Tensorised Modular Architectures for Multi-Omics Generation

Multi-omics data has rich cross-modal biological structure. Standard flat architectures ignore this and treat the feature vector as flat and unstructured. We investigate a modular architecture approach where genes and proteins are grouped into biological modules with specific encoders. A coupling layer models higher-order inter-module interactions. We compare two coupling layer approaches: a dense matrix and a Tensor-Train decomposition, which naturally represents higher-order relationships between modules. Our findings show that modular architecture improves correlation metrics substantially over flat baselines at comparable parameter budgets. Further, our preliminary parameter efficiency experiments indicate that Tensor-Train reaches comparable performance with fewer parameters, a promising direction for capturing multi-omics relationships.