It has long been theorized that differential response of subclones to therapy is important to the development of tumor drug resistance1,2,3,4. Genomics approaches have now accelerated the ability to study intratumoral population evolution, e.g. as reviewed in5. Still, understanding of the dynamics by which tumor subclonal distributions evolve, and in particular in response to therapy, remains limited. An increasingly common approach has been to sequence fragments of patient tumors from multi-region sampling or primary/recurrence/metastasis comparisons6. However, regional samples taken synchronously from a patient tumor will not distinguish the effect of treatment, and determining why sequences differ among regional samples is difficult because their relative growth histories are unknown. Comparison of recurrences, metastases, and primaries is similarly limited by the lack of information during the many months that often separate clinical samples7,8,9.

Patient-derived xenografts (PDXs) are a model system in which human tumors can be grown in mice for controllable time intervals under specified treatments, providing a solution to these challenges10,11,12. PDXs recapitulate therapy response of patient tumors while also retaining some intratumoral heterogeneity. This makes them a valuable system for studying the evolution of tumor subclonal populations, both as a pool for resistant cells13 and as an ecosystem with potentially important interpopulation interactions14.

Another challenge to understanding tumor heterogeneity is that common measurement protocols such as exome-seq have biases that can influence estimates of allele frequency15 and hence subclone cellularity levels. Cellularity inference methods are also sensitive to copy number16, and computational copy number estimates from sequencing or array data are strongly dependent on model assumptions17. Due to these complexities, prior heterogeneity studies have usually focused on whether samples share mutations, rather than interpreting quantitative cellularity values18,19,20. In patient-derived xenografts, state-of-the-art heterogeneity studies have used exome-seq and copy number array measurements with Dirichlet process subclone inference algorithms to estimate intratumoral composition, with a main qualitative finding being that there are strong subclonal changes at first engraftment21,22. Other groups have measured subclonal evolution in xenografts derived from cell lines14,23, though their heterogeneity differs from patient-derived populations. Still, elucidation of the evolutionary response during therapy remains a critical challenge, and accurate quantification will be important for designing new treatment strategies24.

In this study we comparatively investigate how subclonal populations respond to four distinct treatments using patient-derived xenografts for triple negative breast cancers (TNBCs), an aggressive subtype with a high propensity to progress25. Using xenografts and conditionally reprogrammed progenitor cells derived from two patients, we finely quantify how the chemotherapies cisplatin, docetaxel, doxorubicin, and cyclophosphamide impact cancer cell populations. As a control, we also investigate the effects of spatial and temporal parameters. In total, we perform sequencing on 131 samples, providing a deep interrogation of the behaviors within these patients’ cancers. To resolve measurement uncertainties, we make extensive use of droplet digital PCR, a high precision technology to determine matched mutations and copy number changes. These data allow us to perform, to our knowledge, the most exhaustive deconstructions to date of the evolution of phenotypically-relevant tumor populations across treatments and biological replicates. Finally, we analyze these results to identify intratumoral ecological behaviors that may be targetable by dynamic treatment strategies.


Triple-negative breast cancer xenografts exhibit intratumoral heterogeneity

We established TNBC xenografts from two female breast cancer patients using fragments (5–10 mm3) obtained from the patient tumor at resection through engraftment into Nod scid gamma mice (see Methods). Tumors were grown and further dissected over 3–4 passages to generate multiple xenografts from each tumor for a variety of experiments as summarized in a schematic in Fig. 1A. To investigate the effect of chemotherapy on tumors and their subclonal populations, for each of the two patient tumors we generated cohorts of 8–9 xenografts and treated each with one of cisplatin, docetaxel, doxorubicin, cyclophosphamide, or a vehicle control (Fig. 1B). For both models, tumors continued to grow during doxorubicin and cyclophosphamide treatment and responded to docetaxel. Tumor shrinkage was observed in response to cisplatin for one model (TM00099) while for the other model (TM00096) a low level of growth was observed. We generated conditionally reprogrammed progenitor cell cultures from both xenografts as well. For the TM00099 model, we also performed a retreatment experiment, in which we shrank tumors by treating with cisplatin, allowed a drug holiday, and then treated with cisplatin a second time. For the TM00096 model, as a control to understand tumor evolution in untreated samples we obtained multiple untreated spatiotemporally-separated samples. The samples from these diverse experiments for the TM00099 and TM00096 models were used for a variety of sequencing and analysis approaches to determine the subclonal evolution within each model, which we detail below.

Figure 1
figure 1

Exome-seq quantifications of mutations in triple-negative breast cancer xenografts. (A) Schematic of experiments for xenografts generated from two triple negative breast cancer patients (TM00099 and TM00096) (B) Treatment response curves for cohorts of patient-derived xenografts. For each patient tumor, cohorts of 8 mice each were generated for cisplatin, doxorubicin, cyclophosphamide, docetaxel, and a vehicle control. (C) Heatmap of allele frequencies for point mutations identified in residual tumor samples from treated xenografts. (D) Distribution of BRCA2 mutant allele frequencies for 6 TM00099 xenograft samples as measured by exome-seq cancer panel. The mean and standard deviation of mutant allele frequency are 0.545 ± 0.015. (E) Spatial dissection of the P2 passage of tumor TM00096. (F) Mutations predicted by MuTect to differ among samples in the TM00096 tumor.

As a first attempt to assay subclonal changes, for each model we initially sequenced one residual sample per treatment arm as well as untreated samples using an exome-capture platform covering 358 cancer-associated genes at 400× sequencing depth (see Methods). We computationally identified cancer mutations and their allele frequencies (AFs) after filtering out mouse sequence (Supplementary Fig. S1, see Methods). Most recurrent variants exhibited stable AF across all samples and often at values ~50% or >90%, suggesting truncal mutations or germline polymorphisms (Fig. 1C). These provided a quantification of exome-seq measurement uncertainty, e.g. a variant in BRCA2 in the TM00099 tumor yielded AFs of 0.545 ± 0.015 across 6 samples, compared to the AF = 0.5 heterozygous expectation (Fig. 1D). Some loci exhibited stable AF values at lower levels (e.g. HIST1H1C in the TM00096 tumor exhibited 0.18 < AF < 0.27 across samples) suggesting modulation by copy number (CN) changes, and computational CN estimates qualitatively supported this conclusion (see Methods). Other loci showed sample-specific variation, e.g. an ERBB4 mutation in the TM00099 tumor was confirmed by Sanger sequencing to be private to certain samples (Supplementary Fig. S2). As a comparison for tumor evolution in untreated samples, we also performed panel sequencing for the untreated spatiotemporally-separated samples derived from the TM00096 model. These samples included 6 fragments from a dissected xenograft grown for 118 days (Fig. 1E, Supplementary Fig. S3) and 7 fragments from further passaged TM00096 xenografts (See Methods). We observed AF differences among these samples, indicating drift processes were non-negligible on the 2 month time scale (Fig. 1F). Despite these findings, overall we found that sample-specific quantifications of subclone cellularities were challenging from the exome panel data. This was mainly due to high uncertainties in CN inferences (Supplementary Fig. S4), similar to reports from prior studies26,27.

To resolve these uncertainties in measuring subclones, we used droplet digital PCR sequencing (ddPCR, Fig. 2A) to obtain high resolution AF and CN characterizations for a subset of loci and samples (see Methods). We chose mainly loci with higher variation of allele frequencies across the exome-seq datasets, as these would have greater potential for being markers of subclonal populations, plus a few loci with stable allele frequencies to serve as controls. CN loci were chosen based on the AF loci, based on high read count variation across exome-seq datasets, or based on stable read counts across the exome-seq datasets. In total we performed AF or CN measurements at 23 loci for the TM00099 tumor samples and 18 loci for the TM00096 tumor samples. ddPCR measurements were performed on multiple individuals within the treatment cohorts for the two patient models, as well as on conditionally reprogrammed progenitor cell culture samples. This yielded a total of 1633 ddPCR measurements for either allele frequency or copy number.

Figure 2
figure 2

Assessment of mutations and copy number by droplet digital PCR. (A) ddPCR schematic. (B) Distribution of allele frequencies for the BRCA2 locus of Fig. 1D as measured by ddPCR. Bulk tumor samples exhibit mean and standard deviation 0.500 ± 0.011. Cultured cells exhibit mean and standard deviation 0.478 ± 0.055. (C) Distribution of ddPCR-measured copy number at two loci expected to be diploid in TM00099.

As expected, ddPCR AF estimates were more precise and accurate compared to the panel sequencing. For example, for the heterozygous BRCA2 mutation in the TM00099 samples, ddPCR data across the bulk xenograft samples showed AF = 0.50 ± 0.011, improving on the exome-seq uncertainty (Fig. 2B). Under a simple model in which there is one subclone containing a heterozygous mutant allele at a locus while all other cells are diploid non-mutant homozygous, the cellularity of the mutant subclone will 2*AF, which gives a cellularity uncertainty of 2.2%. For ddPCR measurements obtained from conditionally reprogrammed progenitor cells (CRPCs) derived from the same PDX model, we observed a wider distribution of AFs (0.478 ± 0.055). The deviation of the CRPC samples from AF = 0.5 was driven by outliers with AF = 0.32–0.37, suggesting discrete changes consistent with population bottlenecks during in vitro culturing. CN measurements were robust and absolute. For example, two TM00099 control loci expected to be diploid from the exome-seq data exhibited ddPCR CN close to expectations across 47 samples (EXOC4 CN = 1.92 ± 0.10; AKR1B1 CN = 1.92 ± 0.06. See Fig. 2C). The EXOC4 and AKR1B1 copy numbers were not correlated (r = −0.19, p-val = 0.21), indicating that the deviations were due to measurement uncertainties rather than sample-specific quality issues. CN measurements at separate loci within a gene were also highly correlated (two LRP1B loci: r = 0.82, p-val = 1e-12).

Symbiotic growth of recurrent subclones

ddPCR AF data provided evidence for both culturing-dependent bottlenecks and subclonal mixing in the TM00099 samples. For example, some loci (e.g. BRCA2) exhibited relatively stable AF values across samples while other loci exhibited a broader AF range (e.g. ERBB4) as shown in Fig. 3A. At stable loci, the variance in AF was dominated by culturing bottlenecks–the bulk xenograft samples had BRCA2 AF values centered at 0.5 (black); CRPC cultures grown from bulk were centered at 0.5 but with more outliers (blue); and CRPC colonies grown from single cells showed the most extreme outliers and at quantized values (red). In contrast, at ERBB4 we observed a continuous range of AF values, with some of the lowest values coming from bulk xenograft samples. This continuous behavior suggested that the observed ERBB4 values were the result of mixing of subclones of different genotypes, with any value for the mixing ratio allowed.

Figure 3
figure 3

Comparisons of ddPCR mutant allele frequencies across samples. (A) Comparison of allele frequencies of variants observed at the BRCA2 and ERBB4 loci in TM00099. Black dots are bulk tumor samples. Blue dots are conditionally reprogrammed progenitor cell (CRPC) cultures. Red dots are CRPC cultures established by growing from single cells. (B) Comparison of ERBB4 and LRP1B mutations. (C) (Upper right triangle): Pairwise plots of ddPCR-measured mutant allele frequencies for xenograft tumor samples (black) and CRPC samples (red). (Diagonal): Distribution of AF values for each mutation. (Lower left triangle): Density plots of AF values for pairs of mutations. Pearson correlations for each pair of mutations are also shown.

Comparison of other mutations also provided evidence for recurrent mixing between two subclones. For example, we observed a strong correlation between the AF values of ERBB4 and LRP1B (r = 0.67, p-value = 2.6e-14) with dense occurrence of samples along the correlation line (Fig. 3B). These data can be explained by the presence of a subclone having both ERBB4 and LRP1B mutations and another subclone lacking the two mutations, where each sample’s location along the Fig. 3B diagonal is determined by its ratio of the two subclones. Analysis of the full set of pairwise comparisons supported this conclusion (Fig. 3C). While several variants exhibited stable AFs consistent with truncal status in the tumor phylogeny (BRCA2 ~0.5; PTCH1 0.6–0.7; SAAL1 0.3–0.4; TP53 > 0.97), we observed an additional strong correlation between a mutation in MET and each of ERBB4 and LRP1B (MET vs. ERBB4: r = −0.65, p-val = 2.4e-13; MET vs. LRP1B: r = −0.78, pval = 2.2e-16). The negative sign of the correlation indicated that the subclone that lacks the ERBB4 and LRP1B mutations contains the MET mutation.

Copy number data further supported the conclusion of recurrent mixing between two subclones. We measured 9 loci in the TM00099 tumor, and the majority exhibited variance across samples larger than the ~0.1 CN measurement uncertainty observed for the control loci (Fig. 4A, control loci AKR1B1 and EXOC4 not shown). Similar to the AF data, we observed striking CN correlations between loci. In particular, pairwise comparisons within the group of EPHA3, EPHA5, PTPRD, LRP1B, and ERBB4 and all had Pearson correlations r ≥0.73. We also observed strong anti-correlation between AR and the other loci, all with correlation values r ≤ −0.54. Like the ERBB4 AF data, the CN data covered a smooth range of values, supporting their control by a continuous set of subclone mixing ratios. A heatmap of the correlations among all measured mutations and copy numbers is shown in Fig. 4B, indicating the strong associations between mutations and copy number events in the two subclones. Plots of all pairwise comparisons are given in Supplementary Fig. S5, and raw data are available in Supplementary Table S1. A notable aspect of subclone 1 is its extremely high level of internal homogeneity. Within subclone 1 all pairwise correlations are r > 0.69. In contrast, within subclone 2 MET AF and AR CNV are strongly correlated (r = 0.71) but the other correlations are all ≤0.45. This indicates subclone 1 is more homogeneous than subclone 2, likely due to a recent population bottleneck.

Figure 4
figure 4

Resolution of subclones from ddPCR copy number and mutation measurements. (A) Pairwise plots of copy number variations in TM00099. Upper right triangle, diagonal, and lower left triangle are analogous to Fig. 3C. (B) Pearson correlations among all ddPCR-measured mutant AFs and CNVs. The data comprise 47 TM00099 xenograft samples with data for all variables. (C) Comparison of LRP1B allele frequency and LRP1B copy number.

The large number of samples allowed us to identify the genotypes of the subclones despite ubiquitous population mixing in each sample. We determined genotypes by extrapolating from a linear fit of AF values and absolute copy numbers, as illustrated in Fig. 4C for LRP1B AF vs. CN (r = 0.85, p-val = 5.6e-14). These data lie close to a line connecting the points (CN = 2, AF = 0) and (CN = 4, AF = 0.25), indicating that subclone 1 has 1 mutant copy of LRP1B out of 4 total copies while subclone 2 has 2 wildtype copies of the LRP1B locus. We confirmed these inferences by measuring CN and AF in one sample expected to contain mostly subclone 1 (LRP1B AF = 0.25, sample 1 in Supplementary Table S1). For LRP1B we observed copy number number 3.6, approximately consistent with all cells in the sample having 1 mutant copy out of 4 alleles. Other loci also appeared to be homogeneous in this sample as their CN and AF values were plausibly consistent with all cells sharing a genotype, e.g. for ERBB4 (AF = 0.33 and CN = 2.9: consistent with 1 of 3 alleles being mutant), MET (AF = 0.48, CN = 2.00: 1 of 2 copies are mutant), PTCH1 (AF = 0.66, CN = 2.86: 2 of 3 copies are mutant), TP53 (AF = 0.99, CN = 3.01: 3 of 3 copies are mutant), and BRCA2 (AF = 0.50, CN = 4.01: 2 of 4 copies are mutant). This consistency further supported the conclusion of a recent population bottleneck for subclone 1.

For comparison, we applied PyClone16, a widely used subclone inference method to determine if it could also identify the subclones. Although we observed some similarity in the subclones found by PyClone and our subclones 1 and 2, PyClone clustered markers into different subclones for the exome-seq vs. the ddPCR data, demonstrating that generic subclone inference methods are still sensitive to data abundance and quality (Supplementary Fig. S6). Moreover, a limitation of PyClone and related methods is that they rely on implicit assumptions about cancer evolutionary processes that are still poorly understood. For example, PyClone requires the user to specify copy number genotypes of subclones as part of its input, but these are inherently unknown before subclone decomposition. PyClone suggests using bulk WGS or array data to estimate copy number genotypes, but such methods are designed for homogeneous samples and will fail if there is substantial subclonal heterogeneity within individual samples, which is in fact what we observed in the TM00099 samples. Our concerns about such assumptions of the generic methods is why we chose to measure both AF and CN at high precision across large numbers of samples, yielding datasets suitable for direct interpretation.

Recurrent subclones are differentially selected by treatment

We next investigated whether the two subclones might differ in resistance to treatment. For this we analyzed ddPCR data from PDX post-treatment residuals. These residual samples included those used for the initial exome-seq as well as samples from a second set of xenograft cohorts treated for three weeks with cisplatin, cyclophosphamide, docetaxel, doxorubicin, or vehicle. Additionally, 6 mice that showed good response in the second cisplatin cohort were allowed to regrow without drug until they reached the pre-treatment starting volume of 150 mm3 (10–30 days) and were then retreated with a subsequent cycle of cisplatin before sequencing. ddPCR measurements were obtained for multiple xenografts from each cohort, allowing us investigate the relationship of tumor size and subclonal prevalence on a mouse-to-mouse basis.

ddPCR measurements indicated that the efficacy of cisplatin treatment was related to the post-treatment subclone distribution (Fig. 5A). For the samples treated with cisplatin, LRP1B copy number, which increases with subclone 1, exhibited a strong correlation with tumor volume change after treatment (r = 0.77, p-val = 0.0008). This correlation was visible though statistically marginal when restricted to the samples given only one cycle of treatment (Fig. 5A light green dots, r = 0.51, p-val = 0.19). The correlation was stronger for the samples allowed to regrow and be retreated for a second cycle (Fig. 5A dark green dots, r = 0.92, p-val = 0.0092). The sample with the largest volume (Fig. 5A green asterisk) was one subjected to one cycle of cisplatin and then regrown without further treatment, and it had the second highest value of LRP1B copy number. These data indicate that subclone 1 is more susceptible to cisplatin than subclone 2 but that in the absence of treatment subclone 1 grows faster. Neither LRP1B copy number nor tumor volume was correlated with the time at which the tumor was excised (Fig. 5B, LRP1B: r = −0.17, p-val = 0.54; Fig. 5C %volume change: r = −0.25, p-val = 0.36), i.e. tumor volume and relative subclone prevalence were each more impacted by individual-specific variations than by the total time or number of treatment cycles.

Figure 5
figure 5

Relationship of intratumoral heterogeneity and tumor growth. (A) Comparison of LRP1B copy number and tumor volume after cisplatin retreatment. The y-axis indicates % volume change in the tumor at sample collection relative to the start of treatment. LRP1B copy number, a proxy for the relative cellularity of the two major subclones, is correlated with tumor volume change (r = 0.77, p-val = 0.0008). (B) Comparison of time vs. relative cellularity as quantified by LRP1B copy number. (C) Comparison of time vs. tumor volume during cisplatin treatment studies. (D) Comparison of tumor volume and LRP1B copy number for all treated TM00099 samples.

Comparison of treatments indicated that cisplatin was unique in selecting for subclone 2. The non-cisplatin cohorts exhibited larger LRP1B copy number values (ranging from 2.9–4.0) than the cisplatin-treated xenografts (2.0–3.5) as shown in Fig. 5D, providing evidence that cisplatin treatment selected for subclones with low LRP1B CN. For the non-cisplatin cohorts, correlations of subclone ratio vs. tumor volume were weaker. Cyclophosphamide-treated xenografts had a marginal correlation between volume change and LRP1B CN (r = 0.71, p-val = 0.05), and none of the other treatment cohorts exhibited significant correlations (doxorubicin r = −0.26, p-val = 0.53; docetaxel: r = −0.54, p-val = 0.20; vehicle: r = −0.17, p-val = 0.68). Thus the two subclones have differential growth specific to cisplatin therapy, with a possible minor differential susceptibility to cyclophosphamide.

Evolutionary bottlenecks occur commonly during growth

Although the dominant effect in the TM00099 data was the recurrent mixing of subclones 1 and 2, we also observed a minority of samples with outlier values not explainable by such mixing. Such outliers can only arise if population bottlenecks macroscopically increase the cellularity of a subclone with an alternate genotype. We identified outlier samples by comparing AF values to either a constant (BRCA2, PTCH1, TP53, CSMD3) or to the range of values possible from mixing of the subclones (ERBB4, MET, SAAL1–LRP1B AF was used as the predictive variable) (Fig. 6A; see Methods). These samples were required to be outliers by at least 3 times the AF measurement uncertainty for the BRCA2, PTCH1, TP53, and CSMD3 comparisons (p < 4.4e-5) or 7 times the measurement uncertainty for the comparisons to possible mixtures of the subclones (p < 4.5e-19). 18 of the 101 samples exhibited outlier behavior along at least one the 7 AF dimensions, indicating that bottlenecks affect a significant minority of samples. Most samples had outlier behavior in only one dimension, though some samples deviated in up to 3 dimensions. As expected from culturing bottlenecks, a higher fraction of CRPC samples (25%) exhibited outlier behavior. However a non-negligible fraction of xenograft samples (12%) were also outliers. Different outlier samples were biased toward different cell populations, as we observed 12 distinct AF deviation patterns among the outliers (Fig. 6B). Thus these bottlenecks were not explainable by shared selection pressures for a small set of subclones across different samples.

Figure 6
figure 6

Samples with outlier allele frequencies. (A) Pairwise plots AF of 7 mutations vs. LRP1B AF. Samples with outlier behavior are denoted by red, green, and blue, where colors indicate the number of dimensions (i.e. AF values) in which each sample shows outlier behavior. (B) Samples with outlier behavior and the dimensions in which they deviate from expectations. Red indicates an excess AF value and blue indicates a deficient AF value. CRPC samples are labeled as “cells” and other labels indicate xenograft treatments.

To better understand the therapy-induced selection in the TM00099 model, we sought to compare how subclone distributions evolve under reduced selection pressures. However, the stable symbiosis of the two subclones suggested that there might be additional ecological constraints that impact TM00099 subclone evolution. As a simpler comparison we therefore interrogated the TM00096 tumor, for which our cancer panel data did not yield markers of commonly recurring subclones during treatment (Supplementary Fig. S7 and Table S1). For this analysis, we compared 8 untreated samples from the spatial dissection and propagation experiments (Fig. 1D and Supplementary Fig. S3).

Analysis of ddPCR data at loci in the TM00096 samples revealed that spatiotemporally proximate sections have more similar AF values, consistent with expectations from spatially controlled genetic drift (Fig. 7A). For example, we measured AF values for mutations in TTN, ERBB4, and DNMT3B, which were detectable at low levels in many samples. The presence of the mutations in multiple samples indicates that they arose in an early subclone which then spread spatially, as opposed to independent evolution in each sample which would be extremely unlikely (Supplementary Fig. S3). These mutations each spanned a range of AF values (TTN: 0–16.4%; ERBB4: 0–19.4%; DNMT3B: 0–2.3%) broader than the ddPCR measurement uncertainty (1.1%), and values were more similar for proximate samples. However, we also observed substantial differences in cellularity across samples inconsistent with Wright-Fisher drift (see Discussion). For example, a derived P3 sample contained a subclone with a TTN mutation in ~30% of its cells (assuming heterozygous diploidy; see Table S1), even though this subclone was undetectable in 4 of the other samples. Another P3 sample contained a subclone with an ERBB4 mutation making up ~40% of its cells, even though AF was <3% in all other samples. Thus even in the absence of therapy we observed bottlenecks sufficient to replace more than one-third of the cells in a 20 mg fragment during a 2 month period.

Figure 7
figure 7

Spatiotemporal dependence of intratumoral heterogeneity. (A) Allele frequencies for mutations in TM00096 P3 tumors grown from the P2 positions A1, A5, C5, D6, F4, and F6 (left) and a P3 grown from the P2 B4 positions dissected into 2 further samples P3 C4 and F5 (right). (B) Heatmap of correlations for ddPCR measurements of TM00096 spatiotemporal samples.

The evolutionary patterns among spatial samples were qualitatively different from those in the TM00099 treated samples. For the spatially dissected samples, no significant correlations were observed (p < 0.05) between any pair of measured AFs or CNs (Fig. 7B; see also Supplementary Fig. S7). In contrast, within each of subclones 1 and 2 in TM00099, every pair of markers was correlated with p < 0.05. The lack of correlations within the spatial samples was consistent with the presence of multiple subclones whose cellularity levels varied independently.


Using extensive xenograft and cell culture studies with high resolution ddPCR, we have presented to our knowledge the most exhaustive treatment-guided tumor subclonality deconstructions that have been performed in any tumor type to date. In addition, our use of xenografts allowed us to distinguish stromal from tumor sequence based on mouse vs. human origin, removing uncertainties about tumor purity that affect direct clinical samples. Importantly, ddPCR improved subclone quantification relative to the uncertainties that have been frequently described for exome sequencing approaches28, allowing us to identify complex population evolution including symbiosis between recurrent subclones, frequent population bottlenecks, and differential drug sensitivity. These studies demonstrate the value of precise copy number measurements when analyzing intratumoral heterogeneity from bulk measurements, providing a caveat to prior analyses that have relied on mutant allele frequencies alone29.

A major unexpected finding from the TM00099 tumor was that there are two stable populations that coexist, and we were curious what genomic differences might explain their differential cisplatin sensitivity. We analyzed if any mutant AFs or CN aberrations differed between treatment cohorts, and we found that all cohort pairs with a significant AF difference (Welch’s t-test, α = 0.05, Supplementary Fig. S8) involved the cisplatin cohort, though none were significant after multiple testing (Benjamani-Hochberg) correction. However for CN comparisons, there were 9 significant differences after multiple testing correction, and 8 of 9 involved the cisplatin cohort. The two loci with the strongest differences between vehicle and cisplatin were LRP1B CN (average LRP1B CN across 2 measurements, p = 0.00068) and ERBB4 CN (p = 0.00038). Direct comparisons with tumor volume also suggested that CN drives the differential sensitivity of the subclones (Supplementary Table S2). Tumor volume was strongly correlated with the CN amplifications of subclone 1–ERBB4, LRP1B, PTPRD, and EPHA5 CN all exhibited r [0.76,0.79] and p-value [4.8e-4, 9.4e-4]. In contrast, the strongest correlation between AF and volume was for PTCH1 AF (r = −0.79, p-value = 4.4e-4), though PTCH1 AF varied little across samples (AF range = 0.65–0.7) making it unlikely to be functionally important. Although our measurements cover only a limited set of the genome, these results suggest that copy number aberrations are more likely to distinguish the drug sensitivity of the subclones than point mutations, which may be related to the tumor’s positive tandem duplicator phenotype status30.

The large number of samples generated from diverse conditions allowed us to trace the population history within these tumors at unprecedented resolution, analogous to reconstructions of human population history. As illustrated by the correlations in Fig. 4B, subclone 1 is internally more homogeneous than subclone 2, providing evidence that it is a more recently derived subclone. Thus the history of this tumor is similar to the human “Out-of-Africa” event31–one subclone (subclone 1) emerged from a population bottleneck in the original population (subclone 2) and then expanded to make up a majority of the population. Interestingly, the absolute quantification of copy number in ddPCR provided additional evidence for this conclusion. As shown in Fig. 4C, LRP1B has approximately 4 copies per cell in subclone 1 and approximately 2 copies per cell in subclone 2, showing that subclone 1 and subclone 2 are each relatively homogeneous at this locus. However, other loci behave differently. For ERBB4, a linear fit indicates that there are 2.8 copies on average in subclone 1 and 1.3 copies in subclone 2 (see Methods). Thus the subclone 1 value is slightly closer to an integer than the subclone 2 value. Because CN values must be integers in each individual cell, this suggests that subclone 2 is more heterogeneous than subclone 1, consistent with a population bottleneck that led to subclone 1. This effect, while modest, indicates how absolute CN data can reveal heterogeneity in a manner not accessible by relative CN measurements.

We also observed population bottlenecks that caused outlier subclonal distributions in 18% of all samples, including bottlenecks that distinguished spatially separated samples within the same xenograft. Surprisingly, even without therapy, small pre-existing cell populations were able to take over one-third of the cells of a ~20 mg region of a tumor in a period of only 2 months. It remains to be seen how much of these effects are due to selection versus drift processes such as spatial dispersal and neutral cell turnover. While the irregularity of the outlier populations supports the importance of drift32, these allele frequency changes are too large to be explainable by neutral Wright-Fisher drift at the population sizes within xenograft fragments (~106 in a 1 mm3 fragment used for sequencing) and given the length of time (a few dozen cell replications). We speculate that these bottlenecks may be due to additional selection of subclones based on their epigenetics, and that the outlier populations may have more coherent epigenetic characteristics.

Finally, we consider the clinical importance of determining the population histories within these tumors (Fig. 8). In the TM00099 tumor, subclone 1 dominated the non-cisplatin xenografts (70% of cells on average, as interpolated from the LRP1B CN values) while the cisplatin-treated xenografts had reduced levels of subclone 1 (43% of cells on average). Thus therapy was more effective on the more recently-derived, faster-growing subclone. However, even during tumor shrinkage the more resistant population was difficult to kill completely. This corresponds to an ecosystem in which time-dependent treatment strategies such as metronomic therapy33 or adaptive therapy34 to extend the time of non-lethal tumor burden are suitable. For example, in adaptive therapy, treatment dosing is chosen based on dynamic changes in tumor volume, and this approach preferentially kills fast-growing sensitive populations while limiting takeover of the tumor by slow-growing resistant populations. Although ecosystems susceptible to adaptive therapy have been synthetically constructed35,36, to our knowledge ours is the first study to determine the genetics of such an ecosystem arising spontaneously in vivo.

Figure 8
figure 8

Schematic of evolutionary dynamics in the TM00099 tumor. Initially, the tumor contains a diverse ancestral population (blue), where diversity is indicated visually by the variability in cell sizes. A new mutation arises leading to a new population of cells (red). In the absence of treatment, this new population grows more rapidly, but it is also more susceptible to cisplatin treatment. Because the new population has arisen more recently, it is internally more homogeneous, indicated by the consistent sizes of the red cells. The relative growth behavior of the two populations makes the tumor susceptible to adaptive therapy cycles. The red and blue populations correspond to subclones 1 and 2, respectively, in Fig. 4.

Recent reports have suggested that metronomic topotecan/pazopanib therapy is applicable to triple negative breast cancer37 and other studies have recently demonstrated prostate cancer patients treatable by adaptive abiraterone therapy38, making determination of the prevalence and genomics of such ecosystems across cancer types a vital area for further inquiry. While there are non-negligible uncertainties in subclonal estimation for exome-seq-focused datasets such as TCGA39, careful whole genome-based studies have shown that breast cancers often contain subclones with a high cellularity level40. Thus it is possible that many breast and other cancers have ecosystems similar to ours, i.e. they contain high cellularity subclones whose relative proportions can be targeted for treatment. We note that, although we have focused our analysis on the cisplatin cohort from one patient, there may be mutations outside the exome panel that could distinguish balancing between common subclones in our other cohorts as well. Looking forward, we have seen that multi-sample data are essential for accurately distinguishing subclones, and circulating tumor DNA measurements may provide an approach to obtain such information for patients without the extensive xenografting efforts we have described here. ctDNA content has been suggested as a way to quantify tumor burden, but we speculate that quantitative subclonal decomposition from ctDNA may eventually have greater treatment relevance, as the observed ratio of subclones could be used in adaptive therapy mathematical models that stipulate treatment dose from tumor size34.

An interesting last point is that the two subclones from the TM00099 tumor are likely to have symbiotic interactions with one another. We conclude this because all samples contained mixtures of the same major subclones, which is only possible if the two subclones are granularly intermixed. Moreover, the correlation between tumor volume and mixing ratio indicates that each xenograft must be relatively spatially homogeneous with respect to these subclones, otherwise randomness in the spatial locations of sequenced samples would have distorted the correlation. These observations suggest the presence of symbiotic interactions that synchronize the local mixing ratio of the two subclones. Previously, Marusyk et al.14 demonstrated ecological commensalism in a xenograft system generated from cell lines, making it plausible that there are commensal or mutualistic interactions in our system. The ratio of subclones does vary somewhat among the non-cisplatin samples (Fig. 5C), but their composition is systematically different from the cisplatin samples, and both subclones continue to be present at non-negligible levels in the vast majority of samples. In the future, understanding of the dynamics and symbioses among naturally occurring subclones will benefit from finer mathematical modeling of selection and neutral evolution41,42. Such studies may reveal other ecological relationships that can be exploited to improve treatment strategies for cancer.


Tumor Samples

For model TM00096, the tumor sample was obtained from a 45-year-old female patient with triple negative breast cancer subtype determined by immunohistochemistry. For this patient, anthracyclines and taxanes had been used as chemotherapeutic agents prior to surgery. The sample was obtained from a metastasis of the breast cancer to the patient’s lung. To establish the PDX, a fragment of the patient’s tumor was engrafted into a Nod scid gamma (NSG) mouse until stable growth, forming the P0 passage. A subsequent set of P1 mice xenografts were formed by dissecting the derived P0 tumor and reimplanting these into new mice. Fragments of a grown P1 passage were then used to begin the experiment in the P2 passage. During dissection of the P2, all fragments were weighed. For model TM00099, a similar procedure was used to establish xenografts. The tumor sample was obtained from a 44-year-old female patient with triple negative breast cancer diagnosed as invasive ductal carcinoma. The patient was treated with neoadjuvant taxol prior to resection, and the sample used for xenografting was obtained from the primary malignancy. Informed consent was obtained from all patients. The protocol to receive coded clinical information and cancer specimens from providing institutions was approved by The Jackson Laboratory IRB (Protocol #121200011). This protocol requires that providing institutions obtain written informed consent within separate IRB-approved protocols before any specimens or clinical information are provided for PDX development.

Xenograft spatial dissection and temporal propagation

The patient-derived xenograft triple negative breast cancer tumor (JAX PDX resource model TM00096) from the P1 passage was dissected and a fragment (5–10 mm3) was implanted subcutaneously by trocar in the right flank of a female NOD.Cg-Prkdcscid Il2rgtm1Wjl/SzJ (NOD scid gamma or, NSG) mouse. This P2 tumor was then allowed to grow for 118 days at which point its volume was >2000 mm3, after which it was spatially dissected into a 6 × 6 grid. Grid locations were annotated alphanumerically (A1 … F6). 6 of the P2 grid fragments (A2, A6, B5, D3, F1, F5) were sequenced via cancer panel sequencing. The total weight of the 36 P2 fragments was 883 mg. Fragments specified in Fig. 1E were implanted into P3 mice. Implanted fragment weights ranged from 10.3–45.6 mg. After 58 days, four of these P3 xenografts (each with tumor volume >1,000 mm3) were sacrificed. These were the xenografts generated from original positions B4, D6, F4, and F6 in the P2 tumor. Xenografts generated from the D6, F4, and F6 fragments were sequenced. We further dissected into a grid the P3 tumor that had been generated from the B4 fragment of the P2. Two spatial fragments corresponding to positions C4 and P5 in the P3 tumor were analyzed and sequenced. In addition, 3 other P3 xenografts were harvested at 72 days after engraftment. These were the P3 xenografts generated from fragments at positions A1, A5, and C5 in the P2 tumor.

Xenograft treatment experiments

TNBC PDX models were established at The Jackson Laboratory In Vivo Services campus and tested for cisplatin, docetaxel, doxorubicin, and cyclophosphamide sensitivity. Briefly, patient tumor material acquired from biopsy or surgical resection was implanted subcutaneously into the flank of NOD-scid IL-2r gamma-chain null female mice (8–10 weeks old). Models were considered established when log-phase growth in a second passage was evident. Individual tumor-bearing mice were randomized into treatment cohorts of at least six animals each on an accrual basis when tumors reached a volume of 150 mm3 (day 0), at which point each tumor model was assessed for its response to treatment. The treatments used for each drug were: Doxorubicin – once weekly, 2 mg/kg iv, for 3 weeks; Cyclophosphamide – once weekly, 40 mg/kg iv, for 3 weeks; Docetaxel – once weekly, 10 mg/kg iv, for 3 weeks; Cisplatin – once weekly, 2 mg/kg iv, for 3 weeks. Changes in tumor volumes were measured using digital calipers for 4 weeks from the beginning of the treatment or until tumor volumes reached the 1,500 mm3 end point. For the retreatment study, TM00099 and TM00096 mice cohorts were established and subjected to a first cycle of treatment as described. However, instead of being harvested at the end of the treatment, tumors were maintained in their host mice and allowed a drug-holiday period until they grew back to their original pre-treatment volume (~150 mm3), at which point a second cycle of treatment was administered following the same protocol described above.

Conditionally Reprogrammed Progenitor Cells

We cultured Conditionally Reprogrammed Progenitor Cells (CRPCs) in vitro from PDX tumors adapting an approach published by Yuan et al.43. CRPC cultures are heterogeneous, and single cell clonal lines can be derived from them. Tumors from PDX models were dissociated into single cells with collagenase and cultured on irradiated 3T3J2 mouse feeder cells in media to form isolated epithelial colonies. Individual colonies were harvested and expanded as single cell clonal lines (red dots in Fig. 3A,B). We also cultured some PDX tumor samples analogously on feeder cells but without separating into single cells, yielding bulk samples propagated in the in vitro environment. These correspond to the culture samples (blue dots in Fig. 3A,B).

Targeted exome sequencing panel

The Jackson Laboratory cancer panel is a targeted exome sequencing panel consisting of 358 genes. Inclusion of a gene on the panel was based on a three-tier criteria ladder, which consists of genes targeted by a drug (tier 1), genes having only causality in cancer but not drug targeted (tier 2), and genes only implicated in cancer as determined by Jackson Laboratory researchers (tier 3). Ten-percent (41/358) of the genes on the panel are identified as causal, 53% (172/358) are targeted by drugs and/or causal, and the remaining 37% (145/358) are research grade. Genes in tiers one and two were further annotated to signaling pathways involved in efficacy of targeted therapies using the Rat Genome Database Pathway Ontology. The highest percentages of genes on the panel reside in EGFR, MAPK, mTOR, and p53 signaling pathways. The full list of genes and loci is available in Supplementary Table S3.

Mutation calls from cancer panel

FASTQ files were analyzed by a custom pipeline that includes tools to assess read quality, alignment, and variant calling. Human and mouse reads were separated using Xenome44, after which only human reads were considered for further analysis. Human reads were quality trimmed and filtered using the NGS QC toolkit to remove reads with >30% low quality bases (Q < 30)45. The resulting reads were aligned to the human reference genome (hg19) using BWA46. Duplicates were removed using Picard ( and the resulting alignments were processed to minimize artifacts by realigning around indels and recalibrating the base quality score using the Genome Analysis Tool Kit (GATK) software suite47. Single nucleotide variants (SNVs) were called using GATK with HaplotypeCaller in GVCF mode47, and stored to variant call format (VCF)48 files. Variants of low quality or allele frequency <0.01 were removed, while variants with low sequencing depth (<140) were flagged. Genes belonging to the Mucin gene family (MUC16, MUC17, MUC4, MUC5B) were removed from the heterogeneity analysis, since they contain variable number tandem repeats making them polymorphic even in normal genomes. Common dbSNP variants were removed to filter germline effects. The dbSNP definition of a common variant is one polymorphic in at least one population in the 1000 Genomes project49, having minor allele frequency >0.01, and present in at least two 1000 Genomes samples50.

For the spatial and temporal analyses of cancer panel data, we performed mutation calling with MuTect51 using the P0 (the first stable xenograft after implantation from the patient tumor fragment) as control. MuTect using the vehicle as control gave very similar results (data not shown). For these analyses, the MuTect pipeline was chosen over the GATK pipeline because it provided improved sensitivity for detecting low AF mutations that differed between samples. This was to maximize the chances that the chosen markers would distinguish intratumoral evolution. The ddPCR findings subsequently confirmed this success of this approach. When the GATK pipeline was used, most identified mutations were truncal, as reflected by their consistent presence in all samples. Truncal mutations provide no information about subclonality, since they are equivalent in all subclones.

CNV calls from cancer panel

Exon level copy number profiles from the cancer panel were estimated using CONTRA52. The statistical significance of segment copy number levels were recalibrated using ConReg-R to improve the false discovery rate estimates53. We established a “normal” baseline comprised of 3 unrelated HapMap samples: NA12877, NA12878, and NA18507). These HapMap samples were sequenced as controls to minimize depth and sequence-related biases from cancer panel sequencing. The stability p-value of each exon-level copy number was defined based on the standard deviations of 20 adjusted log-ratio values obtained from 20 bins for each exon. When an exon was too small to divide it into 20 bins, a smaller number was used based on the constraints: minimum read depth = 50, minimum bin width = 20 nt. Deletion or amplification was determined based on the behavior of all bins for each exon. Additionally for the TM00096 sample we compared to copy number computed using ASCAT54 from Affymetrix SNP6.0 array data of the tumor P0 passage. Similar to the CONTRA analysis, we observed that subclonal copy number inferences under this method were parameter-sensitive.

Detection of copy number variation and gene mutations by ddPCR

The QX200 Droplet Digital PCR (ddPCR) System (Bio-Rad Laboratories) was used for detection of copy number variation (CNV) and/or gene mutations. Human gene-specific primers and probes were designed based on the human genome reference hg19 (GRCh37). In this study, CUL1 was used as a reference gene for CNV assays since its absolute copy numbers were close to 2.0 for models TM00096 and TM00099. Absolute copy numbers were estimated by Affymetrix SNP6.0 array for P0 samples of TM00096 and TM00099. The loci for ddPCR were chosen based on variants computed from GATK and MuTect on the exome-seq data. Loci with higher variation of allele frequencies across samples had higher priority since we were interested in drug effects on heterogenous cell populations. Some loci with stable mutation allele frequencies of approximately 0.5 or 1.0 were also selected as candidates to verify truncal heterozygous or homozygous mutations. CNV loci were mainly chosen to match the AF-based loci. A few CNV loci were chosen as potential markers of subclonality based on high variation in exome-seq read counts across samples, and a few were chosen as potential diploid controls based on stable read counts. For point mutations, we designed the primers and a FAM-labeled probe that targeted the mutation in the target gene as well as a HEX-labeled probe that targeted the wild type allele, respectively. ddPCR reactions were performed according to the manufacturer’s protocol. Briefly, 10 ng DNA template was mixed with the PCR Mastermix, primers, and probes to a final volume of 20 μL, followed by mixing with 60 μL of droplet generation oil to generate the droplet by the Bio-Rad QX200 Droplet Generator. After droplets were generated, they were transferred into a 96-well PCR plate and then heat-sealed with a foil seal. PCR amplification was performed using a C1000 Touch thermal cycler and once completed, the 96-well PCR plate was loaded on the QX200 Droplet Reader. All ddPCR assays performed in this study included two normal human controls (NA12878 and NA10851) and two mouse controls (NSG and XFED/X3T3) as well as a no-template control (NTC, no DNA template). All samples and controls were run in duplicates. Data was analyzed utilizing the QuantaSoft™ analysis software. Fractional abundance (mutation) or copy number of the duplicate wells was calculated and a t-test was performed to determine statistical significance compared to the controls.

Estimation of copy number values in subclones

Based on the logic of Fig. 4B, we estimated the copy number at ERBB4 in subclones 1 and 2. To do this, we used all samples with measurement data and performed a linear regression of ERBB4 CN versus LRP1B allele frequency. The estimated value for ERBB4 CN in subclone 1 was the linear interpolation at LRP1B AF = 0.25. The estimated value for CN in subclone 2 was the linear interpolation at LRP1B AF = 0. See also Supplementary Fig. S5.

Outlier detection

For the TM00099 mutations not impacted by mixing of the two major subclones (BRCA2, CSMD3, PTCH1, TP53), we identified outlier AF values by calculating a median-centered z-score for each sample: z(gene, sample) = (AF(gene, sample) – Median AF(gene, all samples))/σ(gene, all samples). Samples with |z| > 3 were considered outliers, as these would be extremely unlikely to arise by chance (p < 4.4e-5). For the variables impacted by subclonal mixing of the two major subclones (ERBB4, MET, SAAL1), we calculated outliers based on their perpendicular Euclidean distance D to the line determined by each variable plotted versus LRP1B AF. These lines were those connecting the pairs of points: (LRP1B, ERBB4): (0,0) and (0.25,0.33); (LRP1B, MET): (0,0.67) and (0.25,0.5); (LRP1B, SAAL1): (0,0.25) and (0.25,0.4), respectively. A sample was considered to be an outlier for D > 0.07. This is much larger than the ddPCR AF measurement uncertainty of 0.011 (Fig. 2B). We chose this mixing-based constraint to be even more stringent (p < 4.5e-19 assuming z-score statistics) than that used for the BRCA2, CSMD3, PTCH1, and TP53 deviations in order to compensate for our uncertainty in what subclone mixtures were the precursors for each of the outlier samples.

Ethics Statement

The animal care rules used by The Jackson Laboratory are compatible with the regulations and standards of the U.S. Department of Agriculture and the National Institutes of Health. The protocols used in this study were approved by the Institutional Animal Care and Use Committee (IACUC) of The Jackson Laboratory under protocol number 12027. Mice used in this project were euthanized by carbon dioxide asphyxiation in a manner consistent with the 2013 recommendations of the American Veterinary Medical Association Guidelines on Euthanasia. All methods used in this study were performed in accordance with these guidelines and regulations.