Application of Machine Learning to Development of Copy Number Variation-based Prediction of Cancer Risk

Triphasic Development of the Genetic Code

The genetic code contains an alphabet of genetically encoded amino acids. The ten Phase 1 amino acids, including Gly, Ala, Ser, Asp, Glu, Val, Leu, Ile, Pro and Thr, were available from the prebiotic environment, whereas the ten Phase 2 amino acids, including Phe, Tyr, Arg, His, Trp, Asn, Gln, Lys, Cys, and Met, became available only later from amino acid biosyntheses. In the archaeon Methanopyrus kandleri, the oldest organism known, the standard alphabet of 20 amino acids was "frozen" and no additional amino acid was encoded in the subsequent 3 Gyrs. Four decades ago, it was discovered that the code was frozen because all the organisms were so well adapted to the standard amino acids that oligogenic barriers, consisting of genes that are thoroughly dependent on the standard code, would cause loss of viability upon the deletion of any one amino acid from the code. Once the reason for the freezing of the code was ascertained, procedures were devised by scientists worldwide to enable the encoding of novel noncanonical amino acids (ncAAs). These encoded Phase 3 ncAAs now surpass the 20 canonical Phase 2 amino acids in the code.

Machine learning approaches to predict drug efficacy and toxicity in oncology

In recent years, there has been a surge of interest in using machine learning algorithms (MLAs) in oncology, particularly for biomedical applications such as drug discovery, drug repurposing, diagnostics, clinical trial design, and pharmaceutical production. MLAs have the potential to provide valuable insights and predictions in these areas by representing both the disease state and the therapeutic agents used to treat it. To fully utilize the capabilities of MLAs in oncology, it is important to understand the fundamental concepts underlying these algorithms and how they can be applied to assess the efficacy and toxicity of therapeutics. In this perspective, we lay out approaches to represent both the disease state and the therapeutic agents used by MLAs to derive novel insights and make relevant predictions.

Prognostic and survival impact of BCL9 and RPS6KB1 copy number variation detected from circulating free DNA in hepatocellular carcinoma

BACKGROUND

Circulating cell-free DNA (cfDNA) is a noninvasive substitute to liver biopsy for hepatocellular carcinoma (HCC) molecular profiling. This study aimed to use cfDNA to investigate copy number variation (CNV) in the BCL9 and RPS6KB1 genes and its impact on prognosis in HCC.

METHODS

Real-Time Polymerase Chain Reaction was used to determine the CNV and cfDNA integrity index in 100 HCC patients.

RESULTS

CNV gain in BCL9 and RPS6KB1 genes was detected in 14% and 24% of patients, respectively. Gain in CNV of BCL9 associated with risk of HCC in alcohol drinkers and hepatitis C seropositivity. In patients with RPS6KB1 gain, HCC risk increased with a high body mass index, smoking, schistosomiasis, and Barcelona clinical liver cancer stage (BCLC) A. Gain in both genes showed a high risk of HCC with elevated liver enzymes, Schistosomiasis, BCLC C, and PS > 1. The integrity of cfDNA was higher in patients with CNV gain in RPS6KB1 than those harboring CNV gain in BCL9. Lastly, BCL9 gain and BCL9 + RPS6KB1 gain led to higher mortality rates and reduced survival times.

CONCLUSION

cfDNA was used to detect BCL9 and RPS6KB1 CNVs, which influence prognosis and can be used as independent predictors of HCC patient survival.

Multiclass Cancer Prediction Based on Copy Number Variation Using Deep Learning

DNA copy number variation (CNV) is the type of DNA variation which is associated with various human diseases. CNV ranges in size from 1 kilobase to several megabases on a chromosome. Most of the computational research for cancer classification is traditional machine learning based, which relies on handcrafted extraction and selection of features. To the best of our knowledge, the deep learning-based research also uses the step of feature extraction and selection. To understand the difference between multiple human cancers, we developed three end-to-end deep learning models, i.e., DNN (fully connected), CNN (convolution neural network), and RNN (recurrent neural network), to classify six cancer types using the CNV data of 24,174 genes. The strength of an end-to-end deep learning model lies in representation learning (automatic feature extraction). The purpose of proposing more than one model is to find which architecture among them performs better for CNV data. Our best model achieved 92% accuracy with an ROC of 0.99, and we compared the performances of our proposed models with state-of-the-art techniques. Our models have outperformed the state-of-the-art techniques in terms of accuracy, precision, and ROC. In the future, we aim to work on other types of cancers as well.

Copy number variation profile-based genomic typing of premenstrual dysphoric disorder in Chinese

Premenstrual dysphoric disorder (PMDD) affects nearly 5% women of reproductive age. Symptomatic heterogeneity, together with largely unknown genetics, have greatly hindered its effective treatment. In the present study, analysis of genomic sequencing-based copy-number-variations (CNVs) called from 100-kb white blood cell DNA sequence windows by means of semi-supervised clustering led to the segregation of patient genomes into the D and V groups, which correlated with the depression and invasion clinical types respectively with 89.0% consistency. Application of diagnostic CNV features selected using the correlation-based machine-learning method enabled the classification of the CNVs obtained into the D group, V group, total-patient group and control group with an average accuracy of 83.0%. The power of the diagnostic CNV features was 0.98 on average, suggesting that these CNV features could be employed for the molecular diagnosis of the major clinical types of PMDD. This demonstrated concordnce between the CNV profiles and clinical types of PMDD supported the validity of symptom-based diagnosis of PMDD for differentiating between its two major clinical types, as well as the predominanly genetic nature of PMDD with a host of overlaps between multiple susceptibility genes/pathways and the diagnostic CNV features as indicators of involvement in PMDD etiology.

Genetic-variant hotspots and hotspot clusters in the human genome facilitating adaptation while increasing instability

Background

Genetic variants, underlining phenotypic diversity, are known to distribute unevenly in the human genome. A comprehensive understanding of the distributions of different genetic variants is important for insights into genetic functions and disorders.

Methods

Herein, a sliding-window scan of regional densities of eight kinds of germline genetic variants, including single-nucleotide-polymorphisms (SNPs) and four size-classes of copy-number-variations (CNVs) in the human genome has been performed.

Results

The study has identified 44,379 hotspots with high genetic-variant densities, and 1135 hotspot clusters comprising more than one type of hotspots, accounting for 3.1% and 0.2% of the genome respectively. The hotspots and clusters are found to co-localize with different functional genomic features, as exemplified by the associations of hotspots of middle-size CNVs with histone-modification sites, work with balancing and positive selections to meet the need for diversity in immune proteins, and facilitate the development of sensory-perception and neuroactive ligand-receptor interaction pathways in the function-sparse late-replicating genomic sequences. Genetic variants of different lengths co-localize with retrotransposons of different ages on a “long-with-young” and “short-with-all” basis. Hotspots and clusters are highly associated with tumor suppressor genes and oncogenes (p < 10⁻¹⁰), and enriched with somatic tumor CNVs and the trait- and disease-associated SNPs identified by genome-wise association studies, exceeding tenfold enrichment in clusters comprising SNPs and extra-long CNVs.

Conclusions

In conclusion, the genetic-variant hotspots and clusters represent two-edged swords that spearhead both positive and negative genomic changes. Their strong associations with complex traits and diseases also open up a potential “Common Disease-Hotspot Variant” approach to the missing heritability problem.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40246-021-00318-3.

Evaluation of a genetic risk score for severity of COVID-19 using human chromosomal-scale length variation

INTRODUCTION

The course of COVID-19 varies from asymptomatic to severe in patients. The basis for this range in symptoms is unknown. One possibility is that genetic variation is partly responsible for the highly variable response. We evaluated how well a genetic risk score based on chromosomal-scale length variation and machine learning classification algorithms could predict severity of response to SARS-CoV-2 infection.

METHODS

We compared 981 patients from the UK Biobank dataset who had a severe reaction to SARS-CoV-2 infection before 27 April 2020 to a similar number of age-matched patients drawn for the general UK Biobank population. For each patient, we built a profile of 88 numbers characterizing the chromosomal-scale length variability of their germ line DNA. Each number represented one quarter of the 22 autosomes. We used the machine learning algorithm XGBoost to build a classifier that could predict whether a person would have a severe reaction to COVID-19 based only on their 88-number classification.

RESULTS

We found that the XGBoost classifier could differentiate between the two classes at a significant level (p = 2 · 10-11) as measured against a randomized control and (p = 3 · 10-14) as measured against the expected value of a random guessing algorithm (AUC = 0.5). However, we found that the AUC of the classifier was only 0.51, too low for a clinically useful test.

CONCLUSION

Genetics play a role in the severity of COVID-19, but we cannot yet develop a useful genetic test to predict severity.

Pan-Cancer Classification Based on Self-Normalizing Neural Networks and Feature Selection

Cancer is a one of the severest diseases and cancer classification plays an important role in cancer diagnosis and treatment. Some different cancers even have similar molecular features such as DNA copy number variant. Pan-cancer classification is still non-trivial at molecular level. Herein, we propose a computational method to classify cancer types by using the self-normalizing neural network (SNN) for analyzing pan-cancer copy number variation data. Since the dimension of the copy number variation features is high, the Monte Carlo feature selection method was used to rank these features. Then a classifier was built by SNN and feature selection method to select features. Three thousand six hundred ninety-four features were chosen for the prediction model, which yields the accuracy value is 0.798 and macro F1 is 0.789. We compared our model to random forest method. Results show the accuracy and macro F1 obtained by our classifier are higher than those obtained by random forest classifier, indicating the good predictive power of our method in distinguishing four different cancer types. This method is also extendable to pan-cancer classification for other molecular features.

Highly Recurrent Copy Number Variations in GABRB2 Associated With Schizophrenia and Premenstrual Dysphoric Disorder

OBJECTIVE

Although single-nucleotide polymorphisms in GABRB2, the gene encoding for GABAA receptors β2 subunit, have been associated with schizophrenia (SCZ), it is unknown whether there is any association of copy number variations (CNVs) in this gene with either SCZ or premenstrual dysphoric disorder (PMDD).

METHODS

In this study, the occurrences of the recurrent CNVs esv2730987 in Intron 6 and nsv1177513 in Exon 11 of GABRB2 in Chinese and German SCZ, and Chinese PMDD patients were compared to controls of same ethnicity and gender by quantitative PCR (qPCR).

RESULTS

The results demonstrated that copy-number-gains were enriched in both SCZ and PMDD patients with significant odds ratios (OR). For combined-gender SCZ patients versus controls, about two-fold increases were observed in both ethnic groups at both esv2730987 (OR = 2.15, p = 5.32E-4 in Chinese group; OR = 2.79, p = 8.84E-3 in German group) and nsv1177513 (OR = 3.29, p = 1.28E-11 in Chinese group; OR = 2.44, p = 6.17E-5 in German group). The most significant copy-number-gains were observed in Chinese females at nsv1177513 (OR = 3.41), and German females at esv2730987 (OR=3.96). Copy-number-gains were also enriched in Chinese PMDD patients versus controls at esv2730987 (OR = 10.53, p = 4.34E-26) and nsv1177513 (OR = 2.39, p = 3.19E-5).

CONCLUSION

These findings established for the first time the association of highly recurrent CNVs with SCZ and PMDD, suggesting the presence of an overlapping genetic basis with shared biomarkers for these two common psychiatric disorders.

Genomic subtyping of liver cancers with prognostic application

BACKGROUND

Cancer subtyping has mainly relied on pathological and molecular means. Massively parallel sequencing-enabled subtyping requires genomic markers to be developed based on global features rather than individual mutations for effective implementation.

METHODS

In the present study, the whole genome sequences (WGS) of 110 liver cancers of Japanese patients published with different pathologies were analyzed with respect to their single nucleotide variations (SNVs) comprising both gain-of-heterozygosity (GOH) and loss-of-heterozygosity (LOH) mutations, the signatures of combined GOH and LOH mutations, along with recurrent copy number variations (CNVs).

RESULTS

The results, obtained based on the WGS sequences as well as the Exome subset within the WGSs that covered ~ 2.0% of the WGS and the AluScan-subset within the WGSs that were amplifiable by Alu element-consensus primers and covered ~ 2.1% of the WGS, indicated that the WGS samples could be employed with the mutational parameters of SNV load, LOH%, the Signature α%, and survival-associated recurrent CNVs (srCNVs) as genomic markers for subtyping to stratify liver cancer patients prognostically into the long and short survival subgroups. The usage of the AluScan-subset data, which could be implemented with sub-micrograms of DNA samples and vastly reduced sequencing analysis task, outperformed the usage of WGS data when LOH% was employed as stratifying criterion.

CONCLUSIONS

Thus genomic subtyping performed with novel genomic markers identified in this study was effective in predicting patient-survival duration, with cohorts of hepatocellular carcinomas alone and those including intrahepatic cholangiocarcinomas. Such relatively heterogeneity-insensitive genomic subtyping merits further studies with a broader spectrum of cancers.

A snapshot neural ensemble method for cancer-type prediction based on copy number variations

An accurate diagnosis and prognosis for cancer are specific to patients with particular cancer types and molecular traits, which needs to address carefully. The discovery of important biomarkers is becoming an important step toward understanding the molecular mechanisms of carcinogenesis in which genomics data and clinical outcomes need to be analyzed before making any clinical decision. Copy number variations (CNVs) are found to be associated with the risk of individual cancers and hence can be used to reveal genetic predispositions before cancer develops. In this paper, we collect the CNVs data about 8000 cancer patients covering 14 different cancer types from The Cancer Genome Atlas. Then, two different sparse representations of CNVs based on 578 oncogenes and 20,308 protein-coding genes, including genomic deletions and duplication across the samples, are prepared. Then, we train Conv-LSTM and convolutional autoencoder (CAE) networks using both representations and create snapshot models. While the Conv-LSTM can capture locally and globally important features, CAE can utilize unsupervised pretraining to initialize the weights in the subsequent convolutional layers against the sparsity. Model averaging ensemble (MAE) is then applied to combine the snapshot models in order to make a single prediction. Finally, we identify most significant CNVs biomarkers using guided-gradient class activation map plus (GradCAM++) and rank top genes for different cancer types. Results covering several experiments show fairly high prediction accuracies for the majority of cancer types. In particular, using protein-coding genes, Conv-LSTM and CAE networks can predict cancer types correctly at least 72.96% and 76.77% of the cases, respectively. Contrarily, using oncogenes gives moderately higher accuracies of 74.25% and 78.32%, whereas the snapshot model based on MAE shows overall 2.5% of accuracy improvement.

AluScanCNV2: An R package for copy number variation calling and cancer risk prediction with next-generation sequencing data

The usage of next-generation sequencing (NGS) to detect copy number variation (CNV) is widely accepted in cancer research. Based on an AluScanCNV software developed by us previously, an AluScanCNV2 software has been developed in the present study as an R package that performs CNV detection from NGS data obtained through AluScan, whole-genome sequencing or other targeted NGS platforms. Its applications would include the expedited usage of somatic CNVs for cancer subtyping, and usage of recurrent germline CNVs to perform machine learning-assisted prediction of a test subject's susceptibility to cancer.

Identifying Patients with Atrioventricular Septal Defect in Down Syndrome Populations by Using Self-Normalizing Neural Networks and Feature Selection

Atrioventricular septal defect (AVSD) is a clinically significant subtype of congenital heart disease (CHD) that severely influences the health of babies during birth and is associated with Down syndrome (DS). Thus, exploring the differences in functional genes in DS samples with and without AVSD is a critical way to investigate the complex association between AVSD and DS. In this study, we present a computational method to distinguish DS patients with AVSD from those without AVSD using the newly proposed self-normalizing neural network (SNN). First, each patient was encoded by using the copy number of probes on chromosome 21. The encoded features were ranked by the reliable Monte Carlo feature selection (MCFS) method to obtain a ranked feature list. Based on this feature list, we used a two-stage incremental feature selection to construct two series of feature subsets and applied SNNs to build classifiers to identify optimal features. Results show that 2737 optimal features were obtained, and the corresponding optimal SNN classifier constructed on optimal features yielded a Matthew’s correlation coefficient (MCC) value of 0.748. For comparison, random forest was also used to build classifiers and uncover optimal features. This method received an optimal MCC value of 0.582 when top 132 features were utilized. Finally, we analyzed some key features derived from the optimal features in SNNs found in literature support to further reveal their essential roles.

Feature co-localization landscape of the human genome

Although feature co-localizations could serve as useful guide-posts to genome architecture, a comprehensive and quantitative feature co-localization map of the human genome has been lacking. Herein we show that, in contrast to the conventional bipartite division of genomic sequences into genic and inter-genic regions, pairwise co-localizations of forty-two genomic features in the twenty-two autosomes based on 50-kb to 2,000-kb sequence windows indicate a tripartite zonal architecture comprising Genic zones enriched with gene-related features and Alu-elements; Proximal zones enriched with MIR- and L2-elements, transcription-factor-binding-sites (TFBSs), and conserved-indels (CIDs); and Distal zones enriched with L1-elements. Co-localizations between single-nucleotide-polymorphisms (SNPs) and copy-number-variations (CNVs) reveal a fraction of sequence windows displaying steeply enhanced levels of SNPs, CNVs and recombination rates that point to active adaptive evolution in such pathways as immune response, sensory perceptions, and cognition. The strongest positive co-localization observed between TFBSs and CIDs suggests a regulatory role of CIDs in cooperation with TFBSs. The positive co-localizations of cancer somatic CNVs (CNVT) with all Proximal zone and most Genic zone features, in contrast to the distinctly more restricted co-localizations exhibited by germline CNVs (CNVG), reveal disparate distributions of CNVTs and CNVGs indicative of dissimilarity in their underlying mechanisms.

Copy number variation analysis based on AluScan sequences

BACKGROUND

AluScan combines inter-Alu PCR using multiple Alu-based primers with opposite orientations and next-generation sequencing to capture a huge number of Alu-proximal genomic sequences for investigation. Its requirement of only sub-microgram quantities of DNA facilitates the examination of large numbers of samples. However, the special features of AluScan data rendered difficult the calling of copy number variation (CNV) directly using the calling algorithms designed for whole genome sequencing (WGS) or exome sequencing.

RESULTS

In this study, an AluScanCNV package has been assembled for efficient CNV calling from AluScan sequencing data employing a Geary-Hinkley transformation (GHT) of read-depth ratios between either paired test-control samples, or between test samples and a reference template constructed from reference samples, to call the localized CNVs, followed by use of a GISTIC-like algorithm to identify recurrent CNVs and circular binary segmentation (CBS) to reveal large extended CNVs. To evaluate the utility of CNVs called from AluScan data, the AluScans from 23 non-cancer and 38 cancer genomes were analyzed in this study. The glioma samples analyzed yielded the familiar extended copy-number losses on chromosomes 1p and 9. Also, the recurrent somatic CNVs identified from liver cancer samples were similar to those reported for liver cancer WGS with respect to a striking enrichment of copy-number gains in chromosomes 1q and 8q. When localized or recurrent CNV-features capable of distinguishing between liver and non-liver cancer samples were selected by correlation-based machine learning, a highly accurate separation of the liver and non-liver cancer classes was attained.

CONCLUSIONS

The results obtained from non-cancer and cancerous tissues indicated that the AluScanCNV package can be employed to call localized, recurrent and extended CNVs from AluScan sequences. Moreover, both the localized and recurrent CNVs identified by this method could be subjected to machine-learning selection to yield distinguishing CNV-features that were capable of separating between liver cancers and other types of cancers. Since the method is applicable to any human DNA sample with or without the availability of a paired control, it can also be employed to analyze the constitutional CNVs of individuals.