LABMA: Latent-bottleneck attention-based multimodal architecture for integration of transcriptomics, proteomics, and...

Integrating anatomical MRI with molecular data offers a promising path for understanding complex neurological disorders such as Alzheimer's disease and Parkinson's disease. However, no prior work has successfully combined MRI with both proteomic and transcriptomic modalities due to extreme feature dimensionality imbalances-ranging from 1000 to 6 million features across modalities. Standard fusion approaches fail as high-dimensional modalities dominate gradient-based optimization, effectively marginalizing molecular signals. We propose a Latent-Bottleneck attention-based Multimodal Architecture (LABMA) that addresses this challenge through: (1) modality-specific sparse encoders with differentiable tokenization and selection via Gumbel-Top-k relaxation, (2) gated cross-attention with latent bottlenecks that enforce balanced modality contributions, and (3) integrated gradient-based methods for discovering interpretable cross-modal biomarker interactions. Validated on two independent cohorts (ANMerge: n=376, PPMI: n=1848), our approach achieves consistent improvements of 2-8% in AUROC over best competing methods while identifying reproducible MRI-omics interaction patterns aligned with known disease mechanisms. Ablation studies confirm all three modalities contribute unique information, with tri-modal integration significantly outperforming any subset. To our knowledge, this work represents the first architecture specifically designed to address the extreme dimensional imbalance arising from the joint integration of transcriptomic, proteomic, and MRI for neurological disease analysis.