Memory Sequencing Reveals Heritable Single-Cell Gene Expression Programs Associated with Distinct Cellular Behaviors

Abstract

Non-genetic factors can cause individual cells to fluctuate substantially in gene expression levels over time. It remains unclear whether these fluctuations can persist for much longer than the time of one cell division. Current methods for measuring gene expression in single cells mostly rely on single time point measurements, making the duration of gene expression fluctuations or cellular memory difficult to measure. Here, we combined Luria and Delbrück's fluctuation analysis with population-based RNA sequencing (MemorySeq) for identifying genes transcriptome-wide whose fluctuations persist for several divisions. MemorySeq revealed multiple gene modules that expressed together in rare cells within otherwise homogeneous clonal populations. These rare cell subpopulations were associated with biologically distinct behaviors like proliferation in the face of anti-cancer therapeutics. The identification of non-genetic, multigenerational fluctuations can reveal new forms of biological memory in single cells and suggests that non-genetic heritability of cellular state may be a quantitative property.

Keywords: cancer drug resistance; gene expression memory; single-cell.

Figure 1:. MemorySeq can identify genes with high transcriptional memory.

A. Rare-cell gene expression patterns, both with and without heritability. Histograms of single-cell expression levels are unable to discriminate between these two alternatives. B. Schematic of MemorySeq experiment. We started with a single melanoma cell (WM989-A6), grew it to ~100 cells, then seeded 48 wells from those cells and allowed the cells to proliferate to around ~100,000 cells before subjecting the entire MemorySeq clone to RNA sequencing to determine expression levels. In the case of non-heritable expression, the levels of expression would not vary dramatically between MemorySeq clones, whereas in the heritable case, some clones would exhibit much higher levels of expression when a cell moved into the high expression level state early in the family tree of the clone. To determine how much variability in expression would arise for purely technical reasons, we also performed control experiments by plating around ~100,000 cells directly into individual wells and performing RNA sequencing. C. Expression histograms across n=43 MemorySeq clones for genes identified as non-heritable (left) or heritable (right). D. Coefficient of variation versus mean expression levels for all 23,669 genes that we analyzed across all MemorySeq clones. Points labeled with blue dots (on MemorySeq clones plot) or pink dots (on Noise controls plot) passed the threshold for being identified as a heritable gene. These genes were identified by first fitting a Poisson regression model to the data, and then selecting genes with residuals in the top 2%. This approach identified 227 heritable genes from the MemorySeq clones, but only 30 genes passed this threshold in the Noise control condition. Particular genes from the panel in Fig. 1C are labeled on both plots.

Figure 2:. Time-lapse microscopy verifies rare, high expression states that persist for several cell divisions.

We generated a cell line (WM989-A6-G3 C10-C2 clone E9) that expresses a large but incomplete (and thus nonfluorescent) portion of the mNeonGreen2 fluorescent protein with the remaining piece of mNeonGreen2 fused to NGFR at the endogenous locus. When the NGFR fusion protein expresses, the remaining portion of mNeonGreen2 binds to the NGFR fusion protein and becomes fluorescent. We then performed time-lapse microscopy imaging of the NGFR protein (nucleus labeled with H2b-iRFP670) at 6 hour intervals for 8.75 days. A. We tracked cells through several cell divisions, thus building cellular lineages, and quantified fluorescence intensity for each cell. The plot shows two branches from the same parent cell with fluorescence intensity of mNeonGreen2 over time. B. Series of fluorescent micrographs of the two cells highlighted in panel A. Scale bar is 8μm long. C. Correlations between sibling cells, first cousins, and random pairs of cells (n = 486, 292, 905, respectively). D. We stained cells with antibodies targeting AXL and EGFR, then sorted positive cells, plated them on a glass dish, and took images of their immunofluorescence signal. Subsequently, we acquired transmitted light images every hour for 8.67 days to facilitate tracking of cell lineages, and then we performed immunofluorescence again to measure EGFR and AXL levels at the end of the tracking period. From the timelapse images, we tracked selected lineages that initially contained cells with high levels of EGFR and AXL. The red dots on the left images correspond to the red arrow on the histogram for an example initial cell subjected to tracking. Upon division, we colored the tracks of the sibling cells green and blue respectively. The EGFR and AXL levels for these cells in their final state is indicated by the green and blue arrows on the histograms on the right. Scale bars are 10μm long.

Figure 3:. Single-molecule RNA FISH verifies the quantitative nature of MemorySeq for measuring heritability in single cells.

A. Schematic of spatial RNA FISH experiment. We plated WM989-A6 melanoma cells sparsely on a dish and allowed them to grow for 10 days. We then fixed the cells and performed iterative RNA FISH to measure the expression of 19 genes. Closely related cells will remain in close proximity, thus heritable rare-cell expression would manifest as “patches” of on cells, whereas non-heritable rare-cell expression would display a more salt-and-pepper pattern of expression. Right: micrographs of RNA FISH for 4 genes, EGFR (heritable), NGFR (heritable), EEF2 (housekeeping) and GAPDH (housekeeping). B. Each spot is a cell from an RNA FISH image scan of 12,192 cells (subset of 2,103 cells shown). Cells above a threshold (6 for EGFR, 36 for NGFR and 320 for EEF2) were considered to be in the high expression state and colored green. C. Quantitative comparison of heritability as measured by MemorySeq (x-axis: skewness across MemorySeq clones) and spatial RNA FISH analysis (y-axis). We used the Fano factor measured for spatial bins of 20 nearest cells as a spatial clustering metric; randomly placed high-expression-state cells would display a Poisson distribution and thus give a Fano factor of 1. Cell populations with a Fano factor greater than 1 would display some degree of spatial clustering. Of note, this plot and the plot in panel D contain 18 of the 19 genes that we quantified with RNA FISH because 1 gene (CYR61) did not pass the minimum mean transcripts per million cutoff for analysis in the MemorySeq data. D. We plotted MemorySeq heritability versus the Gini coefficient (from RNA FISH). The Gini coefficient measures expression inequality, and thus indicates the rareness of expression, with 0 being completely equal and 1 being completely unequal. E. Rare cells within clonal WM989-A6 populations marked by high levels of NGFR protein were sorted, cultured for 8–16 hours and then subjected to trametinib treatment at 10nM (MEK inhibitor) for 3 weeks. Image shows the number of resistant colonies (circled) along with number of cells within the resistant colony as indicated. Biological replicate available on Dropbox.

Figure 4:. MemorySeq reveals a rare subpopulation of MDA-MD-231-D4 cells associated with drug resistance.

A. Most MDA-MD-231-D4 cells die upon treatment with paclitaxel for 5 days, but a small subpopulation of cells (cell marked with “?”) survive and become resistant (red cell). B. We performed MemorySeq analysis on MDA-MD-231-D4 cells (n=39 clones, left; n=46 control clones, right). The blue colored dots correspond to genes that we statistically identified as being highly heritable by fitting a Poisson regression model and selecting genes with residuals in the top 2%, as was done with WM989. C. We stained cells with antibody targeting the CA9 surface marker and then sorted out the top 0.5% of cells and the lowest 5%. After culturing for 5 days, we re-stained the cells and measured CA9 levels by flow-cytometry. Potential outcomes for the levels of CA9 staining depending on the degree of gene expression memory. Observed outcome below for CA9 staining 5 days after initial sorting. D. We stained cells with antibody targeting the CA9 surface marker and then sorted out the top 2–4% of cells, the lowest 2–4% of cells, and the total “mix” population into chamber wells, after which we applied paclitaxel 1 day after sorting for 5 days. Transmitted light micrographs show the number of cells remaining after drug treatment for the different populations, and the quantification of the number of cells was performed using cell counting based on nuclear identification by imaging the DAPI nuclear stain and identifying computational techniques. All scale bars are 50μm long.

Figure 5:. MemorySeq enables the identification of coordinated rare-cell expression programs.

A. We measured correlations between genes across MemorySeq clones derived from WM989-A6 melanoma cells. Shown is an example correlation between MMP1 and SERPINB2 across 43 MemorySeq clones. B. Correlations between all pairs of genes exhibiting heritability as determined by the threshold described in Fig. 1. Cook’s distance analysis to test for outliers is given in Supp. Fig. 1D. C. Comparison of coherence between MemorySeq bulk RNA-seq analysis and single-cell correlations as measured by single-molecule RNA FISH. We performed RNA FISH on 20 genes in WM989-A6 cells, keeping for further analysis genes whose RNA FISH Gini coefficient was greater than 0.6 (13 genes remaining). The correlation between bulk MemorySeq RNA-seq levels is on the left, RNA FISH on the right. Callout shows raw RNA FISH counts for 12,192 cells between MMP1 and SERPINB2 in single cells. D. Community detection within the network defined by the correlation matrix of co-expression patterns among the heritable genes. Gray circles indicate genes that did not comprise a network community. Green and Red indicate the two communities detected; KEGG pathway and GO Biological Process analysis results shown for both communities. E. Comparison of rare-cell expression programs identified by MemorySeq and those identified by sorting EGFR-high (top) or NGFR (bottom) high cells (using fluorescent antibody labeling followed by RNA-seq on the high versus mix populations).