STRIVE: interpretable multimodal subtype discovery in Women’s cancers using an attention-weighted joint variational autoencoder
TL;DR
Stratification via Integrated Variational Embeddings demonstrates that attention-weighted multimodal integration surfaces molecularly coherent subtypes in which cluster-defining signatures frequently included epigenomic features, a structure that single-modality approaches would obscure.
Abstract
Background Breast, endometrial, and cervical cancers share a common challenge: molecular heterogeneity that single-modality profiling cannot fully resolve. Gene expression and DNA methylation capture distinct and partially non-overlapping regulatory programs, yet most subtyping frameworks treat them independently or merge them without accounting for modality-specific noise and scale. Objective To develop an interpretable multimodal framework that integrates transcriptomic and epigenomic signals for principled molecular subtype discovery and molecular patient stratification with exploratory assessment of outcome associations across women’s cancers. Methods We developed STRIVE (Stratification via Integrated Variational Embeddings), an attention-based joint variational autoencoder that learns modality-aware embeddings, fuses them into a shared latent space, and supports post hoc feature attribution. Subtypes were defined by K-means clustering on latent embeddings, cluster-defining signatures were derived using standardized mean differences (q < 0.05), and within-subtype survival prediction was assessed via permutation feature importance across multiple classifiers. Results In TCGA-BRCA (n = 866), STRIVE identified six molecular subtypes with coherent multimodal signatures, with DNA-methylation features prominently represented in specific cluster signatures. Within-subtype survival-status prediction showed variable performance across classifiers and subtypes. Replication in TCGA-UCEC (n = 454) and TCGA-CESC (n = 296) recapitulated key signature patterns, with DNA-methylation features prominently represented in specific cluster signatures and more consistent predictive performance in UCEC than CESC. Conclusion STRIVE demonstrates that attention-weighted multimodal integration surfaces molecularly coherent subtypes in which cluster-defining signatures frequently included epigenomic features, a structure that single-modality approaches would obscure. These findings position STRIVE as a framework for multimodal molecular stratification and future evaluation of subtype-specific clinical associations.