Spatially Discontinuous Mutation Topographies in Ductal Carcinoma in Situ Reveal Non-Competitive Growth Dynamics

Abstract

Pre-invasive breast cancer, or ductal carcinoma in situ (DCIS), shares many morphologic and genomic features with invasive breast cancer, yet most DCIS tumors remain indolent over decades. In this study, we performed spatial analyses of somatic mutation patterns in 18 DCIS tumors to infer the underlying growth dynamics. Experimental data combined with mathematical models and Bayesian computation revealed a growth model that mimics the cellular dynamics of branching morphogenesis, which governs normal mammary gland development during puberty. Unlike a traditional model of clonal evolution, this non-competitive growth model allowed for extensive DCIS spread without the selective pressures of subclone competition that promote the emergence of invasive phenotypes. This finding provides a framework to understand the biological basis for the limited progression of DCIS and adds rationale for ongoing clinical efforts to reduce DCIS overtreatment.

INTRODUCTION

Mammography screening has been successful in reducing breast cancer mortality,(1–3) yet its benefits are accompanied by harms such as false positive findings and overdiagnosis.(4) The overdiagnosis of indolent tumors that would not cause any harm in the woman’s remaining lifetime is of particular concern for patients diagnosed with ductal carcinoma in situ (DCIS), a non-invasive breast tumor.(5) DCIS is considered a precursor of invasive breast cancer, yet studies support that as many as 70–80% of DCIS found on mammography would not progress to invasive cancer if left untreated.(6,7) Because it is currently not possible to accurately distinguish indolent from aggressive disease, nearly all DCIS patients undergo immediate treatment.(8) This strategy leads to widespread overtreatment, affecting as many as 40,000 women each year in the US alone.(9)

Over the past decade, DCIS has received considerable attention, with most efforts focused on predicting the risk of invasive progression after diagnosis.(5,6,10,11) Yet at a more fundamental level, the indolent nature of DCIS per se poses a conundrum. Indeed, DCIS shares many morphologic, proteomic and genomic alterations with invasive breast cancer, (7,12–15) and frequently the only histologic distinction between DCIS and invasive cancer is abnormal tumor cell migration beyond the basement membrane. Given these similarities between DCIS and invasive breast cancer, one might expect abnormal cell movement to be an inherent feature of DCIS growth, which is difficult to reconcile with the low rate of progression to invasive cancer.

A seemingly unrelated yet equally perplexing conundrum concerns the spatial subclone topographies in DCIS tumors. Two independent studies—based on spatial single-cell copy number analyses(13) and in situ transcriptomics(16)—characterized DCIS as multiclonal expansions with spatially discontinuous ‘skip’ patterns that span macroscopic length scales. Such disperse skip patterns are difficult to reconcile with the canonical model of clonal cancer evolution which is expected to yield more continuous subclone patterns; or, in the words of Lomakin and colleagues:(16) “These appearances seem at odds with the traditional model of clonal competition in which a fitter clone generates localized monoclonal sweeps.“

Here, we reveal an alternative growth model for DCIS that offers a unified explanation for both tumor indolence and scattered mutation topographies. Locating DCIS on the spectrum between normal breast development and malignancy (Figure 1), we posit that DCIS growth is governed by the dynamic principles of non-competitive normal breast development rather than the competitive clonal evolution of invasive cancer. Breast development during puberty is governed by a process called branching morphogenesis, during which breast ducts grow, branch, and penetrate the surrounding stroma.(17–19) The cellular dynamics of branching morphogenesis have been quantified through lineage tracing in animal models, revealing that normal duct growth is orchestrated by advancing growth buds that each contain multiple stem cell subclones.(17) These stem cells intermittently contribute to ductal growth, leading to multiclonal ducts whose subclones form spatially discontinuous ‘skip’ patterns over macroscopic scales, similar to those observed in DCIS.(13,16)

Figure 1: The female breast: From normal development to invasive cancer.

(A) At birth, the mammary gland consists of the simple embryonic rudiment. (B) During pubertal development, the embryonic rudiment undergoes branching morphogenesis and develops into the adult ductal tree. (C) Ductal carcinoma in situ (DCIS) consists of neoplastic cells that are contained within the ducts and lobules of the adult mammary gland. (D) During invasive progression, DCIS cells penetrate the basement membrane of the ducts and lobules and invade the breast stroma.

The proposed non-competitive growth model of DCIS not only provides a canonical explanation for the discontinuous subclone topographies, but it also provides a compelling mechanism for the indolence of many DCIS. Similar to clonal evolution, this model allows for rapid dissemination of epithelial cells over macroscopic distances. However, unlike clonal expansion driven by subclone competition, the proposed dynamics lack the selective pressures that foster increasingly aggressive phenotypes, thereby limiting the potential for invasion and metastasis.

In this study, we evaluate different growth models of DCIS based on the three-dimensional mutation topographies of 18 tumors over macroscopic length scales of up to 7cm. We first confirm the multiclonal nature of DCIS expansions,(13,16) and then combine the spatially resolved genomic data with mathematical models and Bayesian computation to infer the most likely growth dynamics. We show that the empirical data are difficult to reconcile with canonical clonal evolution yet are consistent with a non-competitive growth model that mimics the cellular dynamics of developmental branching morphogenesis. We propose that normal cell movement conferred by this growth process reveals a biological basis for why many DCIS grow to a macroscopic size and can remain stable for decades without progression to invasive cancer.

MATERIALS AND METHODS

Patient cohort and biological samples

The study was conducted in accordance with the Declaration of Helsinki guidelines and following approval by the Institutional Review Board of the Duke University Medical Center (Pro00054515). The Institutional Review Board approved a waiver of written informed consent because the retrospective analysis of existing clinical data was deemed minimal risk. We identified patients diagnosed with screen-detected breast cancer who underwent breast-conserving surgery or mastectomy at Duke University Medical Center between 1999 and 2016. During the selection process, formalin-fixed paraffin-embedded (FFPE) tissue blocks for cases with a complete spatial block map were obtained from the Duke Pathology archives. Each block was pathology reviewed (A.H.) for diagnosis according to the WHO classification of tumors.(20) A total of 21 cases with tumor tissue present in two or more FFPE blocks were identified through this process, including 11 patients with pure DCIS tumors, and 10 patients with DCIS tumors with synchronous ipsilateral invasive breast cancer (synchronous DCIS). DCIS nuclear grade and estrogen- and progesterone-receptor status were abstracted from the patients’ medical records. As described below, a total of 3 patients were excluded prior to final data analyses, because of technical issues (n=2) or insufficient information content (n=1). The final analytic cohort thus comprised 18 patients, 9 with pure DCIS and 9 with synchronous DCIS (Table S1). Finally, we collected matched normal samples for all patients, in the form of blood (n=4), uninvolved lymph nodes (n=5), or adjacent, morphologically normal breast tissue (n=9).

Whole exome sequencing

To design tumor-specific mutation panels, whole exome sequencing (WES) was performed on bulk tissue samples as follows. For each patient, two or more spatially separated (≥8mm) FFPE blocks were identified, and areas containing DCIS (but no invasive cancer) were macro-dissected from between 10 and 25 hematoxylin-stained tissue sections (5 microns thick). The first and last sections were stained with hematoxylin-eosin (H&E) and reviewed by a study pathologist (A.H.) to confirm that tumor cellularity was at least 70%. DNA was extracted using the GeneRead DNA FFPE Kit (QIAGEN, cat. no. 180134) according to manufacturer instructions. DNA quantity was determined using a QubitTM 1X dsDNA HS Assay Kits (ThermoFisher, cat. n. Q33230), and DNA quality was assessed using the Agilent 2100 Bioanalyzer (RRID:SCR_018043). WES was performed on ≥40ng of genomic DNA from each sample. Each aliquot was sheared to a mean fragment length of 250 bp (Covaris LE220, RRID:SCR_026895), and Illumina sequencing libraries were generated as dual-indexed, with unique bar-code identifiers, using the Accel-NGS 2S PCR-Free library kit (Swift Biosciences, cat. n. 20,096). We pooled groups of 96 equimolar libraries (100 ng/library) for hybrid capture of the human exome as well as a targeted panel of the exons of 83 breast cancer genes, using IDT’s xGen Exome Research Panel v1.0; see Fortunato et al.(21) for details. After hybridization, capture pools were quantitated via KAPA Library Quantification Kit for Illumina platforms (Roche Sequencing Solutions; formerly KAPA Biosystems), following the manufacturer’s protocol, and the final product was sequenced using an Illumina HiSeq 2500 1T instrument (RRID:SCR_016383), multiplexing nine tumor samples per lane. After binning the data based on its index identifier and aligning it to the Genome Reference Consortium Human Build 37 (GRCh37) using the BWA-MEM algorithm (RRID:SCR_010910),(22) sequencing duplicates were identified using Picard’s MarkDuplicates (RRID:SCR_006525), executed via GATK; (RRID:SCR_001876). The resulting BAM files were then used to design the tumor-specific mutation panels as described in the next section. The WES protocol was performed at the Washington University School of Medicine Genome Technology Access Center Core Facility in St Louis (RRID:SCR_001030).

Tumor-specific mutation panels

For each patient, we designed a tumor-specific target panel of single nucleotide variants (SNVs) based on the BAM files obtained from WES of tumor and matched normal tissue. Variants were called using the software MuTect (Broad Institute, RRID:SCR_000559),(23) using default settings. Starting from a combined set of SNVs that had “judgment=KEEP” in at least one of the two samples, we excluded SNVs not mapped to chromosomes 1 through 22 or the X chromosome, SNVs identified as single nucleotide polymorphisms in dbSNP (RRID:SCR_002338)(24) and SNVs that were within 300bp of another SNV. For patients where more than 100 SNVs remained after these exclusions, we decreased the final panel size to 100 or less by first removing mutations at a variant allele frequency (VAF) below 10% in both bulk samples and taking a simple random sample if necessary. SNVs identified in COSMIC (https://cancer.sanger.ac.uk/cosmic; RRID:SCR_002260)(25) were included independently of the above filter settings.

Saturation microdissection

From each tumor, between 2 and 5 spatially separated FFPE blocks that contained individual DCIS ducts or lobules suitable for microdissection were identified by the study pathologists (A.H., D.S.). In mixed tumors, the study pathologists (A.H. and D.S.) further identified DCIS-adjacent areas of IBC suitable for microdissection. From each block, between 5 and 10 consecutive 5-micron tissue sections were prepared on plastic slides and lightly stained with H&E. A study pathologist (D.S.) then microdissected small tissue areas, or spots, using selective ultraviolet light fractionation (SURF) as previously described(26) and implemented by our group.(27). In brief, a micromanipulator was used to place small ink dots over individual duct cross-sections and, in the case of synchronous DCIS tumors, over equivalently sized areas of invasive breast cancer. The absolute number of tumor cells in each microdissected spot was estimated to be between 100 and 500 cells. After the destruction of unprotected DNA through 3–4 hours of short-wave ultraviolet light irradiation, individual ink dots were removed from the slides using a pipette tip and placed in a microfuge tube for DNA extraction.

Targeted mutation sequencing

After proteinase K and TE treatment at 60°C for 4 hours, and then at 98°C for 10 minutes, AMPure XP beads (Beckman Coulter) were added (1.2x) to extract the DNA. Polymerase chain reaction (PCR) was performed directly on the dried beads (35–40 cycles) using a custom AmpliSeq primer (ThermoFisher) for the tumor-specific SNV panels as described above. PCR repeatedly failed for two tumors and led to their exclusion from further analysis (DCIS-118, DCIS-158). Barcoded libraries (OneStep Barcode Library Kit, Qiagen) were then run on Illumina MiSeq (RRID:SCR_016379) or NextSeq (RRID:SCR_016381) sequencers, with an average coverage of >500x and a minimum coverage of 20x for each mutation. The FASTQ files from the sequencers were uploaded to the Galaxy (RRID:SCR_006281) web platform and analyzed using the public Galaxy server (http://usegalaxy.org/).(28) Briefly, our Galaxy pipeline included FASTQ grooming, adapter trimming using TrimGalore (RRID:SCR_011847), short read alignment to GRCh37 via BWA (RRID:SCR_010910), Naive Variant Caller and Variant Annotator. For each locus, we defined the reference and alternate alleles based on the WES results and recorded their respective read counts from the targeted sequencing runs.

Low-pass whole genome sequencing for spatial copy number profiling

A total of 19 spots were microdissected from DCIS-286; 9 spots corresponded to a spot with available somatic mutation calls. Whole genome libraries were prepared with NEBNext^® Ultra™ II DNA Library Prep Kit and sequenced on Illumina NovaSeq-6000 (RRID:SCR_016387) using paired end reads extending 150 bases and demultiplexed into pairs of FASTQ files for each sample. The FASTQ files were aligned to GRCH37 using the BWA-MEM algorithm (RRID:SCR_010910),(22) and the resulting BAM files were used in the CNV analysis pipeline implemented in the R package QDNAseq (RRID:SCR_003174).(29) Count data were obtained, smoothed, and normalized using default settings with bin annotations of size 30 kbp derived from reference genome GRCH37 as provided in the package. CNV calls were obtained using the multi-state mixture model CGHcall (RRID:SCR_001578).(30)

Spatial registration of spots

To construct three-dimensional maps of spot locations within each tumor (Figure S1), we first used the clinical pathology maps that show the spatial relationship of each paraffin block within the excised tissue. These block maps were used to locate pathologic features with respect to surgical margins and to determine the positions of each of the paraffin blocks included in the study along the long axis of the tissue/tumor (referred to as the z-axis). Once positioned along the z-axis, we oriented the thin sections from these blocks based on colored ink stains along the tissue margins. Once the slides were properly oriented, we determined the in-plane location (x- and y-coordinates) of individual spots which had been recorded during microdissection. The origin of the x- and y-coordinates were anchored at the center of each slide, and spot coordinates were recorded after accounting for microscopic magnification. Combining the in-plane x- and y-coordinates with the z-coordinate along the tumor’s long axis thus completed the process of spatial spot registration.

Phenotypic annotation of spots

The histologic phenotype of each spot was determined in three steps. First, two board-certified breast pathologists (A.H. and J.G.) independently reviewed the H&E slides, classified each spot as ‘normal’, ‘benign’, ‘DCIS’ or ‘invasive’, and used a free text field to provide a comprehensive description of all ‘benign’ spots. Spots where the two pathologists agreed on the main category (normal, benign, DCIS, invasive) were considered complete (n=445, or 85%); the remaining spots (n=79, or 15%) were adjudicated by a third board-certified breast pathologist (D.W.). A board-certified pathologist (D.S.) used the free text annotations of all ‘benign’ spots to refine their classification as either ‘normal breast tissue,’ ‘benign breast disease without atypia’, or ‘benign breast disease with atypia’. Finally, a board-certified breast pathologist (A.H.) assigned to each DCIS spot a pathologic subtype (solid, cribriform, micropapillary) and determined whether comedo-like features were present.

Mutation calls

Variant calling based on the tumor-specific SNV panels was performed in each spot separately, using a previously described Bayesian inference method.(31) Briefly, for any given sequencing target, the posterior distribution of the target’s VAF $f$ was calculated by combining a data likelihood and a prior distribution according to Bayes’ theorem. For the data likelihood, we used a binomial model for the variant read count $K$ and the total read count $N$, accounting for a sequencing error rate $e$ as follows

$$P\left(N,K|f\right)=\left(\begin{array}{c}N\\ K\end{array}\right){\left(f\left(1‐e\right)+\left(1‐f\right)e\right)}^K+{\left(fe+\left(1‐e\right)\left(1‐f\right)\right)}^N.$$

Our prior belief about the VAF was modeled as a mixture

$$\pi \left(f\right)=c_0\delta \left(0\right)+\left(1‐c_0\right)Beta\left(1,{\gamma }_s^{‐1}\right),$$

where $c_0$ is the prior probability of the mutation being absent (as reflected by the point mass $\delta \left(0\right))$, and, if present, the mutation’s VAF was assumed to have a prior distribution $Beta\left(1,{\gamma }_s^{‐1}\right),$ where ${\gamma }_s$ is the sample purity. Applying Bayes’ theorem, the posterior distribution of the VAF, or $P\left(f|N,K;e,c_0,{\gamma }_s\right)$, can be calculated explicitly. Finally, the posterior probability that a mutation is absent ($q$) or present ($p$) is then given by

$$q=P\left(f<f_{abs}{\gamma }_s|K,N;e,c_0,{\gamma }_s\right),p=1‐q,$$

where $f_{abs}$ is a pre-defined sequencing threshold. For applications where binary mutation calls were needed, we called individual SNVs absent if $q>95\%$ and present if $p>95\%$. The handling of mutations with $q,p\in [5\%,95\%]$ was determined in situ, depending on the analyses performed.

Unless otherwise specified we used the following parameter values: $e=0.01$ which reflects the empirical error rate of the sequencing platform(32); $f_{abs}=5\%$ to avoid false positives mutation calls(31); $c_0=0.5$ to reflect a lack of prior knowledge about the absence vs presence of a mutation; and ${\gamma }_s=0.8$ as a conservative estimate of the sample purity achieved by SURF.

To assess the sensitivity of mutation calls to the minimum read depth of 20x, we performed downstream sensitivity analyses for several outcome measures under a minimum depth of 100x (Figure S2).

Mutational signatures

To analyze the DNA mutation patterns in our cohort, we compiled a list of targeted mutations that were present in at least one microdissected spot. Using the R package MutSignatures,(33) we categorized mutations into 96 types based on 6 possible single base pair substitution categories (C>A, C>G, C>T, T>A, T>C and T>G) and 16 combinations of 3’ and 5’ nucleotide neighbors. We performed de novo extraction of mutational signatures using the non-negative matrix factorization method (n=1,000 bootstrap iterations, k=2 signatures) and estimated the exposure of each tumor sample to the two signatures. In separate analyses, we performed de novo extraction for k=3 and k=4 signatures; since these resulted in the same two high-quality signatures as extracted for k=2, accompanied by additional low-quality signatures, we chose k=2 for the final analysis. We then compared the two extracted signatures to the COSMIC database (https://cancer.sanger.ac.uk/cosmic; RRID:SCR_002260) using the cosine distance, and further assessed whether matching signatures were breast cancer related(34) or possible sequencing artefacts.

Driver mutation annotation

Single nucleotide variants were annotated using the SIFT annotation tool (https://sift.bii.a-star.edu.sg/; RRID:SCR_012813), which predicts mutation effect and functional impact on the protein. Briefly, SNVs were organized into a VCF file format, specifying chromosome, genomic position, and reference and alternate alleles according to the GCRh37. The VCF file was input into SIFT, which output annotated SNVs, labeling Ensembl transcript and gene IDs, gene name, coding region (CDS, UTR_3, UTR_5), variant type (noncoding, nonsynonymous, stop-gain, substitution, synonymous), and functional prediction (deleterious, tolerated). SNVs with no gene label were labeled as intergenic, and SNVs with a gene label but not in a coding or UTR region were labeled as intronic. Protein coding changes in genes that have been functionally associated with breast cancer in either the TCGA (https://www.cancer.gov/tcga; RRID:SCR_003193) or COSMIC (https://cancer.sanger.ac.uk/cosmic; RRID:SCR_002260) databases were considered putative driver mutations (Table S2). All others were categorized as passenger mutations.

Final study cohort

After eliminating two tumors due to PCR issues, the remaining 19 tumors comprised a total of 508 individual spots and 1,052 targeted loci (Figure S3). Among the 30,369 sequencing targets (each target is a spot-SNV pair), there were 9,990 (23.0%) low-quality targets (LQTs) where either no sequencing results were obtained or the total absolute read count was less than 20. After removing 6 spots with undefined histology, 208 mutations that were LQTs in more than 40% of assayed spots, 21 germline mutations (which were present in the matched normal with a probability ≥99%), 39 spots that contained more than 40% of LQT, and 150 spots with a histology other than DCIS, we excluded one more tumor (DCIS-221) because of low information content (11 of the 14 detected mutations were germline mutations, and the remaining 3 mutations were detected in only one spot each). The resulting analytic cohort comprised 18 tumors with 14,430 targets (5% LTQs) across 313 DCIS spots.

Uncertainty quantification

For tumor statistics based on binary mutation calls, we leveraged the Bayesian framework to propagate posterior uncertainty through Monte Carlo sampling. More precisely, for a tumor with $N$ spots and $M$ mutations, we sampled $T$ independent and identically distributed binary spot-mutation arrays $S=\left(s_{ij}\right)\in {\mathbb{R}}^{N\times M},$ where $s_{ij}\sim Bernoulli\left(p_{ij}\right)$ and $p_{ij}$ is the posterior probability of mutation $j$ being present in spot $i$. For LQTs, because there was no data available, we used the prior probability instead. The statistic of interest was then computed for each of the $T$ realizations of $S$, and the posterior predicted mean and 95% prediction interval were recorded. Unless otherwise specified, the default was $T=\mathrm{1,000}$.

Spatial-genetic correlation

A tumor-level spatial-genetic correlation measure was introduced to assess the degree of spatial intratumor heterogeneity. Uncertainty was quantified as described above, and we focus here on the derivation of the statistic for a single realization of the binary array $S=\left(s_{ij}\right)\in {\mathbb{R}}^{N\times M},$ for a tumor with $N$ spots and $M$ mutations. First, we defined the spatial distance $d_s\left(i,j\right)$ between two spots $i$ and $j$ as their Euclidian ($L^2)$ distance in ${\mathbb{R}}^3$. Next, we introduced the notion of spot $i$’s genotype as the vector $g_i=\left(s_{i1},s_{i2},\dots ,s_{iM}\right)\in {\mathbb{R}}^M$ and defined the genetic distance $d_g\left(i,j\right)$ between two spots $i$ and $j$ as the Manhattan ($L^1$) distance between $g_i$ and $g_j$. Finally, we calculated the spatial and genetic distances between all $N\left(N‐1\right)$ spot pairs and computed their correlation (Pearson’s R).

Expansion index

The expansion index (EI) was introduced to distinguish, at the tumor level, between spatially discontinuous ‘skip’ lesions and continuous `patch` lesions (Figure S4). Again, uncertainty quantification was performed as described above, and we focus here on the derivation of the statistic for a single realization of the binary array $S=\left(s_{ij}\right)\in {\mathbb{R}}^{N\times M},$ for a tumor with $N$ spots and $M$ mutations. The definition of the EI is based on a bivariate characterization $\{{\left(f_i,d_i\right)\}}_{i=1}^M$ of the tumor’s mutations, where $f_i$ is the fraction of DCIS spots in which mutation $i$ is present, and $d_i$ is the normalized diameter of mutation $i$, defined as the maximum Euclidian distance between any two DCIS spots containing the mutation, divided by the maximum Euclidian distance between any two DCIS spots in the tumor. The $EI$ is then obtained by integrating the piecewise constant curve over the $M$ bivariate points

$$EI={\int }_0^{1}h(x)dx,$$

where

$$h\left(x\right)=\underset{i:f_i\le x}{\sup }{d}_i.$$

By definition, $EI\in \left[\mathrm{0,1}\right].$ If mutation diameter grows approximately linearly with the fraction of occupied spots, then $EI\approx 0.5$, indicative of a continuous patch lesion. If there are mutations with a large diameter at a low fraction of occupied spots, $EI\gg 1$, indicative of a disperse skip lesion.

Mutation energy

This statistic was introduced to quantify the mutational diversity of the tumor. Again, uncertainty quantification was performed as described above, and we focus here on the derivation of the statistic for a single realization of the binary array $S=\left(s_{ij}\right)\in {\mathbb{R}}^{N\times M},$ for a tumor with $N$ spots and $M$ mutations, each of which was detected in ≥1 spot(s). First, we applied hierarchical column clustering (using the Manhattan distance) to obtain the spot genotype-clustered array $\tilde{S}=\left({\tilde{s}}_{ij}\right)\in {\mathbb{R}}^{N\times M}.$ Next, in analogy with the Ising model from statistical mechanics,(35) we defined the mutation energy $I_k$ of mutation $k$ as

$$I_k=\frac{1}{N‐1}{\sum }_{j=2}^N|{\tilde{s}}_{kj}‐{\tilde{s}}_{k\left(j‐1\right)}|,$$

where the normalizing factor accounts for the $N‐1$ possible flips and ensures that $I_k\in \left[\mathrm{0,1}\right]$ irrespective of the number of spots in the tumor. Intuitively, $I_k$ measures, for each mutation, the number of “flips” from “absent” to “present” along the rows of the spot-mutation array (e.g., Figure S5). If there is only a single spot genotype in the tumor, then $I_k=0$ for all $k$; and $I_k$ increases as both the number of different spot genotypes and their degree of dissimilarity increase. To quantify the mutation energy at the tumor rather than individual mutation level, we used the median and interquartile range (IQR) across all detected mutations in the tumor.

Mathematical models of DCIS growth

(For details, see Supplementary Methods). To model DCIS growth, we combined a stochastic model of the binary ductal tree structure with a stochastic model of the cellular DCIS growth dynamics. The ductal tree model was based on the experimentally delineated dynamics of branching ductal morphogenesis, that is ductal elongation followed by either branching into two daughter ducts, or branch termination, with equal probability.(17)

Tumor growth along the ductal tree architecture was initiated by random seeding of the first DCIS cell. Growth from this first cell to the macroscopic tumor was modeled as a two-stage process, consisting of an initial exponential expansion subject to the mutation bursts of punctuated evolution (with mutation rate $\mu$ per cell division), followed by an expansive growth along the branching tree structure. To describe the expansive growth phase, we considered three competing models as follows.

Model 1, or Comet model, mimics the con-competitive cellular dynamics that govern pubertal branching morphogenesis of the mammary gland.(17,19) In this model, the DCIS end buds are nucleated by a pool of $N$ long-lived cancer cells, half of which undergo intermittent asymmetric division followed by $n_{TA}$ generations of transit-amplification, and half of which remain quiescent. As the end buds of the growing tumor thus move along the pre-existing ducts, the transit-amplifying progenies of the dividing end bud cells contribute to the growing tumor. Upon reaching a ductal branching point in the pre-existing tree, the long-lived end bud cells are randomly divided between the two daughter branches, and after a round of duplication, the two newly created end buds begin to grow along the respective daughter ducts.

Model 2 is a variation of Model 1, whereby all cells in the DCIS end bud are assumed to undergo intermittent asymmetric division. This variation of the Comet model was introduced to assess its sensitivity to the separation of proliferating and quiescent end bud cells.

Model 3 is a canonical cancer evolution model characterized by uncontrolled proliferation and competition among DCIS cells. To account for spatial crowding and resource constraints behind the actively growing tips of the tumor, we formulated a boundary growth model where only the $N$ cells immediately behind the growing tips contribute to the net growth of the elongating DCIS duct. The same branching dynamics as in Models 1 and 2 were applied.

Model fitting and model selection

(For details, see Supplementary Methods). We used a rejection sampling-based version of approximate Bayesian computation to fit the models to the experimental data, estimate the posterior parameter distributions ($N,\mu ,n_{TA})$, and identify the best fitting model.(36,37) For a given model, we sampled a set of parameters from the prior distributions, simulated a ductal tree and DCIS tumor as described above, and compared the simulated tumor against the experimental tumors in our study cohort using a distance function. By keeping only parameter sets resulting in simulated tumors that were sufficiently similar to the experimental data—that is, the distance between simulation and experiment was below a specified threshold—we thus approximated the posterior parameter distributions. Finally, we used a joint model-parameter space approach(38) to compute the posterior marginal model probabilities and calculate the Bayes’ factors for model selection and we performed posterior predictive checks for the different summary statistics by repeated sampling (M=1,000) from the joint posterior distribution.

Statistical analyses

All statistical analyses were performed using R version 4.2.0 (R Foundation for Statistical Computing, Vienna, Austria; RRID:SCR_001905). All statistical tests were two-sided. Data visualizations were made with R, using the packages ggplot2 (v3.3.6; RRID:SCR_014601), ggbeeswarm (v0.6.0; RRID:SCR_026875), and circlize (v0.4.16; RRID:SCR_002141)(39).

Data and materials availability

The whole exome sequencing data is deposited in the Sequence Read Archive database (https://www.ncbi.nlm.nih.gov/sra; RRID: SCR_004891) under accession code SRP298346. Targeted DNA sequencing data are included as Supplementary Datasets S1-S2; gene annotations are included Supplementary Dataset S3. Computer code used to generate the results is available at https://github.com/mdryser/D5_DCIS (MIT License). All other raw data are available upon request from the corresponding authors.

RESULTS

Spatial mutation topographies

We identified 18 women who had undergone surgery for a diagnosis of screen-detected DCIS (Table S1), including 9 patients with DCIS alone (pure DCIS), and 9 patients with DCIS tumors adjacent to invasive breast cancer (synchronous DCIS). From each surgical specimen we obtained between 2 and 5 spatially separated formalin-fixed and paraffin-embedded (FFPE) tissue regions, and in each tissue section we microdissected (40) and spatially registered small regions, or spots, each containing approximately 100 to 500 epithelial cells (Figure 2A, Figure S1). For each DCIS spot, we determined the genotype through targeted sequencing of tumor-specific mutation panels derived from whole exome sequencing (WES) of macro-dissected DCIS foci.

Figure 2: Multiregional sequencing reveals spatial mutation topographies of DCIS tumors.

(A) Between 2 and 5 spatially separated microscope sections were obtained from 18 DCIS tumors. From each microscope slide, small tissue areas (spots) were microdissected, spatially registered, histologically annotated, and genotyped. Genotyping was based on targeted sequencing of tumor-specific mutation panels that had been derived from whole exome sequencing analyses of macro-dissected DCIS areas. (B) Summary of the genetic spot data for all 18 DCIS tumors. Each sector groups together spots of the same tumor, and tumor labels are shown at the periphery. Differences in height of the outermost track (mutation calls) reflect the varying mutation panel sizes for each tumor. (C) Spatial pattern of a select mutation in DCIS-91 (gene: RAB17, chr2:238484162). Colors the mutation status.

After eliminating germline mutations and low-quality targets (Figure S3), the final study cohort comprised 313 genotyped DCIS spots across 60 tissue sections (Figure 2B). A total of 823 (median per tumor: 45, range: 24–66) mutation targets were identified by WES, of which 544 (66%; median per tumor: 31, range: 8–59) mutations were detected by targeted sequencing (Figure S5). Across all 544 mutations we identified two de novo mutational signatures that matched established consensus signatures implicated in carcinogenesis (Figure S6). Across the 18 DCIS tumors we identified a total of 21 putative driver mutations (median per tumor: 1, range: 0–3) (Table S2). In synchronous DCIS, the putative driver mutations were invariably present in adjacent invasive cancer spots (Table S3). Combining the genotypic spot characterizations with the spatial tumor maps, we constructed geospatially annotated somatic mutation topographies for each DCIS (Figure 2C).

Multiclonal ducts and intra-tumor heterogeneity

The variant allele frequencies (VAFs) of somatic mutations within individual spots contain valuable information about the structure of local cell populations (Figure 3A). Given the high tumor purity of microdissected duct cross-sections, VAFs of 50% or greater indicate locally clonal mutations that are present in all resident cells, whereas VAFs below 50% are suggestive of locally subclonal mutations carried by a subpopulation of resident cells only. Across the DCIS spots in our cohort, the within-spot VAF spectra of detected mutations were generally subclonal and dispersed, as evidenced by low median values and high inter-quartile ranges, respectively (Figure 3B). Because clonal cell populations can exhibit VAFs below 50% in the presence of copy number changes, we performed additional low-pass whole genome sequencing of individual duct sections in a representative tumor (DCIS-286) and calculated ploidy-adjusted cancer cell fractions (CCFs; Figure S7). All 9 duct sections contained multiple mutations with a CFF<0.9, confirming the multiclonal nature of these DCIS spots. In summary, and consistent with several prior studies,(13,16,41) these data suggest that most DCIS ducts in our cohort contain an admixture of distinct genetic subclones.

Figure 3: DCIS tumors are multiclonal and spatially heterogeneous.

(A) The variant allele frequency (VAF) spectra of detected mutations are shown for 4 select spots in DCIS-173; the VAF of two select mutations in the genes SFXN1 (blue) and NAF1 (red) are highlighted. (B) Bivariate summary statistics for spot-level VAF spectra are shown across all DCIS spots (n=313) of the 18 tumors, with median VAF on the x-axis, and interquartile range (IQR) of the VAF on the y-axis. Red color scheme visualizes spot density. (C) Mutation patterns for all DCIS spots in DCIS-168 are organized by hierarchical clustering of mutations (rows) and spatial clustering of spots (columns); spatial clustering was based on one-dimensional t-distributed stochastic neighbor embedding (t-SNE) of the spots’ spatial coordinates. (D) Mutation patterns for all DCIS spots in DCIS-173, see panel C for details and color legend. (E) For each tumor, the spatial correlations of DCIS spot genotypes were quantified using Pearson’s R; DCIS-222 was excluded because it had only 2 DCIS spots. Monte Carlo sampling was used to account for posterior uncertainty of mutation calls, resulting in predicted means (circles) and 95% prediction intervals (bars). Median predicted mean correlation was −0.01, without detectable differences between pure DCIS and synchronous DCIS with adjacent invasive cancer (p=.81, Wilcoxon rank-sum test).

To quantify the degree of genetic intratumor heterogeneity (ITH) in individual tumors, we defined spot genotypes as vectors of binary mutation calls (Figures 3C-D, Figure S5). While some DCIS comprised only few distinct spot genotypes (e.g., Figure 3C), most contained a substantial number of distinct genotypes (e.g., Figure 3D), which is indicative of pervasive ITH. Notably, the mutation panels suggest a lack of spatial clustering of similar spot genotypes, suggesting limited spatial correlations of duct genotypes. We confirmed this observation by computing the correlations of spatial and genetic spot distances (Figure 3E) and found that most tumors exhibited low spatial-genetic correlations (median: -.01), without detectable differences between pure and synchronous DCIS (p=.81, Wilcoxon rank-sum test).

In summary, these findings support the presence of multiclonal ducts and extensive spatial heterogeneity,(13–16) but do not address when and how such ITH arises during tumorigenesis. To investigate this, we turned our attention to the spatial topographies of individual somatic mutations.

Expansive skip lesions favor a model of early evolution

We categorized mutations as public (present in ≥90% of DCIS spots in the tumor) or restricted (present in <90% of DCIS spots); the latter are particularly informative because they allow for tracking of individual subclones in space. Across the 17 tumors with more than 2 DCIS spots, we identified a total of 356 restricted mutations (Table S1). Interestingly, restricted mutations often spanned expansive but discontinuous tumor regions of up to 7cm in diameter, and in 13 of 17 tumors, one or more restricted mutations covered the entire DCIS portion (Figure 4A; sensitivity analysis in Table S4). This finding of expansive mutational skip lesions is consistent with two recent studies that performed spatial subclone mapping in DCIS tumors.(16,42)

Figure 4: DCIS mutations form expansive skip lesions.

DCIS-222 was excluded because it only had 2 DCIS spots. (A) The diameter of restricted mutations (found in <90% of spots; black dots) relative to the extent of the DCIS tumor itself (bar). (B) Scattered mutations are characterized by a lack of spatial separation between spots that do and do not contain the mutation. An example from DCIS-91 is shown. Grey rectangles represent the microscope sections (x-y plane) along the tumor’s long (z-) axis. (C) Contiguous mutations are characterized by a spatial separation of spots that do and do not contain the mutation. An example from DCIS-168 is shown; see also description of panel B. (D) Individual mutations are classified as disperse (e.g., P) or contiguous (e.g., Q) depending on the distance they span (normalized diameter) relative to the fraction of spots they cover. For contiguous mutations, the diameter increases no faster than the fraction of spots covered (yellow shaded area); for scattered mutations, the normalized diameter increases faster (blue shaded area). The expansion index (EI) is a tumor-level measure of the extent of mutation dispersion and ranges from contiguous ($E\le 0.5$) to disperse ($EI>0.5$). (E) Summary of EIs across the cohort. Monte Carlo sampling was used to account for posterior uncertainty of mutation calls, resulting in predicted means (circles) and 95% prediction intervals (bars). Median EI was 0.74 across all tumors, without detectable difference between pure DCIS (median: 0.71) and synchronous DCIS with adjacent invasive cancer (median: 0.74; p=0.88, Wilcoxon rank-sum test).

Mutational skip lesions can arise by two distinct mechanisms, depending on whether evolution takes place early or late in the growth process. In the early evolution scenario, the subclonal mutations arise during the early expansion from the first DCIS cell and then disperse across the ductal tree during expanding tumor growth. In the late evolution scenario, the mutations arise late during tumor expansion and disseminate across the tree through extensive sweeps, in competition against less fit subclones.

Delineation of these two scenarios is possible because they predict different types of spatial mutation patterns. In the early evolution scenario, the passive dissemination of early mutations is expected to produce scattered mutation topographies, or ‘skip’ lesions (Figure 4B). In the late evolution scenario, late mutations that expand through subclonal sweeps are expected to produce more contiguous mutation patches (Figure 4C). To test these predictions against the data, we introduced the expansion index (EI), which ranges from 0 to 1 and measures whether a tumor is dominated by disperse ($EI>.5$) or contiguous ($EI\le 0.5$) mutations. In brief, for each mutation we determined the number of spots covered, and the maximum distance (diameter) spanned; at the same number of spots covered, mutations that have a large diameter are disperse, whereas those with a small diameter are more contiguous (Figure 4D). At the tumor level, we summarized the extent of dispersed mutations using the expansion index (Methods, Figure S4). The median EI across all tumors was 0.74, and 12/17 (71%) tumors were disperse, with an $EI\ge .6$ (Figure 4E). There was no detectable difference in EI between pure DCIS (median: 0.71) and synchronous DCIS (median: 0.74, Wilcoxon rank-sum test: p=0.88). The elevated expansion indices are indicative of mutational skip lesions and suggest that the widespread ITH is likely due to the passive dissemination of early subclones, that is the early evolution scenario.

Two additional observations provide evidence against the late evolution scenario of mutation dissemination. First, expansive subclonal sweeps are expected to yield locally homogeneous ducts,(43) which is at odds with the observation of subclonal VAFs at the spot level (Figure 3B, Figure S8).(16) Second, expansive subclonal sweeps would require the acquisition of a substantial cellular fitness, yet we only found a limited number (n=21) of putative driver mutations in our cohort (Table S2), and there was no evidence that driver mutations were more disperse than passenger mutations (Figure S9).

We note that pervasive copy number changes producing spatially localized losses of mutant alleles could also account for disperse mutation patterns. To explore this possibility, we performed spatial copy number profiling across spots of a large (DCIS-286) in our cohort. Copy-number profiles across DCIS ducts were stable, and in spots where both copy number and DNA mutation data were available, none of the absent mutations coincided with an allelic loss (Figure S7).

In summary, our data support a model of early evolution where genetic subclones arise during the initial expansion from the first DCIS cell and then disperse across the ductal tree through expansive tumor growth. What remains unclear, however, are the cellular dynamics that govern this expansive growth phase.

Mutation data supports a non-competitive growth model of DCIS

The scattered mutation topographies of DCIS tumors are strikingly similar to the subclone patterns that characterize the normal pubertal development of murine mammary ductal trees.(17,19) Indeed, during breast development, individual mammary stem cells contribute to ductal expansion only intermittently to produce dispersed subclone patterns along the branching ductal tree. Based on these similarities, we propose the ‘Comet model’ of DCIS tumorigenesis which recapitulates the stochastic fate rules of ductal elongation and binary branching that characterize pubertal branching morphogenesis.(17,19)

The Comet model posits that DCIS growth, which is confided to the ductal tree, is driven by the expanding end buds of the tumor front which contain populations of long-lived neoplastic cells that arise early in evolution (Figure 5A). These long-lived cells stochastically undergo episodic expansion to produce the subclone populations that populate the mammary ducts with DCIS. When an expanding tumor bud reaches a branching point in the ductal tree, the long-lived cells are randomly divided between the two daughter ducts and then duplicate. Such backward seeding of subclones—reminiscent of a comet’s tail—naturally results in multiclonal DCIS ducts and expansive mutational skip lesions across the involved portions of the mammary tree (Figure 5B). Simulations of the Comet model illustrate the expansive dispersion of subclonal mutations and high levels of ITH (Figure 5C).

Figure 5: The Comet model of DCIS tumorigenesis.

(A) A modified Muller plot illustrating the typically observed data in our cohort. After initial expansion of early subclones, the growth patterns are characterized by multiclonal ducts and disperse skip lesions. (B) The Comet model of DCIS growth mimics the cellular dynamics that govern pubertal branching morphogenesis. During DCIS expansion along existing ductal segments (top), the long-lived neoplastic cells of the DCIS end bud undergo intermittent proliferation; after transit-amplification, the clustered progenies of the long-lived cells become embedded in the growing multiclonal DCIS duct. As the growing tumor reaches a branching point in the ductal tree (bottom), the end bud cells are randomly distributed between the two daughter branches where they duplicate, and the two resulting end buds start growing along their respective daughter branches. (C) Mutation patterns resulting from the Comet model. Left: DCIS growth is initiated at the starting node and propagated across the ductal tree, with pie charts indicating the local variant allele frequencies (VAFs) of a select mutation. Right: the hierarchically clustered mutation pattern corresponding to the simulation in the left panel, illustrating the local presence/absence of mutations (rows) across the examined duct cross-sections (columns). (D) A modified Muller plot illustrating the expected subclone frequencies that arise from a canonical model of cancer evolution along the ductal tree. Initial expansion of the first DCIS cell and subsequent branching growth are governed by quasi-neutral clonal evolution. Due to the thin tube-like geometry of the ducts, individual subclones are expected to rapidly go extinct or fixate, resulting in monoclonal ducts. (E) As in C, but instead using a canonical model of cancer evolution, see Methods for details. (F) Posterior predictive checks for the Comet model (yellow) and canonical clonal evolution model (blue) compared to the data from our cohort (red), shown for each of the 8 summary statistics used for approximate Bayesian computation (Methods). For comparison, and where possible, corresponding summary statistics were derived from the spatial subclone data reported by Lomakin and colleagues(16) (gray: section P1-D1; blue; P1-D2). Horizontal bars represent median values.

The mutation topographies produced by the Comet model contrast with those that arise through the uncontrolled cellular proliferation and subclone competition of clonal evolution. Indeed, when combined with the branching topology and thin tube-like geometry of the ductal tree, these dynamics result in rapid stochastic fixation or extinction of individual mutations along the ductal tree(43) (Figure 5D). Simulations show that, compared to the Comet model, these dynamics lead to a smaller number of subclones and limited ITH (Figure 5E).

For a formal comparison of the two growth mechanisms, we developed a computational platform that mimics our experimental design, see Methods and Supplementary Methods for details. In brief, we simulated a stochastic ductal tree, randomly seeded the first tumor cell, simulated DCIS growth dynamics across the tree, and recorded the respective simulated VAFs of ducts sampled from the final tumor. Fitting the simulations to our experimental data using approximate Bayesian computation, we found that the non-competitive Comet model provided an overall better fit to the data compared to the competitive clonal evolution model (Figure 5F). We formalized the model comparison using Bayesian model selection and confirmed that the Comet model produced a superior fit to the data compared to the clonal evolution model (Bayes’ factor(44) of 11.7; Table S5, Figure S10).

DISCUSSION

Based on the mutation topographies of 18 large human DCIS tumors, we propose the Comet model of DCIS growth. The Comet model posits that multiple genetic subclones arise shortly after the first DCIS cell and then disperse across the ductal tree through a non-competitive growth process that mimics the cellular dynamics of branching morphogenesis from normal pubertal breast development.

Because of its histologic and genomic similarity with invasive breast cancer, DCIS is often considered “just one step” away from invasion. Yet this characterization is at odds with a growing recognition that most DCIS tumors remain non-invasive for decades if left untreated.(6,45) The Comet model offers a potential solution to this clinical incongruency. Indeed, the branching morphogenesis of normal breast development is a regulated expansion where mobile progenitor cells proliferate, differentiate, and branch to form new ductal elements but remain organized within the basement membrane that defines the structure of breast ducts. By mimicking the cellular dynamics of this developmental program of mobile expansion, many DCIS tumors can exhibit normal-like growth into macroscopic yet stable neoplasms without reliance on the uncontrolled proliferation and abnormal organization of invasive cancer. Importantly, in contrast to neoplastic growth governed by clonal evolution, the Comet dynamics are not subject to incessant subclone competition that tends to produce increasingly aggressive phenotypes. This non-competitive model thus provides a simple explanation for the common occurrence of indolent DCIS tumors and provides biologic rationale for an evolving clinical paradigm that seeks to de-escalate treatment in low-risk DCIS patients.(5,46,47)

The Comet model is consistent with previously reported multiclonality and intratumor heterogeneity of DCIS tumors,(12–15) and expands this knowledge with an explanation for the co-occurrence of duct-level multiclonality and global subclone dispersal. While the origins of multiclonality per se can be explained(13) by an early punctuated burst of genomic instability,(48–50) the simultaneous occurrence of duct-level multiclonality and global subclone dispersal have been difficult to reconcile.(16) Indeed, under a canonical model of cancer growth—characterized by uncontrolled proliferation and clonal evolution—the thin tube-like mammary ducts are expected to accelerate local sweeps, resulting in contiguous patches of monoclonal DCIS ducts.(43) In contrast, the Comet model posits that the tumor’s expanding end buds contain multiple, episodically proliferating subclones that produce local multiclonality and global subclone dispersal.

Similar studies performed in colorectal cancers (CRC)(27,48,51) provide a direct comparison of cancer growth patterns between the two organs. In both sites, growth is driven by long-lived progenitor cells, situated in the growing end buds of DCIS and at the base of CRC glands,(52) respectively. Furthermore, the branching of DCIS ducts is analogous to the fission of cancer glands during CRC growth.(52) Yet while the transit-amplifying progenies of CRC stem cells exit the gland within a few days, their DCIS counterparts are embedded in the expanding duct and provide a genomic record of the end buds’ proliferative activity during growth. This difference can explain why CRC glands are generally monoclonal populations dominated by a single fixated subclone, whereas duct cross-sections contain multiple subclones. This comparison highlights the likely role of tissue architecture in shaping the mode of evolution.(43,53,54)

While Bayesian model selection favored the Comet model over a canonical model of cancer evolution, we acknowledge two alternative scenarios that could also account for the observed mutation patterns. First, a recent study of five der(1;16)-positive breast cancers suggests that a burst of genomic instability may occur during early puberty,(42) potentially leading to the passive dispersion of subclones as the ductal tree undergoes developmental branching morphogenesis.(17) These dispersed subclones may subsequently undergo competitive local expansions during repeated estrous cycles,(55) thus contributing to the discontinuous mutational landscapes observed in DCIS. While such field cancerization effects may contribute to multifocal DCIS, they are difficult to reconcile with the unifocal and spatially cohesive lesions analyzed in this study. In particular, there is no evidence that spatially dispersed subclones can remain dormant for decades and then synchronously generate a single continuous DCIS lesion. Second, we cannot exclude the possibility that long-range seeding of individual cells may give rise to the mutational skip patterns observed in our data. However, given the lack of direct evidence for such cellular migration across macroscopic distances within the densely packed architecture of DCIS ducts, the Comet model provides the most parsimonious explanation for the coexistence of duct-level multiclonality and global subclone dispersal observed in our data. DCIS is a heterogeneous disease, and mechanisms of spread may differ between individuals. Future studies leveraging recent technological advances in combined 3D imaging and spatial genomics,(56,57) applied in large and representative DCIS cohorts, will enable a more direct investigation of these alternative growth models.

Our study has limitations. First, due to sequencing constraints in FFPE samples, spot selection was biased toward larger ducts that could yield sufficient genetic material for analysis. While this may have led to an underestimation of overall heterogeneity, our conclusion of local multiclonality and global subclone dispersion would not be significantly altered by the inclusion of smaller ducts. In addition, the naturally sparse and irregular three-dimensional DCIS embeddings led to variability in genetic material across tissue blocks, with some lacking sufficient DNA for sequencing, resulting in centimeter-wide gaps between assayed sections. However, the presence of spatially discontinuous mutation patterns is inherently invariant to additional sampling, ensuring the robustness of our main observations. Second, since our cohort was composed of intermediate to high grade and mostly hormone receptor positive DCIS with solid or cribriform growth patterns, the Comet dynamics may not be applicable to other pathologic subtypes, such as micropapillary and low-grade DCIS. Third, the patient-specific geometry of the latent ductal tree structure may have influenced the spatial organization of the DCIS tumors and thus the observed mutation patterns. To account for this uncertainty, our modeling approach incorporated Monte Carlo sampling of many possible underlying tree structures when fitting the models to the mutation data. In future work, the combination of three-dimensional mammary tree reconstruction alongside the spatial genomic characterizations could further resolve this source of uncertainty.(56,57) Fourth, while it is commonly assumed that DCIS cells grow along the pre-existing mammary ductal tree, an alternative model of neoductogenesis(58) proposes that DCIS may branch off the pre-existing tree to grow its own subtrees. However, as long as the subtrees resulting from neoductogenesis retain the topological structure of the normal mammary tree, our mathematical models remain applicable.

An expansive and structured penetration of the breast stroma in the absence of invasion and metastasis is an inherent feature of branching morphogenesis during normal pubertal breast development. In this study, we provide evidence that DCIS cell migration may be governed by cellular dynamics that mimic this non-competitive developmental process, resulting in indolent tumors that are susceptible to mammographic overdiagnosis. Interestingly, developmental branching morphogenesis is not unique to the breast and is equally implicated in the development of the prostate, thyroid, and lung.(19,59,60) The intriguing observation that cancer overdiagnosis is common in these organs(61–63) raises the possibility that non-competitive growth dynamics reminiscent of developmental branching morphogenesis could be a contributing factor to the etiology of indolent tumors in the prostate, thyroid, and lung as well.

SIGNIFICANCE.

The spatial mutation landscape in ductal carcinoma in situ supports a non-competitive growth model mirroring normal branching morphogenesis, which has implications for managing low-risk patients and for cancers in other branching organs.

Abbreviations:

DCIS

ductal carcinoma in situ

IBC

invasive breast cancer

FFPE

formalin-fixed, paraffin-embedded

SURF

(selective ultraviolet light fractionation)

WES

whole-exome sequencing

SNV

single nucleotide variant

VAF

variant allele frequency

CNV

copy number variation

CCF

cancer cell fraction

ITH

intratumor heterogeneity

EI

expansion index

PCR

polymerase chain reaction

NGS

next-generation sequencing