Toward Trustworthy Multi-View Representation with Fine-Grained Explainability Embeddings

Multiomics co-learning is a powerful analytical paradigm that has benefited biomedical studies substantially. However, due to the diverse information and complex relationships of multiomics data, naive multi-view learning methods usually run into spurious correlations and biased signatures irrelevant to the diseases of interest. Therefore, the learned representations and cross-omics associations cannot translate into clinical knowledge for disease prediction. This issue becomes particularly severe when clinical data are limited and scarce. To handle this issue, we propose a novel and powerful scheme, referred to as the Causality-driven Trustworthy Multi-View maPping approach (Cad-TMVP). Specifically, we design a fined multi-directional mapping module to extract co-expression patterns across different modalities and capture fine-grained interpretability factors. We also meticulously design dynamic mechanisms to facilitate adaptive loss-term reweighting and trustworthy integration of multiple modalities. Cad-TMVP enhances downstream tasks by developing a cooperative learning module that simultaneously performs automated diagnosis and result interpretation. Furthermore, we develop an efficient search strategy and support computation to reduce the high computational burden, making our approach practicable. We conduct extensive experiments on different types of multiomics data. The proposed method establishes new state-of-the-art results in various settings while maintaining excellent interpretability. Thus, it sets a potentially newparadigm in trustworthy multi-modal learning and verifies its flexibility and versatility in real biomedical applications.