Randomness in Cancer Breakpoint Prediction.

Cancer genomes are susceptible to multiple rearrangements by deleting, inserting, and translocating genomic regions. Recently, the problem of finding determinants of breakpoint formations was approached with machine learning methods; however, unlike cancer point mutations, breakpoint prediction appeared to be a more difficult task, and various machine learning models did not achieve high prediction power often slightly exceeding the threshold of random guessing. This raised the question of whether the breakpoints are random noise in cancer mutagenesis or there exist determinants in structural mutagenesis. In the present study, we investigated randomness in cancer breakpoint genome distributions through the power of machine learning models to predict breakpoint hot spots. We divided all cancer types into three groups by degree of randomness in their breakpoint formation. We tested different density thresholds and explored the bias in hot spot definition. We also compared prediction of hot spots versus individual breakpoints. We found that hot spots are considerably better predicted than individual breakpoints; however, some individual breakpoints can also be predicted with a satisfactory power, and thus, it is not proper to filter them from analyses. We demonstrated that positive-unlabeled learning can provide insights into insufficiency of cancer data sets, which are not always reflected by data set sizes. Overall, the present results support the view that cancer breakpoint landscape can be represented by predictable dense breakpoint regions and scattered individual breakpoints, which are not all random noise, but some are generated by detectable mechanism.

Comprehensive analysis of cancer breakpoints reveals signatures of genetic and epigenetic contribution to cancer genome rearrangements.

Understanding mechanisms of cancer breakpoint mutagenesis is a difficult task and predictive models of cancer breakpoint formation have to this time failed to achieve even moderate predictive power. Here we take advantage of a machine learning approach that can gather important features from big data and quantify contribution of different factors. We performed comprehensive analysis of almost 630,000 cancer breakpoints and quantified the contribution of genomic and epigenomic features-non-B DNA structures, chromatin organization, transcription factor binding sites and epigenetic markers. The results showed that transcription and formation of non-B DNA structures are two major processes responsible for cancer genome fragility. Epigenetic factors, such as chromatin organization in TADs, open/closed regions, DNA methylation, histone marks are less informative but do make their contribution. As a general trend, individual features inside the groups show a relatively high contribution of G-quadruplexes and repeats and CTCF, GABPA, RXRA, SP1, MAX and NR2F2 transcription factors. Overall, the cancer breakpoint landscape can be represented by well-predicted hotspots and poorly predicted individual breakpoints scattered across genomes. We demonstrated that hotspot mutagenesis has genomic and epigenomic factors, and not all individual cancer breakpoints are just random noise but have a definite mutation signature. Besides we found a long-range action of some features on breakpoint mutagenesis. Combining omics data, cancer-specific individual feature importance and adding the distant to local features, predictive models for cancer breakpoint formation achieved 70-90% ROC AUC for different cancer types; however precision remained low at 2% and the recall did not exceed 50%. On the one hand, the power of models strongly correlates with the size of available cancer breakpoint and epigenomic data, and on the other hand finding strong determinants of cancer breakpoint formation still remains a challenge. The strength of predictive signals of each group and of each feature inside a group can be converted into cancer-specific breakpoint mutation signatures. Overall our results add to the understanding of cancer genome rearrangement processes.

Tissue-specific impact of stem-loops and quadruplexes on cancer breakpoints formation.

BACKGROUND: Chromosomal rearrangements are the typical phenomena in cancer genomes causing gene disruptions and fusions, corruption of regulatory elements, damage to chromosome integrity. Among the factors contributing to genomic instability are non-B DNA structures with stem-loops and quadruplexes being the most prevalent. We aimed at investigating the impact of specifically these two classes of non-B DNA structures on cancer breakpoint hotspots using machine learning approach. METHODS: We developed procedure for machine learning model building and evaluation as the considered data are extremely imbalanced and it was required to get a reliable estimate of the prediction power. We built logistic regression models predicting cancer breakpoint hotspots based on the densities of stem-loops and quadruplexes, jointly and separately. We also tested Random Forest models varying different resampling schemes (leave-one-out cross validation, train-test split, 3-fold cross-validation) and class balancing techniques (oversampling, stratification, synthetic minority oversampling). RESULTS: We performed analysis of 487,425 breakpoints from 2234 samples covering 10 cancer types available from the International Cancer Genome Consortium. We showed that distribution of breakpoint hotspots in different types of cancer are not correlated, confirming the heterogeneous nature of cancer. It appeared that stem-loop-based model best explains the blood, brain, liver, and prostate cancer breakpoint hotspot profiles while quadruplex-based model has higher performance for the bone, breast, ovary, pancreatic, and skin cancer. For the overall cancer profile and uterus cancer the joint model shows the highest performance. For particular datasets the constructed models reach high predictive power using just one predictor, and in the majority of the cases, the model built on both predictors does not increase the model performance. CONCLUSION: Despite the heterogeneity in breakpoint hotspots' distribution across different cancer types, our results demonstrate an association between cancer breakpoint hotspots and stem-loops and quadruplexes. Approximately for half of the cancer types stem-loops are the most influential factors while for the others these are quadruplexes. This fact reflects the differences in regulatory potential of stem-loops and quadruplexes at the tissue-specific level, which yet to be discovered at the genome-wide scale. The performed analysis demonstrates that influence of stem-loops and quadruplexes on breakpoint hotspots formation is tissue-specific.