Feature attribution via DeepSHAP identified critical genomic contributors to predicted risk, prominently highlighting adjacent genes situated on the 1q21 cytogenetic band: PSMD4 and RFX5. Copy number gain of PSMD4, which encodes a 26S proteasome regulatory subunit, correlated strongly with the well-documented high-risk Gain(1q) lesion. In subsequent multivariate survival modeling controlling for stage and 1q21 status, PSMD4 retained an independent association with progression-free survival with a hazard ratio of 1.70.

Attribution analyses also recovered established myeloma mutational mechanisms and structural rearrangements, reinforcing the interpretability of the deep learning framework. Single base substitution signatures SBS2 and SBS13, linked to the mutagenic activity of AID and APOBEC cytidine deaminases, emerged as influential determinants of heightened predicted risk. Conversely, clock-like mutational signatures SBS1 and SBS5 were associated with decreased predicted risk despite patient age correlating positively with adverse outcomes. Canonical translocations such as t(4;14), t(14;16), and MYC translocations were assigned higher risk scores, whereas t(11;14) exhibited slightly favorable survival trajectories. The authors emphasized several limitations of their work, observing that multi-omics data did not consistently surpass single-omics predictions, a phenomenon observed in broader cancer benchmarks. They concluded that future staging enhancements may require advanced multimodal alignment techniques to prevent modality dominance, alongside the incorporation of routine clinical indicators, specific somatic mutation matrices, autologous stem cell transplantation status, and imaging metrics.

Subject of Research: Deep learning integration of multi-omics data for multiple myeloma risk profiling

Article Title: MyeVAE: a multi-modal variational autoencoder for risk profiling of newly diagnosed multiple myeloma

Article References: Chang, J. G., Chen, J., Chew, G.-L., & Chng, W. J. (2025). MyeVAE: a multi-modal variational autoencoder for risk profiling of newly diagnosed multiple myeloma. BMC Artificial Intelligence, 1(1), Article 8. https://doi.org/10.1186/s44398-025-00009-2

Image Credits: AI Generated

DOI: 10.1186/s44398-025-00009-2

Keywords: multiple myeloma, deep learning, variational autoencoder, multi-omics, survival analysis, MyeVAE, cancer genomics, RNA sequencing, multi-modal, variational, autoencoder, risk