Introduction

Cancer cell proliferation promotes tumor growth with continuously increasing diversity in gene expression profiles, resulting in tumor heterogeneity that allows tumors to include various clonal proportions adapting to the tumor microenvironment, resistance to anticancer treatments, and migration to distant metastatic lesions [3, 37]. Recent large-scale cancer genome analysis has revealed a hierarchical evolution in cancer progression and identified unique mutations in distant metastases from a primary tumor, suggesting that metastatic cancer cells acquire de novo mutations during migration or proliferation in the metastases [7, 39]. However, genetic mutations are not the only cause of the diverse phenotypes of cancer cells, but epigenetic regulation also contributes to the development of tumor diversity [6, 32].

In cancer progression, resistance to anticancer therapies and metastatic potential are the major factors contributing to cancer-related deaths [18]. Thus, it is critical to identify whether cancer cells include a pre-existing population with resistance to anticancer therapies and metastatic abilities or whether these cancer cells appear during cancer progression [9, 10]. Cancer stem-like cells (CSCs) are thought to initiate tumorigenesis and resist environmental stress to promote cancer progression [1]. Pluripotent transcription factors, including NANOG, SOX2, and OCT3/4, contribute to the establishment of stem cell-like properties in cancer cells, similar to embryonic stem cells [19, 24, 25]. Some cancer cell lines indicate the existence of elite cells that initiate tumorigenesis, generate metastases, and survive anticancer drug treatment [2, 19, 25, 29]. However, anticancer drugs do not always select specific clonal populations [24]. Moreover, the contribution of rare clonal proportions with distinct capabilities, such as CSCs, to cancer progression is unclear.

We examined whether a cancer cell population with the distinct capabilities of proliferation, metastasis, and resistance to anticancer drugs is pre-existing or develops post treatment in multiple cancer cell lines using DNA barcoding and gene expression recording systems. To this end, we utilized mouse melanoma cells B16F10 and B16/BL6 (BL6) and the mouse mammary carcinoma cell line 4T1 to identify the distinct cancer cell population promoting cancer progression. These cell lines are often used to analyze metastasis and the human prostate cancer cell line DU145 for tracing CSCs under environmental stress. We previously found that DU145 cells contain NANOG-positive CSCs that escape NK cell attack [13, 31]. In this study, we used a lentivirus-based DNA barcoding system to distinguish cancer cells during cancer progression. We also established DU145 with the insertion of Cas9 immediately before the stop codons of the NANOG, SOX2, and POU5F1 genes. The expression of these transcription factors in the self-targeting guide RNA (stgRNA) was recorded as the frequency of mutations in the sequence, reflecting Cas9 activity [27]. Our results revealed multiple clones and a few clone-composed metastases, the stochastic appearance of anticancer drug-resistant cells, and POU5F1 expression in specific cells responding to either sphere formation or anticancer drugs. Thus, therapeutic strategies are required to target pre-existing and post-treatment cancer cells with metastatic and anticancer drug resistance capabilities.

Results

DNA barcoding system identifies each cell, allowing analysis of cell lineage, proportion, and dynamics at a single-cell level. We designed a lentivirus plasmid with random sequences comprising 30 bases each. The plasmids included 700,807 variations in the depth of ~ 20 M reads, suggesting that 7.1% variation in barcode occupied 95% of the library (Fig. 1A). Sequencing error contributes to a nucleotide mismatch, appearing as hamming distance of 1–2. We sought to reduce the bias causing sequencing errors by aggregating barcodes according to Hamming distance and selecting barcodes detected in plasmids and all barcoded-B16F10, BL6, and 4T1 cells (see Materials and Methods). The obtained 4,367 barcodes did not contain barcodes with excessive amounts (Fig. 1B) and showed an average abundance (Fig. 1C). Thus, we analyzed the dynamics of these 4367 barcodes during cancer progression in B16F10, BL6, and 4T1 cells.

Fig. 1
figure 1

The switch from homogeneous to heterogeneous cancer cell proliferation during tumor growth in DNA-barcoded B16F10 cells. A Cumulative distribution fraction (CDF) of barcodes in plasmid library. X-axis represents the number of barcodes. B Histogram of barcodes commonly detected in the plasmid and cells. X-axis represents the number of reads, whereas Y-axis represents the number of barcodes. Bin size is 1000. C Histogram and CDF of a proportion of barcodes commonly detected in the plasmid and cells. X-axis represents the ranking of barcode in the plasmid, whereas Y-axis represents the proportion of barcode. D The tumor growth curve of barcoding B16F10 cells in C57BL/6N. The tumors with 5, 10, and 15 mm diameters were collected at the indicated days. The day was counted from the inoculation of cancer cells (n = 5 for 5 and 10 mm diameter tumors and n = 6 for 15 mm diameter tumors). E–H Box plots of cell number collected from tumors (E), detected barcode number (F), Shannon entropy (Loge) (G), and corrected barcode number (H). p value was calculated by Welch’s t-test with multiple comparisons Bonferroni correction. I Bubble plot of top 10 barcodes from each of the 17 samples (n = 1 for in vitro, n = 5 for 5 mm, n = 5 for 10 mm, and n = 6 for 15 mm). The total bubble number is 93 in each line since we aggregated the duplicated barcodes detected in the top 10 barcodes of different samples. Bubble size indicates the proportion of the barcode. J Heatmap of Jensen–Shannon divergence calculated using top 10 barcodes in each sample or all barcodes

Full size image

First, we determined whether cancer cells show diversity in cell proliferation in vivo. We inoculated B16F10 cells on the backs of C57BL/6N mice and collected tumors with diameters 5, 10, and 15 mm, corresponding to 4, 10, and 14 days after inoculation (Fig. 1D). Cell numbers increased ~ 200-fold from 5 to 10 mm tumors and ~ fivefold from 10 to 15 mm tumors (p = 0.0018 and p = 0.00019, respectively; Fig. 1E). We found a significant difference in the detected barcode numbers among the samples, although the difference was slight (Fig. 1F). Because the barcode data were skewed, we calculated the corrected barcode number using Shannon entropy to obtain the effective barcode numbers by removing skewing from the data [38] (Fig. 1G and H). The original barcoded-B16F10 cells contained ~ 3250 barcodes. B16F10 tumors drastically increased in cell numbers from 5 to 10 mm tumors, and the number of barcoded cells decreased slightly after cell proliferation (2930–2723, p = 0.025). In contrast, an ~ fivefold increase in cell number from 10 to 15 mm significantly decreased the number of barcode variations (2723–2000, p = 1.2e–06, Fig. 1H). We obtained the top 10 barcodes with the highest number of reads from each sample and created a list of those barcodes. The amounts of barcodes in that list in each sample were used to calculate JS divergence. We then analyzed the change in the abundance of individual barcodes of the top 10 in each sample. Results suggested that the composition of barcodes significantly changed during tumor growth from 10 to 15 mm but not during tumor growth from 5 to 10 mm compared to in vitro barcoded-B16F10 cells (Fig. 1I). We then applied Jensen–Shannon (JS) divergence to compare the similarities between samples [20]. A high JS divergence score indicates low similarity between samples, whereas a low score indicates high similarity. We detected high similarity in the composition of barcodes among in vitro B16F10 cells (5 and 10 mm tumors). In comparison, 15 mm tumors showed different compositions from other samples based on the top 10 barcodes from each sample or all barcodes (Fig. 1J). These data indicate the transition of cell proliferation profiles from homogeneous to heterogeneous during tumor growth in B16F10 tumors.

Next, we examined whether the repression of tumor growth by anticancer drugs decreases the variety of barcodes and enriches a specific clone in B16F10 tumors. We analyzed two anticancer drugs with different molecular mechanisms to repress tumor growth: hemagglutinating virus of Japan envelope (HVJ-E), a UV-irradiated Sendai virus that almost lost replicating ability, and dacarbazine (DTIC) [15, 17, 23]. HVJ-E and DTIC significantly suppressed B16F10 tumor growth (p = 0.0015 and p = 0.0106, respectively; Fig. 2A). Notably, HVJ-E and DTIC-suppressed tumor growth did not decrease the corrected barcode number in B16F10 tumors, compared to the PBS-treated group (p > 0.77, Fig. 2B). We did not find the enrichment of a specific barcode across HVJ-E and DTIC-treated tumors (Fig. 2C). We obtained the top 10 barcodes as described in Fig. 1. DTIC tended to affect the composition of the top 10 barcodes more than PBS and HVJ-E (Fig. 2D). In contrast, the two anticancer drugs were similar to PBS in affecting the composition of all barcodes (Fig. 2D). These results suggest that B16F10 cells do not contain pre-determined anticancer drug-resistant cells and that the cells surviving in the tumor are stochastically determined during tumor growth even under anticancer drug treatment.

Fig. 2
figure 2

Stochastic resistance to HVJ-E and DTIC in barcoded B16F10 tumors. A The tumor growth curve of barcoding B16F10 cells in C57BL/6N. HVJ-E (2000 HAU) and PBS were intratumorally injected into the tumors on days 0, 2, and 4. DTIC (100 mg/kg) was intraperitoneally injected on days 0, 2, and 4. p value was calculated by Welch’s t-test with multiple comparisons Bonferroni correction (n = 4 for PBS, n = 3 for HVJ-E, and n = 5 for DTIC). B Box plot of the corrected barcode number by Shannon entropy. P value was calculated by Welch’s t-test with multiple comparisons Bonferroni correction. C The bubble plot of top 10 barcodes from each of the 13 samples (n = 1 for in vitro, n = 4 for PBS, n = 3 for HVJ-E, and n = 5 for DTIC). The total bubble number is 117 in each line since there is a redundancy. Bubble size indicates the proportion of the barcode. D The heatmap of JS divergence calculated using top 10 barcodes in each sample or all barcodes

Full size image

Next, we determined the cellular composition of lung metastases comprising B16F10 cells. We created a B16F10 lung metastasis model by injecting PBS into the primary tumor and intravenously injecting B16F10 cells to examine whether these methods affect the cellular composition of lung metastases. Three injections of PBS into B16F10 primary tumors promoted the formation of lung metastasis. We separately collected the primary tumor and metastases and then analyzed the barcode composition (Fig. 3A). The corrected barcode number detected in the lung metastases was comparable to that detected in the primary tumor (p = 0.077, Fig. 3B). We also did not find enrichment of specific barcodes in lung metastases compared to the primary tumor (Fig. 3C). The intravenous injection of B16F10 cells resulted in lung metastases (Fig. 3D). A diverse composition of barcodes was observed in each lung metastasis, and the number of barcodes was less than that of barcodes in the original cells (p < 0.00076, Fig. 3E). We did not find barcodes enriched in metastases across all mice (Fig. 3F). The cellular composition of lung metastases was highly dependent on the mouse. Metastases in mouse #1 were composed of a few or multiple different cells, whereas those in mouse #4 and #5 comprised ~ 2000 different clones. Mice #2 and #3 showed metastases composed of a few diverse cell types (Fig. 3G). These data indicated the method-dependent cellular composition of B16F10 lung metastases. These results suggested that most B16F10 cells can contribute to metastasis formation.

Fig. 3
figure 3

A method-dependent cellular composition of the lung metastases in B16F10 tumors. A Schematic diagram depicting the establishment of C57BL/6N mice with lung metastases by three intratumor injections of PBS. The primary tumor and lung metastases were separately collected on day 10 after the first PBS injection. B Box plot of the corrected barcode number by Shannon entropy in the PBS-treated mice. PBS sample in Fig. 2B was used for primary lesion data. p value was calculated by Welch’s t-test. C 100% stacked bar plot of the proportions of barcodes in the PBS-treated mice. D Schematic diagram illustrating the establishment of C57BL/6N mice with lung metastases by intravenous injection of B16F10 cells. The lung metastases were collected separately on day 8 after the injection. E Box plot representing the corrected barcode number by Shannon entropy in the B16F10 cells-intravenously-injected mice. P value was calculated by Welch’s t-test with multiple comparisons Bonferroni correction. F The bubble plot of top 10 barcodes from each of the 54 samples (n = 1 for in vitro, n = 4 for the injected cells, n = 9 for mouse #1, n = 10 for mouse #2, n = 10 for mouse #3, n = 10 for mouse #4, and n = 10 for mouse #5) in the B16F10 cells-intravenously-injected mice. The total bubble number is 165 in each line since there is a redundancy. Bubble size indicates the proportion of the barcode. G 100% stacked bar plot of the proportions of barcodes in the B16F10 cells intravenously injected mice

Full size image

To examine the cellular composition of spontaneous lung metastases, we barcoded BL6 and 4T1 cells since they can spontaneously establish lung metastases from the primary tumor. The barcoded BL6 cells contained ~ 3200 types of barcodes. We inoculated 1 ×  106 barcoded BL6 cells onto the backs of C57BL/6N mice (Fig. 4A). The BL6 primary tumors comprised 683 barcodes on average (median, 747); in contrast, the lung metastases comprised 58 barcodes on average (median, 4) (p = 0.0027, Fig. 4B). Most lung metastases were composed of a few clones; however, some metastases contained ~ 1000 clones, suggesting that it caused a bias when the whole lung was analyzed to determine the cellular composition (Fig. 4C). Each metastasis was enriched in different clones in the same mice and across all the BL6-burdened mice (Fig. 4D). The primary tumors contained 114 barcodes on average (107 on median) although we inoculated the same number of barcoded-4T1 cells into BL6 cells (Fig. 4E and 4F). Most of the 4T1 spontaneous lung metastases contained two barcodes on average (median, two), which was significantly less than that in the primary tumor (p = 0.0015, Fig. 4F). All 4T1 lung metastases analyzed contained one or two clones (Fig. 4G). Mice #1 and #2 formed lung metastases from the different clones, whereas mouse #4 developed five metastases from the same clone (Fig. 4H). Moreover, one lung metastasis in mouse #4 and two metastases in mouse #5 were composed of the same clone, suggesting that 4T1 cells contain pre-existing cells with high metastatic ability. These data suggest that most spontaneous lung metastases are composed of a limited number of stochastically selected cells and that some cell lines, such as 4T1, have a rare population, which is consistent with a previous report [40].

Fig. 4
figure 4

The formation of spontaneous lung metastases by a limited number of cell types. A Schematic diagram illustrating the establishment of C57BL/6N mice with the BL6 spontaneous lung metastases. Barcoded BL6 cells were inoculated on the back of C57BL/6N mice. The primary tumor and lung metastases were separately collected on day 27 after the inoculation. B Box plot of the corrected barcode number by Shannon entropy in BL6 primary tumor and lung metastases. p value was calculated by Welch’s t-test. C 100% stacked bar plot of the proportions of barcodes in the BL6-burdened mice. D The bubble plot of top 10 barcodes corrected from each of the 27 samples (n = 1 for in vitro, n = 7 for primary lesion, and n = 19 for lung metastasis) in the BL6-burdened mice. The total bubble number is 166 in each line since there is a redundancy. Bubble size indicates the proportion of the barcode. E Schematic diagram depicting the establishment of BALB/c mice with the 4T1 spontaneous lung metastases. Barcoded 4T1 cells were inoculated on the back of BALB/c mice. The primary tumor and lung metastases were separately collected on day 42 after the inoculation. F Box plot representing the corrected barcode number by Shannon entropy in 4T1 primary tumor and lung metastases. p value was calculated by Welch’s t-test. G 100% stacked bar plot of the proportions of barcodes in 4T1-burdened mice. H The bubble plot of top 10 barcodes corrected from each of the 21 samples (n = 1 for in vitro, n = 5 for primary lesion, and n = 15 for lung metastasis) in 4T1-burdened mice. The total bubble number is 75 in each line since there is a redundancy. Bubble size indicates the proportion of the barcode

Full size image

The DNA barcoding system can identify diverse cell proliferation, anticancer drug-resistant cells, and cell metastasis. However, it cannot trace the lineage of cells with specific gene expression. CSCs are thought to contribute to cancer progression. However, the descendants of CSCs comprise a significant population of cancer cells after anticancer drug treatment and under cell stress conditions is not clear. Hence, we used Cas9 and stgRNA to establish transcription-recording cancer cells [27]. We previously found that DU145 cells (human prostate cancer cell line) require the expression of NANOG, a stem cell-related transcription factor, to resist anticancer drugs and escape immune surveillance [13, 31]. We inserted the Cas9 sequence along with three repeats of mClover3 immediately before the stop codon of stem cell-related transcription factors, including NANOG, POU5F1, and SOX2 (NANOG-, POU5F1-, and SOX2-DU145 cells) (Fig. 5A). DU145 cells with the Cas9 cassette in each gene formed spheres similar to wild-type cells (Fig. 5B). Although we did not detect significant mClover3-positive cells in NANOG-DU145, POU5F1-DU145, and SOX2-DU145 cells by flow cytometer (Fig. 5C), we detected Cas9 protein in control or Docetaxel (DTX)-treated NANOG-DU145, POU5F1-DU145, and SOX2-DU145 cells by western blotting (Fig. 5D). DU145 cells with the Cas9 cassette in each gene expressed transcription factors comparable to those in wild-type cells (Fig. 5D). DTX- treatment did not significantly affect the protein expression of each transcription factor but sphere formation decreased the protein expression (Fig. 5D and E). These results indicate that the Cas9 cassette acts as a surrogate for the expression of each transcription factor in DU145 cells with Cas9 cassette.

Fig. 5
figure 5

Establishing DU145 cells with stem cell-related transcription factors expression-recording system. A Schematic diagram illustrating Cas9 cassette insertion immediately before the stop codon of NANOG, POU5F1, and SOX2. The Cas9 sequence was ligated to the gene with the T2A peptide sequence. The Cas9 cassette included Cas9, three repeated mClover3, delta thymidine kinase (∆TK), and PGK promoter-driven blasticidin S-resistant gene (BSD). Cas9 and mClover3 were separated at the 2A peptide sequence after translation. B The representative images of wild-type and DNA-barcoded Cas9-introduced DU145 cells. C The dot plot of mClover3 expression analyzed by flow cytometer. D The western blots of NANOG, POU5F1 (OCT3/4), SOX2, and Cas9-TY1 (H3, loading control). Proteins obtained from 1 or 2 × 105 cells were loaded in each lane (NT, non-treatment; DTX, docetaxel treatment; Sphere, sphere formation). E Schematic diagram illustrating the collection of DTX-resistant DU145 cells. F The line plot of cell proliferation of DNA-barcoded Cas9-introduced DU145 cells (n = 3). G The line plot of the corrected barcode number in DNA-barcoded Cas9-introduced DU145 cells on days 0, 7, and 14 (n = 3)

Full size image

To determine whether the expression of these transcription factors promotes resistance to anticancer drugs and stress conditions, we traced the expression of transcription factors in each cell during DTX treatment or sphere formation. We labeled the Cas9 cassette-introduced DU145 cells with a lentivirus-encoding DNA barcode and stgRNA. We identified 116,722 barcodes corrected using Shannon entropy and Hamming distance 2. DTX treatment and sphere formation suppressed cell proliferation in all DU145 cells with the Cas9 cassette and DNA barcoding (Fig. 5F). The number of DNA barcodes in each cell line was decreased on DTX treatment compared to that in untreated cells (Fig. 5G). Sphere formation reduced the number of DNA barcodes in NANOG #112 and POU5F1 cells but not in NANOG #128 and SOX2 cells (Fig. 5G). Next, we analyzed the sequence changes that occurred in stgRNA. We confirmed a mutation at position 16 in the stgRNA sequence in DU145 cells with Cas9 compared to the plasmid control (Fig. S1), indicating that DU145 cells with Cas9 and stgRNA work to record transcription factor expression levels in the stgRNA. We used deletion and insertion mutations for the calculation because mismatch mutations were found throughout the stgRNA sequence, suggesting a Cas9-independent error. Thus, we compared deletion and insertion as percent mutations in stgRNA on days 0, 7, and 14 in each cell. The percentage of mutated stgRNA increased in DTX-treated NANOG-DU145 cells (Fig. 6A). In contrast, we detected a slight increase in mutated stgRNA in DTX-treated POU5F1-DU145 cells (Fig. 6B) but not in SOX2-DU145 cells (Fig. 6C). However, sphere formation did not significantly increase the percentage of stgRNA mutations in any of the cells (Fig. 6A–C).

Fig. 6
figure 6

Tracing a progeny of cells expressing stem cell transcription factors, including NANOG, POU5F1, and SOX2, in anticancer drug treatment or sphere formation. A–C Line plot depicting the percent mutation in stgRNA on days 0, 7, and 14 in NANOG-DU145 (A), POU5F1-DU145 (B), and SOX2-DU145 (C) cells. P value was calculated using Welch’s t-test with Bonferroni adjustment (n = 3). D–F Heatmap representing the average score in each cell of top 10 cells with a high number of mutated reads from each sample in NANOG-DU145 (n = 196 for #112; n = 176 for #128) (D), POU5F1-DU145 (n = 195 for #112; n = 203 for # 114) (E), and SOX2-DU145 (n = 152 for #98; n = 198 for #101) (F) cells. G Schema of the composition of cancer cells during tumor progression. H Schema of method- and cell-type-dependent variations in lung metastasis formation. I Schema of the lineage-specific or -stochastic expression of transcription factors. NT non-treatment, DTX docetaxel treatment; and Sphere sphere formation

Full size image

To examine whether the expression of these transcription factors is associated with an increase in the proportion of cells, we computed Spearman's correlations between the mean of mutation size of stgRNA and the number of barcode reads, weighted by the number of barcode reads. We then detected higher and significant values of the correlations, especially in Day14 DTX-treated POU5F1 cells (Spearman’s correlation 0.455 for #112 and 0.433 for #114), compared to NANOG (Spearman’s correlation 0.410 for #112 and 0.388 for #128) and SOX2 (Spearman’s correlation 0.350 for #98 and 0.370 for #101) cells (Fig. 6D–F, Table S1). The results suggest that cells with POU5F1 expression are associated with an increase in proportion compared to NANOG and SOX2 expression. Next, we examined whether a cell with a specific lineage expresses these transcription factors, particularly in the top 10 cells with a high number of mutated reads. POU5F1-DU145 cells included more cells that are susceptible to recurrently acquiring the mutation during DTX treatment or sphere formation (Fig. 6E). Compared to the POU5F1 expression profile, NANOG and SOX2 tended to be more stochastically expressed in DU145 cells (Fig. 6D and F). The cumulative distribution fraction of the mutation in all cells showed no significant difference between samples in each cell, as mutations in stgRNA were not often detected (Fig. S2). We also did not detect significant changes in the cellular composition during DTX treatment or sphere formation (Fig. S3). These results suggest that POU5F1 is expressed in a specific cell population upon distinct stimulation in DU145 cells.

Discussion

Anticancer therapy effect on tumor composition

Tumors are composed of various types of cancer cells. Whether the composition of cancer cells is consistent or increases in heterogeneity during tumor growth remains unclear. Our DNA-barcoded B16F10 cells showed a homogeneous proliferation within 10 days after inoculation of cancer cells and subsequently switched to a heterogeneous proliferation state 14 days after inoculation and decreased the number of cell variations (Fig. 6G). These results suggest three possibilities: (1) Rapid tumor growth induces hypoxia, resulting in stochastic cell death at a later stage of tumor growth; (2) cell–cell competition includes stochastic cell death in a loser cell since there is limited space in the mouse skin; and (3) Immune cells begin to eliminate cancer cells at a later stage of tumor growth.

Although DTX and HVJ-E treatments significantly suppressed tumor growth, neither decreased the barcode number in any tumor (Fig. 2). The remaining cancer cell profile remains controversial; several studies suggest that chemotherapy eradicates most cancer clones [2, 14, 16, 19, 30, 40]. Taken together, our findings suggest that 4T1 cells have the ability to form cell populations with different properties and comprise elite cells with metastatic capacity.

CSC properties in DU145 human prostate cancer cell line

WE traced the progeny of cells expressing stem cell transcription factors, including NANOG, POU5F1, and SOX2, during anticancer drug treatment or sphere formation. We observed that cells retrospectively expressing NANOG were increased by DTX treatment at some time point, but those retrospectively expressing POU5F1 and SOX2 were not (Fig. 6I). However, DTX treatment or sphere formation induced the expression of POU5F1 but not NANOG and SOX2 in identical cells. Notably, distinct cells showed increased POU5F1 expression in response to DTX treatment or sphere formation, suggesting that POU5F1 is expressed in a scarce population and that the population includes various cells that respond differently to external stimuli. POU5F1 expression is controlled by several transcription factors, such as HIF1 and MYC, in cancer cells. In renal cell carcinoma (RCC), POU5F1 expression is activated by a HIF1-pathway-responsive promoter in the long terminal repeat (LTR) element that is located upstream of POU5F1 [34]. DNA methylation at the MYC-binding elements in the POU5F1 locus is involved in regulating POU5F1 expression in cancer cells [5]. These suggest that epigenetic status at the POU5F1-regulating element region is critical for the regulation of POU5F1 expression. DTX- or sphere-responding POU5F1-DU145 cells might have a distinct epigenetic profile in the POU5F1-regulating element region. However, DTX- or sphere-responding POU5F1-DU145 clones are 0.52% and 0.37% for #112 and 0.82% and 0.37% for #114 in DU145 cells, meaning it is challenging to analyze a property of the rare population even using current single-cell technologies. NANOG expression is regulated by several transcription factors, such as p53, STAT3, Gli1/2, and HIF [8]. SOX2 expression is regulated by SOX4, OCT4, STAT3, E2F, DNMT, and FoxO1 [Cell barcoding

B16F10 (2 × 104 cells/mL), 4T1 (5 × 104 cells/mL), B16BL6 (5 × 104 cells/mL), and DU145-Cas9 knock-in cells (1 × 105 cells/mL) were seeded into 10 or 15 cm culture dishes and incubated for 24 h. Cells were then infected with the barcode library at an multiplicity of infection (MOI) of < 0.3 to be labeled with a single barcode per cell. The medium was replaced with fresh medium 24 h after infection, and the cells were incubated further for 24 h. MOI was assessed by the proportion of blue fluorescent protein (BFP)-expressing cells using flow cytometry (BD FACS Canto II). The cells were maintained in the presence of puromycin for 7 days to remove uninfected cells 48 h after infection.

Sphere-forming culture

The spheres of barcoded-DU145-Cas9 knock-in cells were formed as previously described [31]. Briefly, each cell line (1 × 105 cells) was plated on EZ-Bind Shut™ SP 100 mm Dish (IWAKI) or Prime Surface 60 mm dish (Sumitomo Bakelite) and cultured in DMEM/F12 (Gibco, NY, USA) supplemented with B27 (Gibco), 4 μg/mL insulin (Sigma, MO, USA), 20 ng/mL epithelial growth factor (Gibco), and 20 ng/mL basic fibroblast growth factor (ORF, Kopavogur, Iceland) for 7, 10, or 14 days at 37 °C and 5% CO2.

In vitro treatment with DTX

DU145 Cas9 knock-in cells were seeded into 6-well plates at a density of 1 × 105 cells/well on day 0 and incubated for 24 h. Media were subsequently replaced with fresh media containing 4 nM DTX (Tokyo Chemical Industry) and incubated for 48 h. Cells were retained in the absence of DTX after 48 h of incubation and collected on days 7 and 14.

Mice

Six-week-old female C57BL/6N and BALB/c mice were purchased from CLEA Japan and maintained in a temperature-controlled pathogen-free room. All animals were handled according to the approved protocols and guidelines of the Animal Committee of Osaka University (Osaka, Japan).

Preparation of HVJ-E

HVJ (VR-105 parainfluenza Sendai/52 Z strain) was acquired from the American Type Culture Collection (Manassas, VA, USA), amplified by injection into 10–14-day-old embryonated eggs and purified as previously described [26]. The purified live HVJ virus was inactivated by UV irradiation (189 mJ/cm2). The inactivated HVJ was used as HVJ-E.

Tumors with different size

C57BL/6N mice were inoculated intradermally with barcoded-B16F10 cells (5 × 105 cells/mouse). When the tumor diameter reached approximately 5, 10, or 15 mm, the mice were sacrificed and tumors were collected. Cell suspensions from tumors were obtained by mincing and digestion with collagenase type I (FUJIFILM Wako Pure Chemical Corporation), as previously described [26]. Viable cells were counted using trypan blue (Nacalai Tesque). Tumor size was measured using a slide caliper, and tumor volume was calculated using the following formula: tumor volume (mm3 = 0.5 × length (mm × [width (mm)]2)).

Tumor treatment

C57BL/6N mice were inoculated intradermally with barcoded-B16F10 cells (5 × 105 cells/mouse). When the tumor diameter reached 4.3–4.8 mm, PBS (Nacalai Tesque), HVJ, or DTIC (FUJIFILM Wako Pure Chemical Corporation) was administered. PBS (50 μL/shot) and HVJ (2000 HAU/50μL/shot) were injected directly into the tumor thrice every alternate day. DTIC (100 mg/kg) was administered intraperitoneally thrice every alternate day. Ten days after the first administration, the mice were sacrificed and tumors were collected. Cell suspensions from the tumors were obtained by mincing and digestion with collagenase type I, as previously described [26]. The number of viable cells was counted using trypan blue.

Intravenous lung metastasis model

Barcoded-B16F10 cells (5 × 105 cells) were suspended in 200 μL of PBS (Nacalai Tesque) and injected into the tail vein of C57BL/6N mice. Eight days after the injection of cells, the mice were sacrificed, and macroscopically visible lung metastases were collected.

Spontaneous lung metastasis model

To establish the 4T1 lung metastasis model, barcoded-4T1 cells (1 × 106 cells) were implanted subcutaneously into BALB/c mice. The primary tumors and macroscopically visible lung metastases were collected 42 days after tumor inoculation.

To produce the B16BL6 lung metastasis model, barcoded-B16BL6 cells (1 × 106 cells) were implanted subcutaneously in C57BL/6N mice. The primary tumors and macroscopically visible lung metastases were collected 27 days after tumor inoculation.

Cell suspensions from primary tumors were obtained by mincing and digestion with collagenase type I as previously described [26]. The number of viable cells was counted using trypan blue.

Next-generation sequencing of barcode library

Genomic DNA was isolated using the DNeasy Blood & Tissue Kit (Qiagen) or the Quick-DNA™ Miniprep Plus Kit (Zymo Research). The region containing the barcode and stgRNA was amplified using Q5 Hot Start High-Fidelity 2 × Master Mix (New England BioLabs). For the PCR of primary tumor of Barcode-B16F10 cells, genomic DNA isolated from 2 × 106 cells was used. When the number of harvested barcoded-B16F10 cells was less than 2 × 106 cells, all harvested cells were used. All genomic DNA isolated from each lung metastasis was used for the PCR of lung metastasis. For the PCR of the primary tumors of barcoded-4T1 and barcoded-B16BL6 cells, genomic DNA isolated from 4 × 106 cells was used. For the PCR of DU145 cells, half of the harvested genomic DNA was used. The following cycling conditions were applied for PCR: 30 s at 98 °C; 30 cycles of 10 s at 98 °C, 15 s at 61 °C, and 20 s at 72 °C; and a final extension at 72 °C for 2 min. The following primer pairs were used for the first PCR:


gLibrary-Hiseq-50 bp-SE-U1: ACACTCTTTCCCTACACGACGCTCTTCCGATCTTGTGGAAAGGACGAAACA


gLibrary-HiseqX-L1: GACTGGAGTTCAGACGTGTGCTCTTCCGATCTCTAAAGCGCATGCTCCAGAC

The PCR products were purified using Monarch PCR & DNA Cleanup Kit (New England BioLabs) or NucleoSpin Gel and PCR Clean-up Midi (MACHEREY–NAGEL GmbH & Co., KG).

Eight hundred picograms of purified first PCR products was used as templates for the second PCR with NEB Next Q5 Hot Start HiFi PCR Master Mix (New England BioLabs) in a total volume of 50 μL. In the second PCR, each sample was labeled with 6-bp-long index sequences. The PCR cycling conditions were as follows: 30 s at 98 °C; 12 cycles of 10 s at 98 °C and 1 min 15 s at 65 °C; and a final extension for 5 min at 65 °C. The following primer pairs were used for the second PCR:


CRISPR-Library-F: AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGCTCTTCCGATOT


iPCRtag-X-L2: CAAGCAGAAGACGGCATACGAGATNNNNNNGTGACTGGAGTTCAGACGTGTGCTCTTCCGATOT

PCR products were purified using AMPure XP (Beckman Coulter). DNA concentrations of the PCR products were assessed by qPCR using the KOD SYBR qPCR Mix (TOYOBO). Subsequently, PCR products were pooled at equimolar ratios, and sequencing was performed with 30% PhiX on an Illumina HiseqX with 150 bp paired-end reads.

Western blot analysis

Cell lysates were subjected to SDS-PAGE, and the separated proteins were transferred onto Immobilon-P transfer membranes (Merck Millipore, Tokyo, Japan). To detect the proteins, anti-NANOG [Cell Signaling Technology (CST), MA, USA; #4903], anti-OCT4 (CST; #2750), anti-SOX2 (CST; #3579), anti-Ty1 (Sigma-Aldrich, SAB4800032), and anti-histone H3 (CST; 4499) primary antibodies were used. Enhanced chemiluminescence horseradish peroxidase-conjugated donkey antirabbit IgG (GE Healthcare Japan, Tokyo, Japan) was used as the secondary antibody for the detection of NANOG, POU5F1, SOX2, and histone H3. Signals were detected using Chemi-Lumi One (Nacalai Tesque) or Chemi-Lumi One Super (Nacalai Tesque) and an ImageQuant LAS 4000 mini system (GE Healthcare).

Bioinformatic analysis of data

Tools

Awk v4.1.3 (https://www.gnu.org)

Cutadapt v1.9.1 [22]

FASTX-Toolkit 0.0.13 (http://hannonlab.cshl.edu/fastx_toolkit/)

trim_galore v0.6.6 (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/)

bbmap v38.90 (https://www.osti.gov/servlets/purl/1241166)

umi_tools v1.0.1 [35]

R 4.0.4 (https://www.r-project.org)

Rstudio 1.4.1106 (https://www.rstudio.com)

WeightedCorr v2.1 (https://github.com/matthijsz/weightedcorr)

R package

Dplyr v1.0.5 (https://cran.r-project.org/web/packages/dplyr/index.html)

ggplot2 v3.3.3 (https://cran.r-project.org/web/packages/ggplot2/index.html)

tidyverse v1.3.0 (https://cran.r-project.org/web/packages/tidyverse/index.html)

gplots v3.1.1 (https://cran.r-project.org/web/packages/gplots/index.html)

NGS data of barcodes in B16F10, B16BL6, and 4T1 tumors

The adapter sequence was removed from the read2 sequence using cutadapt with the option -e 0.2. Reads with low quality and < 25 bases were removed using fastq_quality_trimmer with the options -Q33 -t 20 -l 25 and -Q33 -q 20 -p 80. Fastq was converted to Fasta using fastq_to_fasta with the option -Q 33. Next, we counted the reads of each barcode variety in each sample, accepting only length-30 barcode sequences. Although not all of them had the exact form of WSN (W: A/T, S: C/G, N: A/T/G/C) × 10, we did not impose restrictions on that point (in fact, exact WSNx10 reads account for only 10% of the total, and limiting to them would destroy essential parts of our data). The error might be caused by the limitation of a company's ability to generate the precise DNA oligo. Reads with unidentified bases were excluded from the analyses. Since each sample included many similar barcodes, due to reading errors or mutations, we merged them if the Hamming distance was ≤ 2. The algorithm is outlined in (a). The resulting dataset was used in the subsequent analyses. This reduced the variation to approximately 10–50%, depending on the origin of the samples.

figure a

DU145 Expression-Records

  1. 1.

    Expression scores

    Low-quality reads were removed from the paired reads using trim_galore with the option –fastqc -j 6 –paired. Paired reads were combined using bbmerge. stgRNA and barcode sequences were extracted using umi_tools with the option –bc-pattern to determine a user-defined barcode pattern. To estimate the cumulative expression levels of the target gene, we adopted mutation scores from the original stg sequences [27]. We scored a mismatch base as 1, a start of deletion or insertion as 2.5, and an extension of them as 0.5. Alignment with the minimum mutation score was computed using the conventional dynamic planning algorithm. For even-score alignments, mutations were preferred in the following order: mismatch, insertion, and deletion. The obtained alignments gave separate scores for mismatches, insertions, and deletions. Among them, the mismatch scores presented rather noisy behaviors in our data; therefore, we employed only insertion and deletion scores as indicators of the expression levels.

  2. 2.

    Valid stg selection

    The stg sequences should ideally be unique in the barcoded-plasmids inserted into DU145 cells. However, the real sequence data reported additional 1386 stg variations, accounting for 5.2% reads in addition to the correct template (94.8% reads). Some of these variations were very different from the template and produced large noise in the stg sequences of the resulting cells. To delete them, we picked up variations with 4.0 or larger mutation scores from the template, including 143 variations counting 215 reads (0.00068% of the total). For each of the 71,361 stg variations that were observed in any of the samples, we computed the mutation scores from each of the 143 variations, and if any of them was smaller than that from the template, the stg variation was regarded as originating from a wrong-stg plasmid and the corresponding reads were simply ignored. In this way, 46,134 were deleted and 25,227 stg variations survived. The ignored reads were 0.3–7.4% in each sample.

  3. 3.

    Barcode selection and merger

    Similar to the tumor data, including the stg sequences, a set of (barcode, stg, and reads) triplets were obtained for each sample. The barcodes in the original plasmids were selected and merged using the same algorithm as for the tumor data. For example, (A, X, M) and (A′, X, N) were merged into (A, X, M + N), whereas (A, X, M) and (A′, Y, N) yielded (A, X, M) and (A, Y, N), when A and A′ were similar barcodes (Hamming distance ≤ 2). For all other samples, we retained only reads with barcodes included in the resulting set. After this process, the barcode variations were reduced to approximately 25% (plasmid), 50% (total), 40% (Day 0), 20–30% (Day 7), and 10–25% (Day 14).

Data processing using R

Further analysis of barcode data was performed using R. To remove barcodes derived from sequence error, barcodes detected in all plasmids, barcoded-B16F10, BL6, and 4T1 cells were analyzed using R package dplyr. The corrected barcode number of each sample was calculated as previously reported and plotted using the R ggplot2 package [4].

Analysis of barcode composition

For the comparison of the barcode composition between samples, the JS divergence was calculated using all the barcodes or the top 10 barcodes of each sample, and a heatmap was generated using the R ggplot2 package. A bubble plot of the top 10 barcodes for each sample was constructed using the R ggplot2 package. A cumulative bar plot of the barcodes was constructed using the R ggplot2 package.

Analysis of stgRNA

Hierarchical clustering heatmaps of mean mutation scores were constructed using the top 10 barcodes with mutated reads using the R gplot package. A heatmap of the mutation rate (deletion, insertion, and mismatch) in each position in stgRNA was constructed using the R ggplot2 package. The cumulative distribution fractions of insertions and deletions in each sample were plotted using the R ggplot2 package.

Statistical analysis

Statistical analysis was performed using R. Welch’s t-test with Bonferroni adjustment was used to compare the two samples. Weighted Spearman correlation was calculated WeightedCorr. Statistical significance was set at p < 0.05.