Figure 1.

Overall architecture diagram of scAMAC. scAMAC consists of two parts: denoising deep multi-scale autoencoder and self-supervised clustering network. Firstly, the preprocessed gene expression matrix is fed into the denoising deep multi-scale autoencoder, obtaining the latent feature representation of the hidden layer, as well as the reconstructed data . Then, the outputs of the encoding layer and decoding layer are fed into two fully connected layers to obtain and , where the dimensions of and are the same as . Finally, , , and are concatenated and fed into the self-supervised clustering network. The self-supervised clustering network uses the MSA mechanism to capture the relationship between cells and the contribution of each layer of the autoencoder to obtain . The membership matrix of the FKM algorithm is calculated based on and optimized for the self-supervised clustering network. Meanwhile, is used to construct a cell similarity matrix to supervise the parameter update of the autoencoder.