scPheno: A deep generative model to integrate scRNA-seq with disease phenotypes and its application on prediction of COVID-19 pneumonia and severe assessment

Abstract

Cell-to-cell variability is orchestrated by transcriptional variations participating in different biological processes. However, the dissection of transcriptional variability in specific biological process at single-cell level remains unavailable. Here, we present a deep generative model scPheno to integrate scRNA-seq with disease phenotypes to unravel the invisible phenotype-related transcriptional variations. We applied scPheno on COVID-19 blood scRNA-seq to separate transcriptional variations in regulating COVID-19 host immunity and transcriptional variations in maintaining cell-type identity. In silico, we found CLU+IFI27+S100A9+ monocyte as the efficient cellular marker for the prediction of COVID-19 diagnosis. Inspiringly, using only 4 genes upregulated in CLU+IFI27+S100A9+ monocytes can predict the COVID-19 diagnosis of individuals from different country with an accuracy up to 81.3%. We also found C1+CD163+ monocyte and 8 C1+CD163+ monocyte-upregulated genes as the efficient biomarkers for the prediction of severity assessment. Overall, scPheno is an effective method in dissecting the transcriptional basis of phenotype variations at single-cell level.

OmniClustifyXMBD: Uncover putative cell states within multiple single-cell omics datasets

“…We have presented OmniClustify XMBD , an innovative clustering methodology specifically designed for the identification of putative cell states within the intricate landscape of diverse single-cell omics datasets. This method represents an extension of our recent work in adaptive signal isolation [14, 15], now coupled with variational Gaussian mixture modeling. OmniClustify XMBD stands as an exceptional advancement in its capacity to robustly and accurately delineate distinct variations while generating dependable cell clusterings.…”

Clinical Phenotype Prediction From Single-cell RNA-seq Data using Attention-Based Neural Networks

“…CloudPred (He et al, 2021) models the individual points as samples from a mixture of Gaussians, probabilistically assigns points to clusters, then estimates prevalence of the subpopulations and use it to predict the phenotype of that patient. scPheno (Zeng et al, 2022) constructs gene expression profiles by a joint distribution of cell states and disease phenotypes based on a deep generative probabilistic model, and feeds the distribution as the predictive features to support vector machine (SVM) for the phenotype prediction. One of the main weaknesses of these methods is that neither of them uses deep neural networks, which indicates limited model capacity, although it also allows these methods to work under a limited number of labeled training data.…”

“…Here we present ScRAT, a clinical phenotype prediction framework that can learn from limited numbers of scRNA-seq samples with minimal dependence on celltype annotations. Compared to most available scRNA-seq analysis algorithms that model gene expression profiles of different cell clusters by separate Gaussian distributions (He et al, 2021;Zeng et al, 2022), the first contribution in ScRAT is that we utilize the attention mechanism to measure interactions between cells as their correlations, or attention weights. For each cell, we incorporate all of its interaction patterns and attention weights to establish its connections with the corresponding phenotypes.…”

Multimodal weakly supervised learning to identify disease-specific changes in single-cell atlases

“…A significant challenge remains in linking cell-level signals to patient-level phenotypes in an interpretable manner, allowing researchers to understand the underlying cellular processes and mechanisms driving disease phenotypes. Several computational approaches have been developed to predict disease phenotypes at the cellular level [3][4][5][6][7] and at the patient level [8][9][10]. Concurrently, other approaches prioritize cells exhibiting differential transcriptomic signals [11,12] or differential compositional signals compared to a reference phenotype (e.g., healthy vs. diseased) [13].…”

“…Concurrently, other approaches prioritize cells exhibiting differential transcriptomic signals [11,12] or differential compositional signals compared to a reference phenotype (e.g., healthy vs. diseased) [13]. However, these approaches are limited as they model single-cell data based solely on transcriptomics and cannot handle multimodal datasets [3,6]. Although they provide predictions at the patient level, they fail to effectively link these predictions to the cellular processes driving the disease phenotype [8].…”