VARLOMS: A Multimodal CVAE Framework for Synthetic 3D MRI Generation from Multi-Omics Data

Integrating magnetic resonance imaging (MRI) with multi-omics data provides a non-invasive and biologically informed approach to understanding genotype-to-phenotype relationships. While MRI captures anatomical phenotypes, multi-omics data reveal underlying molecular mechanisms. However, the scarcity of paired datasets and the challenge of aligning heterogeneous modalities hinder effective multimodal learning. To address this, we introduce VAR-LOMS, a Conditional Variational Autoencoder (CVAE)-based framework that generates high-resolution synthetic 3D breast MRI volumes conditioned on patient-specific multi-omics profiles. VARLOMS encodes exon and gene expression, miRNA, DNA methylation, somatic mutations, and copy number variations into a 17-dimensional latent space using a deep biologically informed autoencoder. Unlike existing methods that treat modalities independently, VARLOMS captures complex cross-modal interactions, emphasizing the regulatory roles of non-coding RNAs and epigenetic modifications. The model outperforms baseline generators in image fidelity, Fréchet Inception Distance and achieves strong performance in downstream tasks such as oncogenic mutation prediction and molecular subtype classification. Validation through t-SNE and ablation analysis confirms the biological relevance of selected omics features. VARLOMS provides a scalable, interpretable, and data-efficient solution for synthetic MRI generation, enabling data augmentation, digital twin construction, and integrative radiogenomics for precision medicine. The source code and datasets are publicly available at: https://github.com/ahsanrizvi/VARLOM.git.