Fig 1.
Multiple levels of heterogeneity are believed to operate within tumors.
(left) The genetic “axis” defines accumulating mutational differences that have an effect on phenotype, e.g., drug sensitivity. Although depicted linearly for simplicity, note that mutational accumulation is often nonlinear, occurring instead in a branching manner. (middle) Each genetic clone has an associated epigenetic landscape, where cells are distributed across basins, known as “attractors.” The topography of the epigenetic landscape is defined by the dynamical biochemical network that controls cell fate and function. Quasi-potential energy, U(x), lies along the y-axis and quantifies the relative stabilities of basins; molecular state lies along the x-axis and refers to position within the high-dimensional molecular state space on which the quasi-potential energy is defined (see S1 Text for further discussion). Note that positions of basins across genetic states will not necessarily align (since mutations alter the epigenetic landscape) and they are purposely not aligned in the illustration. (right) Cell states fluctuate within epigenetic basins due to intrinsic (e.g., gene expression) and extrinsic sources of noise. Most fluctuations are minor and do not significantly change the cell state but occasionally a large fluctuation results in a barrier crossing, i.e., a phenotypic state change.
Fig 2.
Phenotypic differences among PC9 cell line versions and discrete sublines quantified in terms of drug response.
(A) Population growth curves for 3 cell line versions treated with 3 μM erlotinib for approximately 3 weeks, plus vehicle (DMSO) control. (B) DIP rate distributions compiled from single-colony growth trajectories under erlotinib treatment (3 μM) in a cFP assay. DIP rates are calculated from the growth curves 48 h postdrug addition to the end of the experiment. Dashed black lines signify zero DIP rate, for visual orientation. (C) Seven DS derived from PC9-VU were treated with 3 μM erlotinib for 3 weeks, along with vehicle control. Parental PC9-VU is included for reference. (D) DIP rate distributions from a cFP assay of the sublines in 3 μM erlotinib. Parental PC9-VU is included for reference. In A and C, dots are the means of 6 experimental replicates at each time point; solid lines are best fits to the drug response trajectories with point-wise 95% confidence intervals. In B and D, DIP rate distributions are plotted as kernel density estimates. The data underlying this figure can be found in github.com/QuLab-VU/GES_2021. cFP, clonal fractional proliferation; DIP, drug-induced proliferation; DS, discrete subline.
Fig 3.
Genomic and transcriptomic characterizations of PC9 cell line versions.
(A) Mean-centered mutation count by chromosome for all cell line versions. For each chromosome, versions with fewer mutations than the mean have a bar pointing inwards, while those with more mutations than the mean point outwards. Chromosome numbers are noted on the outside edge of the circle. Average CV across all chromosomes is noted. (B) Proportions of unique mutations for all cell line versions. (C) Numbers of IMPACT mutations unique to each cell line version, stratified by IMPACT classification (“low,” “moderate,” and “high”). Percentage of unique IMPACT mutations relative to the total number of unique mutations for each cell line version is denoted above each bar. (D) Quantification of mutational differences between cell line versions based on a signature of genes with a high nonsynonymous mutational load. Rows are ordered the same as in Fig 4D and numbered in increments of 10 for ease of reference (see Table B in S1 Text). Heatmap elements are colored based on type of mutation. Total numbers of mutations (stratified by mutation type) across genes and cell line versions are shown as bar plots to the right and above the heatmap, respectively. (E) Copy number variant detection for cell line versions. Red corresponds to amplifications, blue to deletions. The average signal across all cells in the cell line versions was used to define the baseline reference. (F) UMAP visualization of single-cell transcriptomes for cell line versions. For comparison purposes, the UMAP space is defined over all 8 PC9 samples (including cell line versions and PC9-VU sublines; see Materials and methods). (G) GO comparison analysis of unique IMPACT mutations and DEGs for cell line versions. A correlation coefficient (Pearson) was calculated for each sample. Terms with -log10(p) > 2 on either axis are displayed on the plots. The data underlying this figure can be found in github.com/QuLab-VU/GES_2021. CV, coefficient of variation; DEG, differentially expressed gene; GO, Gene Ontology; UMAP, Uniform Manifold Approximation and Projection.
Table 1.
Simplified representation of hallmark gene signature scores across different members of the PC9 cell line family.
Single-cell transcriptomics data from all 8 PC9 cell line family members were subjected to a VISION functional interpretation analysis to calculate scores for single cells from 50 MSigDB hallmark gene signatures. Distributions of scores were calculated for each cell line family member-gene signature pair. Signature score distributions that show larger (↑) or smaller (↓) values compared to other cell line family members are noted. Cell line family members with the largest transcriptomic differences (PC9-MGH, PC9-BR1, DS8, and PC9-VU) were chosen for comparisons. “PC9-VU (no DS8)” includes the parental PC9-VU and DS3, DS6, DS7, and DS9 subline distributions, since they all have a large degree of transcriptomic overlap.
Fig 4.
Genomic and transcriptomic characterizations of PC9-VU discrete sublines.
(A) Mean-centered mutation count by chromosome for 5 (of the 7) sublines. Average CV across all chromosomes is noted. (B) Proportions of unique mutations in each subline. (C) Numbers of IMPACT mutations unique to each subline, stratified by IMPACT classification (“low,” “moderate,” and “high”). Percentage of unique IMPACT mutations relative to the total number of unique mutations in each subline is denoted above each bar. (D) Quantification of mutational differences between sublines based on the same gene signature as for the cell line versions (Fig 3D). Rows are ordered the same as in Fig 3D and numbered in increments of 10 for ease of reference (see Table B in S1 Text). Heatmap elements are colored based on type of mutation. (E) CNV detection for sublines. Red corresponds to amplifications, blue to deletions. PC9-VU was used as the baseline reference to compare sublines. (F) UMAP visualization of single-cell transcriptomes for sublines plotted in the common UMAP space of all 8 PC9 samples (including cell line versions and PC9-VU sublines; see Materials and methods). (G) GO comparison analysis of unique IMPACT mutations and DEGs for sublines. A correlation coefficient (Pearson) was calculated for each sample. Terms with -log10(p) > 2 on either axis are displayed on the plots. Samples without a correlation line or coefficient (DS3, DS6, and DS7) did not meet the minimum threshold of data points to compute a correlation. The data underlying this figure can be found in github.com/QuLab-VU/GES_2021. CNV, copy number variant; CV, coefficient of variation; DEG, differentially expressed gene; GO, Gene Ontology; UMAP, Uniform Manifold Approximation and Projection.
Fig 5.
Stochastic simulations of a simple birth-death model reproduce DIP rate distributions from PC9-VU sublines.
(A) Experimental cFP time courses for 2 representative sublines (DS3 and DS4) in response to 3 μM erlotinib (same data used to generate the corresponding DIP rate distributions in Fig 2D). Each trace corresponds to a single colony, normalized to 72 h postdrug treatment. Only colonies with initial cell counts greater than 50 at the time of treatment are shown. (B) In silico cFP time courses from a one-state model with division and death rate constants that closely reproduce the experimental time courses in A. Trajectories are normalized to the time at which the simulated drug treatment was initiated, simulated cell counts are plotted only at experimental time points, and only colonies with initial cell counts greater than 50 at the time of simulated drug treatment are shown. (C) Comparison of experimental and simulated DIP rate distributions from time courses in A and B. Distributions are compared statistically using the AD test [66]. Dashed black line signifies zero DIP rate, for visual orientation. (D) Parameter scan of division and death rate constants for the 2 sublines in A–C. For each pair of rate constants, the same number of model simulations were run as the associated experimental cFP time courses in A. DIP rates were calculated and compiled into a distribution and then statistically compared against the corresponding experimental DIP rate distribution using the AD test. All parameter pairs with p < 0.05 (see Materials and methods) are colored white, indicating lack of statistical correspondence to experiment. “×” denotes a division and death rate constant pair used in B. (E) Same as A but for subline DS8. (F) Same as B but for DS8 using a two-state model. (G) Same as C but for DS8. (H) Same as D but for DS8 using a four-dimensional (2 division-death rate constant pairs) parameter scan and projected into 2 dimensions. “×” denotes parameter values used to generate the simulated DIP rate distribution in F. The data underlying this figure can be found in github.com/QuLab-VU/GES_2021. AD, Anderson–Darling; cFP, clonal fractional proliferation; DIP, drug-induced proliferation.
Fig 6.
Schematic interpretation of the results of our analyses on the PC9 cell line family.
Cell line versions PC9-MGH (light green) and PC9-VU (blue) represent different genetic clones within our in vitro tumor model that emerged from a common PC9 founder clone (gray). Each genetic clone has an associated epigenetic landscape, including PC9-VU, which has at least 3 distinct basins corresponding to the sublines DS3, DS7/DS9, and DS6. Cell line version PC9-BR1 (red) was derived from PC9-VU via drug-induced genetic clonal selection in EGFRi, while DS8 (dark green) independently acquired a genetic resistance mutation in the absence of selective drug pressure. However, it remains unclear from which basin DS8 arose (signified by question marks) and whether the resistance mutation was acquired before or after the subline was established (see S1 Text and S11 Fig). Note that PC9-VU is depicted as lying closer in the phylogenetic tree to the PC9 founder line than PC9-MGH because PC9-VU appears to be a closer genetic descendent of the founder, based on our genomic data. Likewise, DS8, which is believed to have emerged only recently, is depicted closer than PC9-BR1 to the common ancestor PC9-VU. EGFRi, EGFR inhibitor.