sciCAN: Single-cell chromatin accessibility and gene expression data integration via Cycle-consistent Adversarial Network

Abstract

The booming single-cell technologies bring a surge of high dimensional data that come from different sources and represent cellular systems from different views. With advances in single-cell technologies, integrating single-cell data across modalities arises as a new computational challenge and gains more and more attention within the community. Here, we present a novel adversarial approach, sciCAN, to integrate single-cell chromatin accessibility and gene expression data in an unsupervised manner. We benchmarked sciCAN with 3 state-of-the-art (SOTA) methods in 5 scATAC-seq/scRNA-seq datasets, and we demonstrated that our method dealt with data integration with better balance of mutual transferring between modalities than the other 3 SOTA methods. We further applied sciCAN to 10X Multiome data and confirmed the integrated representation preserves information of the hematopoietic hierarchy. Finally, we investigated CRSIPR-perturbed single-cell K562 ATAC-seq and RNA-seq data to identify cells with related responses to different perturbations in these different modalities.

The performance of deep generative models for learning joint embeddings of single-cell multi-omics data

“…Similar to the sciCAN model presented by Xu et al [60], scAEGAN [24] also embraces the concept of cycle consistency, integrating the adversarial training mechanism of a cycle GAN [66] into an autoencoder framework. Specifically, for each modality, a discriminator and a generator are defined.…”

“…In multi-omics data integration, such adversarial approaches are typically integrated into (V)AE models as additional components to regularize the latent representation and/or the decoder reconstructions [22, 60, 65], while a purely GAN-based approach is given by, e.g., Amodio and Krishnaswamy [2].…”

“…This corresponds to the idea of cyclical adversarial training as described in Section 2.4 and Figure 1 E. More generally, this concept is based on a cycle GAN [66] and is also present in, e.g., Khan et al [24], Wang et al [56], Xu et al [60], Zhao et al [65] and Zuo et al [68].…”

“…In addition to the standard GAN loss for each modality, a cycle loss is calculated by mapping a cell from one modality to the second modality with the second modality's generator and mapping it back to the first modality with the first modality's generator and comparing that to the original observation. Unlike for Xu et al [65], the model does not rely on a common feature set but first trains an autoencoder model independently for each modality before training a cycle GAN on the two latent spaces to enforce their consistency.…”

“…In the sciCAN model presented by Xu et al [65], modality-specific autoencoders map the input data to a latent space for each modality, and a discriminator is employed to distinguish between the two modalities based on their latent representations. Additionally, a cross-modal generator is employed that generates synthetic scATAC-seq data based on the scRNA-seq latent representation, and a second discriminator is employed to distinguish between generated and real scATAC-seq samples.…”