Paired transcriptome and 3D genome from one cell: scHiR fuses two sparse modalities into a single cell embedding
Synopsis
Technologies that simultaneously sequence single-cell transcriptome and 3D genome already yield paired measurements from the same cell, but the two modalities differ sharply in noise, scale and biological meaning, and were previously analyzed separately with latent spaces that need not align. scHiR maps gene expression and chromatin contact matrices into one hyperspherical latent space through modality-specific encoders and reconstructs both signals. On paired data from HiRES, GAGE-seq and LiMCA, the joint embedding matched or exceeded single-omics methods in clustering, and on mouse main olfactory epithelium it raised the Spearman correlation between pseudotime and developmental stage from 0.357 and 0.205 for single-omics embeddings to 0.517.
Interpretation
Paired single-cell transcriptome and 3D genome data can now enter one cell-level embedding instead of being compressed separately and concatenated. Prior practice ran scVI, scMGCA and scGNN on expression profiles and scHiCluster and Fast-Higashi on chromatin contact maps, analyzing the two omics separately or concatenating features whose latent spaces are not necessarily aligned. On five paired datasets from three technologies, HiRES, GAGE-seq and LiMCA, k-means clustering on the joint embedding gave ARI and NMI competitive with representative single-omics methods; on the HiRES mouse embryonic stem cell data the joint embedding reached ARI 0.542, above scVI at 0.488, scGNN at 0.343, scMGCA at 0.244, Fast-Higashi at 0.179 and scHiCluster at 0.150.
The model encodes each cell's expression profile and chromatin contact map separately, then fuses them into one shared latent variable. The expression branch feeds raw counts of the top 2,000 highly variable genes into a multilayer perceptron; the chromatin branch splits the whole-genome contact map into intra-chromosomal matrices at 1 Mb resolution, runs a LeNet-based convolutional encoder per chromosome, and averages across chromosomes. The two feature sets are concatenated to parameterize a von Mises–Fisher posterior on the unit hypersphere, from which the shared latent variable is sampled; expression is reconstructed with a zero-inflated negative binomial likelihood and chromatin with a Poisson-form reconstruction loss, plus a KL divergence to a uniform prior and a term that prevents mean directions from collapsing.
The joint embedding preserves and refines cell-type and developmental-trajectory structure. scHiR(HiC), trained on chromatin data alone, was clearly weaker than the joint embedding in most settings, while the joint embedding matched or exceeded the expression-only scHiR(RNA). On mouse main olfactory epithelium with diffusion pseudotime, the joint embedding reached Spearman correlation 0.517 and Kendall's tau 0.398 against developmental stage, above scHiR(RNA) at 0.357 and 0.257 and scHiR(HiC) at 0.205 and 0.152; on HiRES mouse brain, Leiden clustering split seven original cell types into 12 clusters.
Perspective
Evaluation covers paired datasets from three technologies, HiRES, GAGE-seq and LiMCA, including mouse brain, mouse embryonic stem cells, mouse cortex, human bone marrow CD34+ cells, human cell lines and mouse main olfactory epithelium; the model currently handles intra-chromosomal contact matrices at a fixed resolution.
The model currently focuses on intra-chromosomal contact matrices at a fixed resolution, which may limit modeling of finer-scale contacts and inter-chromosomal interactions; regulatory relationships between chromatin architecture and gene expression would still need additional experimental evidence.