Fig. 1: Overview of the scCRAFT pipeline.

Created in BioRender. He, C. (2025) https://BioRender.com/x29d2y6a Schematic diagram of scCRAFT's architecture. The algorithm includes a variational autoencoder at its core, which integrates a batch-ignorant encoder and dual generator-decoders, trained through an Evidence Lower Bound (ELBO) loss. A batch classifier undergoes adversarial training alongside the encoder, utilizing cross-entropy loss. Additionally, the model incorporates triplet loss, calculated within the embedding space Z, guided by labels derived from unsupervised clustering performed on X. b Simplified diagram of the conceptual blueprint of scCRAFT, illustrating the synergistic contribution of the VAE, batch classifier, and triplet loss towards the overarching goals of biological conservation and batch correction. c I and II illustrate how the triplets are chosen based on the low and high-resolution clustering and the mechanism of triplet loss training, respectively. IIII demonstrates the application of scCRAFT on a selected subset from the pancreatic islet dataset, which includes three batches and three cell types. The results compare modeling outcomes with and without the implementation of the triplet loss and show the positions of an example triplet before and after training.