ImmuGT-ConRes: a consistency learning residual network for predicting pan-cancer immunotherapy response from gene expression data.

Immunotherapy is leading the transformation in cancer treatment, but its response rate remains low, and there is a scarcity of immune targets that can produce lasting effects. Multi-omics data is widely used in immunotherapy. However, traditional analysis methods often struggle to uncover hidden correlations within the data, limiting the exploration of potential immunotherapy targets. Here, we pooled immunotherapy data from over 40 cohorts and developed ImmuGT-ConRes (Genomic Image Transformation with Consistency Learning and Residual Networks). ImmuGT-ConRes uses a contrastive learning approach, incorporating a dual-branch data augmentation strategy and a multi-scale stride convolution structure. By integrating residual networks and attention mechanisms, ImmuGT-ConRes adapts to multi-scale gene feature capture, reduces information loss between shallow details and deep semantics, and enhances the model's robustness to input perturbations. We demonstrated that ImmuGT-ConRes exhibited superior predictive performance and provided interpretability by ranking gene weights based on the attention mechanism, providing an effective approach for immunotherapy target mining. These results suggest that ImmuGT-ConRes offers a promising framework for immunotherapy target discovery and merits further investigation.