Fig. 1: The overview of INSTINCT.

a The model structure of INSTINCT. INSTINCT adopts a two-stage strategy. In stage 1, the model of INSTINCT consists of a GAT encoder, a MLP decoder, and a discriminator. Taking the preprocessed data matrix of multiple samples and the concatenated neighbor graph as input, it simply learns to project all the spots from feature space (PCs space) into a low-dimensional shared latent space, while the discriminator is pretrained. During stage 2, a noise generator is added to the model. After the extraction of biological variations performed by the encoder, the noise generator captures the original pattern of batch effects for each spot in the latent space, and a PNG process simulates a different pattern of batch effects for each spot. These two simulated batch effects are then added to the corresponding spot separately, enabling reconstruction back to its origin domain or conversion to a new domain, referred to as the stochastic domain translation process. After training, the latent representation matrix \({{{\bf{Z}}}}\) can be extracted and utilized for various applications and downstream analyses. b The sketch figure of the stochastic domain translation procedure, using the integration process of two samples as an illustration. The GAT encoder, MLP decoder, and discriminator respectively complete the processes of biological variations extraction, reconstruction or generation, as well as discrimination and classification. The batch effects addition procedure is completed by the noise generator and the PNG process. Specifically, the encoder extracts the biological variations within the data and projects it to the latent space. Subsequently, two different patterns of batch effects are added to the low-dimensional representation of each spot, enabling reconstruction to its original domain or generation to the target domain through the decoder. The raw data and the generated data are then discriminated and classified by the discriminator. c The applications of the integration results of INSTINCT.