<p>Gene biomarkers are invaluable for understanding medical mechanisms and targeting drugs. An effective interpretability method is to identify the key factors of gene biomarkers. Many interpretable methods in deep learning are based on layer-wise relevance propagation (LRP) that is particularly effective in MLP-based backbone networks, but it can only extract features for individual samples rather than group samples. While, what we want to get is the biomarkers for a kind of disease group, which can represent a disease mechanism, not those for a certain individual. So, we propose to use prototype of the ith category to denote the group features of the ith disease to find gene biomarkers of the category by LRP algorithm. Whereas LRP algorithm can find features, its accuracy is related to the classification efficiency. The higher classification accuracy is, it can find more accurate features based on LRP. Building on the aforementioned theory, in this paper, we propose a two-branch network architecture. The first branch is MLP architecture and primarily designed for identifying group gene biomarkers based on LRP, while the second branch focuses on enhancing classification efficiency, thereby supporting the first objective. In addition, to improve classification accuracy furtherly, we propose a cosine similarity loss based on two-branch network that can promote the two networks to extract more diverse and fine-grained features. To validate this idea, we construct two variants of the two-branch network: ConvAttMLP and scBERTMLP. Finally, we evaluate the networks and find 20 specificity biomarkers for IgAN disease subtype using our group LRP algorithm.</p>