Fig. 8
Gated axial-attention detailed structure. Firstly, this model implements feature embedding for all gene sets, using horizontal and vertical gated attention mechanisms, respectively. By calculating the horizontal (\(Q_{h} \times K_{h}\)) and vertical (\(Q_{v} \times K_{v}\)) attention scores, the model is able to identify key gene sets. With the help of these key gene sets, we can conduct further studies such as difference analysis, enrichment analysis, and gene characterisation, which provide the basis for subsequent analysis and interpretation. Subsequently, the model fuses the outputs of horizontal and vertical gated attention and calculates scores for each category through a linear layer. Ultimately, these scores were converted into corresponding probabilities via a softmax function.
