GREMI: an Explainable Multi-omics Integration Framework for Enhanced Disease Prediction and Module Identification

Abstract

By capturing complementary information from multiple omics data, multi-omics integration has demonstrated promising performance in disease prediction. However, with the increasing number of omics data, accurately quantifying the characterization ability of each omics and avoiding mutual interference becomes challenging due to the intricate relationships among omics data. Here, we propose a novel multi-omics integration framework that improves diagnostic prediction. Specifically, our framework involves constructing co-expression and co-methylation networks for each subject, followed by applying multi-level graph attention to incorporate biomolecule interaction information. The true-class-probability strategy is employed to evaluate omics-level confidence for classification, and the loss is designed using an adaptive mechanism to leverage both within- and across-omics information. Extensive experiments demonstrate that the proposed framework outperforms state-of-the-art methods on classification tasks, and indicate that the integration of three omics yields superior performance compared to employing only one or two data types. Code and data are available at https://github.com/Yaolab-fantastic/GRAMI-NET.

Comprehensive review of Transformer‐based models in neuroscience, neurology, and psychiatry

“…Furthermore, Jing et al [156] suggested a Transformer-driven addiction-perception generative adversarial network to detect brain connectivity linked to nicotine addiction. In the realm of molecular interaction networks, models like GREMI [158] and TEMINet [160] are used to integrate heterogeneous multi-omics data in an adaptive manner. They employ a graph attention mechanism to capture disease-specific nuances and complexities, enabling a better understanding of molecular interactions.…”

“…Similar to Multilayer Perceptron and CNNs, Transformers face challenges in effectively handling the complex, nonlinear, and nonsequential nature of brain networks. Brain networks, especially those involving multimodal and heterogeneous data, [145,152,158] exhibit intricate interconnections and hierarchical structures that go beyond the capabilities of standard Transformer architectures. These networks require a nuanced approach that goes beyond the traditional Transformer framework in order to properly capture and analyze their complex dynamics.…”

“…Additionally, the interpretability of Transformer models in network neuroscience is a significant concern. Despite the use of various post-hoc interpretability strategies, [152,155,156,158] the outputs from these models can vary in reliability and significance. While some results may seem plausible, others may be misleading or even contain counterfactual elements due to limitations in the models' ability to effectively discern and filter out irrelevant data.…”

“…In the context of Transformer models, Shapley values offer quantifiable insights into how each feature contributes to a particular prediction. Unlike traditional importance scores, [158] which lack a solid theoretical foundation, Shapley values provide a mathematically grounded approach to discerning the marginal contribution of each feature. For instance, Kokalj et al [224] used Kernel SHAP to explain Transformer-based classifiers.…”