ADFusion: Multi-modal adaptive deep fusion for cancer subtype prediction

The identification of cancer subtypes is crucial for personalized treatment. Subtype prediction can be achieved by using multi-modal data collected from patients. Multi-modal cancer data contains hidden joint information that cannot be adequately tapped by current vector-based fusion methods. To address this, we propose a multi-modal adaptive deep fusion network ADFusion, which utilizes a hierarchical graph convolutional network HiGCN for high-quality representation of multi-modal cancer data. Subsequently, an adaptive deep fusion network based on deep equilibrium theory is designed to capture effectively multi-modal joint information, which is then fused with multi-modal feature vectors to produce the fused features. HiGCN includes co-expressed genes and sample similarity networks, which provide a more nuanced consideration of the relationships between genes, and also between samples, achieving superior representation of multi-modal genes data. Adaptive deep fusion network, with flexible non-fixed layer structure, is designed for mining multi-modal joint information, automatically adjusting its layers according to real-time training conditions, ensuring flexibility and broad applicability. ADFusion was evaluated across 5 public cancer datasets using 3 evaluation metrics, outperforming state-of-arts methods in all results. Additionally, ablation experiments, convergence analysis, and interpretability analysis also demonstrate the performance of ADFusion.