Chromosome arm aneuploidies shape tumour evolution and drug response

Abstract

Chromosome arm aneuploidies (CAAs) are pervasive in cancers. However, how they affect cancer development, prognosis and treatment remains largely unknown. Here, we analyse CAA profiles of 23,427 tumours, identifying aspects of tumour evolution including probable orders in which CAAs occur and CAAs predicting tissue-specific metastasis. Both haematological and solid cancers initially gain chromosome arms, while only solid cancers subsequently preferentially lose multiple arms. 72 CAAs and 88 synergistically co-occurring CAA pairs multivariately predict good or poor survival for 58% of 6977 patients, with negligible impact of whole-genome doubling. Additionally, machine learning identifies 31 CAAs that robustly alter response to 56 chemotherapeutic drugs across cell lines representing 17 cancer types. We also uncover 1024 potential synthetic lethal pharmacogenomic interactions. Notably, in predicting drug response, CAAs substantially outperform mutations and focal deletions/amplifications combined. Thus, CAAs predict cancer prognosis, shape tumour evolution, metastasis and drug response, and may advance precision oncology.

Fig. 1. CAA frequencies provide insights into tumour evolution.

a Box plot comparing CAA burden per tumour for haematological and solid cancers. Shown are mean (+), median with 95% confidence intervals (notches), interquartile ranges and all data points. P value: Mann–Whitney U test. b Box plot as in a showing CAA burden per cancer type. Abbreviations of each cancer type are shown in Supplementary Data 1. c Box plot as in a showing the number of chromosome arms lost or gained in haematological and solid cancers. P values: Mann–Whitney U test (unpaired), Wilcoxon signed-rank test (paired). d Contingency tables showing expected (E) and observed (O) numbers (n) and percentages (%) of CAA-positive haematological and solid tumours with indicated arm-level gain:loss ratios. Bar graphs show the respective expected and observed fractions. P values: Chi-square tests. e Shooting star plots showing fractions of tumours with G > L as a function of the total number of CAAs per sample. Odd and even numbers are shown separately. Dot sizes are proportional to the fractions of haematological (orange) and solid tumours (blue). P values: binomial tests. f Tumour evolution model showing that both haematological and solid cancers initially gain few chromosome arms, whereas only solid cancers subsequently preferentially lose chromosome arms. g Distributions of CAA-positive solid tumours with indicated intra-tumour chromosome arm gain (G):loss (L) ratios according to clinical stage. Median CAA burden and sample sizes are shown for each stage. P values: p = 7.1 × 10⁻⁵, p = 1.4 × 10⁻⁴, p = 0.0013, respectively, Chi-square tests relative to stage I. P value abbreviations are defined in the Methods section. Source data are provided as a Source Data file.

Fig. 2. CAAs in primary and metastatic samples shape tumour evolution.

a Violin plots comparing CAA burden in primary and metastatic solid tumours. Medians and interquartile ranges are shown. LN, lymph node; Met, metastases at any metastatic site. P value: Mann–Whitney U test. b Heatmap of the frequency differential in metastatic disease relative to primary cancers. Numbers in tiles refer to q values, i.e., FDR-adjusted p values of Fisher’s exact tests. Empty tiles, not significant (q > 0.05); 1, q < 0.05; 2, q < 10⁻²; 3, q < 10⁻³, etc. c Heatmap as in b, but per metastatic site. Yellow boxes highlight a degree of specificity. The black/yellow heatmap below shows the frequencies of indicated CAAs in primary brain cancers. d Stochastic tumour evolution model. The network models the order in which CAAs are acquired during breast cancer development. Nodes represent CAAs and their sizes are proportional to CAA frequencies. Edges represent estimated transition probabilities and their thicknesses are proportional to the probabilities. Note that the edges are directed, from left to right. Source data are provided as a Source Data file.

Fig. 3. Identification of 72 CAAs that predict good or poor cancer patient survival outcome.

a, b Examples of Kaplan–Meier curves showing disease-free survival of patients with/without indicated CAA. P values: log-rank test and multivariate Cox proportional hazard analysis. Hazard ratios (HR) and 95% confidence intervals are shown. The effect of whole-genome doubling (WGD) on survival outcome was also assessed in univariate and multivariate analyses. See also Supplementary Data 4 and 5. Source data are provided as a Source Data file.

Fig. 4. Identification of 88 co-occurring CAA pairs that synergistically predict good or poor cancer patient survival outcome.

a Matrices and table of selected results from pan-cancer CAA probabilistic cooccurrence analyses. Tile colours indicate whether CAA combinations occur significantly more (‘positive’) or less (‘negative’) frequently than expected. b Network of CAA cooccurrences in the LGG dataset. Nodes represent CAAs, including losses (−, dark red) and gains (+, light green). Sizes are proportional to frequencies. Edge colours indicate statistically significant positive (dark blue) or negative (orange) probabilities as in (a) with thickness inversely proportional to probability. c Volcano plot showing frequencies and probabilities of co-occurring CAAs in the BRCA dataset. Pairs co-occurring at p > 0.05 or involving > 10% of patients are shown in blue (positive) or red (negative). Green circles highlight co-occurrences significantly predicting patient survival outcome in multivariate analyses. d, e Example Kaplan–Meier survival curves of co-occurring CAA pairs predicting poor (d) or good prognosis (e). Statistics as in Fig. 3a, b. Source data are provided as a Source Data file.

Fig. 5. Summary of CAA-informed survival outcome, with negligible impact of whole-genome doubling.

a Sunburst charts for disease-free and overall survival. In total, n = 97 and n = 63 statistically significant individual CAAs (‘1 CAA’) or CAA combinations (‘2 CAAs’) were identified. Inner rings indicate fractions of 1 CAA and 2 CAAs. Outer rings indicate how many of the two individually cooccurring CAAs significantly predict patient survival outcome. b Sunburst chart showing number of patients analysed (n = 6,977) and fractions for whom CAAs predict good (green, left) or poor prognosis (red, right) based on individual (‘1 CAA’) or co-occurring CAAs (‘2 CAAs’). c Left: Ring pie charts showing the number of cancer types for which whole-genome doubling (WGD) significantly affects poor overall or disease-free survival in univariate Cox proportional hazard analyses. Right: Ring pie charts depicting the multivariate Cox proportional hazard p values of the 44 single or co-occurring CAAs for which WGD showed p values < 0.05 in univariate analyses (left). These multivariate analyses included WGD as a covariate. Source data are provided as a Source Data file.

Fig. 6. CAAs shape drug response and outperform other genomic events in response prediction.

a Heatmap of Spearman correlations between CAA burden and a pharmacogenomic predictor of pathologic complete response (pCR) to preoperative paclitaxel and fluorouracil-doxorubicin-cyclophosphamide (T/FAC) chemotherapy. Numbers in tiles show q values, i.e., FDR-corrected significance from Fisher’s exact tests, as Fig. 2b. Ns, not significant (q > 0.05). b Sunburst plot showing the distribution of pharmacogenomic alterations used in our machine learning model, including 285 high-confidence cancer genes (GCs), 425 recurrently copy number-altered chromosomal segments (RACSs), collectively referred to as cancer functional events (CFEs), and 78 CAAs. c Bubble volcano plot showing how specific CFEs and CAAs alter response to anti-cancer drugs, as determined by elastic net regression. The impact is shown as Glass’ Δ log₁₀(IC₅₀) effect size between cell lines with and without the alteration. Bubble colours correspond to cancer types. Bubble sizes are proportional to the numbers of cell lines. Selected CFEs and CAAs are highlighted in blue and red, respectively. Two CAA-drug interactions can be explained by focal CNAs (*). d Beeswarm plots showing the extent to which several CAAs significantly increase drug resistance or sensitivity. Cell lines negative (−) and positive (+) for indicated CAAs are shown. Horizontal lines represent the mean IC₅₀ value. See Supplementary Data 10 for full data. e Tukey boxplot showing the fractions of CFEs or CAAs that significantly alter drug response. Shown are the medians with interquartile ranges and all data points. Lines connect the respective fractions for each of the 22 cancer types. P value: Paired Wilcoxon signed-rank test. f Summary table showing the performance of machine learning models based on CFEs alone, CAAs alone and CFEs and CAAs combined. g Beeswarm plot showing the differences between the F₁ performance scores from the CFE-based and the CAA-based models for each of the tested 39 drugs. Mean and 95% confidence interval are shown. P values: one sample t-test (top), Fisher’s exact test (right). Source data are provided as a Source Data file.