Evolutionary Pressures Shape Undifferentiated Pleomorphic Sarcoma Development and Radiotherapy Response

Abstract

Radiotherapy is an integral component in the treatment of many types of cancer, with approximately half of cancer patients receiving radiotherapy. Systemic therapy applies pressure that can select for resistant tumor subpopulations, underscoring the importance of understanding how radiation impacts tumor evolution to improve treatment outcomes. We integrated temporal genomic profiling of 120 spatially distinct tumor regions from 20 patients with undifferentiated pleomorphic sarcomas (UPS), longitudinal circulating tumor DNA (ctDNA) analysis, and evolutionary biology computational pipelines to study UPS evolution during tumorigenesis and in response to radiotherapy. Most unirradiated UPS displayed initial linear evolution followed by subsequent branching evolution with distinct mutational processes during early and late development. Metrics of genetic divergence between regions provided evidence of strong selection pressures during UPS development that further increased during radiotherapy. Subclone abundance changed following radiotherapy with subclone contraction tied to alterations in calcium signaling, and inhibiting calcium transporters radiosensitized sarcoma cells. Finally, ctDNA analysis accurately measured subclone abundance and enabled non-invasive monitoring of subclonal changes. These results demonstrate that radiation exerts selective pressures on UPS and suggest that targeting radioresistant subclonal populations could improve outcomes after radiotherapy.

Introduction

Tumors are complex, dynamic, and heterogeneous entities comprised of different populations of cells arising from a common progenitor (1). Over the course of their natural history, cancers are exposed to selection pressures imparted by spatial constraints, the immune system, and microenvironment conditions (2). Initial tumorigenesis, growth, and metastasis all represent unique evolutionary conditions that shape intratumoral heterogeneity. Furthermore, systemic therapy has been shown to select for resistant tumor subpopulations that precipitate relapse (3), underscoring the importance of understanding tumor heterogeneity and evolution to improve treatment outcomes.

Although half of patients with cancer receive radiotherapy, the effect of radiotherapy on tumor evolution remains to be defined (4). A prior study in patients undergoing chemoradiation for rectal cancer identified substantial changes in subclone abundance after treatment (5). However, it can be challenging to distinguish changes in subclone abundance secondary to selection versus stochastic genetic drift (6), and the effects of radiotherapy alone on subclonal dynamics have not been explored in human tumors. Although clonal mutations have been shown to alter the intrinsic radiosensitivity of tumors (7,8), it is unclear if radioresistant subclonal tumor populations contribute to relapse after radiotherapy. Previous clinical studies have reported inferior local control after re-irradiation compared with radiation-naïve tumors, suggesting radiotherapy selects for resistant tumor cells (9). However, lower radiation doses are typically delivered in the re-irradiation setting due to concern for normal tissue toxicity, and modern radiation techniques can successfully control locally recurrent tumors after prior radiotherapy in many clinical scenarios (10,11). Furthermore, a preclinical study did not observe a difference in the radiosensitivity of human cancer cell line xenografts that recurred after fractionated radiotherapy in mice (12), raising the possibility that local relapse after radiotherapy represents a stochastic event.

Soft tissue sarcomas are histologically and molecularly complex tumors that arise from connective tissues throughout the body (13). Undifferentiated pleomorphic sarcomas (UPS), which make up approximately 15% of soft tissue sarcomas, harbor a higher mutational burden than most other subtypes (14,15). Recent studies have demonstrated that a subset of patients with metastatic UPS respond to immune checkpoint inhibitors (16,17), and the addition of pembrolizumab to standard of care preoperative radiation therapy improves outcomes in patients with localized UPS (18). Substantial intratumoral genomic and transcriptomic heterogeneity that can exist within sarcomas (19,20). Pre-clinical data from mouse models suggests that UPS are predominantly shaped by neutral evolution during early development (21). However, the subclonal architecture within human leiomyosarcomas and rhabdomyosarcomas suggests evolutionary pressures may lead to subclonal selection during sarcoma development (20,22). Localized UPS are primarily managed surgically with radiotherapy added pre- or post-operatively to improve local control in large, high grade tumors (23). Soft tissue sarcomas are often considered to be radioresistant with low rates of local control after radiotherapy alone (24). As a result, UPS represent a unique opportunity to study tumor evolution in the absence and presence of radiotherapy.

Analysis of clonal evolution over time in human tumors is limited by the availability of tumor tissue from multiple time points and errors from sampling a single region. Circulating tumor DNA (ctDNA) released from tumor cells during cell division and death can be non-invasively and longitudinally analyzed, providing a unique insight into the heterogeneous genetic landscape across an individual patient’s cancer (4,25). Previous studies have applied ctDNA analysis to track tumor evolution during progression and in response to systemic therapy (26–29). However, few studies have explored the correlation between ctDNA analysis and detailed spatial genomic profiling of tumor tissue (26), and the ability to track clonal evolution using ctDNA analysis during radiotherapy or in patients with sarcomas has not been demonstrated.

Here, we integrate spatial and temporal genomic profiling of tissue, circulating tumor DNA analysis, and computational modeling to investigate intratumoral heterogeneity and clonal evolution within UPS during tumor development and radiotherapy. We observe evidence of subclonal selection during sarcomagenesis that increases during radiotherapy, associated with dramatic changes in subclonal abundance following treatment. Furthermore, we demonstrate that ctDNA analysis represents a promising approach to non-invasively track tumor evolution during radiotherapy.

Materials and Methods

Human subjects

Written informed consent was obtained from all patients. Study protocols were approved by the Institutional Review Boards at Stanford University and University Medical Center of Johannes Gutenberg University Mainz (UMC-Mainz). All studies were conducted in accordance with the Declaration of Helsinki.

Tumor and blood samples

Peripheral venous blood samples and spatially distinct regions from opposite sides of pre-treatment and surgical resection tumor specimens were prospectively collected by the Stanford Cancer Institute Tissue Procurement Shared Resource. Blood samples were collected in K2EDTA tubes (Becton Dickinson) and centrifuged at 1600 × g for 10 minutes at room temperature. Plasma and plasma-depleted whole blood were collected and stored at −80C prior to DNA isolation. We retrospectively identified patients with non-metastatic high-grade undifferentiated pleomorphic sarcoma, at least two spatially distinct tumor regions, and matched blood samples available for analysis. Viable tumor was confirmed within each specimen on a hematoxylin and eosin-stained slide by a board-certified pathologist. When possible, each tumor was subsampled into regions at least 0.3 cm apart to be consistent with prior studies (30). This resulted in a total of 120 tumor regions that were processed as individual samples for DNA extraction and sequencing as described below.

DNA extraction, library preparation, and sequencing

To maximize the number of regions available for analysis, we included a combination of fresh frozen and formalin-fixed paraffin embedded (FFPE) tumor samples. Using the bioinformatic strategies described below to minimize the identification of artefactual variants and ultra-deep re-sequencing of all regions to validate called variants, we did not observe a difference in the number of mutations called or ITH metrics between fresh frozen and FFPE samples (Supplementary Fig. S1). Tumor DNA was isolated from fresh frozen samples using the QIAamp Fast DNA Tissue Kit (Qiagen) and FFPE samples using the DNAstorm FFPE Kit (CELLDATA) according to the manufacturers’ instructions. Leukocyte DNA from plasma-depleted whole blood was extracted using the DNeasy Blood and Tissue kit (Qiagen), and cell-free DNA (cfDNA) was extracted from 3–5 ml of plasma using the QIAmp Circulating Nucleic Acid Kit (Qiagen) according to the manufacturer’s instructions. Tumor DNA and leukocyte DNA were fragmented using a Covaris S2 sonicator and purified using the QIAquick PCR Purification Kit (Qiagen). DNA was quantified using the Qubit dsDNA High Sensitivity Kit (Thermo Fisher Scientific).

Tumor, normal, and cfDNA libraries were prepared with the KAPA HyperPrep Kit (Roche) with minor modifications to the manufacturer’s instructions and custom sequencing adapters incorporating dual-index sample barcodes for demultiplexing and molecular barcodes for deduplication and error-suppression as described previously (31,32). Target enrichment was performed via hybrid-capture using the xGen Hybridization Kit (Integrated DNA Technologies) according to the manufacturer’s instructions with 1.5 mM hypotaurine (Sigma-Aldrich, H1384) added to the hybrid capture reaction. Samples were sequenced on a NovaSeq 6000 (Illumina) with 150 bp paired-end reads.

Variant calling and targeted resequencing

Tumor DNA and matched leukocyte DNA were initially sequenced using the xGen Exome Research Panel v2 (Integrated DNA Technologies). Sequencing data were demultiplexed, mapped, and deduplicated as described previously (31). Copy number was estimated using TitanCNA (33) with a maximum ploidy of 4 and maximum number of clusters set to 5. Segmented copy number values were analyzed using GISTIC 2.0.22 (RRID: SCR_000151) to identify recurrent copy number alterations (34).

SNVs were identified with a consensus calling-approach using VarScan2(35), Mutect2 (36), and Strelka (37). Variants called by at least 2 callers were further filtered to remove variants with an allele fraction <5% in the tumor, ≥1 read in the matched leukocyte DNA, <20x positional depth in either the tumor or leukocyte sample, population allele frequency >1% in the gnomAD database (38), in repetitive regions of the genome (http://www.repeatmasker.org), and variants not located in an exon, UTR, or splice site. To address the challenges of formalin-fixation and oxidative damage during library preparation, we applied previously characterized empiric filters aimed at identifying artefactual mutations (39,40). For samples affected by these artefacts, we also excluded variants with an allele fraction <10% for the specific base changes associated with formalin-fixation (C>T/G>A) and oxidative damage (C>A/G>T). The resulting filtered SNVs were used to generate Discovery Pools (Integrated DNA Technologies) for each patient to enrich for the entire landscape of SNVs called in at least one region, and each tumor and leukocyte DNA library was re-captured for ultra-deep resequencing to a mean deduplicated depth of 3600x. The final variant list was generated by filtering for SNVs with an allele fraction ≥ 1% in at least one tumor region and ≤ 1 read in the matched leukocyte DNA.

Estimating cancer cell fraction (ccf)

The deduplicated variant allele frequency (vaf) from ultra-deep resequencing along with the purity (pu), total copy number (nt), major copy number (na), minor copy number (nb), and prevalence of the CNA (pa) determined by TitanCNA from the whole exome sequencing data were utilized to calculate the ccf for each mutation using a modified method based on previous work (6,41). For mutations in diploid regions of the genome without a CNA, the ccf was calculated as follows:

$${\mathit{\text{ccf}}}_{\mathit{\text{NoCNA}}}=\frac{\mathit{\text{vaf}}\times 2}{\mathit{\text{pu}}}$$

For mutations in regions with CNAs, we first calculated the proportion of cells with the CNA (sAGP) and the effective copy number given the purity of the sample and prevalence of the CNA (nc2).

$$\mathit{\text{sAGP}}=\mathit{\text{pu}}\times \mathit{\text{pa}}$$

$$\mathit{\text{nc}}2=\mathit{\text{nt}}\times \mathit{\text{sAGP}}+2\times \left(1-\mathit{\text{sAGP}}\right)$$

The ordering of the SNV relative to the CNA and the allele affected were determined as described in the Supplementary Methods by calculating the probability of the following scenarios given the observed number of mutant reads: Early Major: the SNV occurred on the major allele before the CNA, Early Minor: the SNV occurred on the minor allele before the CNA, Late: the SNV occurred after the CNA, and Independent: the SNV and CNA occurred in different lineages. Finally, the ccf was calculated based on the highest probability scenario.

$${\mathit{\text{ccf}}}_{\mathit{\text{Early}}\mathit{\text{Major}}}=\mathit{\text{max}}(\frac{\mathit{\text{vaf}}\times \mathit{\text{nc}}2-\mathit{\text{na}}\times \mathit{\text{sAGP}}}{\mathit{\text{pu}}}+\mathit{\text{pa}},\mathit{\text{nc}}2*\mathit{\text{vaf}}/(\mathit{\text{pu}}*\text{na}))$$

$${\mathit{\text{ccf}}}_{\mathit{\text{Early}}\mathit{\text{Minor}}}=\mathit{\text{max}}(\frac{\mathit{\text{vaf}}\times \mathit{\text{nc}}2-\mathit{\text{nb}}\times \mathit{\text{sAGP}}}{\mathit{\text{pu}}}+\mathit{\text{pa}},\mathit{\text{nc}}2*\mathit{\text{vaf}}/(\mathit{\text{pu}}*\text{nb}))$$

$${\mathit{\text{ccf}}}_{\mathit{\text{Late}}\mathit{\text{or}}\mathit{\text{Independent}}}=\frac{\mathit{\text{vaf}}\times \mathit{\text{nc}}2}{\mathit{\text{pu}}}$$

Mutational clustering and phylogenetic trees

The final filtered SNVs and allele fractions from ultra-deep resequencing were used as input to define mutational clusters with PyClone-VI (42) using default parameters and tumor purity and local copy number determined by TitanCNA. The PyClone clusters were input to Pairtree (43), and the rprop variant of gradient descent was utilized to generate phylogenetic trees. Only clusters with >1 SNV and a cancer cell prevalence of >2.5% in at least one region were included. Phylogenetic trees and fish plots were visualized using ClonEvol and the ‘fishplot’ R package, respectively (44,45). For patients with paired pre-treatment and post-radiotherapy samples, mutational clustering and generation of phylogenetic trees was repeated twice. First, the regions at each timepoint were analyzed separately to enable comparison with the unpaired samples. Timepoints with only one region were not included in this analysis, and SNVs were considered clonal if they were in the founder node of the phylogenetic tree. Second, all regions were combined to generate a unified tree and enable comparison of subclone prevalence before and after radiotherapy.

Mutational signature analysis

The contributions of COSMIC mutational signatures 1, 2, 5, and 13 were analyzed using the ‘mutSignatures’ R package (46) based on their importance in soft tissue sarcomas (14). All the mutations from the clones on the trunk of the phylogenetic tree were combined and compared to the mutations from clones on the branches. Patients with less than 10 trunk or branch mutations were not included in this analysis.

Computational modeling of sarcoma evolution

To simulate sarcoma intratumoral heterogeneity under different evolutionary modes ranging from neutral evolution to strong selection, we utilized a previously-characterized computational model of spatial tumor growth that enables simulated multi-region sampling (6). Ten virtual tumors were generated per selection coefficient s (0, 0.01, 0.02, 0.03, 0.05, and 0.10), and 8 regions were virtually sampled. The average sequencing depth was set to 1000, and previously reported parameters were utilized (final tumor size = 10⁹ cells, deme size = 10⁴ cells, birth rate = 0.55, death rate = 0.45, and driver mutation rate per cell division = 10⁻⁶) except for the mutation rate per cell division, which was adjusted to 0.04 to match the mutational burden in the patient UPS from our cohort.

Intratumoral heterogeneity metrics

Between region genetic divergence was quantified using five previously described metrics of ITH (6) with minor modifications to accommodate our higher sequencing depth and the higher rate of CNAs within soft tissue sarcomas compared with other cancers (14). The computational pipeline is described in detail in the Supplementary Methods. Briefly, the ccf for subclonal mutations across regions was used to calculate the fraction of high-frequency subclonal SNVs (fHsub), the fraction of high frequency region-specific subclonal SNVs out of all region-specific subclonal SNVs (fHRS), dissimilarity of the site frequency spectrum between regions measured by Kolmogorov-Smirnov distance (KSD), Wright’s fixation index (FST) (47), and the area of under the curve for the cumulative site frequency spectrum divided by the area under the curve for a theoretical neutral site frequency spectrum (rAUC). Timepoints with only one region were excluded from this analysis.

Functional enrichment analysis

To identify significantly expanding and contracting subclones, the cancer cellular prevalence and standard deviation estimated by Pyclone for each subclone with >1 SNV were decomposed and combined across all regions at each timepoint using the number of mutations assigned to each clone as the number of observations (48). The resulting means were compared using two-sided t-tests, and p-values were adjusted for multiple hypothesis testing using the Benjamini-Hochberg procedure. Unique genes with missense or nonsense mutations in expanding and contracting clones were used as input for functional enrichment analysis for the Kyoto Encyclopedia of Genes and Genomes (KEGG) biological pathway gene sets (RRID: SCR_012773) using g:Profiler (RRID: SCR_006809) (49).

RNA-sequencing and differential gene expression analysis

Published RNA-sequencing expression counts from primary mouse sarcomas collected after 20 Gy or no irradiation were downloaded from the Gene Expression Omnibus (GEO, RRID: SCR_005012, accession GSE148856) (50). Patients treated with preoperative chemoradiation therapy for soft tissue sarcomas at UMC-Mainz with FFPE tumor tissue from paired pre-treatment biopsies and post-treatment surgical resection specimens were retrospectively identified. Regions with the highest sarcoma content were identified for sampling based on hematoxylin and eosin staining. RNA was extracted using RNAstorm FFPE RNA Extraction kits (Cell Data Sciences), and libraries were prepared for sequencing using SMARTer Stranded Total RNA-Seq v2-Pico Input Mammalian kits (Takara Bio). Libraries were sequences on a NovaSeq 6000 (Illumina) with 150 bp paired-end reads, and FASTQ files were quasi-aligned using Salmon. Samples with fewer than 1 million mapped reads were excluded from further analysis. Differential gene expression analysis was performed using DESeq2 (RRID: SCR_015687) (51) with patient included in the design for the chemoradiation cohort. Volcano plots were generated using the ‘EnhancedVolcano’ R package. Gene set enrichment analyses for the KEGG “Calcium signaling pathway” and “Cholinergic synapse” gene sets were performed using the ‘fgsea’ R package (52) with genes ranked by log₂ fold change.

Clonogenic assays

HT1080 (RRID: CVCL_0317) and SW872 (RRID: CVCL_1730) cells were obtained in April 2021 from the American Type Culture Collection (#TCP-1019) where cell lines were authenticated by STR analysis and tested for mycoplasma via Hoechst DNA stain, agar culture, and PCR. After receipt, cells were passaged 1 to 10 times prior to use and tested annually for mycoplasma using the MycoAlert PLUS Mycoplasma Detection Kit (Lonza, # LT07–701), most recently in October 2024. Cells were seeded in triplicate with or without 0.25 or 0.5 mM Caloxin 2A1 TFA (MedChemExpress, #HY-P3278A) diluted in water two hours before irradiation with an IC-250 X-ray biological irradiator (Kimtron). Cells were irradiated at 64 cm using 225 kV X-rays filtered by 0.5 mm Cu with a dose rate of 1.02 Gy/minute. Cells were then cultured in Eagle’s MEM (HT1080) or DMEM/F-12 (SW872) medium at 37°C with 5% CO₂ for 12 days to form colonies. Colonies were fixed with glutaraldehyde (6.0% v/v), stained with crystal violet (0.5% w/v), rinsed with deionized water, dried, and counted. A population of more than 50 cells was counted as one colony, and surviving fractions were calculated relative to unirradiated controls as described previously (53).

Circulating tumor DNA analysis

Cell-free DNA libraries from plasma samples were captured using the custom hybrid capture panels described above for ctDNA analysis, and the sequencing data was processed as described above. Error suppression using molecular barcodes and in silico background polishing were performed as previously described (54). The presence of each clone defined by Pyclone was analyzed by monitoring for all of the variants assigned to the clone using a previously described Monte Carlo-based ctDNA detection index (55). Clones were considered to be detected in the plasma at a detection index less than 0.05. Plasma ccf was calculated as described in the Supplementary Methods.

Statistical analysis

The primary objective of this study was to compare selection pressure in unirradiated and irradiated sarcomas. We performed a power analysis to detect evidence of selection by comparing virtual tumors developing under neutral evolution with virtual tumors under a selection coefficient s of 0.1. For fHsub, fHrs, KSD, and FST, 11 samples per group would provide greater than 75% power to detect an increase in selection (56). Two-sided student’s t-tests were used to compare the means between unpaired groups, paired t-tests were used to compare means between paired groups, and one-way ANOVAs were used to compare the means between more than two groups. Proportions were compared using two-sided Fisher’s exact tests. Correlation between variables was assesses using the Pearson correlation coefficient. Statistical significance was assumed at P<0.05. Statistics were performed with Prism 9 (GraphPad Software, RRID: SCR_002798) or R version 3.6.2 through the RStudio environment (RRID: SCR_000432).

Data and code availability

Anonymized clinical data, sample metrics, and somatic mutation data for the patients in this study are provided in the Supplementary Tables. The code utilized to calculate cancer cell fraction and intratumoral heterogeneity metrics is available at https://github.com/ejmoding/. The RNA sequencing data generated in this study are publicly available in GEO at GSE279683. The DNA sequencing data generated in this study are publicly available in the Database of Genotypes and Phenotypes at phs003830.v1.p1. Previously generated data analyzed in this study were obtained from GEO at GSE148856. Any additional information is available from the lead author upon request.

Results

Spatial and temporal genomic profiling of soft tissue sarcomas

To explore the intratumoral heterogeneity within soft tissue sarcomas before and after radiotherapy, we retrospectively identified 20 patients with localized high-grade UPS treated with definitive surgery with or without preoperative radiation (Fig. 1A and Supplementary Table S1). We focused on UPS because they have previously been shown to have a higher mutational burden than other soft tissue sarcoma histologies (14), enabling the identification and tracking of subclonal tumor populations defined by single nucleotide variants (SNVs) shared within the same cancer cell. Among these 20 patients, 12 patients were treated with preoperative radiotherapy 4–6 weeks prior to surgical resection (60%) and 8 patients received upfront surgery with or without adjuvant radiotherapy (40%). For patients treated with preoperative radiotherapy, paired pre-treatment biopsies and surgical resection specimens were available for 7 patients. The median preoperative radiation dose and fractionation was 50 Gy in 25 fractions.

Figure 1. Evolutionary modes during early and late sarcoma development.

A, Schematic outlining the patient samples analyzed in this study. 120 spatially distinct tumor regions from undifferentiated pleomorphic sarcoma (UPS) pre-treatment biopsies and surgical resection specimens were analyzed with whole exome sequencing to identify single nucleotide variants (SNVs) and copy number alterations. Personalized panels were designed for each patient, and ultra-deep sequencing was performed on tumor samples to validate the SNVs and plasma samples for ctDNA analysis. B, Violin plots displaying the number of subclonal populations in unirradiated (No RT) and irradiated (Post RT) soft tissue sarcomas. Solid lines represent the median and dashed lines show the interquartile range. P value was calculated using a two-sided t-test. C, Representative phylogenetic trees from unirradiated and irradiated UPS. Genes frequently altered by SNVs are displayed on the trees. Phylogenetic trees for the full cohort are displayed in Supplementary Fig. S3A. D, Stacked bar plot comparing the phylogenetic tree pattern in unirradiated and irradiated sarcomas. P value was calculated using a two-sided Fisher’s exact test. E, Donut charts displaying the percentage of clonal and subclonal mutations for the most frequently mutated genes across the whole cohort. Mutations were considered clonal if they were within the founder cluster of the phylogenetic trees. The total number of samples with mutations in each gene are shown in the center of the plot. Additional genes are shown in Supplementary Fig. S3B. F, Normalized contribution of sarcoma-associated COSMIC mutational signatures 1 and 5 to mutations found within clusters on the trunk versus the branches of phylogenetic trees. P values were calculated using two-sided paired t-tests.

We profiled a total of 120 spatially distinct tumor regions and matched leukocyte DNA using whole exome sequencing to identify clonal and subclonal somatic SNVs and copy number alterations (CNAs) within each tumor. Customized hybrid-capture panels were designed to re-sequence each tumor and leukocyte sample at high depth to validate all SNVs and precisely measure the allele fraction of each mutation for accurate subclonal clustering. Due to the high abundance of CNAs within UPS (14), we accounted for clonal and subclonal CNAs when clustering mutations and calculating cancer cell prevalence (see Methods and Supplementary Methods).

Combining across all regions for each sarcoma, we identified a mean of 85 SNVs across unirradiated and irradiated sarcomas (Supplementary Fig. S2A and Supplementary Table S2). Although radiotherapy leads to DNA damage and prior studies have observed that radiation-induced tumors and recurrent tumors after radiotherapy harbor a higher mutational burden (57,58), we did not observe a significant difference in the number of SNVs or CNAs in unirradiated and irradiated sarcomas (Supplementary Fig. S2B and C). In patients with paired samples, there was no significant difference in the tumor mutational burden before and after radiotherapy (mean 3.4 vs. 3.3 mutations per megabase, Supplementary Fig. S2D). Furthermore, there was not a significant change in the genes altered by SNVs or CNAs (Supplementary Table S3) or the contribution of COSMIC mutational signatures (Supplementary Fig. S2E). These results suggest that radiotherapy did not alter the mutational profile of UPS on the short time scale of our study (4–6 weeks after completion of radiotherapy). This observation may be explained by the fact that sarcoma cells with new mutations caused by radiotherapy would need to expand for these mutations to be detected by bulk tumor sequencing.

Distinct evolutionary modes during early and late sarcoma development

To explore the subclonal structure and evolutionary history of UPS, we utilized our high depth spatial genomic profiling to infer subclonal clusters and construct phylogenetic trees based on the cancer cell prevalence of each SNV. We observed an average of 6 distinct mutational clusters per UPS, and there were similar numbers of mutational clusters in unirradiated and irradiated sarcomas (Fig. 1B and Supplementary Table S4). Phylogenetic trees from both unirradiated and irradiated UPS frequently displayed initial linear evolution followed by subsequent branching (Fig. 1C and Supplementary Fig. S3A). These findings are consistent with a mixed model of evolution in UPS in which smaller population sizes drive initial stepwise evolution followed by a transition to branching evolution later in sarcoma development as the population size continues to expand (59). We did not observe a significant difference in the branching pattern between unirradiated and irradiated sarcomas (Fig. 1D).

Consistent with an important role in early sarcomagenesis, SNVs in TP53 were uniformly observed within the founder cluster on the phylogenetic trees (Fig. 1E). In contrast, the other frequently mutated genes in our cohort such as MUC16 and COL2A1 were most frequently subclonal (Supplementary Fig. S3B). Interestingly, ATRX, which has previously been reported as a driver mutation in soft tissue sarcomas (14), was only mutated in subclonal clusters (n=3). To explore whether mutational processes change over the course of sarcoma development, we compared the contribution of COSMIC mutational signatures to the mutations found within the early (trunk) clusters versus the later (branch) clusters on the phylogenetic trees. We observed a significantly higher contribution of the COSMIC 1 mutational signature to trunk clusters and a significantly higher contribution of the COSMIC 5 mutational signature to branch clusters (Fig. 1F and Supplementary Fig. S3C). Although both signatures have been described as clock-like mutational processes that accumulate over time, these results suggest that the underlying mutational processes within sarcomas shift from early to late tumor development.

Strong evolutionary selection shapes sarcoma development and radiotherapy response

Because longitudinal changes in subclone abundance can be caused by both evolutionary selection and stochastic genetic drift, we measured evolutionary selection in UPS using five intratumoral heterogeneity (ITH) metrics derived from population genetics (fHsub, fHRS, KSD, FST, and rAUC, see Methods). These metrics of genetic divergence between tumor regions have been validated to measure evolutionary selection in human tumors, with higher values indicative of higher levels of selection pressure (6). To confirm that these ITH metrics can measure evolutionary selection in UPS, we used a previously described computational model of spatial tumor growth (6) to simulate tumors under neutral evolution and increasing selection pressure with a similar mutational burden to the UPS in our cohort (Fig. 2A). In this model, advantageous mutations alter the cell birth and death rate based on the selection coefficient s.

Figure 2. Measuring selection pressure in computationally simulated sarcomas and patient sarcomas.

A, Schematic of spatially simulated sarcoma growth under different modes of evolution. A previously described computational model (6) was used to simulate sarcoma transformation, acquisition of mutations, and 3-dimensional growth under neutral evolution or increasing selection pressure. Simulated sarcomas were then virtually sampled to mirror multi-region sampling of human soft tissue sarcomas. B, Violin plots comparing intratumoral heterogeneity metrics in simulated undifferentiated pleomorphic sarcomas (UPS) developing under neutral evolution (s=0) and increasing selection pressure by adjusting the selection coefficient s. Plots for other metrics are shown in Supplementary Fig. S4A. P values were calculated using one-way ANOVAs. C, Violin plots comparing intratumoral heterogeneity metrics for simulated UPS under neutral evolution and unirradiated patient UPS (No RT). P values were calculated using two-sided t-tests. D, Violin plots comparing intratumoral heterogeneity metrics based on spatial genomic profiling in unirradiated (No RT) and irradiated (Post RT) patient undifferentiated pleomorphic sarcomas (UPS). P values were calculated using two-sided t-tests. Solid lines on the violin plots represent the median and dashed lines show the interquartile range. E, Comparison of intratumoral heterogeneity metrics in paired pre-treatment (Pre RT) and post-treatment (Post RT) samples from the same patient. P values were calculated using two-sided paired t-tests. Timepoints with only one region were excluded from this analysis.

Although the ITH metrics varied in their dynamic range, we observed a significant increase in all the ITH metrics with higher selection except for rAUC, which remained stable even at high levels of selection and was excluded from additional analysis (Fig. 2B, Supplementary Fig. S4A, and Supplementary Table S5). To test the hypothesis that UPS develop under selection pressure in the absence of radiation, we compared the same ITH metrics for unirradiated patient sarcomas and simulated sarcomas developing under neutral evolution. We observed a significant increase in 3 of 4 ITH metrics in unirradiated sarcomas compared to neutral evolution (Fig. 2C and Supplementary Table S6), suggesting strong evolutionary pressures shape sarcoma development.

We next explored whether radiotherapy increases selection for resistant subclonal populations, which would be reflected by a further increase in the ITH metrics. Due to the varying dynamic range across the various metrics, we first sought to identify the metrics capable of detecting an increase in selection compared with unirradiated sarcomas (i.e. not at the plateau of the dynamic range). Two metrics (KSD and FST) increased in simulated sarcomas under strong selection (s=0.10) compared with unirradiated patient sarcomas (Supplementary Fig. S4B). Consistent with an increase in selection after radiotherapy, both KSD and FST were significantly higher in irradiated sarcomas than unirradiated sarcomas (Fig. 2D). To further explore the evolutionary effects of radiotherapy on sarcomas, we next looked at the subset of patients for which we had paired samples from multiple regions of both biopsies and post-radiotherapy surgical resection specimens (n=5). We observed a similar significant increase in KSD and FST after radiotherapy (Fig. 2E), confirming the selection pressure imposed by radiotherapy in UPS.

Temporal subclonal expansion and contraction during radiotherapy

Having established that radiotherapy increases selection for resistant subclones as measured by genetic divergence between regions, we next analyzed subclonal architecture over time in our paired pre-treatment and post-radiotherapy sarcomas. We repeated mutation clustering and generated phylogenetic trees including all pre-treatment and post-radiotherapy regions for each patient. Although the majority of subclones were present at both timepoints, we observed dramatic changes in the cancer cell prevalence of subclones following radiotherapy (Fig. 3A and Supplementary Fig. S5). Across all paired sarcoma samples, 44% of clones significantly increased in prevalence, 44% significantly decreased, and 12% were not significantly changed (Fig. 3B).

Figure 3. Subclonal contraction and expansion during radiotherapy.

A, Representative phylogenetic tree (left panel) and fish plot of mutational cluster cancer cell prevalence (right panel) pre-treatment (Pre RT) and post-radiotherapy (Post RT). Branches on the phylogenetic tree are colored based on whether the mutational cluster was present pre-treatment, post-radiotherapy, or at both time points. Mutational clusters are filled with the same color on the phylogenetic tree and fish plot. B, Fold-change in subclone cancer cell prevalence from pre-treatment to post-radiotherapy for all patients with paired samples available for analysis. Significantly expanding clones are colored in red, significantly contracting clones are colored teal, and subclones without a significant change are colored gray. C-D, Most significant pathways altered in (C) expanding and (D) contracting clones by functional enrichment analysis using KEGG gene sets. Vertical dotted lines represent the threshold for significance (P = 0.05). E-G, Differential gene expression and gene set enrichment analysis in primary murine and human sarcomas after radiotherapy. E, Schematics of each cohort. F, Volcano plots displaying differentially expressed genes. Dashed lines indicate genes with an adjusted P value < 0.05 and Log₂(Fold Change) > 0.5 or < −0.5. Significantly enriched calcium transporters are annotated. G, Enrichment plots for the KEGG “Calcium signaling pathway” in irradiated vs. unirradiated sarcomas. H-I, Clonogenic survival in (H) HT1080 and (I) SW872 human soft tissue sarcoma cell lines treated with vehicle or 0.5 mM Caloxin 2A1 alone or in combination with 6 Gy radiotherapy. Results were normalized to control cells. P values calculated using two-sided t-tests are displayed on the plots.

Although there were few recurrently mutated genes between the subclones that expanded and contracted across patients (Supplementary Table S7), we explored potential pathways involved in radioresistance and radiosensitivity of subclonal sarcoma populations using functional enrichment analysis (49). Considering genes with nonsynonymous and nonsense mutations, we did not observe any significantly enriched pathways in expanding clones (Fig. 3C and Supplementary Table S8). However, genes mutated in contracting clones were significantly enriched for the “Cholinergic synapse” and “Calcium signaling pathway” KEGG gene sets (Fig. 3D and Supplementary Table S9). Notably, calcium signaling has previously been associated with progression and treatment resistance in solid cancers (60–62), providing evidence that mutations in this pathway may decrease calcium signaling and lead to radiosensitivity.

To validate the role of calcium signaling in sarcoma response to radiotherapy, we analyzed gene expression in previously published irradiated versus untreated primary murine UPS (50) and a new cohort of paired human soft tissue sarcomas collected after chemoradiation therapy versus pre-treatment biopsies (Fig. 3E–F and Supplementary Table S10). Using gene set enrichment analysis (63), we identified significant enrichment in the “Calcium signaling pathway” after radiotherapy in both cohorts (Fig. 3G). The “Cholinergic synapse” gene set was significantly enriched in human sarcomas after chemoradiation (P=2×10⁻⁴), but not primary murine sarcomas after radiotherapy (P=0.19).

Across the cohorts, we observed significant upregulation of several calcium transporters after radiotherapy (Fig. 3F), including ATP2A1, ATP2A3, ATP2B2, and ATP2B1, which code for calcium ATPases responsible for transporting calcium ions out of the cytoplasm (62). To investigate whether blocking calcium transport can radiosensitize sarcomas, we treated two human soft tissue sarcoma cell lines with the plasma membrane calcium ATPase (PMCA) inhibitor Caloxin 2A1 (64) alone or in combination with radiotherapy. In both cell lines, Caloxin 2A1 significantly reduced clonogenic survival when combined with radiotherapy (Fig. 3H and I and Supplementary Fig. S6). These results support an important role for calcium signaling in radiation response within soft tissue sarcomas and suggest that targeting calcium transporters could enhance the soft tissue sarcoma response to radiotherapy. Taken together with our functional enrichment analysis in human UPS, these findings demonstrate that the SNVs within subclonal UPS cells can have a functional impact on radiosensitivity.

Tracking subclonal evolution using ctDNA analysis

Analysis of peripheral blood using ctDNA analysis enables longitudinal non-invasive monitoring of tumor genomics, and recent studies have suggested that ctDNA analysis may better capture the full heterogeneity within solid tumors than tissue sampling (65,66). To explore the ability of ctDNA analysis to non-invasively measure and track changes in subclonal architecture, we performed ultra-deep sequencing of matched plasma samples to monitor the subclonal tumor populations identified by multi-region tumor sequencing (Supplementary Table S11). We were able to able to detect the majority (52.4%) of subclones by ctDNA analysis (Fig. 4A and Supplementary Table S12). Subclones detected by ctDNA analysis had a significantly higher cancer cell prevalence measured by multi-region tumor sequencing (Fig. 4B), consistent with the hypothesis that subclones with higher tumor prevalence also have higher prevalence in plasma. In addition, the tumor cell prevalence of subclones measured by ctDNA analysis was highly correlated with the prevalence measured by multi-region tumor sequencing (R²=0.67, P<0.0001, Fig. 4C), suggesting that ctDNA analysis can accurately determine subclonal abundance. Despite this strong correlation, it is worth noting some subclones that were highly prevalent in tumors were not detected via ctDNA analysis, potentially reflecting other biologic factors beyond tumor prevalence that may influence ctDNA detection (25).

Figure 4. Non-invasive tracking of subclonal architecture with circulating tumor DNA (ctDNA) analysis.

A, Donut chart displaying the percentage of subclones identified by spatial genomic profiling of tumor tissue detected in matched plasma samples using ctDNA analysis. B, Plot of subclone cancer cell prevalence measured by sequencing tumor tissue for subclones detected and undetected by ctDNA analysis. C, Correlation of cancer cell prevalence measured by spatial genomic profiling of tumor tissue and ctDNA analysis. Only clones detected in the ctDNA were included in the analysis. D, Stacked bar plots summarizing ctDNA detection of subclones that were not observed by tumor sequencing at the time of plasma collection. E, Longitudinal measurement of subclone cancer cell prevalence using ctDNA analysis for a patient with plasma samples collected pre-treatment (Pre RT), during radiotherapy (During RT), and after radiotherapy (Post RT). Each line represents a different subclone. A subclone with a missense mutation in the calcium transporter ATP2B2 is highlighted on the plot.

We next explored whether ctDNA analysis could identify subclonal tumor populations that may have been missed due to sampling bias on multi-region sequencing of tumor tissue. As described above, the majority of subclones (70%) in patients with paired pre-treatment and post-radiotherapy tumor samples were observed by multi-region tumor sequencing at both time points. Using ctDNA analysis, half of the subclones only observed in the pre-treatment tumor sample were detected in the post-treatment plasma (Fig. 4D). Similarly, half of the subclones only observed in the post-radiotherapy tumor sample were detected in the pre-treatment plasma, demonstrating that they were present but missed by tissue sequencing prior to treatment. These results demonstrate the ability of ctDNA analysis to account for sampling errors associated with tumor sequencing.

Finally, we performed longitudinal ctDNA analysis in a patient (SRC127) with plasma samples collected pre-treatment, one week into radiotherapy, and post-radiotherapy at the time of surgery. Remarkably, we observed evidence of subclonal expansion and contraction only a week into treatment that persisted through radiotherapy (Fig. 4E). For example, a subclone harboring an R32H missense mutation in ATP2B2, which codes for the calcium transporter PMCA2, decreased from a cancer cell prevalence of 71% prior to treatment to 40% one week into radiotherapy. These findings demonstrate the potential for ctDNA analysis to longitudinally track subclonal dynamics during treatments such as radiotherapy.

Discussion

Accumulating evidence suggests that intratumoral heterogeneity and subclonal tumor cell populations are clinically important for progression and response to therapy (2). In this study, we integrated spatial and temporal genomic profiling, circulating tumor DNA analysis, and computational modeling to dissect the intratumoral heterogeneity and evolutionary forces at play during UPS development and radiotherapy. We observed strong evidence of evolutionary selection during sarcomagenesis that was further increased by radiotherapy.

Prior studies have reported different modes of evolution across human tumor types (6,59). Using metrics of genetic divergence between regions, we observed evidence of positive selection within human UPS. Although most prior studies have identified a predominant pattern of evolution across solid tumor types (i.e. linear or branching), it has been proposed that tumors may change their mode of evolution over time (59). A modeling study suggested that small population sizes during early tumor development may drive linear evolution with branching evolution occurring later in tumor development (67), and the phylogenetic trees from UPS provide evidence for that model. Consistent with a change in evolutionary dynamics from early to late sarcoma development, we also observed a change in the mutational signatures from early truncal mutations to late branch mutations.

Our finding that SNVs in TP53 were uniformly observed in founder clusters suggests that TP53 alteration predominantly occurs early during UPS development. In contrast, although ATRX mutations are common within sarcomas and other solid tumors (68), we only observed subclonal SNVs in ATRX. Interestingly, Atrx deletion has been shown to increase the response of sarcomas to radiation in preclinical models (69), but the cancer cell prevalence of these mutations will likely impact their relevance as a biomarker of radiation response and drug target in sarcomas. In contrast to prior studies that observed a higher mutational burden in locally recurrent gliomas after radiotherapy (57) and distinct mutational signatures in radiation-induced cancers (58), we observed no difference in the number of mutations or mutational signatures in UPS treated with preoperative radiotherapy due to the short time interval between completing radiotherapy and tumor collection. Future studies examining locally recurrent sarcomas after radiotherapy could be helpful to investigate the potential role of radiation-induced mutations in tumor recurrence.

The response of tumors to radiotherapy can vary dramatically across histologic subtypes and even amongst patients with the same type of tumor. Prior efforts to identify genomic alterations and tumor biology associated with radiation response have focused on clonal driver mutations present within all cancer cells or utilized xenograft cell line models that may lack the subclonal diversity of primary patient tumors (7,8,12). Here, we provide evidence that radiotherapy exerts selection pressure on human tumors that leads to expansion of resistant tumor subpopulations. Modeling experiments have suggested that even small populations of resistant subclones could dramatically affect the ability of tumors to be cured with radiotherapy (70). As a result, a better understanding of the mechanisms that lead to subclonal radioresistance will be critical to design new therapies that enhance radiotherapy response. We observed and validated a role for altered calcium signaling in contracting subclonal populations, suggesting that subclonal SNVs can have a functional impact on cellular radiosensitivity. Notably, a prior study comparing pre-treatment and post-treatment tumor samples from patients undergoing chemoradiation therapy for rectal cancer also identified a role for calcium signaling in the response of subclones to chemoradiation (5). We found that inhibiting PMCAs using Caloxin 2A1 reduced clonogenic survival after radiotherapy in human soft tissue sarcoma cell lines. Although the effect on radiosensitivity was modest, multiple transporters are involved in regulating intracellular calcium levels (62). Future studies should investigate the therapeutic potential of targeting multiple calcium transporters simultaneously to enhance radiation response. We did not find any significantly enriched pathways by analyzing genes altered by SNVs in expanding subclones. It is possible that the SNVs we were able to track in this study co-occurred with other genomic alterations and/or epigenetic changes that help to determine the radiosensitivity of subclonal tumor cell populations. Future analyses using other approaches, including single cell sequencing technologies, could help to explore this possibility in more depth.

Studying tumor evolution has been limited by the lack of longitudinal tissue samples from patients and biases introduced by sampling one or a few tumor regions. Because ctDNA can be released from all the tumor populations within a patient and non-invasively sampled at multiple time points, it provides a unique window into intratumoral heterogeneity. Recent studies have highlighted the ability of ctDNA analysis to track clonal evolution during tumor development and in response to systemic therapy (26–29). In addition, a recent study demonstrated a strong correlation between subclone cancer cell prevalence measured by spatial genomic profiling of tumor tissue and ctDNA analysis in patients with lung cancer (26). Our study provides further proof of concept for using ctDNA as a tool to detect and dynamically monitor subclonal dynamics in UPS. Despite the low levels of ctDNA in patients with localized disease, we were able to detect the majority of subclones using personalized panels that enabled monitoring of all mutations within each clone identified from whole exome tumor sequencing. We showed the ability of ctDNA analysis to detect tumor-occult subclones identified from other timepoints and monitor subclonal expansion and contraction during radiotherapy. Because ctDNA has the potential to be released from all tumor subclones, it may overcome sampling biases from sequencing tumor tissue and better characterize genetic heterogeneity in some cases. These results illustrate the potential for ctDNA to be used as a powerful tool to study fundamental cancer biology and as a biomarker to identify expanding subclonal populations that could be targeted to improve outcomes with radiotherapy.

Unlike the majority of solid cancers that develop from epithelial tissues, sarcomas arise from mesenchymal tissues and are incredibly complex with over 100 histologic subtypes (13). In this study, we sacrificed cohort size to enable dense profiling of intratumoral heterogeneity and focused on UPS with a higher mutational burden to limit variation across patients and maximize our ability to detect changes in subclonal structure after radiotherapy. Because prior studies have suggested that the patterns of evolution may vary across sarcoma subtypes (71), future studies with larger cohorts of patients across other sarcoma and solid tumor histologies will be helpful to confirm the generalizability of our findings. Furthermore, recent studies have demonstrated that the addition of immune checkpoint inhibitors to preoperative radiotherapy leads to excellent pathologic response rates and improves disease-free survival in UPS (18,72). Future studies should investigate the impact of immunotherapy on tumor evolution during radiotherapy. Indeed, immunohistochemistry of tumors collected after neoadjuvant radiotherapy with concurrent nivolumab with or without ipilimumab found heterogeneous immune infiltration across regions within the same tumors (72).

Taken together, our results confirm radiotherapy increases selection pressures within UPS. Furthermore, we provide proof of concept for the utility of using ctDNA analysis to study tumor evolutionary biology during radiotherapy. Our findings suggest that resistant subclonal tumor populations may precipitate relapse after radiotherapy, raising the possibility that targeting these populations could improve patient outcomes.

Statement of Significance.

Radiotherapy mediates tumor evolution by leading to the expansion of resistant subclonal cancer cell populations, indicating that developing approaches to target resistant subclones will be crucial to improve radiotherapy response.