DG-MSGAT: A Biologically-informed Differential Gene Multi-Scale Graph Attention Network for predicting neoadjuvant...

Accurate prediction of the efficacy of neoadjuvant therapy - particularly the likelihood of achieving a pathological complete response (pCR) - is critical to improving outcomes in patients with rectal cancer. The anticipation of therapeutic response prior to surgery enables the development of personalized treatment strategies and reduces unnecessary interventions for non-responders. While genetic profiling has been integrated into predictive models to enhance response estimation, many existing approaches overlook gene-gene interactions. Furthermore, they often struggle with the high dimensionality, noise, and sparsity inherent in gene expression data. To address these limitations, we propose a biologically informed model, the Differential Gene Multi-Scale Graph Attention Network (DG-MSGAT). This model integrates differential expression signals with multi-scale gene interaction patterns to improve the accuracy of treatment response prediction.

By integrating gene expression profiles with differential expression signals, we construct a patient-specific gene graph whose edges are defined based on curated protein-protein interaction data. This graph is then processed by DG-MSGAT, a multi-scale graph attention network that utilizes stacked attention layers and residual connections to model hierarchical gene dependencies and preserve feature integrity. The resulting representation is subsequently used to estimate the probability of achieving a pathological complete response.

In patients with locally advanced rectal cancer, the DG-MSGAT model substantially outperformed conventional algorithms - including support vector machines, decision trees, and random forests - in predicting neoadjuvant therapy efficacy. Network analysis identified key genes (e.g., TP53, EGFR, CTNNB1) and immune-related pathways that are consistent with clinically established determinants of therapeutic response.

The DG-MSGAT model offers a promising advancement in the prediction of neoadjuvant therapy outcomes in rectal cancer. By effectively modeling gene interactions and mitigating the limitations associated with high-dimensional gene expression data, it provides a clinically relevant tool to support personalized treatment decision-making.