Exploring Heterochromatin Gene Dynamics Reveals Major Contributors to Chromosomal Instability in Cancer Student: Dylan Haynes-Simmons Student Number: 2780600 Supervisor: Dr Aniek Janssen Daily Supervisor: Aditya Dixit Department: Centre for Molecular Medicine UMC Utrecht Group: Janssen Group Examiner: Reza Haydarlou MSc: Bioinformatics and Systems Biology Track: Bioinformatics Course Code: XM_0103 Credit Points (ECs): 60 Duration: 10 months 1 1. Abstract Heterochromatin plays a critical role in maintaining genomic stability and regulating gene expression, making it essential for the proper functioning of cells. It ensures genomic integrity by stabilising repetitive DNA sequences and protecting against harmful genetic recombinations. Chromosomal instability, a hallmark of cancer, involves increased rates of chromosomal missegregation leading to aneuploidy, structural chromosomal rearrangements, and copy number variations. Dysregulation of genes coding for heterochromatin proteins is thought to contribute to CIN and cancer progression. However, the underlying functional mechanisms remain poorly understood, particularly regarding the impact of individual genes on CIN. This study addresses this gap by constructing an interactome of genes encoding for heterochromatin proteins and employing a machine-learning model to explore their relation to CIN features in cancer. Our analysis revealed a distinct transcriptional signature of heterochromatin genes in cancerous conditions. The machine-learning model demonstrated an overall accuracy of 0.805 in classifying arm-level aneuploidies. Genes identified as significant contributors to CIN features included CENP-A, linked to multiple arm aneuploidies; SETDB1, associated with pericentromeric instability on chromosome 1; and NUP133 and NUP155, which influenced aneuploidies across several chromosomes, highlighting their potential roles in maintaining genomic stability. These findings emphasise the essential roles of heterochromatin genes in CIN, illustrating how misregulation of specific genes such as CENP-A, SETDB1, NUP133, and NUP155 could contribute to tumorigenesis. This deeper understanding of the molecular mechanisms influencing chromosomal instability provides valuable insights for future research which could further elucidate the role of heterochromatin protein-encoding genes in cancer. 2. Introduction Heterochromatin, a tightly packed form of DNA, is a chromatin type identified by darkly stained regions of the nucleus (Passarge, 1979). It is essential for maintaining genome stability and regulating gene expression in eukaryotic cells. It was initially described through studies on position-effect variegation in Drosophila, where heterochromatin was observed to cause mosaic gene silencing (Lomberk et al., 2006). The structure of heterochromatin is characterised by the presence of specific proteins, such as Heterochromatin Protein 1 (HP1), which binds to di- and tri-methylated histone H3 lysine 9 (H3K9me2/3) and facilitates the formation of a repressive chromatin environment (Hennig, 1999). Importantly, heterochromatin functions as a “guardian” of genomic integrity, ensuring the stability of repetitive DNA sequences and protecting against harmful genetic recombinations (Janssen et al., 2018). Heterochromatin is often associated with the nuclear lamina, a fibrous layer lining the inner nuclear membrane, and the proteins that reside there (Olins et al., 2010; Margalit et al., 2007; Maison et al., 1997; Poleshko et al., 2013; Zeng at al., 2010; Shibuya et al., 2014; Dunce et al., 2018; Li et al., 2018). The nuclear lamina helps organise chromosomes and establish genome spatial organisation through lamina-associated domains (LADs) (Van Steensel & Belmont, 2017; Shevelyov & Ulianov, 2019). LADs typically contain silent or lowly expressed genes and contribute to transcriptional repression and chromatin compaction (Van Steensel & Belmont, 2017). Despite its known roles in gene silencing and chromosomal architecture, many aspects of heterochromatin’s functional mechanisms remain poorly understood (Lomberk et al., 2006; Hennig, 1999). This lack of understanding is significant given heterochromatin’s crucial role in chromosomal stability, a concept central to exploring how chromosomal instability originates in cancer. 2 Chromosomal instability (CIN) is a hallmark of cancer characterised by an increased rate of chromosomal missegregation, leading to aneuploidy, structural rearrangements, and copy number variations (Janssen et al., 2018). The dysregulation of heterochromatin and its associated proteins is thought to play a significant role in CIN. Alterations in the composition of the nuclear lamina and LADs during cancer progression result in the reorganisation of heterochromatin, leading to changes in gene expression and genomic instability (Bellanger et al., 2022). These changes in LADs are often accompanied by epigenetic modifications, such as the loss of H3K9me2 and DNA hypomethylation, which contribute to the loss of heterochromatin integrity and increased susceptibility to chromosomal rearrangements (Van Steensel & Belmont, 2017). Consequently, the restructuring of LADs and the nuclear lamina in cancer cells calls attention to the critical relationship between nuclear architecture and chromosomal stability (Smith et al., 2018). Studying the dynamics of heterochromatin-associated proteins in the context of cancer could provide significant insights into their roles in regulating features of CIN, paving the way for targeted therapeutic strategies. Recent studies have highlighted that the dysregulation of heterochromatinassociated proteins and their encoding significantly contribute to CIN. For instance, proteins such as HP1 are essential for maintaining chromosomal integrity, gene silencing, and the formation of heterochromatin, with reductions in these proteins linked to cancer progression (Dialynas et al., 2008). Similarly, Polycomb-protein encoding genes are frequently upregulated in aggressive cancers, contributing to transcriptional repression and tumour development (Clermont et al., 2015). Defects in enzymes such as topoisomerases, important for heterochromatin structure and function, can lead to the de-repression of transposons, resulting in chromosomal instability (Lee & Wang, 2019; Amoiridis et al., 2024). These examples underline the pivotal roles of heterochromatin-associated proteins in maintaining genomic stability and their significant impact on chromosomal instability in cancer. However, a complete understanding of the role of heterochromatin misregulation in the development of CIN in cancer is currently missing. Therefore, further research is necessary to elucidate the molecular mechanisms by which such genes influence CIN and to explore their potential as therapeutic targets. Given the potentially fundamental role of heterochromatin in maintaining genomic stability and its dysregulation in cancer, our study aimed to delineate the roles of “heterochromatin-associated genes” (i.e. genes encoding for heterochromatin proteins) in and across cancer types by constructing a comprehensive interactome and employing a machine-learning model to assess their impact on chromosomal instability. Previous studies have demonstrated the utility of machine learning in genomic feature selection and association studies. Random forest algorithms have been effectively used to select relevant genes from high-dimensional microarray data, improving the identification of biomarkers associated with cancer progression (Anaissi et al., 2013). Additionally, an ensemble model based on a support vector machine approach proved adept in feature selection and providing robust biomarker identification (Anaissi et al., 2016). Similarly, integrating gene expression profiles to infer CIN signatures has provided significant insights into the underlying molecular mechanisms and has been predictive of clinical outcomes in multiple cancer types (Carter et al., 2006). Here, our approach involved integrating RNA-sSeq data from The Cancer Genome Atlas (TCGA) to build a detailed expression profile of heterochromatin- 3 associated genes across various cancer types and analyse their impact on chromosomal instability (Fig.1). By querying the STRING Protein-Protein Interaction (PPI) database, we expanded our core gene set to construct an extensive heterochromatin interactome. This interactome included numerous genes encoding for proteins linked to heterochromatin structure or function and implicated in cancer dysregulation. To relate the aberrations in the expression of heterochromatin-associated genes to CIN features, we developed a machine learning-based approach using a multioutput stacked ensemble model. This model estimated multiple features indicative of CIN, such as arm-level aneuploidies, homologous recombination deficiency (HRD) signatures, and copy Core Protein List Input Output Component Step 1: expansion of the heterochromatin gene set Process STRING PPI Set-of-Interest Step 2: creation of the expression & CIN dataset for the set-of-interest TCGA PanCan Dataset Set-of-Interest dataset Step 3: modelling the CIN features from the set-of-interest dataset XGBoost ML Model Genomic Feature Importance CIN Feature Interactions Figure 1. Flowchart of the workflow Structure of the complete workflow from the core protein set to the feature importance 4 number variations in pericentromeric regions. A subsequent feature importance analysis identified key genes as significant contributors to CIN features, providing novel insights into the molecular mechanisms underlying chromosomal instability in cancer. An additional interaction analysis provided further context into the network of heterochromatin-associated gene interactions. By employing this detailed approach, we aim to bridge the gap in understanding the complex relationship between heterochromatin dynamics and chromosomal instability in cancer. 3. Results 3.1 A transcriptionally distinct heterochromatin gene set is a signature of cancer Heterochromatin genes are known to serve a multitude of nuclear functions, yet their roles in cancer remain enigmatic (Janssen et al., 2018); as such, an interactome of heterochromatin-associated genes could provide crucial insights into how these genes influence chromosomal instability, gene expression regulation, and ultimately, cancer progression. A core set of 18 protein-coding genes known to be enriched in heterochromatin and the nuclear lamina (Table 1) as identified in recent literature (Olins et al., 2010; Margalit et al., 2007; Maison et al., 1997; Poleshko et al., 2013; Zeng at al., 2010; Shibuya et al., 2014; Dunce et al., 2018; Li et al., 2018), was used to find the candidate heterochromatin interactome and generate the set-of-interest (SOI). This initial set of 18 proteins was expanded (n = 327) by querying the STRING protein-protein interaction (PPI) database (Szklarczyk et al., 2023) with the candidate gene set, specifically for genes encoding for physically interacting proteins (Table S1). This led to a SOI of 327 protein-encoding genes. Initially, the SOI was analysed to gain a general Protein name Gene name Anchor factor Man1 LEMD3 BAF LBR LBR HP1 Emerin EMD BAF LAP1B TOR1AIP1 Lamin A/C LAP2b TMPO BAF LEMD2 (LEM2) LEMD2 BAF PRR14 PRR14 HP1 Lamin A/C LMNA Histone Lamin B1 LMNB1 Histone Lamin B2 LMNB2 Histone IFFO1 IFFO1 XRCC4 & Lamin A/C BAF BAF (BANF1) BAF TERB1 TERB1 TRF1 (Shelterin) TERB2 TERB2 TERB1/TRF1 MAJIN MAJIN TERB1/2/TRF1 HP1a CBX5 H3K9 HP1b CBX1 H3K9 HP1y CBX3 H3K9 Table 1. Heterochromatin & LAD protein set Table of the intitial 18 Heterochromatin and LAD proteins selected to build the heterochromatin interactome. The anchor factor specifies the complex or protein linking the Heterochromatin protein to the Nuclear Lamina. understanding of its transcriptional characteristics and evaluate its potential for studying chromosomal instability in the context of cancer. To achieve this, TCGA’s PanCan dataset was used to obtain RNAseq data for the SOI in cancer samples (n = 10535), as well as healthy samples (n = 725). Additionally, unique control sets (n = 547) composed of genes not included in the SOI but identical in number of genes (n = 327) were generated to provide multiple baselines for comparison across both conditions. When assessed with each other and the SOI, the control sets showed distinct transcriptional profiles. This was validated by Friedman’s test (p <2.2e-16), implying control sets failed to show identical differences in their mean expression levels when compared to the SOI’s mean expression level, in both the cancerous as well as healthy conditions. To validate the uniqueness of the control sets, we conducted a contrast analysis for each set. This involved comparing the difference in mean expression between the set of 5 interest (SOI) and the control set to the global mean difference between the SOI and all other control sets. We repeated this process for each control set to ensure thorough validation. This analysis revealed that the majority of control sets differed when measured against the global trend (Mann-Whitney U, adjusted p < 0.05) for A C Control Set Background Gene Set of Interest Gene of Interest 25 150 CBX3 − Log10 P − Log10 P 20 15 10 SYNE1 CBX7 100 FOS SMARCA2 GABARAPL1 PPP2CB RHOJ 50 DSN1 DNMT1 TOP2A KIF20A EZH2 CENPA CDCA8 AURKB CDK1 CHAF1B HMGA2 QKI EGF 5 0 0 −0.2 −0.1 0.0 0.1 Log2 fold change 0.2 −2 −1 0 1 Log2 fold change 2 3 B Control Set Set of Interest − Log10 P Breast invasive carcinoma Head and Neck squamous cell carcinoma 7.5 7.5 5.0 5.0 2.5 2.5 0.0 0.0 −0.50 −0.25 0.00 0.25 0.50 −0.50 Lung adenocarcinoma −0.25 0.00 0.25 0.50 Lung squamous cell carcinoma 7.5 7.5 5.0 5.0 2.5 2.5 0.0 0.0 −0.50 −0.25 0.00 0.25 0.50 −0.50 Log2 fold change −0.25 0.00 0.25 0.50 Figure 2. Differential gene expression patterns of the interactome set-of-interest in cancerous versus normal tissues. A&B) Differential expression (log2 of the mean differences in expression values between paired samples; x-axis) in the set-of-interest (red) compared to 547 randomly generated control sets (blue). For all healthy and cancerous paired samples (A), and selected cancer types (B). Adjusted p-values (y-axis) calculated using the wilcoxon signed-rank test. C) Differential gene expression, highlighting genes in the set-of-interest (red) compared to background genes from the TCGA dataset (blue) for all paired samples. P-values (y-axis) and fold changes (x-axis) calculated with edgeR package. All p-value cutoffs set to 0.01. 6 both the cancerous subset (99.7%) and the normal subset (96.2%), indicating that the control sets would indeed serve as a broad and distinct range of baselines. To determine whether the SOI’s expression differs significantly between cancerous and healthy conditions compared to the controls, we conducted a differential analysis of all gene sets using only paired cancer-healthy samples. We found that the SOI distinctly showed a significant change in average expression (Paired permutation test, p = 10-20) that differed from the general trend presented by the control sets (Fig.2a). Moreover, when refining the analysis to specific cancer types, this aberration from the trend was also observed in three of the four representative cancer types (breast invasive carcinoma, head and neck squamous cell carcinoma & lung squamous cell carcinoma; Fig.2b). These results seemed to support the suitability of the defined set for the next analyses. We further resolved the expression characteristics of the individual genes within the SOI by performing a differential gene expression analysis (DGEA) on the full TCGA genome across cancerous and healthy conditions (Fig.2c). This analysis demonstrated that several of the genes within our set display significantly altered expression in cancer which could drive the distinctive aggregate expression patterns observed previously (Fig.2a). These genes, which include TOP2A, KIF20A, CENPA, AURKB, EZH2 and CHAF1B mostly code for structural, mitotic and chromatin modifier proteins. The standout genes proved to be fairly invariable when the DGEA was repeated on the representative cancer types (Fig.S1). Thus, the heterochromatin interactome, which constitutes the SOI, offers significant and unique value for analysing heterochromatin-based aberrations and their relation to chromosomal instability in cancer. 3.2 Employing a Machine Learning model to decode the impact of heterochromatin on chromosomal instability features To relate the aberrations in the expression of heterochromatin-associated genes to the presence of CIN features, we developed a machine learning-based approach with a multioutput stacked ensemble model to uncover the underlying links (Fig.3). The model used the cancerous SOI RNAseq data as input and estimated multiple features (n = 56) indicative of CIN. The model was structured for the classification of arm-level aneuploidies (n = 39), and estimation of homologous recombination deficiency (HRD) signatures (n = 3) as well as copy number variation (CNV) values for a selection of pericentromeric genomic regions (n = 14). We selected these specific features for measuring CIN based on their established roles in depicting chromosomal instability. Arm-level aneuploidies were chosen due to their direct representation of chromosome number variations, which are pivotal in understanding the extent of chromosomal missegregation, a hallmark of CIN (Baker et al., 2024; Adell et al., 2023). Homologous recombination deficiency (HRD) signatures, including loss of heterozygosity scores (Abkevich et al., 2012), number of telomeric imbalances (Birkbak et al., 2012), and large-scale state transitions (Popova et al., 2012), are reflective of defects in DNA repair mechanisms, ultimately contributing to genomic instability (Pellegrino et al., 2020; Doig et al., 2023). Finally, copy number variations in pericentromeric regions were included since these heterochromatin-enriched regions are known to undergo chromosomal rearrangements but remain poorly understood in terms of their genomic importance (Ehrlich, 2002; Barra & Fachinetti, 2018). 7 Regression Classification 4q 4p 3q SOI RNAseq Dataset 3p 2q 2p 1q 1p ai1 loh Input lst1 peri1 peri2 peri3 Base Layer Base Predictions 4q 4p 3q 3p 2q 2p 1q 1p ai1 loh lst1 peri1 peri2 peri3 4q 4p 3q 3p 2q 2p 1q 1p ai1 loh lst1 peri1 peri2 peri3 Meta Layer Output Figure 3. Configuration of the Chromosome Instability Model Configuration of the XGBoost-based, stacked, multi-output model. Together, we expect these features to provide a solid representation of structural DNA anomalies associated with CIN. The HRD and arm aneuploidy feature types were readily provided for the PanCan TCGA samples. For the pericentromeric copy number (CN) features, a custom approach was devised which involved leveraging annotations for transposable element repeats throughout the human genome from recent studies (Hoyt et al., 2022; Altemose et al., 2022), made possible by the Telomere-toTelomere reference genome (Nurk et al., 2022), and associating them with the segmented CN records for the TCGA samples (see methods). The configuration of the model employed a stacked design where the target CIN features were initially inferred independently of one another with extreme gradient boosting (XGBoost; Chen & Guestrin, 2016) learners followed by a second layer of predictions with XGBoost learners using the previous predictions as input (Fig.3). After the final predictions, the original genes in the SOI were assessed in terms of their impact on the presence of the distinct CIN features. 3.3 The Chromosome Instability model accurately classifies arm-level aneuploidies The model accurately classified most armlevel aneuploidy features achieving a 8 global mean accuracy of 0.805. It performed slightly better in the base layer (mean Accuracy = 0.807) than in the meta layer (mean Accuracy = 0.802) (Table S2). The worst accuracy occurred for 18p, the p arm of chromosome 18 (mean Accuracy = 0.731), which saw an increase in performance from the base prediction A (base Accuracy =0.727) to the meta prediction (Accuracy = 0.736). The best accuracy occurred for 5q (mean Accuracy = 0.859) despite seeing a decrease in performance from the base to the meta layer (Base Accuracy = 0.862, Meta Accuracy = 0.856) (Fig.4a). Accuracy of the Aneuploidy Learners 1.0 Accuracy 0.9 Learner Type 0.8 Base Meta 0.7 10 p 10 q 11 p 11 q 12 p 12 q 13 q 14 q 15 q 16 p 16 q 17 p 17 q 18 p 18 q 19 p 19 q 20 p 20 q 21 q 22 q 9p 9q 8p 8q 7p 7q 6p 6q 5p 5q 4p 4q 3p 3q 2p 2q 1p 1q 0.6 Learner B Pericentromeric and HRD Learners in terms of R2 1.00 0.75 R2 Learner Type Base 0.50 Meta 0.25 _3 22 20 ri_ pe ri_ pe pe ri_ 17 _2 16 _1 ri_ pe 16 ri_ pe ri_ 10 9 pe ri_ 8 pe pe ri_ 7 pe ri_ 5 pe ri_ 4 pe ri_ 3 pe ri_ 2 pe ri_ 1 ri_ pe h_ hr d t1 lo ls ai 1 0.00 Learner Figure 4. Performance of the general Chromosome Instability model predictions. Predictive performance of the individual XGBoost learners (x-axis) determined using the overall accuracy (y-axis) for the 39 classification learners (A), and the coefficient of determination (R2; y-axis) for the 17 regression-based learners (B), assessed for both the base layer (blue) and meta layer (red). 9 The arm-level aneuploidy features unanimously displayed a high degree of imbalance across the three possible classes, with the “normal” aneuploidy class consistently making up more than two-thirds of the distribution. Hence, we included a probabilistic “class weight” parameter in the arm-level aneuploidy learners during modelling to account for the imbalance. This method seems to have proved efficient since 76.9% of these learners significantly outperformed their corresponding baseline model (p <= 0.05) which solely classifies using the majority class (Table S2). The general accuracy tended to decrease from the base layer to the meta layer, while the F1 score saw an increase when going from the base layer classifications (mean base F1 = 0.718) to the meta layer (mean meta F1 = 0.725) (supplementary data). This increase in overall F1 score was observed for over half of the learners (59%) when averaging across all classes and was particularly relevant for minority class F1 scores (69% for the “loss” class; 56% for the “gain” class) in contrast to the majority class F1 scores which did not follow this trend (21%) (supplementary data). We concluded that the general model’s classification demonstrates a strong capability in accurately identifying armlevel aneuploidy features, with a notable global mean accuracy of 0.805. Further analysis of the cancer-specific subsets revealed that while the learners maintained a performance within a similarly consistent range, the accuracy was somewhat lower (Fig.S2). Thereby indicating that the model’s generalisability across different cancer types may vary. 3.4 The Chromosome Instability model exhibits variable performance in estimating HRD and Pericentromeric features Although the model performed consistently for the classification objectives, its performance on regressionbased tasks was more varied. When projecting the remaining CIN features, the predictiveness of the individual learners showed a high degree of variability, with the coefficient of determination (R2) for the best-performing learner, peri_20 (mean R2 = 0.765) and the worst, peri_8 (mean R2 = 0.225) separated by 0.54 (Table S3). The variation in performance was shown to be less important for the HRD-based learners, however, whose R2 remained above 0.570 (Fig.4b). Interestingly, across all but one regression learner, the performance increased upon the transition from the base layer to the meta layer (mean base R2 = 0.445 vs mean meta R2 = 0.472). In the sole exception, peri_3, the learner performance was identical across both layers (R2 = 0.510). When considering the root mean squared error (RMSE) however, a decrease from baselayer to meta-layer estimations was observed for every feature, emphasising the performance improvement as a result of the stacked approach. With a relatively low global average R2 (0.459), a significant amount of variability among the individual learners, and a reliable improvement in the meta-layer compared to the base layer, the performance for the regression portion of the model proved to be mixed. These results highlight the challenges in accurately predicting certain features with consistency and suggest a potentially more complex relationship between the expression of the genes within the SOI and certain characteristics of chromosome instability. When the model was applied to cancerspecific subsets, the observed trends largely mirrored those of the general model, albeit with lower performance metrics (Fig.S3). Notably, the cancerspecific analysis revealed greater 10 variability in the predictive dominance of meta-learners compared to base-learners, deviating from the consistency observed in the general model. This suggests that while the model maintains its predictive approach across different cancer types, the performance nuances in these subsets underscore the complexity of CIN features’ behaviour in varied oncological contexts. 3.5 Feature Importance analysis reveals significant gene impacts on Chromosome Instability features To examine the impact of specific genes on the prediction of distinct features, we employed a feature importance analysis using the gain metric (G) which calculates the relative improvement in predictive performance attributable to a specific input feature. As a result of having a large number of input genes, only the top contributing genes were included in the evaluation. The results from the base-layer genomic feature analysis revealed that a total of 45 different genes contributed the most to the base-learners, yielding a large number of paired gene-to-learner contributions (56*45) (Fig.5a). MCM3AP displayed the greatest individual impact in the analysis from its contribution to the 21q learner (GMCM3AP = 0.255). Yet, MCM3AP’s average impact was low (mean GMCM3AP = 0.007, Table 2), slightly above the base layer’s uniform gain threshold (UGTbase = 0.003), the theoretical gain given an equal contribution by all genes (UGTbase = 1/327) and below the average impact of the 44 genes (0.008). CENPA possessed a greater overall influence (mean GCENPA = 0.020) and was the top gene for all three HRD-based learners, ai1 (GCENPA = 0.216), lst1 (GCENPA = 0.176) and loh_hrd (GCENPA = 0.123) (Table 2). Overall, the impact of the top genes showed notable variability. A moderate proportion of gene-learner pairs (14.8%) CIN Feature Gene Impact (Gain) Average impact 1p 1q 2p 2q 3p 3q 4p 4q 5p 5q 6p 6q 7p 7q 8p 8q 9p 9q 10p 10q 11p 11q 12p 12q 13q 14q 15q 16p 16q 17p 17q 18p 18q 19p 19q 20p 20q 21q 22q ai1 lst1 loh_hrd peri_1 peri_2 peri_3 peri_4 peri_5 peri_7 peri_8 peri_9 peri_10 peri_16_1 peri_16_2 peri_17 peri_20 peri_22_3 ZNF644 NUP133 PCGF1 XRCC5 ZNF621 RPN1 CDK1 PGRMC2 NUP155 CAMLG BAG6 ORC3 GET4 AGK CHMP7 ESRP1 PSIP1 TOR1A SUV39H2 DNAJB12 RRP8 EED PHC1 SLC25A3 UBAC2 PPP2R5C ZNF280D USP7 VPS4A C1QBP CHMP6 SMCHD1 MBD1 SGTA TRIM28 DDRGK1 VAPB MCM3AP XRCC6 CENPA CENPA CENPA SETDB1 PCGF1 RPN1 YTHDC1 CAMLG GET4 ESRP1 TOR1A DNAJB12 VPS4A VPS4A MPRIP CHMP4B SMARCB1 0.151 0.120 0.101 0.157 0.116 0.195 0.045 0.091 0.200 0.214 0.198 0.229 0.136 0.142 0.216 0.152 0.160 0.158 0.120 0.239 0.157 0.188 0.130 0.199 0.157 0.096 0.152 0.149 0.249 0.136 0.133 0.222 0.151 0.104 0.115 0.150 0.122 0.255 0.142 0.216 0.176 0.123 0.153 0.167 0.196 0.060 0.061 0.120 0.065 0.115 0.148 0.103 0.251 0.068 0.253 0.188 0.005 0.004 0.009 0.006 0.004 0.010 0.020 0.006 0.014 0.007 0.006 0.007 0.011 0.006 0.011 0.008 0.005 0.007 0.004 0.012 0.005 0.005 0.005 0.007 0.004 0.004 0.005 0.006 0.014 0.004 0.007 0.005 0.006 0.004 0.004 0.005 0.008 0.007 0.006 0.020 0.020 0.020 0.007 0.009 0.010 0.005 0.007 0.011 0.008 0.007 0.012 0.014 0.014 0.006 0.009 0.007 0.154 0.008 Average Table 2. Top genomic impact by CIN feature Table of the CIN features and the top genomic contributor to their base-learner in terms of impact (Gain) with the corresponding value, and the gene's average impact across all features. 11 peri_22_3 peri_20 peri_17 peri_16_2 peri_16_1 peri_10 peri_9 peri_8 peri_7 peri_5 peri_4 peri_3 peri_2 peri_1 loh_hrd lst1 ai1 22q 21q 20q 20p 19q 19p 18q 18p 17q 17p 16q 16p 15q 14q 13q 12q 12p 11q 11p 10q 10p 9q 9p 8q 8p 7q 7p 6q 6p 5q 5p 4q 4p 3q 3p 2q 2p 1q 1p Gain 0.25 0.20 0.15 0.10 0.05 0.00 AG B K C AG 1 6 C QB AM P C LG C DK C EN 1 H P M A C P4 H B C MP D HM 6 D P D RG 7 N K AJ 1 B1 E 2 ES ED R G P1 ET M M B 4 C D M 1 3 M AP N PR U IP N P13 U 3 P1 O 55 P RC PG CG 3 R F1 M C PP PH 2 P2 C1 R PS 5C I R P1 P R N1 SE RP TD 8 B SL SG 1 SM C2 TA 5 A A SM RC 3 B SU CH 1 V3 D1 TO 9H2 TR R1 I A U M28 BA U C2 S VA P7 VP PB XR S4A XRCC YT CC5 ZN HD 6 F2 C1 ZN 80 D ZNF62 F6 1 44 Base Learner A Gene Uniform Gain Threshold: 0.003 peri_22_3 peri_20 peri_17 peri_16_2 peri_16_1 peri_10 peri_9 peri_8 peri_7 peri_5 peri_4 peri_3 peri_2 peri_1 loh_hrd lst1 ai1 22q 21q 20q 20p 19q 19p 18q 18p 17q 17p 16q 16p 15q 14q 13q 12q 12p 11q 11p 10q 10p 9q 9p 8q 8p 7q 7p 6q 6p 5q 5p 4q 4p 3q 3p 2q 2p 1q 1p Gain 0.75 0.50 0.25 0.00 1p 1q 2p 2q 3p 3q 4p 4q 5p 5q 6p 6q 7p 7q 8p 8q 9p 9 10q 10p 11q 11p 12q 12p 13q 14q 15q 16q 16p 17q 17p 18q 18p 19q 19p 20q 20p 21q 22q q ai 1 lo lst h_ 1 pe hrd peri_1 peri_2 peri_3 peri_4 peri_5 peri_7 peri_8 r peper i_9 i_ peri_1 10 ri_ 6_ 1 pe16_ r i_ 2 p 1 e pe r 7 ri_ i_2 22 0 _3 Meta Learner B Base Learner Uniform Gain Threshold: 0.018 Figure 5. General Model Feature Importance. The impact of the input features towards the individual model learners in terms of the Gain metric (gradient). Evaluating the impact of the base-learner’s top contributing genes (x-axis) on all base-learners (y-axis; A), and the impact of the base learners (x-axis) on all meta-learners (y-axis; B). Displaying the Uniform Gain Threshold values (bottom right; see methods). 12 provided no apparent contribution (G = 0.000) while a smaller fraction of pairs (2%) supplied more than a tenth of the total contribution (G > 0.100) towards specific learners (Fig.5a). These findings accentuate the complex nature of gene contributions to chromosomal instability and demonstrate the nuances captured by the model, with certain genes exerting significant influence across multiple learners while others show substantial impact on unique features with narrower influence. Analysing the genomic feature importance from the cancer-specific models, we saw they had a similar number of total contributing genes (45-49) (Fig.S4). We also found that only nine contributing genes were shared between the different cancer-specific feature importance analyses and that each model was represented by a unique most impactful gene based on the greatest gain, YBX1 (LUSC, 1p GYBX1 = 0.257), EGFR (LUAD, peri_7 GEGFR = 0.287), MAPRE1 (BRCA, peri_20 GBRCA = 0.331) and EED (HNSC, 11q GEED = 0.400), all with varying degrees of impact and to distinct base-learners (Fig.S5). This variability in gene contributions across different cancer types not only highlights the tailored nature of genomic interactions in cancer but also reflects the model’s capability to adaptively capture these distinct interactions. 3.6 Exploration of feature interactions reveals selective dependencies in Chromosome Instability We then explored dependencies between CIN features within our stacked model configuration. This analysis considered the initial base-layer predictions and their corresponding contributions to each of the refined meta-layer learners. Here, the UGTmeta (0.018) was higher than UGTbase due to the number of base-learners (n = 56) being less than the number of genes. The analysis revealed that feature-specific base-learners were the primary contributors to their corresponding metalearners for most CIN features (95%). Exceptions included three features peri_9, peri_16_2, and peri_22_3. For these cases, the main contributors were the q-arm aneuploidy features for the same chromosomes (9q, 16q, and 22q respectively; Fig.5b). The findings also demonstrated a high degree of variability in the contributions of the most impactful learners. The contribution from 4p’s base-learner to its meta-learner (G4p = 0.254) was approximately one-third of peri_20’s baselearner to its specific meta-learner (Gperi_20 = 0.908). We found a few minor interactions among the combinations occurring between pericentromeric baselearners and arm-specific meta-learners targeted to the same chromosome (Fig.5b). The peri_2 and 2q base-learner to metalearner contribution (Gperi_2 = 0.265) demonstrate this phenomenon. The same types of contributions in the reverse, from the arm-level base-learner to the pericentromeric meta-learner, did not generate the same degree of impact (peri_2 G2q = 0.018). The lack of contributions between different feature-specific learners was further revealed by low median values of meta-learner gains, the median contribution a chosen meta-learner receives, which never exceeded 0.003 (Table S4). This observation was further emphasised by low median base-learner contributions, the median impact contributed by a chosen base-learner, which only exceeded the UGTmeta for the ai1 base-learner (median Gai1 = 0.042). The examination of CIN feature interactions ultimately highlighted nuanced, yet selective dependencies between base and meta-layer contributions. This suggests that distinct CIN features can selectively interact, but that broad, large-scale dynamics are uncommon. 13 Viewing the interactions in the context of the individual cancer models, we found that the trend showing feature-specific base-learners contributing the most to their corresponding meta-learner largely remained true. Yet, the interactions demonstrated some significant differences from the general model, as well as discrepancies between the unique cancer models (Fig.S5). For the LUSC model, the loss of heterozygosity (loh_hrd) baselearner showed an elevated impact on all HRD-based meta-learners (Gloh_hrd = 0.268 to 0.570) and greater than that of ai1 which tended to be more impactful across the HRD features for all other models. The LUSC and BRCA models demonstrated relevant contributions from chromosome arm feature base-learners to pericentromeric feature meta-learners which differed from the general model pattern where the reverse was more prominent. Additionally, pericentromeric feature and HRD feature base-learner contributions were noticeably greater and more wide-ranging, affecting all types of meta-learners for the HNSC model and to an even greater extent for the LUAD model. The cancer-specific feature interaction analyses reveal a marked shift from the general model’s feature interactions and suggest that CIN feature dynamics are uniquely modulated across various cancer types. 4. Discussion The current study aimed to delineate the roles of heterochromatin-associated genes in cancer by constructing a comprehensive interactome and employing an advanced machine-learning model to assess their impact on chromosomal instability (CIN). Our findings emphasise the identification of potentially relevant heterochromatin genes that significantly influence chromosomal instability, providing new insights into the molecular mechanisms underpinning cancer development and progression. The integrity of the analysis depended on both a core gene set and an extended interactome, which together constructed a representative set of heterochromatinassociated genes potentially dysregulated in cancer. The interactome constructed using the STRING query included numerous genes previously linked to heterochromatin, structurally or functionally, several of which had been implicated in cancerous dysregulation. Polycomb group gene EZH2 is pivotal for transcriptional repression and was shown to be upregulated in aggressive cancers, such as neuroendocrine prostate cancer 18. Similarly, topoisomerases, especially TOP2A, have been demonstrated to be essential for maintaining chromatin integrity and potentially contribute to cancer when defects arise (Lee et al., 2019; Amoiridis et al., 2024). Additionally, the centromere-specific CENP-A is vital for maintaining the integrity of centromeric DNA during replication. Its removal causes replication stress, and mitotic defects (De Rop et al., 2012), contributing to chromosomal instability (Giunta et al., 2021). Including these and other welldocumented genes during the extension phase validates the methodology's robustness and illustrates the careful approach taken to ensure the integrity and relevance of the study in understanding the role of heterochromatin-associated genes in chromosomal instability in cancer. Our genomic feature importance analysis identified several heterochromatinassociated genes as significant contributors to CIN features. These genes include CENP-A, NUP133, NUP155, and SETDB1, each playing a distinct role in maintaining genomic stability and influencing cancer progression. The analysis identified CENP-A as a significant contributor to various CIN features including numerous arm 14 aneuploidies. This identification of CENPA in terms of these structural CIN features proves consistent with prior studies which identified defects in CENP-A with missegregation events contributing to the increased presence of aneuploidies (De Rop et al., 2012; Giunta et al., 2021; Black et al., 2018). The analysis in the general model also highlighted SETDB1, a histone methyltransferase, as a significant contributor to the pericentromeric CIN feature on chromosome 1. Robbez-Masson et al. (2017) showed that the loss of SETDB1 function leads to the upregulation of retrotransposons, resulting in genomic instability. Retrotransposons, a type of transposable element (TE), have been extensively mapped to pericentromeric regions of chromosome 1 (Nurk et al., 2022; Altemose et al., 2022), suggesting that the identified importance of SETDB1 in our analysis may be due to its role in maintaining the stability of these regions by repressing retrotransposon activity, thus preventing the genomic instability associated with copy number variations in these critical areas. These results reinforce the central roles of specific heterochromatin-associated genes in contributing to distinct CIN features. Nucleoporins, the building blocks of Nuclear Pore Complexes (NPCs), redistribute to structures such as kinetochores, spindle poles, and the mitotic spindle during mitosis (Chatel & Fahrenkrog, 2011). Prior studies have shown that NPCs are important in DNA damage repair in heterochromatin (Ryu et al., 2015), correlated NUP155 with microsatellite instability (Wang et al., 2024), and linked mutations or altered expression of nucleoporins to missegregation events leading to aneuploidies (Nakano et al., 2011). Our results not only corroborated these findings by underlining NUP133 and NUP155 as significant CIN contributors but also revealed their specific impact on distinct features, associating NUP133 with aneuploidies on the q arm of chromosome 1 and NUP155 with aneuploidies across 5p, 5q, 8p and 13q among others. This further emphasises the fundamental role of nucleoporins in safeguarding genomic integrity and highlights the consequences of their dysregulation in promoting chromosomal missegregation and aneuploidy in cancer cells, offering new insights into potential mechanisms underlying these phenomena. Notably absent from the 45 distinct genes revealed during the genomic feature importance analysis were any of the genes constituting the initial core gene set. In cancer-specific scenarios, the breast cancer model’s feature importance analysis identified CDK1 as the top contributor to the CIN feature associated with the pericentromeric region of chromosome 1. CDK1 plays a critical role in cell cycle regulation and is frequently overexpressed in various cancers, including breast cancer (Ryu et al., 2015; Sofi et al., 2022). Malumbres and Barbacid (2009) describe how CDK1, along with other cyclindependent kinases (CDKs), controls cell cycle progression and ensures proper chromosome segregation during mitosis. Misregulation and overexpression of CDK1 can disrupt mitotic processes, leading to chromosomal instability. The identification of CDK1’s impact on the pericentromeric CIN feature aligns with these findings, supporting its role in driving chromosomal instability in cancer. In the head and neck cancer CIN model, CDC42 was also identified as the top contributor to the CIN feature involving the p arm of chromosome 1. CDC42, known for regulating cell cycle progression, cytoskeletal organisation, and cellular migration, has been implicated in multiple cancers due to its role in oncogenic processes like invasion and metastasis (Hodder et al., 2023). CDC42associated kinase (ACK) promotes cancer progression by regulating protein stability and degradation pathways. The identification of CDC42 in our model is 15 consistent with its role in maintaining chromosomal stability through actin dynamics and cell division. Overall, the genomic feature importance analysis confirms existing knowledge and proposes novel mechanisms by linking various heterochromatin-associated genes to chromosomal instability features. This demonstrates the robust and exhaustive aspect of our methodology, indicating its potential to provide valuable insights into the molecular underpinnings of cancer. The cancer-specific feature importance analyses further enhance our understanding by highlighting unique gene contributions across different cancer types. These findings lay the groundwork for further research to explore these potential connections and their implications for understanding chromosomal instability in cancer. The feature importance analysis in our study relied on identifying the top contributors, based on Gain, to each CIN feature. While this approach highlights key genes, it has potential drawbacks. Specifically, any gene which was not the top contributor to any one feature was excluded from the final analysis, even if it contributed significantly to multiple CIN characteristics. This limitation could have caused us to overlook important genes which play a substantial role in chromosomal instability. To address this, future analyses could include more than just the top gene per feature. Additionally, redefining the impact of genes on the features by using alternative feature importance metrics provided by XGBoost, such as coverage or frequency, or a combination of these metrics, may offer insights from a different perspective. Coverage measures the number of samples affected by a particular gene, while frequency assesses how often a gene is used in a learner’s decision-making process. Integrating these metrics could provide a more nuanced picture of each gene’s contribution to CIN features. Moreover, the inferences drawn from the feature importance analyses must be viewed in the context of the model's overall performance. Gain measures how the performance of a learner improves with the inclusion of a specific gene. Therefore, a low-performing base-learner can assign significant importance to certain genes, which may skew the perceived impact of these genes. Consequently, any conclusions made should consider not only the feature importance metrics but also the learners' overall performance to ensure a balanced and accurate interpretation of the results. The performance evaluation of our machine-learning model, encompassing both classification and regression tasks, accentuates its effectiveness in capturing key aspects of chromosomal instability (CIN). The classification tasks, particularly for arm-level aneuploidies, demonstrated robust accuracy and improved F1 scores from the base to the meta layer, indicating a balanced model more adept at handling diverse class distributions. However, the regression tasks exhibited variability in performance, reflecting the inherent complexity of predicting certain CIN features. The observed improvements in the meta layer for the regression tasks suggest that the stacked model configuration enhances predictive refinement. Yet the variability and limited improvement indicate the need for further optimisation to achieve consistent performance. This mixed performance highlights the model’s strengths in classification while pointing to areas for improvement in regression, emphasising the multifaceted nature of genomic instability and the challenges in providing wide-ranging predictions. Exploring alternative machine-learning models for the individual CIN base-learners, such as Random Forest, which has demonstrated notable performance when handling both class imbalances and high-dimensional 16 biological data (Anaissi et al., 2013; Chafai et al., 2023), or Support Vector Machines which have proven effective with genomic feature selection and classification through ensemble methods (Anaissi et al., 2016; Chafai et al., 2023), could potentially enhance the prediction accuracy and stability for specific CIN features. The second importance analysis, which explored interactions between CIN features, revealed selective dependencies that offer deeper insights into the structural anomalies associated with chromosomal instability. The identification of specific arm aneuploidies significantly impacting pericentromeric meta-learner predictions stresses the interconnected nature of certain CIN features. This nuanced understanding of feature interactions suggests that while broad-scale interactions are rare, specific dependencies play a vital role in the manifestation of genomic instability. The variability in cancer-specific analyses further indicates that CIN feature dynamics could be uniquely modulated across different cancer types, highlighting the model’s ability to adapt to distinct oncological contexts. These findings enrich our broader analysis by emphasising the importance of considering both individual gene impacts and feature interdependencies, ultimately enhancing our understanding of the complex mechanisms underlying chromosomal instability in cancer. 5. Conclusion In conclusion, this study successfully delineated the roles of heterochromatinassociated genes in cancer by employing a broad heterochromatin interactome and an ensemble machine-learning model to assess their impact on chromosomal instability (CIN). The methodology proved robust, combining a core gene set and an extended interactome to identify significant contributors to CIN features. The genomic feature importance analysis not only corroborated existing knowledge but also revealed novel mechanisms, highlighting the contributions of key genes across various cancer types. While the model demonstrated strong performance in classification tasks, particularly for armlevel aneuploidies, it also showed variability in regression tasks, affirming the complexity of predicting CIN features. Additionally, the analysis of interactions between CIN features emphasised the intricate dependencies among them, suggesting that specific structural anomalies are interlinked. These findings enhance our understanding of the molecular mechanisms underlying chromosomal instability in cancer, providing a solid foundation for subsequent research. Future studies could employ knockdowns, knockouts, or induced over-expression of highlighted genes to further explore these connections and their implications regarding CIN in greater detail. 6. Methods Candidate Protein Selection and Interactome Construction To identify a heterochromatin-specific interactome, an initial list of 18 heterochromatic proteins recognised for their association with the nuclear lamina and interactions with heterochromatin, based on literature and annotations from the UniProt database (Wang et al., 2021) was manually curated (Table 1). To explore the interaction landscape of these candidate proteins, a Python script (STRING_Interactome_Generation.py) was developed to interface with the STRING Protein-Protein Interaction (PPI) database version 12.0 (Szklarczyk et al., 2023). This script was designed to translate protein names into STRING database identifiers, subsequently retrieving lists of genes encoding proteins that physically interact with the candidates. 17 Interaction queries were filtered to include only physical network interactions with a minimum confidence score of 0.4 (scale 0 to 1), ensuring a thorough yet relevant interactome dataset consisting of 327 unique genes including the initial candidates. The Python script executed its query via a custom function. Each protein name from the provided list was converted into the corresponding STRING identifier using the database’s API. This conversion process secured the most relevant identifier for each protein, by referencing human taxonomy with the NCBI identifier 9606 (Schoch et al., 2020; Sayers et al., 2019). Subsequently, for each obtained STRING identifier, data on physical interaction partners was sought. To efficiently manage the volume of data, the results were constrained to a maximum of 1,500 interactions per protein. The output from the STRING database composed of a unique list of gene names, encompassing both the original candidates and their interaction partners, formed the interactome for subsequent analyses. 6.1 Acquisition and Processing of TCGA Data RNA-seq Data RNA sequencing data from The Cancer Genome Atlas (TCGA) Pan-Cancer (PANCAN) initiative was utilised for this study, accessed through the TOIL project via the Xena Browser platform (Goldman et al., 2021). This data includes gene expression patterns from the TCGA samples. Two primary datasets were utilised: the tcga_RSEM_gene_tpm and the tcga_RSEM_gene_expected_counts. Each dataset includes data from 10,535 samples and covers approximately 60,499 unique gene identifiers. The tcga_RSEM_gene_tpm dataset comprises transcripts per million (TPM), normalised and log-transformed to log2 (tpm+0.001) which facilitates consistent analysis across low and high expression genes. Gene identifiers in this dataset are aligned with the Gencode v23 annotation, ensuring precise gene mapping and relevance to recognised genomic features (Frankish et al., 2023). The tcga_RSEM_gene_expected_counts dataset provides raw count measures of gene expression. This format was selected to support specific quantitative analyses such as library size normalisation (Vivian et al., 2017). The raw tcga_RSEM_gene_expected_counts dataset was processed into a structured CSV file using the TCGA_Raw_Data_Processing.py script Copy Number Variation Data To examine genomic alterations across various cancers, copy number variation (CNV) data for TCGA samples was collected, using two distinct types of CNV data. The first type, gene-level CNVs, was sourced from the Xena Browser, utilizing the SNP6 array method across all TCGA cohorts. This dataset represents gene-level copy number alterations expressed as log ratios of tumour to normal tissue. The processing of this gene-level data employs the GISTIC2 method as part of the TCGA pipeline, which maps segmented CNV data to gene-level estimates (Mermel et al., 2011). The second type of data, segmented CNVs, was acquired from the Genomic Data Commons (GDC) (Grossman et al., 2016). This dataset includes data for all TCGA cohorts, providing a detailed representation of chromosomal alterations, segmented by chromosome and base pair regions. Chromosome Instability Data Types To assess the genomic stability across all cancer samples, data on chromosome instability was analysed. This included arm-level aneuploidies and aneuploidy scores, along with three measures of 18 homologous recombination deficiency (HRD). Arm-level aneuploidies and aneuploidy scores were sourced from the ABSOLUTE purity/ploidy file, named TCGA_mastercalls.abs_tables_JSedit.fixe d.txt. This dataset quantifies whole-arm or segmental gains and losses of chromosomes by sample, providing a detailed measure of aneuploidy across the TCGA PANCAN cohort. HRD scores, measuring homologous recombination deficiency, encapsulate multiple aspects of chromosomal instability, including LOH (loss of heterozygosity, “loh_hrd”), LST (largescale state transitions, “lst1”), and NtAI (number of telomeric allelic imbalances, “ai1”). LST specifically measures chromosomal breakages producing fragments larger than 10 Mb, while NtAI assesses allelic imbalances extending to telomeres. Each score provides a single value per sample. TCGA Sample Metadata Additionally, TCGA metadata was downloaded in TSV format from the GDC (Grossman et al., 2016). The metadata selected included the Tissue Source Site (TSS) codes, which link samples to specific cancer studies, TCGA Study Abbreviations that identify cancer types, and Sample Type Codes that differentiate between cancerous and normal tissues. 6.2 Exploratory Gene-set Analysis Set-of-Interest and Control Sets Generation This analysis, implemented in R, utilised Log2 TPM RNA-seq data from the TCGA PANCAN project, which was restructured to create specific datasets for evaluation. The primary datasets focused on the 327 interactome genes identified via the STRING, collectively termed the Set-ofInterest (SOI). For the cancerous SOI dataset, sample IDs were filtered to include only those from cancerous tissues. Conversely, the non-cancerous SOI dataset excluded any samples derived from cancerous or metastatic tissue types. To provide a comparative framework, multiple cancerous control datasets were generated. Each control set contained a random selection of genes, matching the SOI gene count but excluding any SOI genes. To ensure the uniqueness of each control dataset, impose an upper limit on the number of control sets and prevent redundancy, no control set was composed solely of genes previously used in other sets. This selection process continued iteratively until all non-SOI genes were included in at least one control set, ensuring full genomic coverage. Corresponding non-cancerous datasets were also generated for each cancerous control, maintaining an identical genomic background for accurate comparisons. This methodology generated 547 unique control datasets for the cancerous and noncancerous conditions which were saved in CSV format for subsequent analyses. Comparative Expression Analysis within Cancerous and Non-Cancerous Tissues To enable a complete comparative aggregate expression analysis of the Setof-Interest (SOI) against the control sets, preprocessing involved calculating the mean log2 expression across all genes within the Set-of-Interest (SOI) for each sample. These calculations were performed for both cancerous and noncancerous datasets, using a function, designed to output a summary data frame. This data frame contained the global mean expression value for each sample. A similar approach was adopted for each control set. The process generated a global expression value for each sample by calculating the mean log2 expression across the genes in each control set. The global expression values from the controls were 19 then directly compared with their corresponding SOI global expression values. The difference in global expression values between the SOI and each control set was computed per sample, capturing distinct expression patterns between the SOI and the control sets. These differences were recorded in differential data frames, one for cancerous and another for noncancerous conditions. Additionally, the average expression values from the control sets were also stored in condition-specific data frames. The distribution of expression differences for each SOI-control set pair was visually assessed using violin plots generated using the ggplot2 package in R (Wickham, 2016). These plots provided a detailed view of the shape of the distributions, highlighting any skew or deviations from normality. Based on this, appropriate nonparametric statistical tests were performed. The Friedman test (non-parametric equivalent of the repeated measures ANOVA) was applied using the coin package in R (Hothorn et al., 2006) to determine if there were significant differences across multiple control sets when paired with the SOI. The test’s results indicated whether any specific control set showed significant differences in expression with the SOI, compared to other control sets within the same condition. For each SOI-control pair, the Wilcoxon Signed-Rank Test was conducted through the stats package in R (R Core team, 2013). This compared the mean difference of the aggregate expression levels for the specific pair to the global mean difference of all other SOI-to-control combinations. This test helped identify specific pairings where the expression differences were significantly above or below the overall mean, highlighting unique expression patterns between the SOI and each control set. A two-sided paired permutation test, implemented with the coin package, was performed for each SOI-to-control pair to further assess the significance of the difference between the general expression level of the SOI to that of each control set. By randomly shuffling the labels of the datasets and recalculating the test statistic for each permutation, a distribution of the statistic under the null hypothesis was constructed. The significance levels from these tests were adjusted for multiple comparisons using the BenjaminiHochberg procedure. To understand if the expression patterns observed in cancerous and non-cancerous datasets were consistent across different types of tissues, a narrow version of the analysis was performed, assessing the aggregate expression difference between the SOI and a single control set, parsed by cancer type. A single control set was selected and the paired permutation test with multiple comparison correction was reapplied with the SOI across different tissue types. Comparative Expression Analysis between Conditions and across Cancer Types Two data frames containing the mean expression value across all genes for each participant were created. One, based on the cancerous SOI dataset, and another using the corresponding non-cancerous set. The two data frames were then merged on the participant ID (sample ID) with cancertype information integrated from the metadata to facilitate condition-specific analyses. The difference between each paired metric was calculated yielding the log2 fold change for each sample. A two-sided Wilcoxon Signed Rank test was then applied to compare the log fold changes between normal and cancerous conditions for each sample. The test was repeated for all control sets, the P values were adjusted for multiple comparison and 20 the results were saved to a new data frame. The results were visualised in a volcano plot (Fig.2a), using the EnhancedVolcano package (Blighe et al., 2013), displaying the log fold changes against the negative logarithm (base 10) of the adjusted Pvalues, to emphasise sets with statistically significant and biologically relevant changes in their overall expression. This analysis was further refined by repeating the Wilcoxon Signed Rank test and volcano plotting for subsets of data categorised by cancer type (Fig.2a). A differential gene expression analysis (DGEA) was also performed between the cancerous and non-cancerous conditions using the complete genomic data available. The expected counts RNAseq data was pre-processed by restricting the set to participants with paired (cancer-healthy) samples, exponentiating the counts (reversing the log transformation), removing samples with no counts and genes with low counts (<10), and normalising using the TMM method (Robinson & Oshlack, 2010) to adjust for library size differences across samples using the edgeR package (Chen et al., 2024; Chen et al., 2016; McCarthy et al., 2012; Robinson et al., 2010). The DGEA was visualised in the form of an EnhancedVolcano plot highlighting the SOI genes (Fig.2c). This process was repeated across the different cancer types (Fig.S1). The configuration of the Set of Interest (SOI) and corresponding control sets, the generation of normal gene expression datasets, the comparative exploratory analysis and the DGEA were performed using custom python and R markdown scripts (SOI_and_Control_Set_Configuration.py, TCGA_Normal_Gene_Set_Generation.py, TCGA_Comparative_exploratory_annalysi s.Rmd, and TCGA_dge_analysis.Rmd respectively) 6.3 Chromosome Instability Model A stacked, multi-output, Extreme Gradient Boosting (XGBoost; Chen & Guestrin, 2016) machine-learning model was designed to utilise RNA-seq data from the set of interest (SOI) to predict chromosome instability feature values derived from the TCGA PANCAN dataset and ultimately to quantify the genomic contributions (Fig.3). All development was done in R. RNA-seq Data Preparation The RNA sequencing data utilised was based on the tcga_RSEM_gene_tpm and tcga_RSEM_gene_expected_counts data from the TCGA. The RNA-seq data, comprising both log2 TPM (transcripts per million) and log2 expected counts, was loaded from the saved files. The metadata files including tissue source site identifiers, disease study categorisations, and gene annotations were integrated to filter and annotate the datasets appropriately. The datasets underwent several processing measures. Using the TCGA metadata, samples not classified as cancerous were excluded from the datasets. Both TPM and expected count datasets were exponentiated. Sample identifiers were matched with corresponding metadata to ensure accurate sample. Library size differences were normalised using the same TMM method as for the DGEA. Where multiple entries for a single gene existed, median expression values were computed and used. Post-normalisation, the dataset was further refined by retaining only those genes that are part of the SOI. Final Dataset Compilation: The processed data was structured into 4 formats: • Raw counts • Normalised counts (scaled) • Log-transformed raw counts (log) 21 • Log-transformed normalised counts (log-scaled) Each format was prepared separately for both the TPM and expected counts datasets, yielding four different SOI RNAseq datatypes per condition, which were saved to separate CSV files. The generated datasets underwent a process to segregate individual samples into test and training sets. All training sets contained the same samples to ensure consistency. The training sets were used for the hyperparameter tuning of the model. Whereas the complete data (training + test) was used for the model training and testing which was performed using a five-fold cross validation process. All sample IDs from the eight datasets were extracted. Common sample IDs across all data types were identified through the intersection of sample ID lists. A random 75% of common sample IDs were selected to form the training set, ensuring that the training sets included only samples present in all datasets. For each RNA-seq data type, the full dataset was then split into training and testing sets based on this randomly selected list of sample IDs. The data was arranged to ensure that all datasets maintained the same order of samples, enhancing consistency across the model input data. The training and testing sets for each data type were saved in CSV format. Pericentromeric CNV Analysis & Feature Engineering To assess cancer-associated genomic alterations, an R-based methodology (Pericentromeric_cnv_segmentation.Rmd) for identifying and quantifying CNVs within pericentromeric regions of each TCGA sample was developed. The TCGA’s PANCAN segmented CNV data (Chrom_seg_CNV.txt) was loaded using dplyr. Pericentromeric regions were defined to include classical human satellites such as HSat, gamma (γ), beta (β), and specific alpha (α) satellite sequences, along with centromeric transition (ct) regions based on recent genomic studies involving data generated by the Telomere-to-Telomere (T2T) consortium (Barra & Fachineti, 2018; Hoyt et al., 2012; Altemose et al., 2022). The T2T data was loaded with the rtracklayer package (Chen et al., 2016). The data was accessed via the UCSC genome browser in the form of a BigBed file (Robinson et al., 2010). Genomic segments within the sourced data were filtered based on these annotations to ensure widespread coverage of pericentromeric regions. These segments were initially filtered to exclude non-pericentromeric genomic features such as rDNA segments. The resulting segments within each chromosome were grouped and sorted by start positions. Adjacent segments, where the start position of one corresponded to the end positions of the other, or overlapping segments were merged to form contiguous pericentromeric regions. For each chromosome, the longest contiguous pericentromeric region starting and ending with an HSat was identified. Additionally, pericentromeric segments starting and ending with an HSat and exceeding a threshold of one megabase-pair (1mbp) were also selected to ensure coverage of significant pericentromeric areas. These regions were then labelled with unique identifiers for each chromosome (“peri_” followed by the chromosome number and an additional number if more than one pericentromeric region was found for the chromosome in question). The TCGA segmented CNV data was then filtered to include only segments which partially or fully aligned with any of the newly defined pericentromeric regions. For each pericentromeric region, the overlap with the CNV segments from the TCGA PanCan dataset was calculated to determine the overlap-weighted mean CNV. This metric considered the 22 proportion of overlap between each TCGA segment and the corresponding T2Tdefined pericentromeric region, with a greater extent of overlap resulting in a higher weight for the CNV value of the pericentromeric region for the given sample. The resulting CNV values were aggregated across all TCGA samples, producing a dataset of the copy number aberrations across fourteen distinct pericentromeric regions by sample. Additional Chromosomal Instability (CIN) Features The model incorporates several other chromosomal instability features, namely: TCGA arm-level aneuploidy data for thirty-nine chromosomal arms (categorised as either a -1, 0 or 1), and three numerical and continuous Homologous Recombination Deficiency (HRD) signatures, for each sample from TCGA datasets. Machine-Learning Model Configuration The model employed a stacked architecture with two distinct layers, designed to predict the fifty-six chromosomal instability (CIN) features. In the base layer, each feature was addressed independently with a unique XGBoost learner (model) optimised for the feature’s characteristics, using one of the generated SOI RNA-seq datasets (training and test) as input in a five-fold cross-validation strategy. The cross-validation strategy allowed for the use of the full data to make predictions while limiting any data leakage. XGBoost learners were selected for their ability to handle categorical and numerical data and, their lack of assumptions regarding the distribution of the relationship between the variables. Out-offold (OOF) predictions, generated during the testing phases of the cross-validation, were aggregated by sample into a dataset. This dataset, aligned with true labels, underwent a secondary five-fold division to train the feature-specific meta-learners across the second meta-layer, to enhance the predictions. The final output comprised refined OOF predictions across all CIN features. Model Hyperparameters The learners across both the base and meta layers utilised several parameters precisely tailored to their features. The majority of these are hyperparameters standardly applied to XGBoost models to prevent under or overfitting including maximum tree depth (max_depth), minimum child weight (min_child_weight), number of trees (nrounds), learning rate (eta) and the regularisation parameter gamma (Chen & Guestrin, 2016). Additionally, the datatype for the SOI RNA-seq dataset used as input in the base layer was specifically selected for based on the target CIN feature. Each base-learner therefore used one of the eight possible choices as input. All the parameters assigned and selected were tuned in a ten-fold cross-validation process across both layers separately using the SOI RNA-seq training sets. This process resulted in hyperparameter tables with different hyperparameter values and their resulting evaluation metric (log loss or RMSE). The selected hyperparameter value for each specific combination of parameter, CIN feature, and layer were selected for manually by assessing the best hyperparameters based on the evaluation metrics. The Tables used to curate the specific hyperparameters were created using the several R scripts depending on the learner type, regression or classification, and the layer, base or meta (base_class_tune.r, base_regress_tune.r, meta_learner_tune.r). Class Imbalance To manage class imbalances in the armlevel aneuploidy (categorical) features, class weights inversely proportional to their frequencies were assigned to each 23 learner. These weights were normalised, keeping the values between 0 and 1, with a minimum value set to 0.1 to prevent assigning a weight of 0 to the majority class. This was implemented in an R script (arm_lev_aneu_weight.r) which returned a table with the appropriate class weights to use. Prediction Evaluation The predictions made across all learners were saved as csv files and evaluated separately in terms of the learner type. Classification learners were assessed using mean log loss during cross-validation and based on their accuracy, balanced accuracy, precision, recall, specificity and F1 score for the final test predictions. Regression learners were evaluated using the root mean squared error (RMSE) during cross-validation and using the R2 metric for the final predictions. Feature Importance The primary feature importance analysis involved the quantification of the influence of the input genes on each of the base layer predictions from the CIN learners using XGBoost’s framework. For the heatmap evaluation, only genes identified as top contributors to a CIN learner were included. A subsequent interaction feature importance analysis was implemented after the meta-layer predictions to evaluate interdependencies between CIN features. The quantification metric was defined in terms of Gain, which indicates the contribution of an input feature to a learner by measuring the relative improvement in predictive performance calculated based on the difference in the loss function. The CIN model training, testing and the genomic and interaction feature importance analyses were implemented in an R markdown script (CIN_model_analysis.Rmd). Cancer-Specific Analysis The compiled TCGA PanCan data was parsed by cancer type and the model was repeatedly run for each cancer-specific subset containing 300 samples or more. CIN features with constant identical values throughout a particular subset were not used in the modelling and were therefore dropped from that cancer-specific analysis. Cancer-specific scripts for hyperparameter tuning (cancer_specific_base_class_tune.r, cancer_specific_base_regress_tune.r, cancer_specific_meta_learner_tune.r), class weight configuration (cancer_specific_arm_lev_aneu_weight.r), and predictions and feature importance analyses (cancer_specific_CIN_model_analysis.Rm d) were developed. 7. Abbreviations CIN - Chromosomal Instability HP1 - Heterochromatin Protein 1 LADs - Lamina-Associated Domains PPI - Protein-Protein Interaction SOI - Set of Interest TCGA - The Cancer Genome Atlas TPM - Transcripts Per Million HRD - Homologous Recombination Deficiency CNV - Copy Number Variation XGBoost - Extreme Gradient Boosting DGEA - Differential Gene Expression Analysis NPC - Nuclear Pore Complex TSS - Tissue Source Site GDC - Genomic Data Commons LUSC - Lung Squamous Cell Carcinoma BRCA - Breast Carcinoma LUAD - Lung Adenocarcinoma HNSC - Head and Neck Squamous Cell Carcinoma AI - Allelic Imbalance LST - Large-Scale State Transitions 8. Data Availability The TCGA PanCan Gene-level RNAseq data, HRD score data, and the Gene-level Copy Number data can be found at https://xenabrowser.net/datapages/?cohort=TCGA %20PanCancer%20(PANCAN)&removeHub=https%3A%2 F%2Fxena.treehouse.gi.ucsc.edu%3A443. The Segmented Copy Number data and Arm-level aneuploidy data can be accessed at https://gdc.cancer.gov/about- 24 data/publications/pancanatlas. The sample meta data for the TCGA PanCan study can be downloaded from https://gdc.cancer.gov/resourcestcga-users/tcga-code-tables https://gdc.cancer.gov/resources-tcga-users/tcgacode-tables. The peri/centromeric satellite sequence annotations from the T2T project can be downloaded from https://genome.ucsc.edu/cgibin/hgTables?db=hub_3671779_hs1&hgta_group= map&hgta_track=hub_3671779_censat&hgta_table =hub_3671779_censat&hgta_doSchema=describe+ table+schema. The scripts are available at https://github.com/DylJHS/Chromosome_Instabilit y. For supplementary data, contact dylanhsimmons@hotmail.fr 9. Acknowledgements Sincere thanks are extended to Aniek Janssen, Aditya Dixit, the Janssen Group, and Stefan Prekovic for their invaluable support and contributions. 10. References 1. 2. 3. 4. 5. 6. Passarge E. Emil Heitz and the concept of heterochromatin: longitudinal chromosome differentiation was recognized fifty years ago. Am J Hum Genet. 1979 Mar;31(2):106-15. PMID: 377956; PMCID: PMC1685768. Lomberk, G., Wallrath, L., & Urrutia, R. (2006). The heterochromatin protein 1 family. Genome biology, 7, 1-8 Hennig, W. (1999). Heterochromatin. Chromosoma, 108(1), 1-9. Janssen, A., Colmenares, S. U., & Karpen, G. H. (2018). Heterochromatin: guardian of the genome. Annual review of cell and developmental biology, 34(1), 265-288. Janssen, A., Colmenares, S. U., & Karpen, G. H. (2018). Heterochromatin: guardian of the genome. Annual review of cell and developmental biology, 34(1), 265-288. Olins, A. L., Rhodes, G., Welch, D. B. M., Zwerger, M., & Olins, D. E. (2010). Lamin B receptor: Multi-tasking at the nuclear envelope. Nucleus, 1(1), 53–70. https://doi.org/10.4161/nucl.1.1.10515 Margalit, A., Brachner, A., Gotzmann, J., Foisner, R., & Gruenbaum, Y. (2007). Barrier-to-autointegration factor--a BAFfling little protein. Trends in cell biology, 17(4), 202–208. https://doi.org/10.1016/j.tcb.2007.02.004 7. Maison, C., Pyrpasopoulou, A., Theodoropoulos, P. A., & Georgatos, S. D. (1997). The inner nuclear membrane protein LAP1 forms a native complex with B-type lamins and partitions with spindleassociated mitotic vesicles. The EMBO journal, 16(16), 4839–4850. https://doi.org/10.1093/emboj/16.16.4839 8. Poleshko, A., Mansfield, K. M., Burlingame, C. C., Andrake, M. D., Shah, N. R., & Katz, R. A. (2013). The human protein PRR14 tethers heterochromatin to the nuclear lamina during interphase and mitotic exit. Cell reports, 5(2), 292–301. https://doi.org/10.1016/j.celrep.2013.09.02 4 9. Zeng W, Ball AR Jr, Yokomori K. HP1: heterochromatin binding proteins working the genome. Epigenetics. 2010 May 16;5(4):287-92. doi: 10.4161/epi.5.4.11683. Epub 2010 May 3. PMID: 20421743; PMCID: PMC3103764. 10. Shibuya, H., Ishiguro, K., & Watanabe, Y. (2014). The TRF1-binding protein TERB1 promotes chromosome movement and telomere rigidity in meiosis. Nature cell biology, 16(2), 145–156. https://doi.org/10.1038/ncb2896 11. Dunce, J.M., Milburn, A.E., Gurusaran, M. et al. Structural basis of meiotic telomere attachment to the nuclear envelope by MAJIN-TERB2-TERB1. Nat Commun 9, 5355 (2018). https://doi.org/10.1038/s41467-01807794-7 12. Li, W., Bai, X., Li, J., Zhao, Y., Liu, J., Zhao, H., Liu, L., Ding, M., Wang, Q., Shi, F. Y., Hou, M., Ji, J., Gao, G., Guo, R., Sun, Y., Liu, Y., & Xu, D. (2019). The nucleoskeleton protein IFFO1 immobilizes broken DNA and suppresses chromosome translocation during tumorigenesis. Nature cell biology, 21(10), 1273–1285. https://doi.org/10.1038/s41556-019-03880 13. Van Steensel, B., & Belmont, A. S. (2017). Lamina-associated domains: links with chromosome architecture, heterochromatin, and gene repression. Cell, 169(5), 780-791. 14. Shevelyov, Y. Y., & Ulianov, S. V. (2019). The nuclear lamina as an organizer of chromosome architecture. Cells, 8(2), 136. 15. Bellanger A, Madsen-Østerbye J, Galigniana NM, Collas P. Restructuring of Lamina-Associated Domains in Senescence and Cancer. Cells. 2022; 11(11):1846. https://doi.org/10.3390/cells11111846 25 16. Smith, E. R., Capo-Chichi, C. D., & Xu, X. X. (2018). Defective Nuclear Lamina in Aneuploidy and Carcinogenesis. Frontiers in oncology, 8, 529. https://doi.org/10.3389/fonc.2018.00529 17. Dialynas GK, Vitalini MW, Wallrath LL. Linking Heterochromatin Protein 1 (HP1) to cancer progression. Mutat Res. 2008 Dec 1;647(1-2):13-20. doi: 10.1016/j.mrfmmm.2008.09.007. Epub 2008 Sep 24. PMID: 18926834; PMCID: PMC2637788. 18. Clermont PL, Lin D, Crea F, Wu R, Xue H, Wang Y, Thu KL, Lam WL, Collins CC, Wang Y, Helgason CD. Polycombmediated silencing in neuroendocrine prostate cancer. Clin Epigenetics. 2015 Apr 3;7(1):40. doi: 10.1186/s13148-0150074-4. PMID: 25859291; PMCID: PMC4391120. 19. Lee SK, Wang W. Roles of Topoisomerases in Heterochromatin, Aging, and Diseases. Genes. 2019; 10(11):884. https://doi.org/10.3390/genes10110884 20. Amoiridis, M., Verigos, J., Meaburn, K. et al. Inhibition of topoisomerase 2 catalytic activity impacts the integrity of heterochromatin and repetitive DNA and leads to interlinks between clustered repeats. Nat Commun 15, 5727 (2024). https://doi.org/10.1038/s41467-02449816-7 21. Anaissi, A., Kennedy, P.J., Goyal, M. et al. A balanced iterative random forest for gene selection from microarray data. BMC Bioinformatics 14, 261 (2013). https://doi.org/10.1186/1471-2105-14-261 22. Anaissi, Ali, et al. "Ensemble feature learning of genomic data using support vector machine." PloS one 11.6 (2016): e0157330. 23. Carter, S. L., Eklund, A. C., Kohane, I. S., Harris, L. N., & Szallasi, Z. (2006). A signature of chromosomal instability inferred from gene expression profiles predicts clinical outcome in multiple human cancers. Nature genetics, 38(9), 1043–1048. https://doi.org/10.1038/ng1861 24. Szklarczyk D, Kirsch R, Koutrouli M, Nastou K, Mehryary F, Hachilif R, Annika GL, Fang T, Doncheva NT, Pyysalo S, Bork P‡, Jensen LJ‡, von Mering C‡. The STRING database in 2023: protein– protein association networks and functional enrichment analyses for any sequenced genome of interest. Nucleic Acids Res. 2023 Jan 6;51(D1):D638-646. 25. Baker, T.M., Waise, S., Tarabichi, M. et al. Aneuploidy and complex genomic rearrangements in cancer evolution. Nat Cancer 5, 228–239 (2024). https://doi.org/10.1038/s43018-02300711-y 26. Adell MAY, Klockner TC, Höfler R, et al. Adaptation to spindle assembly checkpoint inhibition through the selection of specific aneuploidies. Genes Dev. 2023;37(5-6):171-190. doi:10.1101/gad.350182.122 27. Abkevich, V., Timms, K. M., Hennessy, B. T., Potter, J., Carey, M. S., Meyer, L. A., Smith-McCune, K., Broaddus, R., Lu, K. H., Chen, J., Tran, T. V., Williams, D., Iliev, D., Jammulapati, S., FitzGerald, L. M., Krivak, T., DeLoia, J. A., Gutin, A., Mills, G. B., & Lanchbury, J. S. (2012). Patterns of genomic loss of heterozygosity predict homologous recombination repair defects in epithelial ovarian cancer. British journal of cancer, 107(10), 1776–1782. https://doi.org/10.1038/bjc.2012.451 28. Birkbak, N. J., Wang, Z. C., Kim, J. Y., Eklund, A. C., Li, Q., Tian, R., BowmanColin, C., Li, Y., Greene-Colozzi, A., Iglehart, J. D., Tung, N., Ryan, P. D., Garber, J. E., Silver, D. P., Szallasi, Z., & Richardson, A. L. (2012). Telomeric allelic imbalance indicates defective DNA repair and sensitivity to DNA-damaging agents. Cancer discovery, 2(4), 366–375. https://doi.org/10.1158/2159-8290.CD-110206 29. Popova, T., Manié, E., Rieunier, G., CauxMoncoutier, V., Tirapo, C., Dubois, T., Delattre, O., Sigal-Zafrani, B., Bollet, M., Longy, M., Houdayer, C., Sastre-Garau, X., Vincent-Salomon, A., Stoppa-Lyonnet, D., & Stern, M. H. (2012). Ploidy and large-scale genomic instability consistently identify basal-like breast carcinomas with BRCA1/2 inactivation. Cancer research, 72(21), 5454–5462. https://doi.org/10.1158/0008-5472.CAN12-1470 30. Pellegrino B, Musolino A, Llop-Guevara A, et al. Homologous Recombination Repair Deficiency and the Immune Response in Breast Cancer: A Literature Review. Transl Oncol. 2020;13(2):410422. doi:10.1016/j.tranon.2019.10.010 31. Doig KD, Fellowes AP, Fox SB. Homologous Recombination Repair Deficiency: An Overview for Pathologists. Mod Pathol. 2023;36(3):100049. doi:10.1016/j.modpat.2022.100049 32. Ehrlich M. DNA hypomethylation, cancer, the immunodeficiency, centromeric region 26 instability, facial anomalies syndrome and chromosomal rearrangements [published correction appears in J Nutr 2002 Nov;132(11):3432]. J Nutr. 2002;132(8 Suppl):2424S-2429S. doi:10.1093/jn/132.8.2424S 33. Barra, V., Fachinetti, D. The dark side of centromeres: types, causes and consequences of structural abnormalities implicating centromeric DNA. Nat Commun 9, 4340 (2018). https://doi.org/10.1038/s41467-01806545-y 34. Hoyt, S. J., Storer, J. M., Hartley, G. A., Grady, P. G. S., Gershman, A., de Lima, L. G., Limouse, C., Halabian, R., Wojenski, L., Rodriguez, M., Altemose, N., Rhie, A., Core, L. J., Gerton, J. L., Makalowski, W., Olson, D., Rosen, J., Smit, A. F. A., Straight, A. F., Vollger, M. R., … O'Neill, R. J. (2022). From telomere to telomere: The transcriptional and epigenetic state of human repeat elements. Science (New York, N.Y.), 376(6588), eabk3112. https://doi.org/10.1126/science.abk3112 35. Altemose, N., Logsdon, G. A., Bzikadze, A. V., Sidhwani, P., Langley, S. A., Caldas, G. V., Hoyt, S. J., Uralsky, L., Ryabov, F. D., Shew, C. J., Sauria, M. E. G., Borchers, M., Gershman, A., Mikheenko, A., Shepelev, V. A., Dvorkina, T., Kunyavskaya, O., Vollger, M. R., Rhie, A., McCartney, A. M., … Miga, K. H. (2022). Complete genomic and epigenetic maps of human centromeres. Science (New York, N.Y.), 376(6588), eabl4178. https://doi.org/10.1126/science.abl4178 36. Nurk S, Koren S, Rhie A, et al. The complete sequence of a human genome. Science. 2022;376(6588):44-53. doi:10.1126/science.abj6987 37. Chen, T., & Guestrin, C. (2016). XGBoost: A Scalable Tree Boosting System. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 785–794). New York, NY, USA: ACM. https://doi.org/10.1145/2939672.2939785 38. De Rop V, Padeganeh A, Maddox PS. CENP-A: the key player behind centromere identity, propagation, and kinetochore assembly. Chromosomal. 2012 Dec;121(6):527-38. doi: 10.1007/s00412-012-0386-5. Epub 2012 Oct 26. PMID: 23095988; PMCID: PMC3501172. 39. Giunta, S., Hervé, S., White, R. R., Wilhelm, T., Dumont, M., Scelfo, A., Gamba, R., Wong, C. K., Rancati, G., Smogorzewska, A., Funabiki, H., & Fachinetti, D. (2021). CENP-A chromatin prevents replication stress at centromeres to avoid structural aneuploidy. Proceedings of the National Academy of Sciences of the United States of America, 118(10), e2015634118. https://doi.org/10.1073/pnas.2015634118 40. Black EM, Giunta S. Repetitive Fragile Sites: Centromere Satellite DNA As a Source of Genome Instability in Human Diseases. Genes (Basel). 2018 Dec 7;9(12):615. doi: 10.3390/genes9120615. PMID: 30544645; PMCID: PMC6315641. 41. Luisa Robbez-Masson, Christopher H.C. Tie, Helen M. Rowe; Cancer cells, on your histone marks, get SETDB1, silence retrotransposons, and go!. J Cell Biol 6 November 2017; 216 (11): 3429–3431. doi: https://doi.org/10.1083/jcb.201710068 42. Chatel, G., & Fahrenkrog, B. (2011). Nucleoporins: leaving the nuclear pore complex for a successful mitosis. Cellular signalling, 23(10), 1555–1562. https://doi.org/10.1016/j.cellsig.2011.05.0 23 43. Ryu T, Spatola B, Delabaere L, Bowlin K, Hopp H, Kunitake R, Karpen GH, Chiolo I. Heterochromatic breaks move to the nuclear periphery to continue recombinational repair. Nat Cell Biol. 2015 Nov;17(11):1401-11. doi: 10.1038/ncb3258. Epub 2015 Oct 26. PMID: 26502056; PMCID: PMC4628585. 44. Wang ZQ, Wu ZX, Wang ZP, Bao JX, Wu HD, Xu DY, Li HF, Xu YY, Wu RX, Dai XX. Pan-cancer analysis of NUP155 and validation of its role in breast cancer cell proliferation, migration, and apoptosis. BMC Cancer. 2024 Mar 19;24(1):353. doi: 10.1186/s12885-024-12039-6. PMID: 38504158; PMCID: PMC10953186. 45. Nakano H, Wang W, Hashizume C, Funasaka T, Sato H, Wong RW. Unexpected role of nucleoporins in coordination of cell cycle progression. Cell Cycle. 2011;10(3):425-433. doi:10.4161/cc.10.3.14721 46. Sofi S, Mehraj U, Qayoom H, Aisha S, Almilaibary A, Alkhanani M, Mir MA. Targeting cyclin-dependent kinase 1 (CDK1) in cancer: molecular docking and dynamic simulations of potential CDK1 inhibitors. Med Oncol. 2022 Jun 20;39(9):133. doi: 10.1007/s12032-02201748-2. PMID: 35723742; PMCID: PMC9207877. 47. Malumbres, M., & Barbacid, M. (2009). Cell cycle, CDKs and cancer: a changing 27 paradigm. Nature reviews. Cancer, 9(3), 153–166. https://doi.org/10.1038/nrc2602 48. Hodder S, Fox M, Binti Ahmad Mokhtar AM, Mott HR, Owen D. ACKnowledging the role of the Activated-Cdc42 associated kinase (ACK) in regulating protein stability in cancer. Small GTPases. 2023 Dec;14(1):14-25. doi: 10.1080/21541248.2023.2212573. PMID: 37194323; PMCID: PMC10193877. 49. Chafai, N., Bonizzi, L., Botti, S., & Badaoui, B. (2023). Emerging applications of machine learning in genomic medicine and healthcare. Critical Reviews in Clinical Laboratory Sciences, 61(2), 140– 163. https://doi.org/10.1080/10408363.2023.22 59466 50. Wang Y, Wang Q, Huang H, Huang W, Chen Y, McGarvey PB, Wu CH, Arighi CN, UniProt Consortium. A crowdsourcing open platform for literature curation in UniProt Plos Biology. 19(12):e3001464 (2021) 51. Schoch CL, et al. NCBI Taxonomy: a comprehensive update on curation, resources and tools. Database (Oxford). 2020: baaa062. 52. Sayers EW, et al. GenBank. Nucleic Acids Res. 2019. 47(D1):D94-D99. 53. Goldman, M.J., Craft, B., Hastie, M. et al. Visualizing and interpreting cancer genomics data via the Xena platform. Nat Biotechnol (2020). https://doi.org/10.1038/s41587-020-05468 54. Frankish A, Carbonell-Sala S, Diekhans M, Jungreis I, Loveland JE, Mudge JM, Sisu C, Wright JC, Arnan C, Barnes I, Banerjee A, Bennett R, Berry A, Bignell A, Boix C, Calvet F, Cerdán-Vélez D, Cunningham F, Davidson C, Donaldson S, Dursun C, Fatima R, Giorgetti S, Giron CG, Gonzalez JM, Hardy M, Harrison PW, Hourlier T, Hollis Z, Hunt T, James B, Jiang Y, Johnson R, Kay M, Lagarde J, Martin FJ, Gómez LM, Nair S, Ni P, Pozo F, Ramalingam V, Ruffier M, Schmitt BM, Schreiber JM, Steed E, Suner MM, Sumathipala D, Sycheva I, UszczynskaRatajczak B, Wass E, Yang YT, Yates A, Zafrulla Z, Choudhary JS, Gerstein M, Guigo R, Hubbard TJP, Kellis M, Kundaje A, Paten B, Tress ML, Flicek P. Nucleic Acids Res 2023 : 51 ; d1 ; D942-D949. 55. Vivian J, Rao AA, Nothaft FA, Ketchum C, Armstrong J, Novak A, Pfeil J, Narkizian J, Deran AD, MusselmanBrown A, Schmidt H, Amstutz P, Craft B, Goldman M, Rosenbloom K, Cline M, O'Connor B, Hanna M, Birger C, Kent WJ, Patterson DA, Joseph AD, Zhu J, Zaranek S, Getz G, Haussler D, Paten B. Toil enables reproducible, open source, big biomedical data analyses. Nat Biotechnol. 2017 Apr 11;35(4):314-316. doi: 10.1038/nbt.3772. PMID: 28398314; PMCID: PMC5546205. 56. Mermel, C.H., Schumacher, S.E., Hill, B. et al. GISTIC2.0 facilitates sensitive and confident localization of the targets of focal somatic copy-number alteration in human cancers. Genome Biol 12, R41 (2011). https://doi.org/10.1186/gb-201112-4-r41 57. Grossman, Robert L., Heath, Allison P., Ferretti, Vincent, Varmus, Harold E., Lowy, Douglas R., Kibbe, Warren A., Staudt, Louis M. (2016) Toward a Shared Vision for Cancer Genomic Data. New England Journal of Medicine375:12, 1109-1112 58. Blighe K, Rana S, Lewis M (2024). EnhancedVolcano: Publication-ready volcano plots with enhanced colouring and labeling. R package version 1.22.0, https://github.com/kevinblighe/Enhanced Volcano. 59. Robinson, M.D., Oshlack, A. A scaling normalization method for differential expression analysis of RNA-seq data. Genome Biol 11, R25 (2010). https://doi.org/10.1186/gb-2010-11-3-r25 60. Chen Y, Chen L, Lun ATL, Baldoni P, Smyth GK (2024). “edgeR 4.0: powerful differential analysis of sequencing data with expanded functionality and improved support for small counts and larger datasets.” bioRxiv. doi:10.1101/2024.01.21.576131. 61. Chen Y, Lun ATL, Smyth GK (2016). “From reads to genes to pathways: differential expression analysis of RNASeq experiments using Rsubread and the edgeR quasi-likelihood pipeline.” F1000Research, 5, 1438. doi:10.12688/f1000research.8987.2. 62. McCarthy DJ, Chen Y, Smyth GK (2012). “Differential expression analysis of multifactor RNA-Seq experiments with respect to biological variation.” Nucleic Acids Research, 40(10), 4288-4297. doi:10.1093/nar/gks042. 63. Robinson MD, McCarthy DJ, Smyth GK (2010). “edgeR: a Bioconductor package for differential expression analysis of digital gene expression data.” Bioinformatics, 26(1), 139-140. doi:10.1093/bioinformatics/btp616. 28 64. Lawrence M, Gentleman R, Carey V (2009). “rtracklayer: an R package for interfacing with genome browsers.” Bioinformatics, 25, 1841-1842. doi:10.1093/bioinformatics/btp328, http://bioinformatics.oxfordjournals.org/co ntent/25/14/1841.abstract. 65. Nassar, L. R., Barber, G. P., Benet-Pagès, A., Casper, J., Clawson, H., Diekhans, M., Fischer, C., Gonzalez, J. N., Hinrichs, A. S., Lee, B. T., Lee, C. M., Muthuraman, P., Nguy, B., Pereira, T., Nejad, P., Perez, G., Raney, B. J., Schmelter, D., Speir, M. L., Wick, B. D., … Kent, W. J. (2023). The UCSC Genome Browser database: 2023 update. Nucleic acids research, 51(D1), D1188–D1195. https://doi.org/10.1093/nar/gkac1072 66. Nassar, L. R., Barber, G. P., Benet-Pagès, A., Casper, J., Clawson, H., Diekhans, M., Fischer, C., Gonzalez, J. N., Hinrichs, A. S., Lee, B. T., Lee, C. M., Muthuraman, P., Nguy, B., Pereira, T., Nejad, P., Perez, G., Raney, B. J., Schmelter, D., Speir, M. L., Wick, B. D., … Kent, W. J. (2023). The UCSC Genome Browser database: 2023 update. Nucleic acids research, 51(D1), D1188–D1195. https://doi.org/10.1093/nar/gkac1072 67. Wickham H (2016). ggplot2: Elegant Graphics for Data Analysis. SpringerVerlag New York. ISBN 978-3-31924277-4, https://ggplot2.tidyverse.org. 68. Hothorn T, Hornik K, van de Wiel MA, Zeileis A (2006). “A Lego system for conditional inference.” The American Statistician, 60(3), 257–263. doi:10.1198/000313006X118430. 69. R Core Team (2013). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0 29 Supplemental Information 30 CHD5 TMEM201 MAD2L2 AGTRAP CLCN6 CDC42 KDM1A AHDC1 SPOCD1 HDAC1 RBBP4 AK2 CDCA8 ZMPSTE24 GUCA2B YBX1 CC2D1B ZNF644 LRIF1 RHOC SIKE1 SETDB1 POGZ LMNA PMF1 POU2F1 TOR1AIP1 RNF2 KDM5B PPP2R5A NSL1 LBR PARP1 NUP133 HNRNPU AHCTF1 E2F6 DNMT3A CENPA SPDYA SPAST PCBP1 PCGF1 CHMP3 BCL2L11 RND3 CACNB4 HECW2 CANX TXN HMGA2 RPA1 BARD1 CDYL SPTAN1 TMPO C1QBP XRCC5 NUP153 SET SLC25A3 MIS12 OBSL1 POU5F1 TOR1A TCHP ACAP1 ACSL3 BAG6 BRD3 TBX5 POLR2A SP100 EHMT2 WDR5 HNF1A CHD3 PDE6D NELFE EHMT1 RHOF AURKB ATG7 BRD2 PFKP SMAD9 NDEL1 TMEM43 RING1 SUV39H2 RB1 MPRIP ZNF621 DAXX VIM KPNA3 SREBF1 CTNNB1 LEMD2 COMMD3.BMI1 LMO7 SUZ12 TRAK1 HMGA1 BMI1 UBAC2 PCGF2 RHOA CUL7 ZNF239 CUL4A CDK12 BAP1 MLIP ZWINT TFDP1 TOP2A RYBP ORC3 CDK1 CHAMP1 STAT3 CHMP2B HDAC2 SIRT1 PARP2 RND2 SENP7 BCLAF1 DNAJB12 DAD1 BRCA1 RPN1 SYNE1 DUSP13 PRMT5 GFAP TFDP2 QKI PCGF6 CHMP4A KPNB1 KPNA4 SUN1 NUP98 TINF2 CBX1 FXR1 GET4 RHOG RHOJ SRSF1 SOX2 ACTB RRP8 SYNE2 KPNA2 DNAJB11 RAC1 TRIM66 MAX SUMO2 TFRC RPA3 MYOD1 FOS CBX8 YTHDC1 CBX3 PAX6 BCL11B EIF4A3 PAPSS1 HOXA5 DDB2 YY1 CHMP6 EGF EGFR DDB1 PPP2R5C RAC3 PGRMC2 POM121 INCENP RCOR1 SMCHD1 SMARCA5 AKAP9 STX5 BAHD1 PIAS2 SMAD1 SND1 TM7SF2 MGA SMAD2 NPR3 NUP205 BANF1 TP53BP1 MBD1 NIPBL TRIM24 RHOD WDR76 ADNP2 NUP155 KIAA1549 GSTP1 ZNF280D LMNB2 AGGF1 ZC3HAV1 NUMA1 SMAD3 SGTA AP3B1 KDM7A EED PML CHAF1A XRCC4 BRAF IFFO1 USP7 HNRNPM PRDM6 AGK CHD4 PRKCB DNMT1 LMNB1 FAM131B PHC1 MAPK3 ILF3 PPP2CA EZH2 GABARAPL1 ZNF764 SMARCA4 CAMLG CHMP7 GUCY2C PRR14 BRD4 SMAD5 PPP2CB LRRK2 CTCF AKAP8L KIF20A WRN YAF2 CDH1 WIZ MATR3 TERF1 SENP1 VPS4A CEBPA HDAC3 CHMP4C RND1 SF3B3 ZNF382 FAM114A2 ESRP1 CBX5 IST1 SIRT2 NPM1 SMARCA2 IL23A GABARAPL2 ERCC2 FAF2 PSIP1 BAZ2A WWOX CRX Table S1. Heterochromatin Interactome PPP2R1A ZNF579 TRIM28 CHMP2A DDRGK1 RRBP1 BANF2 ASXL1 DNMT3B MAPRE1 CHMP4B DSN1 L3MBTL1 ZMYND8 ADNP VAPB CTSZ TPTE CHAF1B RRP1B MCM3AP MAPK1 SMARCB1 RAC2 SUN2 CBX7 L3MBTL2 XRCC6 NUP50 RBBP7 SUV39H1 KDM5C ZMYM3 ATRX CUL4B MMGT1 FATE1 MECP2 EMD UBL4A TTN ZNF462 LEMD3 ZC3H18 CCDC155 Genes output by the STRNG query composing the heterochromatin interactome or Set-of-Interest (SOI) Background Gene Gene of Interest Breast invasive carcinoma 40 30 HOXA5 CDK1 CDCA8 KIF20A LMNB1 EZH2 AURKB ZWINT CBX7 CLCN6 EGFR CBX3 HNRNPU −2.5 0.0 PARP1 KPNA2 RHOJ MPRIP KDM5B SYNE1 RND3 SRSF1 DSN1 WDR76 GABARAPL1 FOS RPN1 BARD1 TMEM43 SMAD9 TFDP2 NDEL1 ZNF280D QKI SPOCD1 SUN2 SMARCA2 CEBPA VIM RHOF ESRP1 SYNE2 MECP2 RND1 NPR3 ZNF462 SUN1 MLIP EGF CDH1 GFAP 20 10 0 TOP2A CENPA 2.5 5.0 Head and Neck squamous cell carcinoma 40 30 20 CACNB4 TM7SF2 LRRK2 CBX7 − Log10 P 10 MAPK3 DUSP13 FOS RND1 LMO7 GUCY2C MLIP EGF TTN 0 CHMP2B −2.5 DNMT3B CBX3 0.0 SPOCD1 HMGA2 2.5 5.0 Lung adenocarcinoma 40 30 CBX7 ZWINT CDCA8 RHOJ QKI PAPSS1 HOXA5 LMO7 SMAD9 CACNB4 NPR3 SMARCA5 FOS TTN 20 EZH2 AURKB CDK1 IL23A EGF LRRK2 10 RND1 0 KIF20A TOP2A CENPA HMGA2 SOX2 −2.5 0.0 2.5 5.0 Lung squamous cell carcinoma 40 30 20 10 0 LRRK2 LMO7 SYNE1 CENPA CDCA8 EZH2 CHAF1A CDK1 TOP2A LMNB2 NUP155 BRCA1 KPNA2 ZWINT KIF20A AURKB RAC3 LMNB1 SOX2 DNMT3B PFKP IL23A RHOJ HECW2 CBX7 TBX5 MPRIP RND1 FOS PRKCB CACNB4 CCDC155 CHD5 −2.5 RHOA MAPK3 RHOF RND2 NPR3 0.0 Log2 fold change HMGA2 MLIP 2.5 5.0 Supplemental Figure 1. Differential Gene Expression in the set-of-interest across different tissue types. Differential gene expression in cancerous versus normal condition. Highlighted are genes in the set-of-interest (red) compared to background genes from the TCGA dataset (blue) for all paired samples within specific tissue types. All p-value cutoffs are set to 0.01. P-values (y-axis) and fold changes (x-axis) are produced with edgeR package. Plots produced with EnhancedVolcano R package. Base Layer Meta Layer CIN Feature Accuracy Baseline Accuracy p-value Accuracy Baseline Accuracy p-value 1p 1q 2p 2q 3p 3q 4p 4q 5p 5q 6p 6q 7p 7q 8p 0.847 0.843 0.819 0.793 0.836 0.817 0.738 0.815 0.801 0.862 0.825 0.811 0.830 0.824 0.802 0.795 0.697 0.833 0.869 0.702 0.747 0.731 0.755 0.708 0.722 0.800 0.755 0.699 0.739 0.592 9.10E-28 4.36E-166 9.99E-01 1.00E+00 5.53E-141 8.63E-43 1.08E-01 1.40E-31 8.64E-68 1.63E-162 2.28E-07 3.05E-28 4.16E-134 2.10E-60 5.13E-290 0.833 0.846 0.808 0.789 0.828 0.807 0.736 0.810 0.801 0.856 0.817 0.794 0.825 0.823 0.802 0.794 0.703 0.832 0.866 0.703 0.747 0.730 0.756 0.708 0.728 0.799 0.753 0.700 0.742 0.596 1.63E-21 5.09E-214 1.00E+00 1.00E+00 1.55E-161 3.19E-40 8.76E-02 3.28E-34 8.49E-89 2.61E-184 8.32E-06 7.07E-21 2.98E-160 7.00E-72 0.00E+00 0.840 0.844 0.813 0.791 0.832 0.812 0.737 0.813 0.801 0.859 0.821 0.803 0.828 0.823 0.802 0.794 0.700 0.833 0.867 0.703 0.747 0.730 0.756 0.708 0.725 0.799 0.754 0.699 0.741 0.594 8q 9p 9q 10p 10q 11p 11q 12p 12q 13q 14q 15q 16p 16q 17p 17q 18p 18q 19p 19q 20p 20q 21q 22q 0.772 0.752 0.832 0.800 0.835 0.782 0.806 0.746 0.800 0.837 0.826 0.784 0.805 0.824 0.842 0.813 0.727 0.804 0.824 0.784 0.792 0.836 0.779 0.833 0.669 0.689 0.731 0.746 0.773 0.781 0.787 0.754 0.820 0.701 0.767 0.785 0.771 0.686 0.607 0.796 0.705 0.691 0.809 0.796 0.715 0.714 0.736 0.719 3.08E-76 1.32E-29 4.13E-83 2.73E-25 6.26E-36 4.24E-01 8.96E-05 9.52E-01 1.00E+00 1.60E-144 5.45E-32 5.90E-01 4.13E-12 1.01E-143 0.00E+00 1.66E-04 3.13E-05 1.32E-96 6.10E-04 9.92E-01 4.25E-46 1.90E-118 3.25E-16 3.51E-105 0.776 0.754 0.825 0.782 0.833 0.770 0.792 0.745 0.788 0.833 0.823 0.781 0.786 0.828 0.847 0.803 0.736 0.800 0.814 0.786 0.776 0.835 0.771 0.824 0.672 0.688 0.728 0.751 0.774 0.784 0.786 0.759 0.821 0.700 0.766 0.790 0.771 0.686 0.613 0.797 0.706 0.692 0.807 0.793 0.720 0.716 0.739 0.714 2.13E-102 1.45E-42 4.08E-103 1.51E-12 2.57E-42 9.99E-01 9.04E-02 9.99E-01 1.00E+00 3.36E-182 3.66E-40 9.80E-01 2.99E-04 2.17E-203 0.00E+00 5.74E-02 2.49E-10 3.03E-116 3.40E-02 9.40E-01 6.87E-33 4.82E-150 1.64E-12 8.06E-126 0.774 0.753 0.828 0.791 0.834 0.776 0.799 0.745 0.794 0.835 0.825 0.783 0.796 0.826 0.845 0.808 0.732 0.802 0.819 0.785 0.784 0.835 0.775 0.828 0.670 0.688 0.730 0.748 0.774 0.782 0.787 0.756 0.820 0.700 0.767 0.788 0.771 0.686 0.610 0.796 0.705 0.691 0.808 0.794 0.718 0.715 0.737 0.716 Mean 0.808 0.741 0.802 0.742 0.805 0.741 Table S2. Chromosome instability model classification performance Performance of the classification XGBoost learners in the base layer of the general model. Mean Mean Baseline Accuracy Accuracy A Accuracy of the Aneuploidy Learners Cancer type: BRCA 1.0 0.9 Accuracy 0.8 Learner Type Base Meta 0.7 0.6 12 p 12 q 13 q 14 q 15 q 16 p 16 q 17 p 17 q 18 p 18 q 19 p 19 q 20 p 20 q 21 q 22 q p q 11 10 11 p q 9q 10 9p 8q 8p 7q 7p 6q 6p 5q 5p 4q 4p 3q 3p 2q 2p 1q 1p 0.5 Learner B Accuracy of the Aneuploidy Learners Cancer type: LUSC 1.0 0.9 Accuracy 0.8 Learner Type 0.7 Base Meta 0.6 0.5 q 16 p 16 q 17 p 17 q 18 p 18 q 19 p 19 q 20 p 20 q 21 q 22 q q 15 q 14 13 q p 12 12 q p 11 11 q 10 p 10 9q 9p 8q 8p 7q 7p 6q 6p 5q 5p 4q 4p 3q 3p 2q 2p 1q 1p 0.4 Learner C Accuracy of the Aneuploidy Learners Cancer type: LUAD 1.0 0.9 Accuracy 0.8 Learner Type 0.7 Base Meta 0.6 0.5 0.4 10 p 10 q 11 p 11 q 12 p 12 q 13 q 14 q 15 q 16 p 16 q 17 p 17 q 18 p 18 q 19 p 19 q 20 p 20 q 21 q 22 q 9q 9p 8q 8p 7q 7p 6q 6p 5q 5p 4q 4p 3q 3p 2q 2p 1q 1p 0.3 Learner D Accuracy of the Aneuploidy Learners Cancer type: HNSC 1.0 0.9 Accuracy 0.8 Learner Type 0.7 Base Meta 0.6 0.5 22 q 21 q 20 q 20 p 19 q 19 p 18 q 18 p 17 q 17 p 16 q 16 p 15 q 14 q 13 q 12 q 12 p 11 q 11 p 10 q 10 p 9q 9p 8q 8p 7q 7p 6q 6p 5q 5p 4q 4p 3q 3p 2q 2p 1q 1p 0.4 Learner Supplemental Figure 2. Performance of the classification of the Cancer-speicific Chromosome Instability models Classification performance of the individual XGBoost learners determined using the overall accuracy (y-axis) for the 39 classification learners (x-axis), assessed for both the base layer (blue) and meta layer (red) for Breast invasive carcinoma (A), Lung squamous cell carcinoma (B), Lung adenocarcinoma (C) and Head and Neck squamous cell carcinoma (D). Base Layer Meta Layer 2 2 CIN Feature RMSE RMSE R R ai1 0.710 3.923 0.740 3.678 lst1 0.570 4.569 0.600 4.384 loh_hrd 0.590 3.218 0.620 3.050 peri_1 0.240 0.145 0.260 0.143 peri_2 0.400 0.104 0.430 0.100 peri_3 0.510 0.162 0.510 0.159 peri_4 0.220 0.155 0.260 0.149 peri_5 0.280 0.149 0.310 0.142 peri_7 0.460 0.164 0.490 0.156 peri_8 0.210 0.207 0.240 0.201 peri_9 0.400 0.173 0.410 0.168 peri_10 0.440 0.152 0.470 0.147 peri_16_1 0.420 0.130 0.460 0.126 peri_16_2 0.470 0.171 0.500 0.164 peri_17 0.360 0.168 0.390 0.161 peri_20 0.750 0.117 0.780 0.107 peri_22_3 0.540 0.164 0.550 0.162 Mean 0.445 0.816 0.472 0.776 Table S3. Chromosome instability model regression performance 2 Mean R 0.725 0.585 0.605 0.250 0.415 0.510 0.240 0.295 0.475 0.225 0.405 0.455 0.440 0.485 0.375 0.765 0.545 0.459 Performance of the regression XGBoost learners in the base layer of the general model in terms of the coefficient of determination and root mean squared error. Mean RMSE 3.801 4.476 3.134 0.144 0.102 0.161 0.152 0.145 0.160 0.204 0.171 0.149 0.128 0.167 0.165 0.112 0.163 0.796 A Pericentromeric and HRD models in terms of R2 Cancer type: BRCA 1.0 0.9 0.8 0.7 Learner Type 0.6 R2 Base Meta 0.5 0.4 0.3 0.2 0.1 _3 ri_ 22 20 pe pe pe pe ri_ 17 16 ri_ ri_ _2 _1 16 ri_ pe 10 pe ri_ 8 ri_ pe lo h pe ri_ 1 rd _h ls t1 ai 1 0.0 Learner B Pericentromeric and HRD models in terms of R2 Cancer type: LUSC 1.0 0.9 0.8 0.7 Learner Type 0.6 R2 Base Meta 0.5 0.4 0.3 0.2 0.1 _3 20 ri_ 22 ri_ pe pe 17 pe ri_ 10 ri_ pe pe ri_ 9 8 ri_ pe 7 ri_ pe 5 ri_ pe 4 ri_ ri_ pe lo pe 3 d hr h_ t1 ls ai 1 0.0 Learner C Pericentromeric and HRD models in terms of R2 Cancer type: LUAD 1.0 0.9 0.8 0.7 Learner Type 0.6 R2 Base Meta 0.5 0.4 0.3 0.2 0.1 8 i_ pe r 7 i_ pe r 5 r i_ pe 1 i_ pe r rd lo h_ h t1 ls ai 1 0.0 Learner D Pericentromeric and HRD models in terms of R2 Cancer type: HNSC 1.0 0.9 0.8 0.7 Learner Type 0.6 R2 Base Meta 0.5 0.4 0.3 0.2 0.1 20 pe ri_ 8 ri_ pe 7 ri_ pe 3 r i_ pe hr d h_ lo ls t1 ai 1 0.0 Learner Supplemental Figure 3. Performance of the regression of the Cancer-speicific Chromosome Instability models Performance of the individual XGBoost learners determined using the coefficient of determination (R2; y-axis) for the regression learners (x-axis), assessed for both the base layer (blue) and meta layer (red) for Breast invasive carcinoma (A), Lung squamous cell carcinoma (B), Lung adenocarcinoma (C) and Head and Neck squamous cell carcinoma (D). Cancer type: BRCA Cancer type: LUSC Gain 0.0 0.1 0.2 Gain 0.3 0.0 0.1 0.2 0.3 0.4 peri_22_3 peri_20 peri_17 peri_10 peri_9 peri_8 peri_7 peri_5 peri_4 peri_3 loh_hrd lst1 ai1 22q 21q 20q 20p 19q 19p 18q 18p 17q 17p 16q 16p 15q 14q 13q 12q 12p 11q 11p 10q 10p 9q 9p 8q 8p 7q 7p 6q 6p 5q 5p 4q 4p 3q 3p 2q 2p 1q 1p B C AG 1Q 6 C BP B C X C C D 3 O M C ENK1 M H P D MA 3. P BM 7 D CU I1 D DR L4B N G AJ K B1 E2 12 F E 6 EGED G AB E F ARSR R AP H PL1 D H AC1 D M AC2 M AP 3 M AP K1 C R N M3 E1 U A N P1 P U 3 P 3 PPPC 15 P2 GF5 R 6 PS 5C R IP1 P SL R N1 C RP SM25A8 SM A 3 C D2 H SND1 SOD1 S X SUP1 2 0 TM 0 TM ERO2 E F TOM41 TR R 3 IM1A 2 U TX 8 BA N U C2 S V P VPAP 7 S4B XR W A C IZ YT YBC6 ZN H X1 F DC ZN28 1 F60D 44 B C AG 1 C QB 6 C AM P C LG 2D C 1B C DK E C NP1 H A M D DA P7 D D RGD1 N K AJ 1 B1 2 EI EE F D ES 4A 3 H RP L3 DA 1 M M BTC2 AP L R 2 M E1 B N D N IP 1 U B P L PC 13 G 3 PH F1 R C1 R AC R 1 P R 1B R SGP8 SI TA K S SI E1 SMLC RT 2 2 SMAR 5A 3 A C SM RCA2 C A5 H SU SND1 V3 D 1 SU9H 2 TE Z1 2 TF RF 1 TODP 2 U R1 BA A C U 2 SP V 7 VPAP B X S ZN RC 4A F C ZN280 5 ZNF6 D F721 64 Base Learner Base Learner peri_22_3 peri_20 peri_17 peri_16_2 peri_16_1 peri_10 peri_8 peri_1 loh_hrd lst1 ai1 22q 21q 20q 20p 19q 19p 18q 18p 17q 17p 16q 16p 15q 14q 13q 12q 12p 11q 11p 10q 10p 9q 9p 8q 8p 7q 7p 6q 6p 5q 5p 4q 4p 3q 3p 2q 2p 1q 1p Gene Gene Uniform Gain Threshold: 0.003 Cancer type: LUAD Uniform Gain Threshold: 0.003 Cancer type: HNSC 0.0 0.1 Gain 0.2 0.0 peri_8 peri_7 peri_5 peri_1 loh_hrd lst1 ai1 22q 21q 20q 20p 19q 19p 18q 18p 17q 17p 16q 16p 15q 14q 13q 12q 12p 11q 11p 10q 10p 9q 9p 8q 8p 7q 7p 6q 6p 5q 5p 4q 4p 3q 3p 2q 2p 1q 1p 0.1 0.2 0.3 0.4 AH C AU TF 1 C RK 1Q B C B AM P C LG C BX D 1 C C4 D 2 C K C EN 12 H P M A C P4 H B CM D ULP7 D D RG4A N K D AJB 1 N 1 M 2 G AB E T3 AR GF B AP R G L1 H E M DA T4 C C M 3 3 N AP IP O BL PCRC G 3 F POPIA 1 M S2 PS121 R IP1 H R OA IN R G1 P S RRN1 SMLC2 P8 A SM R 5A3 A C SM RCA5 C B1 H SU S D1 V3TX5 TE 9H R 2 T F TR INF1 IM 2 U 28 VPSP S 7 W 4A D R 5 XR WI ZN C Z F2 C5 ZN 80 F7 D 64 Base Learner Base Learner peri_20 peri_8 peri_7 peri_3 loh_hrd lst1 ai1 22q 21q 20q 20p 19q 19p 18q 18p 17q 17p 16q 16p 15q 14q 13q 12q 12p 11q 11p 10q 10p 9q 9p 8q 8p 7q 7p 6q 6p 5q 5p 4q 4p 3q 3p 2q 2p 1q 1p Gene Uniform Gain Threshold: 0.003 AK A A P9 C TG 1Q 7 C BP C B C DCX1 C H 4 O M M C P 2 M H 4B D M 3 D .BMP7 D D RG I1 N D AJ K1 N B AJ 11 B1 E 2 EGED G FR H ET D 4 H A M DAC2 C C M 3 3 N AP P IP PG CBBL R P1 M PHC2 PMC1 PP P F1 P2 M L PRR5C R R14 H R OG IN R G1 P R N1 R SL SIRP8 C T2 SM25A A 3 SMSM D2 A SMAR D3 ARCA SM C 2 C A4 TEHD 1 TF RF D 1 TF P2 U R BA C V C2 VPAP SB W 4A XRDR 5 C XR C C 5 C 6 Gain Gene Uniform Gain Threshold: 0.003 Supplementary Figure 4. Genomic Importance of Cancer-Specific Models. The genomic contribution towards the individual base-learners (y-axis) in terms of the Gain metri (gradient) for the Breast invasive carcinoma (A), Lung squamous cell carcinoma (B), Lung adenocarcinoma (C) and Head and Neck squamous cell carcinoma (D) models and displaying the Uniform Gain Threshold value (see methods). Genes represented on the x-axis. Cancer type: BRCA Cancer type: LUSC Gain Gain 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 peri_22_3 peri_20 peri_17 peri_10 peri_9 peri_8 peri_7 peri_5 peri_4 peri_3 loh_hrd lst1 ai1 22q 21q 20q 20p 19q 19p 18q 18p 17q 17p 16q 16p 15q 14q 13q 12q 12p 11q 11p 10q 10p 9q 9p 8q 8p 7q 7p 6q 6p 5q 5p 4q 4p 3q 3p 2q 2p 1q 1p 1p 1q 2p 2q 3p 3q 4p 4q 5p 5q 6p 6q 7p 7q 8p 8q 9p 9 10q 10p 11q 11p 12q 12p 13q 14q 15q 16q 16p 17q 17p 18q 18p 19q 19p 20q 20p 21q 22q q a lo ls i1 h_ t1 pe hr d peri_ 3 peri_ 4 peri_ 5 peri_ 7 peri_ pe ri 8 _ peri_ 9 r 1 peperi_10 ri_ i_ 7 2220 _3 Meta Learner 1p 1q 2p 2q 3p 3q 4p 4q 5p 5q 6p 6q 7p 7q 8p 8q 9p 9 10q 10p 11q 11p 12q 12p 13q 14q 15q 16q 16p 17q 17p 18q 18p 19q 19p 20q 20p 21q 22q q a lo ls i1 h_ t1 pe hrd peri_1 peper ri_ 8 peri_1i_1 ri_ 6_0 pe16_1 r 2 peperi_1 ri_ i_27 22 0 _3 Meta Learner peri_22_3 peri_20 peri_17 peri_16_2 peri_16_1 peri_10 peri_8 peri_1 loh_hrd lst1 ai1 22q 21q 20q 20p 19q 19p 18q 18p 17q 17p 16q 16p 15q 14q 13q 12q 12p 11q 11p 10q 10p 9q 9p 8q 8p 7q 7p 6q 6p 5q 5p 4q 4p 3q 3p 2q 2p 1q 1p Base Learner Base Learner Uniform Gain Threshold: 0.02 Cancer type: LUAD Uniform Gain Threshold: 0.019 Cancer type: HNSC Gain 0.0 0.2 0.4 Gain 0.6 0.00 0.25 0.50 0.75 peri_8 peri_7 peri_5 peri_1 loh_hrd lst1 ai1 22q 21q 20q 20p 19q 19p 18q 18p 17q 17p 16q 16p 15q 14q 13q 12q 12p 11q 11p 10q 10p 9q 9p 8q 8p 7q 7p 6q 6p 5q 5p 4q 4p 3q 3p 2q 2p 1q 1p 1p 1q 2p 2q 3p 3q 4p 4q 5p 5q 6p 6q 7p 7q 8p 8q 9p 9 10q 10p 11q 11p 12q 12p 13q 14q 15q 16q 16p 17q 17p 18q 18p 19q 19p 20q 20p 21q 22q q ai 1 lo ls h_ t1 h pe rd peri_1 peri_5 peri_7 ri_ 8 1p 1q 2p 2q 3p 3q 4p 4q 5p 5q 6p 6q 7p 7q 8p 8q 9p 9 10q 10p 11q 11p 12q 12p 13q 14q 15q 16q 16p 17q 17p 18q 18p 19q 19p 20q 20p 21q 22q q ai 1 lo ls h_ t1 pe hrd peri_3 p ri_7 peeri_ ri_ 8 20 Meta Learner Meta Learner peri_20 peri_8 peri_7 peri_3 loh_hrd lst1 ai1 22q 21q 20q 20p 19q 19p 18q 18p 17q 17p 16q 16p 15q 14q 13q 12q 12p 11q 11p 10q 10p 9q 9p 8q 8p 7q 7p 6q 6p 5q 5p 4q 4p 3q 3p 2q 2p 1q 1p Base Learner Base Learner Uniform Gain Threshold: 0.022 Uniform Gain Threshold: 0.022 Supplementary Figure 5. The CIN feature impacts of Cancer-Specific Models. The base-learner (x-axis) contributions towards the individual meta-learners (y-axis) in terms of the Gain metric (gradient) for the Breast invasive carcinoma (A), Lung squamous cell carcinoma (B), Lung adenocarcinoma (C) and Head and Neck squamous cell carcinoma (D) models and displaying the corresponding Uniform Gain Threshold value (see methods).