Journal of Pathology J Pathol May 2022; 257: 68–81 ORIGINAL ARTICLE Published online 6 April 2022 in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/path.5870 Copy-number intratumor heterogeneity increases the risk of relapse in chemotherapy-naive stage II colon cancer Sara Lahoz1 , Ivan Archilla2 , Elena Asensio1, Eva Hernandez-Illan1, Queralt Ferrer1, Sandra L opez-Prades2, Ferran Nadeu , Javier Del Rey , Rebeca Sanz-Pamplona , Juan José Lozano , Antoni Castells , Miriam Cuatrecasas2* 3 4 5 6 1 and Jordi Camps1,4* 1 Translational Colorectal Cancer Genomics, Gastrointestinal and Pancreatic Oncology Team, Institut d’Investigacions Biomèdiques August Pi i Sunyer (IDIBAPS), Hospital Clínic de Barcelona, Centro de Investigacion Biomédica en Red de Enfermedades Hepaticas y Digestivas (CIBEREHD), University of Barcelona, Barcelona, Spain 2 Pathology Department, Biomedical Diagnostic Center, Hospital Clínic de Barcelona, Institut d’Investigacions Biomèdiques August Pi i Sunyer (IDIBAPS), Centro de Investigacion Biomédica en Red de Enfermedades Hepaticas y Digestivas (CIBEREHD), University of Barcelona, Barcelona, Spain 3 Molecular Pathology of Lymphoid Neoplasms, Institut d’Investigacions Biomèdiques August Pi i Sunyer (IDIBAPS), Hospital Clínic de Barcelona, Centro de Investigacion Biomédica en Red de Cancer (CIBERONC), Barcelona, Spain 4 Department of Cell Biology, Physiology and Immunology, Faculty of Medicine, University Autonomous of Barcelona, Bellaterra, Spain 5 Unit of Biomarkers and Susceptibility, Oncology Data Analytics Program (ODAP), Catalan Institute of Oncology (ICO), Oncobell Program, Bellvitge Biomedical Research Institute (IDIBELL) and CIBERESP, l’Hospitalet de Llobregat, Spain 6 Bioinformatics Platform, Centro de Investigacion Biomédica en Red de Enfermedades Hepaticas y Digestivas (CIBEREHD), Madrid, Spain *Correspondence to: J Camps or M Cuatrecasas, Gastrointestinal and Pancreatic Oncology Team, Institut d’Investigacions Biomèdiques August Pi i Sunyer (IDIBAPS), Rossello 149–153, 4th Floor, 08036 Barcelona, Spain. E-mail: jcamps@clinic.cat (J Camps) or mcuatrec@clinic.cat (M Cuatrecasas) Abstract Optimal selection of high-risk patients with stage II colon cancer is crucial to ensure clinical benefit of adjuvant chemotherapy. Here, we investigated the prognostic value of genomic intratumor heterogeneity and aneuploidy for disease recurrence. We combined targeted sequencing, SNP arrays, fluorescence in situ hybridization, and immu- nohistochemistry on a retrospective cohort of 84 untreated stage II colon cancer patients. We assessed the clonality of copy-number alterations (CNAs) and mutations, CD8+ lymphocyte infiltration, and their association with time to recurrence. Prognostic factors were included in machine learning analysis to evaluate their ability to predict individ- ual relapse risk. Tumors from recurrent patients displayed a greater proportion of CNAs compared with non-recurrent (mean 31.3% versus 23%, respectively; p = 0.014). Furthermore, patients with elevated tumor CNA load exhibited a higher risk of recurrence compared with those with low levels [p = 0.038; hazard ratio (HR) 2.46], which was con- firmed in an independent cohort (p = 0.004; HR 3.82). Candidate chromosome-specific aberrations frequently observed in recurrent cases included gain of the chromosome arm 13q (p = 0.02; HR 2.67) and loss of heterozygosity at 17q22–q24.3 (p = 0.05; HR 2.69). CNA load positively correlated with intratumor heterogeneity (R = 0.52; p < 0.0001). Consistently, incremental subclonal CNAs were associated with an elevated risk of relapse (p = 0.028; HR 2.20), which we did not observe for subclonal single-nucleotide variants and small insertions and deletions. The clinico-genomic model rated an area under the curve of 0.83, achieving a 10% incremental gain com- pared with clinicopathological markers (p = 0.047). In conclusion, tumor aneuploidy and copy-number intratumor heterogeneity were predictive of poor outcome and improved discriminative performance in early-stage colon cancer. © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd on behalf of The Pathological Society of Great Brit- ain and Ireland. Keywords: stage II colon cancer; cancer genomics; copy-number alterations; mutational profiling; intratumor heterogeneity; machine learning Received 28 July 2021; Revised 17 December 2021; Accepted 13 January 2022 No conflicts of interest were declared. Introduction years after curative intended surgery, which compro- mises survival rates. Although extensive biomarker- Early-stage colon cancer poses a major therapeutic chal- driven research has succeeded in identifying risk lenge due to the lack of strong biomarkers to predict dis- predictors for tumor dissemination, selection of patients ease recurrence [1]. Around 10–15% of patients for adjuvant chemotherapy is still controversial in stage diagnosed with stage II colon cancer relapse within 5 II colon cancer [2]. © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd on behalf of The Pathological Society of Great Britain and Ireland. This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made. Copy-number heterogeneity predicts colon cancer relapse 69 To date, the most relevant clinical and pathological Materials and methods factors for discriminating high-risk stage II individuals are bowel perforation or obstruction, tumor size and high Cohorts of patients and tumor samples histological grade, lymphovascular or perineural inva- sion, and serosal invasion, although most of them have A total of 84 stage II colon adenocarcinomas a modest individual effect on recurrence risk [3]. Micro- (pT3-4N0M0) provided by the Hospital Clínic of satellite instability (MSI) is one of the most solid molec- Barcelona/IDIBAPS Tumor Biobank were retrospec- ular markers with clinical utility in non-advanced colon tively analyzed for this study. Patients were diagnosed cancer, and along with BRAFV600E mutation, is able to between 2005 and 2016, and none of them received adju- define a subgroup of stage II colon cancer patients vant chemotherapy after surgical resection of the primary (~6%) with improved survival when solely treated with tumor. Among them, 38 patients (45%) developed dis- surgery [4]. Tumor budding and the presence of poorly ease recurrence and 46 (55%) did not show recurrence differentiated clusters are pathology-based markers that after a median follow-up time of 7.6 years (range from also contribute to the prognostication of these patients <1 to 14.5 years). The primary endpoint was time to [5]. Moreover, Immunoscore is a robust immuno recurrence (TTR), defined as the time from surgery of histochemistry-based prognostic index for quantifying the primary tumor to disease recurrence, where death tumor-infiltrating CD8+ T cells that has been thoroughly without recurrence was censored at the time of death. validated in stages II–III colorectal cancer (CRC) [6,7]. This study was approved by the institutional ethics com- Notwithstanding, recommendations in clinical guide- mittee of the Hospital Clínic of Barcelona (register lines for risk stratification still rely on under-sensitive HCB/2018/0174), and all patients signed an informed clinical and histopathological criteria. consent form in accordance with the Declaration of Genomic intratumor heterogeneity has been identified Helsinki. For analytical validation purposes, an indepen- as a prognostic predictor in solid malignancies [8,9], due dent cohort of 99 untreated stage II microsatellite-stable to its putative ability of prompting somatic evolutionary (MSS) colon cancer patients (Colonomics) was included processes that can drive cancer progression [10] and ther- (EGAD00010001253) [25]. apeutic failure [11,12]. Much of the genetic heterogeneity observed in solid tumors is triggered by chromosomal instability (CIN) [13], resulting in a pervasive expansion SNP arrays and copy-number analysis of chromosomal aneuploidies and mutational events that Copy-number and loss-of-heterozygosity (LOH) profiling increases tumor subclonality and accelerates cancer evo- of tumors was performed using genome-wide Affymetrix lution [14,15]. Such instability could be responsible for OncoScan SNP arrays on unmatched tumor specimens. fueling the tempo-spatial spread of relapse-proficient sub- All resulting CEL files were loaded onto Nexus Copy clones that might influence the eventual metastatic seed Number software version 9.0 (BioDiscovery, El Segundo, [16,17], potentially explaining mechanisms of disease CA, USA) for data analysis and visualized using CNApp recurrence, such as those described in chronic lympho- [26]. Major and minor allele numbers of each segment cytic leukemia [18]. In this sense, increased subclonal were inferred utilizing the Allele-Specific Copy Number copy-number alterations (CNAs) have been associated Analysis of Tumors (ASCAT) R package version 2.5 with a greater risk of recurrence or death in non-small cell [27]. Copy-number intratumor heterogeneity was inferred lung cancer [19] as well as in hepatocellular carcinoma from whole-genome copy-number profiles derived from [20]. Furthermore, ongoing CIN enables cancer cells to bulk tumor SNP-array data. The proportion of tumor cells, rapidly assemble complex karyotypes with inflated bur- or cancer cell fraction (CCF), carrying each CNA was dens of CNAs [21], conferring the tumor a selective mathematically derived as a function of its B-allele fre- advantage towards a more aggressive phenotype [22]. quency and the average major/minor allele numbers esti- Indeed, the tumor CNA load has been proposed as a bio- mated by ASCAT, correcting by the sample purity, marker of reduced survival in prostate cancer [23] and in using the following equation: metastatic CRC was also predictive of a diminished   response to bevacizumab combinatorial therapy [24]. 2  BAF  1 1 CCFCNA ¼  Nevertheless, genomic complexity and intratumor hetero- ½ðmajor  1Þ  ððmajor  1Þ  BAFÞ  ððminor  1Þ  BAFÞ p geneity are yet to materialize as clinically useful tools to monitor cancer progress for therapeutic decision-making. where BAF corresponds to the major frequency (>0.5) Here, we performed an integrative analysis of somatic of the B-allele, major to the number of copies of the CNAs and mutational events, their subclonal status, B-allele, minor to the number of copies of the A-allele, along with the CD8+ immune infiltration and other clin- and p to the sample purity (0 to 1) obtained from icopathological features in stage II colon cancer. By ASCAT. The clonality of all detected CNAs was esti- applying machine learning-based modeling using data mated by inferring the CCF of each chromosomal seg- from single nucleotide polymorphism (SNP) arrays, ment from SNP-array data. CNAs and copy-number next-generation sequencing (NGS), fluorescence in situ neutral-LOH (cnLOH) segments were considered sub- hybridization (FISH), and immunohistochemistry, we clonal when their CCF was less than 85% and clonal suggest that genomic intratumor heterogeneity and otherwise [28], although more stringent CCF thresh- aneuploidy are determinants of tumor dissemination. olds were also tested. Details on data segmentation © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd J Pathol 2022; 257: 68–81 on behalf of The Pathological Society of Great Britain and Ireland. www.pathsoc.org www.thejournalofpathology.com 70 S Lahoz et al and processing are available in Supplementary mate- patients was assessed by means of its area under the rials and methods. curve (AUC) (see extended description in Supplemen- tary materials and methods.) Targeted next-generation sequencing Details for tissue microarrays, FISH, immunohisto- chemistry, prognostic modeling, and statistical analysis The full-coding region of 48 CRC-related genes (sup- are provided in Supplementary materials and methods. plementary material, Table S1) was sequenced using a MiSeq platform (Illumina, San Diego, CA, USA) on a subset of 44 unpaired tumor samples to identify single nucleotide variants (SNVs) and small insertions and Results deletions (indels). Only those mutations detected by two callers, MuTect2 (v2.2) and Illumina Somatic Clinicopathological characteristics and their Variant Caller (v2.6), were considered for ulterior prognostic value in the study population analyses. A bioinformatic pipeline was designed to filter low-quality variants, exclude synonymous and Clinical and tumor-related histopathological characteris- germline SNPs, and to only select mutations with path- tics are listed in Table 1. Median age at diagnosis was ogenic functional effects. A detailed description is pro- 74 years (range 55–91 years) and 54% of the individuals vided in Supplementary materials and methods. were male. Most tumors were pT3 stage (73%) and had low histological grade (88%). Lymphovascular or peri- neural invasion and an invasive infiltrating margin were Predictive modeling by machine learning analysis significantly more frequent in patients with disease Multivariable classification analyses were computed relapse (p = 0.016 and 0.0008, respectively). Recurrent integrating all master co-variables to predict the proba- tumors showed remarkably lower amounts of tumor- bility of relapse for each patient using the caret frame- infiltrating CD8+ lymphocytes compared with non- work in R. After imputation of missing values using recurrent (p = 0.0003) (supplementary material, the mice package and LASSO feature selection using Figure S1A). High tumor budding and poorly differenti- glmnet on each training dataset, the gradient boosting ated clusters were also associated with disease relapse machine algorithm from gbm was employed to classify (p = 0.019 and 0.008, respectively). patients with 10-fold cross-validation to reduce overfit- To investigate the independent prognostic value of ting of predicted data. Discriminative performance of clinicopathological features, Cox proportional hazards the model to discern recurrent from non-recurrent models were fitted for TTR (Table 2). The mean time Table 1. Patient and tumor-related clinicopathological characteristics. Variable Strata All (N = 84) Recurrent Non-recurrent P value patients (N = 38) patients (N = 46) N % N % N % Age Median (range), years 77 (55–91) 79 (55–87) 76 (55–91) 0.24 Sex Male 45 53.57% 23 60.53% 22 47.83% 0.35 Female 39 46.43% 15 39.47% 24 52.17% pT stage T3 61 72.62% 26 68.42% 35 76.09% 0.59 T4 23 27.38% 12 31.58% 11 23.91% Histological grade High 10 11.90% 2 5.26% 8 17.39% 0.10 Low 74 88.10% 36 94.74% 38 82.61% Microsatellite status MSS 74 88.10% 36 94.74% 38 82.61% 0.10 MSI 10 11.90% 2 5.26% 8 17.39% Lymphovascular or perineural invasion Yes 13 15.48% 10 26.32% 3 6.52% 0.016 No 71 84.52% 28 73.68% 43 93.48% Infiltrating margin Invasive 49 58.33% 30 78.95% 19 41.30% 0.0008 Pushing 35 41.67% 8 21.05% 27 58.70% Tumor budding score Bd1 42 50% 13 34.21% 29 63.04% 0.019 Bd2 16 19.05% 8 21.05% 8 17.39% Bd3 26 30.95% 17 44.74% 9 19.57% Poorly differentiated clusters G1 47 55.95% 15 39.47% 32 69.56% 0.008 G2 24 28.57% 17 44.74% 7 15.22% G3 13 15.48% 6 15.79% 7 15.22% Lymphocytic CD8+ infiltration High 50 59.52% 14 36.84% 36 78.26% 0.0003 Low 34 40.48% 24 63.16% 10 21.74% PDL1 expression score (≥5%) High 21 25% 7 18.42% 14 30.43% 0.31 Low 63 75% 31 81.58% 32 69.57% CDX2 expression Positive 81 96.43% 36 94.74% 45 97.83% 0.59 Negative 3 3.57% 2 5.26% 1 2.17% MSI, microsatellite-instable; MSS, microsatellite-stable; PD-L1, programmed death ligand-1. P values ≤0.05 are displayed in bold. © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd J Pathol 2022; 257: 68–81 on behalf of The Pathological Society of Great Britain and Ireland. www.pathsoc.org www.thejournalofpathology.com Copy-number heterogeneity predicts colon cancer relapse 71 upon relapse in our cohort was 1.7 years (range 0.26– with non-recurrent (supplementary material, Table S2). 5.9 years). The strongest independent predictors in the Genomic imbalances affecting chromosome arm 13q11– multivariable setting were CD8+ lymphocytic infiltration q34 showed the most differential frequency between the [p = 0.00019; hazard ratio (HR) 0.25; 95% CI 0.12–0.52] two clinical groups (65% versus 36% of gains in recurrent (supplementary material, Figure S1B) and lymphovascu- and non-recurrent cases, respectively; p = 0.018, lar or perineural invasion (p = 0.0007; HR 3.70; 95% CI FDR = 0.32). Two focal CNAs were exclusively present 1.73–7.91) (supplementary material, Figure S1C). Addi- in recurrent tumors, namely the loss of 17q22–q24.3 tionally, increased tumor budding was also associated (12% versus 0%; p = 0.029, FDR = 0.51) and 14q gain with shorter TTR (p = 0.026; HR 1.56; 95% CI 1.05– (14% versus 0%; p = 0.017, FDR = 0.31) (Figure 1A). 2.31) (supplementary material, Figure S1D). In addition, cnLOH on 17p13.3–p13.1 was also over- represented in recurrent tumors (23% versus 5%; p = 0.02, FDR = 0.40). Simple linear regression analyses Specific chromosomal aberrations and high CNA load showed a significantly positive correlation between in recurrent tumors median FISH signals and log2 ratios from SNP arrays for Genome-wide profiling of CNAs was performed to iden- the two experimentally analyzed aberrations (R = 0.82 tify potential chromosomal regions differentially altered and R = 0.80 for chromosome arms 13q and 17q, respec- in recurrent versus non-recurrent tumors. The most fre- tively; p < 0.0001 in both cases) (supplementary material, quent (>35%) CNAs across all tumors included gains Figure S2A,B), thus reproducing the frequencies of CNAs on chromosomes 7, 8q, 13, and 20, and losses (>20%) previously detected by SNP arrays (Figure 1B–D). affecting chromosome arms 5q, 8p, 14q, 15q, 17p, and Since aneuploidy often increases during cancer pro- 18q (Figure 1A). The most commonly altered minimal gression, the fraction of aberrant genome (hereafter region (<2 Mb) of gain was located at 20q13.32– referred as to CNA load) was derived from each tumor 20q13.33 (71%), and the most repeatedly minimal peak SNP-array segmented data. The mean CNA load across of loss was 18q21.2–18q21.31 (61%). Of note, the fre- tumors was 27%, ranging from 0% to 66%. Tumors from quencies of genomic imbalances observed in this study patients with relapse exhibited significantly greater CNA are in line with those previously reported [29]. loads in comparison to those without relapse, with mean Several chromosomal regions appeared significantly values being 31.3% in recurrent versus 23% in non- more altered in tumors from recurrent patients compared recurrent tumors (p = 0.014) (Figure 1E). When MSI- Table 2. Univariate and multivariable Cox models for TTR, stratified by annotated clinicopathological and tumor-genomic variables. Variable Strata Univariate analysis Multivariable analysis HR (95% CI) P value HR (95% CI) P value Age – 1.03 (0.99–1.07) 0.16 1.03 (0.99–1.07) 0.17 Gender Female versus male 0.63 (0.33–1.21) 0.17 0.63 (0.31–1.29) 0.21 pT stage T4 versus T3 1.29 (0.65–2.56) 0.46 1.05 (0.51–2.16) 0.89 Histological grade High versus low 0.46 (0.11–1.93) 0.29 0.47 (0.11–1.99) 0.31 Microsatellite status MSS versus MSI 2.38 (0.57–9.91) 0.23 1.33 (0.28–6.39) 0.72 Lymphovascular or perineural invasion Yes versus no 2.94 (1.41–6.10) 0.004 3.70 (1.73–7.91) 0.0007 Infiltrating margin Invasive versus pushing 3.76 (1.72–8.23) 0.0009 3.48 (1.54–7.89) 0.0028 Tumor budding score Bd2 versus Bd1 1.73 (0.72–4.17) 0.22 1.89 (0.73–4.88) 0.19 Bd3 versus Bd1/2 2.34 (1.23–4.47) 0.009 1.97 (0.98–3.98) 0.057 Continuous 1.67 (1.16–4.40) 0.0056 1.56 (1.05–2.31) 0.026 Poorly differentiated clusters G2 versus G1 3.06 (1.52–6.15) 0.002 3.23 (1.56–6.71) 0.002 G3/2 versus G1 2.56 (1.33–4.92) 0.005 2.59 (1.32–5.09) 0.0058 Continuous 1.49 (1.01–2.20) 0.043 1.46 (0.99–2.17) 0.059 Lymphocytic CD8+ infiltration High versus low 0.26 (0.14–0.51) 0.00009 0.25 (0.12–0.52) 0.00019 PDL1 expression score (≥5%) High versus low 0.54 (0.24–1.23) 0.14 0.59 (0.25–1.39) 0.23 CDX2 expression Positive versus negative 0.24 (0.06–0.99) 0.048 0.06 (0.01–0.38) 0.0028 Chr 13q gain Gain versus diploid 2.76 (1.31–5.85) 0.008 2.67 (1.16–6.14) 0.02 Chr 17q LOH LOH versus diploid 2.46 (1.09–5.54) 0.03 2.69 (0.99–7.35) 0.05 CNA load High versus low 2.31 (1.12–4.79) 0.023 2.46 (1.05–5.74) 0.038 CNA load (Colonomics cohort) High versus low 3.70 (1.58–8.68) 0.003 3.82 (1.54–9.52) 0.004 Subclonal CNAs High versus low 2.19 (1.12–4.30) 0.022 2.20 (1.09–4.44) 0.028 Subclonal SNVs and indels High versus low 1.24 (0.51–2.99) 0.64 1.73 (0.65–4.58) 0.30 KRAS mut Mut versus wild-type 0.74 (0.31–1.73) 0.49 0.76 (0.31–1.87) 0.55 SOX9 mut Mut versus wild-type 0.15 (0.06–0.41) 0.03 0.20 (0.03–1.55) 0.12 NOTCH1 mut Mut versus wild-type 3.61 (1.45–8.98) 0.003 4.24 (1.54–11.69) 0.005 SYNE1 mut Mut versus wild-type 2.90 (1.18–7.16) 0.02 2.62 (0.96–7.15) 0.06 Chr, chromosome; CI, confidence interval; CNA, copy-number alteration; HR, hazard ratio; LOH, loss of heterozygosity; MSI, microsatellite-instable; MSS, microsatellite- stable; PD-L1, programmed death ligand-1. P values ≤0.05 are displayed in bold. Multivariable Cox models include age, sex, stage pT3/T4, histological grade, and microsatellite status as adjusting co-variables. © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd J Pathol 2022; 257: 68–81 on behalf of The Pathological Society of Great Britain and Ireland. www.pathsoc.org www.thejournalofpathology.com 72 S Lahoz et al Figure 1. Genome-wide CNA profiles associated with patient relapse outcome and tumor CD8+ immune infiltration levels. (A) Whole-genome plot showing the accumulated frequency of copy-number alterations per chromosome, in recurrent (N = 38) and non-recurrent (N = 46) patients. (B) Proportion of recurrent and non-recurrent tumors with different copy-number status for chromosome arms 13q and 17q detected by FISH. (C, D) Representative FISH images of gain of chromosome arm 13q and loss of 17q utilizing gene locus-specific probes against CDX2 and SOX9 as a surrogate of each chromosome arm, respectively. (E, F) Violin plots displaying measures of the fraction of aber- rant genome, as a proxy of each tumor level of aneuploidy, in the two clinical groups, in both the Hospital Clínic and the Colonomics cohorts. Black lines indicate median values. P value was obtained by a two-tailed Student’s t-test. (G) Correlation paired analysis between the pro- portion of aberrant genome and levels of intratumor CD8+ lymphocytes analyzed by immunohistochemistry. Rho and P values were obtained using Spearman’s association test. positive cases were excluded, the mean values were (p = 0.058) (Figure 1F). In agreement with previous 32.7% versus 26.8%, respectively (p = 0.075) (supple- findings of decreased cytotoxic immune activity mentary material, Figure S2C). These results were con- by CD8+ T cells in highly aneuploid tumors [30], we firmed in the Colonomics cohort, which consisted of also detected a negative correlation between the CNA 99 MSS stage II colon cancer patients with means of load and CD8+ lymphocyte infiltration (R = 0.31; 34.4% in recurrent versus 23% in non-recurrent tumors p = 0.004) (Figure 1G). © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd J Pathol 2022; 257: 68–81 on behalf of The Pathological Society of Great Britain and Ireland. www.pathsoc.org www.thejournalofpathology.com Copy-number heterogeneity predicts colon cancer relapse 73 Recurrent tumors exhibit an increased number of Figure S3A). In addition, subclonal CNAs appeared to subclonal copy-number alterations positively correlate with CNA load (R = 0.52; p < 0.0001) (supplementary material, Figure S3B), sug- Next, we sought to investigate the association of copy- gesting that intratumor heterogeneity might result from number intratumor heterogeneity with disease relapse. increased aneuploidy. We found considerable levels of subclonality amongst Multi-probe FISH was performed to examine the tumors, with 53% of total CNAs appearing as subclonal dynamics of subclonal CNAs for chromosome regions events, ranging from 0% to 100% per sample. Of those 13q12.2, 17q24.3, and 8q24.21. The results indicated CNAs designated as subclonal, median CCF was 56% that tumors with copy-number gain on 13q, 17q or 8q (9.8–84.9%). Relapsed tumors carried significantly exhibited increased subclonal populations compared more subclonal CNAs than non-relapsed, with a median with those with a diploid status (supplementary material, of 10 versus 7 events per group (two-tailed Mann– Figure S3C–E). Remarkably, when quantifying FISH Whitney U-test, p = 0.02) (Figure 2A,B). Statistical signals for chromosome arm 13q, relapsed carcinomas significance persisted unchanged when more stringent displayed an increased number of subclonal populations CCF cut-offs were tested (supplementary material, with copy-number alterations in comparison to non- Figure 2. Intratumor heterogeneity of CNAs for untreated patients with stage II colon cancer. (A) Bar charts exhibit the number of CNAs that were found to be clonal or subclonal (<85% of cancer cells) per tumor inferred from SNP arrays, according to each patient recurrence status, CNA load, and CD8+ immune infiltration. (B) Representative density plots showing the distribution of CNAs in a recurrent (left panel) and a non-recurrent (right panel) tumor. (C) FISH images of a tumor pair with a gained (left) and a diploid (right) 13q, illustrating divergent levels of subclonal heterogeneity. The FISH panel includes fluorescence-based probes for CDX2 (13q) and the centromeric probe CEP10. (D) Histograms depicting the number of cell populations detected by FISH experimental analysis, from two exemplary tumors with different 13q copy- number status and disparate recurrence outcome. Each subclonal population is defined as having a different copy number for chromosome arm 13q. (E) Box plot exhibiting the number of FISH subclonal populations for chromosome arm 13q, including all tumors from recurrent (N = 38) and non-recurrent (N = 46) patients. P value was obtained by a Mann–Whitney U-test for non-parametric data. © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd J Pathol 2022; 257: 68–81 on behalf of The Pathological Society of Great Britain and Ireland. www.pathsoc.org www.thejournalofpathology.com 74 S Lahoz et al relapsed lesions (p < 0.0001) (Figure 2C–E). In contrast, 44 MSS tumors, including 22 from recurrent patients. we did not observe this association for chromosome Sequencing achieved a mean coverage of 96X in tar- 17q or 8q (supplementary material, Figure S3F–H), geted exon regions. Overall, 43 out of the 48 cancer- further supporting the selective impact of the chromo- related genes showed a mutation in at least one patient some 13q gain in fostering copy-number intratumor het- (Figure 3A). Multi-gene profiling revealed a total of erogeneity and genomic complexity in early-stage colon 198 somatic non-synonymous mutations falling in tumors. protein-coding regions, accounting for both single nucleotide variants and small insertions/deletions (indels) (supplementary material, Table S3). Three Subclonal SNV and indel counts are not associated genes showed a frequency of somatic mutation greater with disease relapse than 20%: APC (56.82%), TP53 (54.55%), and KRAS To determine whether aneuploidy and levels of subclo- (47.73%) (Figure 3A). As expected, over 85% of muta- nal CNAs were associated with the status of the most fre- tions affecting APC were nonsense or frameshift, while quently mutated genes in CRC, a targeted capture DNA in TP53 and KRAS the majority were missense changes sequencing approach was performed on a subset of (75% and 100%, respectively). Consistent with this, Figure 3. Mutational profiling and subclonal status of MSS tumors from patients with stage II colon cancer. (A) Oncoprint chart illustrates somatic mutation events (i.e. SNVs and indels) of the studied genes detected by targeted NGS, arranging cases in two groups according to patient recurrence status. Each column represents an individual tumor and each row a gene. Red/blue bar plot (top) shows the number of mutations per tumor, elucidating their clonal or subclonal (<85%) status. Orange/blue bar plot (right) represents the incidence (number of times) at which each gene was observed mutated. Genes with no mutations are not shown. Color bar plots (bottom) provide data on the geno- mic and pathological markers annotated in the legend (right). (B) Proportions of mutations that were identified as clonal (blue) or subclonal (red) for the eight most frequently mutated genes. Values are indicated as percentage (%). These percentages were obtained from the 41 patients included in the targeted sequencing that could also be analyzed by ASCAT for the clonality of SNVs and indels. © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd J Pathol 2022; 257: 68–81 on behalf of The Pathological Society of Great Britain and Ireland. www.pathsoc.org www.thejournalofpathology.com Copy-number heterogeneity predicts colon cancer relapse 75 86% of mutations affecting KRAS altered hotspot codon and subclonal CNAs) (Figure 3A and supplementary 12 or 13, including variant G35A, which represented material, Table S3). In contrast, the mutational fre- 45% of the total RAS mutations. SOX9 also manifested quency of SOX9, SYNE1, and NOTCH1 appeared to a disproportionately high fraction (87.5%) of truncating correlate significantly with recurrence (p = 0.046, events. FDR = 0.14), SOX9 being more frequently mutated No statistically significant differences were in non-recurrent tumors. observed for APC, TP53, and KRAS mutational status To investigate levels of intratumor heterogeneity regarding the recurrence condition, nor were other concerning single nucleotide variants, we estimated clinical or genomic features (including the CNA load the CCF harboring each mutation, allowing for Figure 4. Prognostic value of tumor aneuploidy and genomic heterogeneity for TTR in stage II colon cancer patients. Kaplan–Meier estimates displaying the cumulative proportion (Y axis) of study patients who were relapse-free over an 8-year period (X axis), stratifying by (A) the presence or absence of chromosome arm 13q aberration; (B) 17q22–24.3 LOH (comprising loss or somatic cnLOH); (C) a high or low propor- tion of CNA load in the study population (N = 84), and (D) in an independent cohort of patients (N = 99) diagnosed with MSS tumors; (E) the level of subclonal CNAs identified with SNP arrays; and (F) a high or low level of subclonal SNVs and indels detected by targeted NGS. P values were obtained using the log-rank test and hazard ratios using a Cox regression model with proportional hazards, correcting by age, sex, stage pT3/T4, histological grade, and microsatellite status. Cut-off values to categorize numerical variables were as follows: (i) for CNA load (Hospital Clínic cohort), 24.41%; (ii) for CNA load (Colonomics cohort), 32%; (iii) for subclonal CNAs, 13.5; and (iv) for subclonal mutations, 1 (out of an average 4.5 mutations per sample, as derived from the 48-gene panel sequencing). © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd J Pathol 2022; 257: 68–81 on behalf of The Pathological Society of Great Britain and Ireland. www.pathsoc.org www.thejournalofpathology.com 76 S Lahoz et al discrimination of subclonal (<85%) from clonal vari- Genomic markers improve the prediction of disease ants. Of the total variants identified, 53% were subclo- recurrence compared with clinicopathological nal, showing a median CCF of 49% (10–84%). In our parameters alone analysis, recurrent tumors did not exhibit increased subclonal SNVs and indels compared with non- Predictive modeling was performed to classify patients recurrent ones (p = 0.75) (Figure 3A). Among genes with the primary endpoint of predicting their individual with a mutational frequency above 15%, APC, TP53, risk of relapsing. Genomic and mutational features with and KRAS, as well as SOX9, appeared to be mutated the highest prognostic values (supplementary material, in a clonal state in most cases (62.5%). Conversely, Figure S5), together with all clinicopathological NOTCH1, LRP1B, SYNE1, and CSMD1 were predom- markers, were included in conventional machine learn- inantly subclonal (Figure 3B). ing analysis to assess their integrated discriminatory power in our cohort of 84 patients. Among the eight clas- sifier algorithms tested (see Supplementary materials and methods), the gradient boosting machine method Tumor aneuploidy and intratumor heterogeneity are was selected as it ranked the shortest fraction of misclas- associated with risk of disease relapse sified cases (data not shown). Variables included in each fold classification, as determined by LASSO feature When we assessed the independent prognostic value of selection (see Supplementary materials and methods), the above-described genomic features for TTR (Table 2 are reported in supplementary material, Table S6. The and supplementary material, Table S4 and Figure S4A), median number of predictor features included for each patients with gain of chromosome arm 13q exhibited a classification was 12/22 (54.5%) per fold (range 6–17). shorter TTR and a higher risk for recurrence compared We constructed 10-fold cross-validated models using with those with a diploid 13q (p = 0.02; HR 2.67; 95% three different combinations of risk predictors: clinical- CI 1.16–6.14) in a multivariable analysis correcting for , genomic- and clinico-genomic-based, each one com- age, sex, pT3/T4, histological grade, and microsatellite sta- prising the annotated variables indicated in supplemen- tus (Figure 4A). TTR was also consistently shorter in those tary material, Figure S6. For the clinical-based model, patients with tumors carrying LOH at 17q22–q24.3 the classifier algorithm correctly predicted 58/84 cases (i.e. loss or cnLOH) (p = 0.05; HR 2.69; 95% CI (69% accuracy), achieving 70% sensitivity and 69% 0.99–7.35) (Figure 4B). specificity; for the genomic-based, it correctly spotted We then addressed the association of high levels of 52/84 patients (62% accuracy), with 58% sensitivity tumor aneuploidy and genomic intratumor heterogeneity and 66% specificity; while for the clinico-genomic with adverse clinical outcome. Kaplan–Meier analysis model, predictions yielded 64/84 correctly classified showed that patients with tumors bearing high CNA patients (76% accuracy), attaining 73% sensitivity and loads were at a significantly higher risk of relapse com- 80% specificity. For instance, false-positive rates were pared with those with low aneuploidy, independent of always inferior to their respective false-negative counts patient clinical status (p = 0.038; HR 2.46; 95% CI in the two biomarker models comprising genomic fea- 1.05–5.74) (Figure 4C). The median time until relapse tures (Table 3). was 31.48 months in the higher-risk group versus In terms of discriminatory ability, receiving operating 75.28 months in the lower-risk group. This same associ- characteristic (ROC) curve analyses revealed that the ation was validated in the independent cohort Colo- clinico-genomic combination achieved an AUC of 0.83 nomics (p = 0.004; HR 3.82; 95% CI 1.54–9.52) (95% CI 0.74–0.92), as opposed to inferior AUC values (Figure 4D). of 0.74 (95% CI 0.63–0.84) for the clinical-based and Furthermore, elevated copy-number intratumor het- 0.66 (95% CI 0.54–0.78) for the genomic-based models, erogeneity was associated with higher risk of recurrence respectively (Figure 5 and Table 3). The addition of the (p = 0.028; HR 2.20; 95% CI 1.09–4.44) (Figure 4E). genomic-related variables to baseline clinicopathologi- The median time until relapse was 25.03 months in the cal parameters caused a significant 10% incremental high subclonal CNAs group compared with gain in the AUC (DeLong’s p = 0.047). 70.80 months in the lower group. However, there was no significant correlation of the proportion of subclonal SNVs and indels with risk of relapse (p = 0.30; HR 1.73; 95% CI 0.65–4.58) (Figure 4F). Regarding the Discussion prognostic value of specific mutated genes, patients with SOX9 mutations exhibited a significantly longer TTR in Intratumor heterogeneity fosters cancer evolution, lead- a univariate analysis (p = 0.028; HR 0.15; 95% CI ing to metastatic progress and therapy resistance [8]. In 0.06–0.41) (supplementary material, Figure S4B). Con- the present study, we investigated the relative contribu- versely, NOTCH1 and SYNE1 mutations were associ- tion of genomic intratumor heterogeneity involving ated with an increased risk of disease recurrence CNAs and mutational events as a prognostic determinant (p = 0.003; HR 3.61; 95% CI 1.45–8.98; and for disease recurrence in untreated stage II colon cancer p = 0.02; HR 2.90; 95% CI 1.18–7.16, respectively) patients. Additionally, we constructed a predictive (Table 2 and supplementary material, Figure S4C,D). model integrating the levels of genomic complexity © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd J Pathol 2022; 257: 68–81 on behalf of The Pathological Society of Great Britain and Ireland. www.pathsoc.org www.thejournalofpathology.com Copy-number heterogeneity predicts colon cancer relapse 77 Figure 5. Discriminative ability of the three stage II biomarker models for predicting individual patient probability of recurrence. ROC curves are shown for comparative analysis of the three different combinations of variables. The AUC scores and 0.95 CI for each model are indicated. Table 3. Performance metrics of the biomarker models obtained by machine learning to predict relapse. Endpoint AUC Sensitivity Specificity PPV NPV Accuracy (0.95 CI) (0.95 CI) (0.95 CI) (0.95 CI) (0.95 CI) (0.95 CI) Clinico-genomic-based model 0.83 0.73 0.80 0.75 0.78 0.76 (0.74–0.92) (0.59–0.84) (0.67–0.89) (0.64–0.83) (0.69–0.84) (0.74–0.78) Clinical-based model 0.74 0.70 0.69 0.65 0.73 0.69 (0.63–0.84) (0.53–0.83) (0.57–0.79) (0.56–0.73) (0.63–0.82) (0.67–0.71) Genomic-based model 0.66 0.58 0.66 0.58 0.65 0.62 (0.54–0.78) (0.43–0.71) (0.52–0.78) (0.48–0.68) (0.57–0.73) (0.60–0.64) CI, confidence interval; NPV, negative predictive value; PPV, positive predictive value. along with clinicopathological features, enabling the consideration that the validation cohort consisted of identification of those patients at high risk of recurrence MSS tumors exclusively. Therefore, these findings with notable discriminative ability. might provide evidence on the prognostic contribution Tumor aneuploidy, or the presence of chromosome of tumor aneuploidy in early-stage colon cancer. As imbalances, is a hallmark of human solid malignancies regards chromosome-specific regions with the ability to [31] and has been depicted as an unfavorable prognostic discriminate recurrent lesions, we found several CNAs factor pan-cancer [32,33]. We define the CNA load as a including gain of chromosome arm 13q and loss at static measurement of the levels of numerical chromo- 17q22–q24.3, although their statistical significance somal alterations genome-wide. Increased CNA load decreased after multiple testing correction. In line with has already been shown to predict poor outcome in met- this, a previous study already showed that gains of astatic CRC [24], yet it has not been thoroughly explored CDX2, located at 13q12.2, were exclusively seen in pri- in early-stage disease as a prognostic biomarker. In our mary recurrent adenomas compared with those without cohort of stage II colon cancer, tumors from recurrent recurrence [34]. The analysis of the Colonomics series patients encompass significantly higher CNA load com- by Alonso et al [25] revealed several focal CNA regions pared with non-recurrent. This trend is maintained when associated with disease-free survival, yet showed no broad excluding MSI cases, which in general exhibit near- recurrent CNAs correlated with prognosis in MSS stage II diploid karyotypes and fall in the non-recurrent group. colon cancers. Of note, we identified that sample purities in Importantly, our data support a strong correlation the Colonomics cohort were ~10% inferior than those from between high CNA load and lower probability of being our cohort (median 68%, ranging from 32–100%, in the relapse-free in two independent cohorts, with the Hospital Clínic cohort and median 59.5%, ranging from © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd J Pathol 2022; 257: 68–81 on behalf of The Pathological Society of Great Britain and Ireland. www.pathsoc.org www.thejournalofpathology.com 78 S Lahoz et al 38–100%, in the Colonomics dataset; Mann–Whitney test, overexpression of SOX9 at the invasive front with sig- p = 0.007), which probably hampered the copy-number nificantly higher relapse-free times in stage II colon can- analysis and consequently, subclonal CNAs might have cer [43]. Finally, our observations on SOX9-mutated and remained masked. APC-wild-type tumors displaying more abundant CD8+ To capture an approximative measure of the degree of lymphocyte populations than their counterparts suggest intratumor heterogeneity of each tumor, we determined a putative role for the WNT/β-catenin pathway in modu- the subclonal status of CNAs, cnLOH, and mutations. lating the tumor-infiltrating T-cell compartment [44]. The fact that all patients included in this study were Besides the independent prognostic value of genomic chemotherapy-untreated avoids the genetic bottleneck markers, our data reinforce the idea that tumor microenvi- caused by cancer therapy, enabling better preservation ronment features (e.g. low intratumor CD8+ lymphocytic of the original overall clonal diversity. Increased levels infiltration) and the presence of lymphovascular or peri- of subclonal CNAs have been associated with poor out- neural invasion are still the most potent determinants of come in non-small cell lung cancer [19] and hepatocellu- poor outcome in locally advanced colon cancer. The lar carcinoma [20], while mutational multi-clonality impact of anti-cancer immune cytotoxicity as a major correlated with worse disease-free survival in stages determinant for disease recurrence in stages II–III has I–IV CRC [35]. Likewise, our data show that tumors been widely validated in multiple independent series of with elevated subclonal CNAs exhibit a significantly patients over the past 5 years [4,45,46]. In addition, tumor shorter TTR, making it plausible that early-stage tumors budding and poorly differentiated clusters might consti- with extensive copy-number intratumor heterogeneity tute early histological features potentially leading to an might fall in the high-risk group likely to benefit from ulterior metastatic phenotype [5], which might explain adjuvant chemotherapy. These results are in line with a their strong ability to predict risk of relapse in early-stage recent report showing that chromosomal copy number het- colon cancer. Finally, in a large-cohort study, the lack of erogeneity is able to predict poor patient survival pan- CDX2 expression was associated with shorter relapse- cancer [36]. Besides, we find that tumors with a high free survival in stage II colon cancer [47]. We observed CNA load display increased degrees of subclonal CNAs, the same trend regardless of the low number of tumors which reinforces the idea that intratumor heterogeneity pre- with CDX2 negative expression. dominantly occurs in highly aneuploid tumors through a Despite the plethora of identified prognostic bio- process governed by ongoing CIN. In contrast, the amount markers, prediction of stage II colon cancer recurrence of subclonal SNVs and indels did not appear to be associ- based on molecular information is still an open problem. ated with recurrence in our cohort, strengthening the poten- Here, we created a machine learning-based predictive tial prognostic power of CIN over the tumor mutational model integrating a range of 22 risk predictors. The best burden, possibly because a sole CNA can disrupt the tran- configuration of our clinico-genomic model achieved sig- scription of a multitude of genes simultaneously [37,38]. nificant discriminant capacity, rating an AUC of 83%, Of note, our targeted sequencing approach was designed comparable with previous reported AUC values in vari- to test the mutational status of a limited panel of 48 driver ous stage II colon cancer cohorts using molecular-based genes, which makes it difficult to extrapolate these results signatures [48,49]. In our predictive model, the addition to the entire genome. Our results are consistent with previ- of genomic-related markers to the patient clinicopatholog- ous reports showing that gene copy numbers, but not cod- ical information caused a significant increase in the AUC ing mutations, are highly discordant between colorectal of 9 points, which was also associated with a similar primary tumors and their matched metastases [39]. Consid- increase in the sensitivity and accuracy. Intriguingly, ering these results, the need to assess CNAs using noninva- across the three combination models, specificity values sive approaches, such as liquid biopsy, urges further were always higher than their respective sensitivities by attention in the clinical setting. 5–10 points, stressing the need to identify more sensitive The mutational landscape of CRC has been well char- strategies capable of reducing false-negative rates. An acterized [29]. Our results suggest that tumor levels of obvious limitation of the present study is that 22 predictors subclonal CNAs are independent of the mutational status included in the clinico-genomic model might be difficult of the main CRC driver genes (i.e. APC, TP53, and to manage in routine clinical practice. To reduce the num- KRAS). We find that none of the previous three mutated ber of variables and to address potential multicollinearity, genes correlates with risk of relapse, in concordance a feature selection step was included in the machine- with reported data in stages II–III colon cancer [4,40]. learning process, which selected CD8+ lymphocyte infil- It is noteworthy that the mutational status of SOX9, tration in 100% of folds, SOX9 mutational status in 50% NOTCH1, and SYNE1 displayed prognostic value in of folds, and both the CNA load and subclonal CNAs in our cohort. In CRC, overexpression of NOTCH1 has 25% of folds, these latter two appearing as mutually been shown as a negative predictor of overall survival exclusive. Interestingly, the additive predictive power of [41], consistent with our findings for TTR. Conversely, subclonal CNAs over CD8+ lymphocyte infiltration was patients with mutations in SOX9 exhibited a longer the same as SOX9 mutations over CD8+ lymphocyte infil- TTR in our series, but this association appears to weaken tration (AUC = 71%), suggesting a biological link in the multivariable setting. This result is in line with a between SOX9 and the tumor T-cell infiltration. previously reported association of SOX9 mutations with To summarize, our results reinforce the potential value increased overall survival in metastatic CRC [42] and of intratumor heterogeneity driven by chromosomal © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd J Pathol 2022; 257: 68–81 on behalf of The Pathological Society of Great Britain and Ireland. www.pathsoc.org www.thejournalofpathology.com Copy-number heterogeneity predicts colon cancer relapse 79 instability as a prognostic factor in early-stage colon can- for Biotechnology Information (NCBI) Gene cer. In this scenario, patients with tumors harboring high Expression Omnibus and are accessible through GEO levels of aneuploidy and subclonal CNAs might be at risk Series accession number GSE172191. of relapse and could benefit from early therapeutic inter- vention during disease monitoring. Given the retrospective design of this study, we advocate for these results to be fur- ther validated in randomized, prospective clinical trials References incorporating NGS strategies, intending to optimize patient 1. Benson AB, Venook AP, Al-Hawary MM, et al. NCCN Guidelines stratification at the adjuvant setting. Insights: Colon Cancer, Version 2.2018. J Natl Compr Canc Netw 2018; 16: 359–369. Acknowledgements 2. Tournigand C, de Gramont A. Chemotherapy: is adjuvant chemother- apy an option for stage II colon cancer? Nat Rev Clin Oncol 2011; 8: We would like to thank Dr Joaquim Radua (from 574–576. 3. Meropol NJ. Ongoing challenge of stage II colon cancer. J Clin Oncol IDIBAPS) for statistical assistance, Dr Daniel Aguilar 2011; 29: 3346–3348. (from CIBEREHD) for bioinformatics support, Dr 4. Dienstmann R, Villacampa G, Sveen A, et al. Relative contribution of Thomas Ried (from the National Cancer Institute/ clinicopathological variables, genomic markers, transcriptomic sub- National Institutes of Health) for critical reading, and Dr typing and microenvironment features for outcome prediction in stage Mireia Aragay (from University of Barcelona) for English II/III colorectal cancer. Ann Oncol 2019; 30: 1622–1629. editing. We are also grateful to the Biobank-Tumor Bank 5. Ueno H, Ishiguro M, Nakatani E, et al. Prospective multicenter study and Functional Genomics platforms from IDIBAPS- on the prognostic and predictive impact of tumor budding in stage II Hospital Clínic of Barcelona for preparing and processing colon cancer: results from the SACURA trial. J Clin Oncol 2019; samples and for microarrays and next-generation 37: 1886–1894. sequencing procedures. The work was supported by the 6. Mlecnik B, Bindea G, Angell HK, et al. Integrative analyses of colo- Xarxa de Bancs de Tumors de Catalunya sponsored by rectal cancer show Immunoscore is a stronger predictor of patient sur- Pla Director d’Oncologia de Catalunya (XBTC). The vival than microsatellite instability. Immunity 2016; 44: 698–711. 7. Pagès F, Mlecnik B, Marliot F, et al. International validation of the study has been developed in part in the Centre Esther consensus Immunoscore for the classification of colon cancer: a prog- Koplowitz from IDIBAPS/CERCA Programme/ nostic and accuracy study. Lancet 2018; 391: 2128–2139. Generalitat de Catalunya. This research was funded 8. Greaves M, Maley CC. Clonal evolution in cancer. Nature 2012; 481: by grants from the Instituto de Salud Carlos III and 306–313. co-funded by the European Regional Development 9. Iacobuzio-Donahue CA, Litchfield K, Swanton C. Intratumor Fund (ERDF) (CPII18/00026, PI17/01304, heterogeneity reflects clinical disease course. Nat Cancer 2020; PI20/00863), the CIBEREHD and CIBERONC pro- 1: 3–6. grams from Instituto de Salud Carlos III, the 10. Cross W, Kovac M, Mustonen V, et al. The evolutionary landscape of Agència de Gestio d’Ajuts Universitaris i de colorectal tumorigenesis. Nat Ecol Evol 2018; 2: 1661–1672. Recerca, Generalitat de Catalunya (2017 SGR 11. Diaz LA Jr, Williams RT, Wu J, et al. The molecular evolution of 1035), and Fundacion Científica de la Asociacion acquired resistance to targeted EGFR blockade in colorectal cancers. Española Contra el Cancer (GCB13131592CAST). Nature 2012; 486: 537–540. This article is based on work from COST Action 12. Johnson BE, Mazor T, Hong C, et al. Mutational analysis reveals the origin and therapy-driven evolution of recurrent glioma. Science CA17118, supported by COST (European Coopera- 2014; 343: 189–193. tion in Science and Technology). SL obtained a 13. Sansregret L, Vanhaesebroeck B, Swanton C. Determinants and clin- PFIS grant from Instituto de Salud Carlos III and ical implications of chromosomal instability in cancer. Nat Rev Clin co-funded by the European Regional Development Oncol 2018; 15: 139–150. Fund (ERDF) (FI18/00221). 14. Sottoriva A, Kang H, Ma Z, et al. A Big Bang model of human colo- rectal tumor growth. Nat Genet 2015; 47: 209–216. Author contributions statement 15. Watkins TBK, Lim EL, Petkovic M, et al. Pervasive chromosomal instability and karyotype order in tumour evolution. Nature 2020; 587: 126–132. SL, IA, MC and JC conceptualized the study. SL, IA, 16. Yates LR, Gerstung M, Knappskog S, et al. Subclonal diversification EA, QF, SL-P and JDR. performed experiments. SL, of primary breast cancer revealed by multiregion sequencing. Nat EH-I, FN and JJL analyzed the data. RS-P provided the Med 2015; 21: 751–759. external dataset. AC and MC contributed to accessing 17. Hu Z, Ding J, Ma Z, et al. Quantitative evidence for early metastatic clinical samples and patient data. SL, IA, MC and JC seeding in colorectal cancer. Nat Genet 2019; 51: 1113–1122. wrote the manuscript. MC and JC supervised the study. 18. Landau DA, Carter SL, Stojanov P, et al. Evolution and impact of sub- All the authors read and agreed with the final version clonal mutations in chronic lymphocytic leukemia. Cell 2013; 152: of the manuscript. 714–726. 19. Jamal-Hanjani M, Wilson GA, McGranahan N, et al. Tracking the evolution of non-small-cell lung cancer. N Engl J Med 2017; 376: Data availability statement 2109–2121. 20. Losic B, Craig AJ, Villacorta-Martin C, et al. Intratumoral heteroge- SNP-array and targeted next-generation sequencing neity and clonal evolution in liver cancer. Nat Commun 2020; data have been deposited in the National Center 11: 291. © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd J Pathol 2022; 257: 68–81 on behalf of The Pathological Society of Great Britain and Ireland. www.pathsoc.org www.thejournalofpathology.com 80 S Lahoz et al 21. Taylor AM, Shih J, Ha G, et al. Genomic and functional approaches on gene expression levels and disease outcome. Mol Aspects Med to understanding cancer aneuploidy. Cancer Cell 2018; 33: 2019; 69: 48–61. 676–689.e3. 39. Mamlouk S, Childs LH, Aust D, et al. DNA copy number changes 22. Wangsa D, Quintanilla I, Torabi K, et al. Near-tetraploid cancer cells define spatial patterns of heterogeneity in colorectal cancer. Nat show chromosome instability triggered by replication stress and Commun 2017; 8: 14093. exhibit enhanced invasiveness. FASEB J 2018; 32: 3502–3517. 40. Shen Y, Han X, Wang J, et al. Prognostic impact of mutation profil- 23. Hieronymus H, Schultz N, Gopalan A, et al. Copy number alteration ing in patients with stage II and III colon cancer. Sci Rep 2016; 6: burden predicts prostate cancer relapse. Proc Natl Acad Sci U S A 24310. 2014; 111: 11139–11144. 41. Chu D, Zhang Z, Zhou Y, et al. Notch1 and Notch2 have opposite 24. Smeets D, Miller IS, O’Connor DP, et al. Copy number load predicts prognostic effects on patients with colorectal cancer. Ann Oncol outcome of metastatic colorectal cancer patients receiving bevacizu- 2011; 22: 2440–2447. mab combination therapy. Nat Commun 2018; 9: 4112. 42. Javier BM, Yaeger R, Wang L, et al. Recurrent, truncating SOX9 25. Alonso MH, Auss o S, Lopez-Doriga A, et al. Comprehensive analysis mutations are associated with SOX9 overexpression, KRAS mutation, of copy number aberrations in microsatellite stable colon cancer in and TP53 wild type status in colorectal carcinoma. Oncotarget 2016; view of stromal component. Br J Cancer 2017; 117: 421–431. 7: 50875–50882. 26. Franch-Exposito S, Bassaganyas L, Vila-Casadesús M, et al. CNApp, 43. Marcker Espersen ML, Linnemann D, Christensen IJ, et al. SOX9 a tool for the quantification of copy number alterations and integrative expression predicts relapse of stage II colon cancer patients. Hum analysis revealing clinical implications. Elife 2020; 9: e50267. Pathol 2016; 52: 38–46. 27. Van Loo P, Nordgard SH, Lingjærde OC, et al. Allele-specific copy 44. Sun X, Liu S, Wang D, et al. Colorectal cancer cells suppress CD4+ T number analysis of tumors. Proc Natl Acad Sci U S A 2010; 107: cells immunity through canonical Wnt signaling. Oncotarget 2017; 8: 16910–16915. 15168–15181. 28. Nadeu F, Clot G, Delgado J, et al. Clinical impact of the subclonal 45. Van den Eynde M, Mlecnik B, Bindea G, et al. The link between the architecture and mutational complexity in chronic lymphocytic leuke- multiverse of immune microenvironments in metastases and the mia. Leukemia 2018; 32: 645–653. survival of colorectal cancer patients. Cancer Cell 2018; 34: 29. The Cancer Genome Atlas Network. Comprehensive molecular char- 1012–1026.e3. acterization of human colon and rectal cancer. Nature 2012; 487: 46. Sanz-Pamplona R, Melas M, Maoz A, et al. Lymphocytic infiltra- 330–337. tion in stage II microsatellite stable colorectal tumors: a retro- 30. Davoli T, Uno H, Wooten EC, et al. Tumor aneuploidy correlates with spective prognosis biomarker analysis. PLoS Med 2020; 17: markers of immune evasion and with reduced response to immuno- e1003292. therapy. Science 2017; 355: eaaf8399. 47. Dalerba P, Sahoo D, Paik S, et al. CDX2 as a prognostic biomarker 31. Zack TI, Schumacher SE, Carter SL, et al. Pan-cancer patterns of in stage II and stage III colon cancer. N Engl J Med 2016; 374: somatic copy number alteration. Nat Genet 2013; 45: 1134–1140. 211–222. 32. Carter SL, Eklund AC, Kohane IS, et al. A signature of chromo- 48. Kandimalla R, Linnekamp JF, van Hooff S, et al. Methylation of somal instability inferred from gene expression profiles predicts WNT target genes AXIN2 and DKK1 as robust biomarkers for recur- clinical outcome in multiple human cancers. Nat Genet 2006; 38: rence prediction in stage II colon cancer. Oncogenesis 2017; 6: e308. 1043–1048. 49. Jacob H, Stanisavljevic L, Storli KE, et al. A four-microRNA classi- 33. Bielski CM, Zehir A, Penson AV, et al. Genome doubling shapes the fier as a novel prognostic marker for tumor recurrence in stage II colon evolution and prognosis of advanced cancers. Nat Genet 2018; 50: cancer. Sci Rep 2018; 8: 6157. 1189–1195. 50. Dentro SC, Wedge DC, Van Loo P. Principles of reconstructing the 34. Fiedler D, Heselmeyer-Haddad K, Hirsch D, et al. Single-cell genetic subclonal architecture of cancers. Cold Spring Harb Perspect Med analysis of clonal dynamics in colorectal adenomas indicates CDX2 2017; 7: a026625. gain as a predictor of recurrence. Int J Cancer 2019; 144: 1561–1573. 51. Torabi K, Mir o R, Fernandez-Jiménez N, et al. Patterns of somatic 35. Joung JG, Oh BY, Hong HK, et al. Tumor heterogeneity predicts met- uniparental disomy identify novel tumor suppressor genes in colorec- astatic potential in colorectal cancer. Clin Cancer Res 2017; 23: tal cancer. Carcinogenesis 2015; 36: 1103–1110. 7209–7216. 52. Ueno H, Hase K, Hashiguchi Y, et al. Site-specific tumor grading sys- 36. van Dijk E, van den Bosch T, Lenos KJ, et al. Chromosomal copy tem in colorectal cancer: multicenter pathologic review of the value of number heterogeneity predicts survival rates across cancers. Nat quantifying poorly differentiated clusters. Am J Surg Pathol 2014; 38: Commun 2021; 12: 3188. 197–204. 37. Henrichsen CN, Vinckenbosch N, Zöllner S, et al. Segmental copy 53. Lugli A, Kirsch R, Ajioka Y, et al. Recommendations for reporting number variation shapes tissue transcriptomes. Nat Genet 2009; 41: tumor budding in colorectal cancer based on the International Tumor 424–429. Budding Consensus Conference (ITBCC) 2016. Mod Pathol 2017; 38. Ried T, Meijer GA, Harrison DJ, et al. The landscape of genomic 30: 1299–1311. copy number alterations in colorectal cancer and their consequences References 50–53 are cited only in the supplementary material. SUPPLEMENTARY MATERIAL ONLINE Supplementary materials and methods Figure S1. Kaplan–Meier estimates on TTR stratified by the annotated immunohistochemistry-based variables Figure S2. Correlation of FISH counts with SNP-array copy-number quantification Figure S3. Extended assessments to validate intratumor heterogeneity analysis Figure S4. Kaplan–Meier measures of TTR stratified by the mutational status of three genes tested by NGS © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd J Pathol 2022; 257: 68–81 on behalf of The Pathological Society of Great Britain and Ireland. www.pathsoc.org www.thejournalofpathology.com Copy-number heterogeneity predicts colon cancer relapse 81 Figure S5. Analysis of the prognostic risk of the clinico-genomic variables included in the predictive model Figure S6. Flowchart of the machine learning-based approach to develop our predictive model based on the annotated clinical and genomic variables Table S1. Summary of the targeted genes and exon regions tested in the NGS analysis Table S2. Chromosomal regions with a differential frequency of copy-number or LOH alteration between relapsed and non-relapsed tumors identified by SNP arrays Table S3. Differential frequency of single nucleotide variants or indels in relapsed and non-relapsed tumors for the 48-gene panel analyzed by targeted sequencing Table S4. Alternative multivariable Cox models for TTR, stratified by annotated clinico-pathological and tumor-genomic variables Table S5. Train and test AUCs per fold and DeLong’s pairwise comparisons between biomarker models Table S6. Variables selected per fold using LASSO-based feature selection © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd J Pathol 2022; 257: 68–81 on behalf of The Pathological Society of Great Britain and Ireland. www.pathsoc.org www.thejournalofpathology.com