Modeling Bi-modality Improves Characterization of Cell Cycle on Gene Expression in Single Cells

Advances in high-throughput, single cell gene expression are allowing interrogation of cell heterogeneity. However, there is concern that the cell cycle phase of a cell might bias characterizations of gene expression at the single-cell level. We assess the effect of cell cycle phase on gene expression in single cells by measuring 333 genes in 930 cells across three phases and three cell lines. We determine each cell's phase non-invasively without chemical arrest and use it as a covariate in tests of differential expression. We observe bi-modal gene expression, a previously-described phenomenon, wherein the expression of otherwise abundant genes is either strongly positive, or undetectable within individual cells. This bi-modality is likely both biologically and technically driven. Irrespective of its source, we show that it should be modeled to draw accurate inferences from single cell expression experiments. To this end, we propose a semi-continuous modeling framework based on the generalized linear model, and use it to characterize genes with consistent cell cycle effects across three cell lines. Our new computational framework improves the detection of previously characterized cell-cycle genes compared to approaches that do not account for the bi-modality of single-cell data. We use our semi-continuous modelling framework to estimate single cell gene co-expression networks. These networks suggest that in addition to having phase-dependent shifts in expression (when averaged over many cells), some, but not all, canonical cell cycle genes tend to be co-expressed in groups in single cells. We estimate the amount of single cell expression variability attributable to the cell cycle. We find that the cell cycle explains only 5%–17% of expression variability, suggesting that the cell cycle will not tend to be a large nuisance factor in analysis of the single cell transcriptome.

高通量单细胞基因表达技术的发展，使研究者得以解析细胞异质性（cell heterogeneity）。然而学界普遍担忧，细胞的细胞周期时相（cell cycle phase）可能会对单细胞水平的基因表达表征产生偏倚。本研究通过对3种细胞系、3个细胞周期时相下的930个细胞中的333个基因进行检测，评估了细胞周期时相对单细胞基因表达的影响。我们无需通过化学阻滞即可无创地确定每个细胞的周期时相，并将其作为协变量纳入差异表达（differential expression）检验分析流程。我们观察到了双峰基因表达（bi-modal gene expression）现象——这是一种已有文献报道的现象：即原本丰度较高的基因，在单个细胞中的表达要么呈强阳性，要么完全无法被检测到。这种双峰性可能同时由生物学因素与技术因素共同导致。无论其来源为何，我们的研究均表明，若要从单细胞基因表达实验中获得准确的推断结果，必须对该双峰性进行建模校正。为此，我们提出了一种基于广义线性模型（generalized linear model）的半连续建模框架，并利用该框架对3种细胞系中均存在显著细胞周期效应的基因进行了表征分析。相较于未考虑单细胞数据双峰性的分析方法，我们的新型计算框架能够更精准地检测到已有文献报道的细胞周期相关基因。我们利用该半连续建模框架构建了单细胞基因共表达网络（gene co-expression networks）。这些网络分析结果显示：除了在大量细胞平均后呈现出时相依赖性的表达变化之外，部分（而非全部）经典细胞周期基因（canonical cell cycle genes），在单个细胞中往往会以基因簇的形式实现共表达。我们还估算了单细胞基因表达变异中可归因于细胞周期的比例。我们发现，细胞周期仅能解释5%~17%的基因表达变异，这表明在单细胞转录组（single cell transcriptome）分析中，细胞周期并不会成为影响较大的混杂因素（nuisance factor）。