| At present,research on the use of bioinformatics tools to mine cancer-associated biomarkers has emerged at various levels of life activities(genome,transcriptome,epigenome,proteome,and metabolome).The identified biomarkers has been wildly studied for the application to cancer diagnosis.As an important epigenetic modification,DNA methylation is closely related to the occurrence and development of cancer.Compared to normal cells,DNA methylation patterns in tumor cells typically manifest as a decrease in the overall methylation level of the whole genome and an increase in the methylation level of the promoter region.Therefore,effective identification of DNA differential methylation patterns is of great significance for understanding cancer pathogenesis,cancer diagnosis and drug development.In this paper,a new method for identifying differentially methylated loci of DNA is proposed by using cosine similarity theory.Besides,we integrated the differentially expressed genes and differentially methylated loci/regions to predict cancer using machine learning methods,where the experimental results show that the integration of multi-scale gene features could benefit cancer diagnosis.The main research work and innovations of this paper are as follows.(1)A new method for identifying DNA differentially methylated loci was proposed.First,the methylation data in cases and controls are represented in vector form.Then,we define a reference vector and apply the cosine similarity theory to model the vectorized methylation data.By adding a sinusoidal term to the cosine similarity formula,the accuracy of differentially methylated loci recognition is significantly improved.Next,we introduce the skewness to model the outliers in the methylation data to reduce the interference of outliers on the identification of differentially methylated loci.Finally,we combine the improved cosine similarity formula and the skewness model.The model defines a score for the differentially methylated loci,which not only indicates the methylation status of the CpG site but also tells the degree of differential methylation.Moreover,we used statistical methods to estimate the significance of methylation sites.The innovations of this experiment are as follows:First,vectorize the methylation data,and then apply the cosine similarity theory to identify differentially methylated loci.Our method model not only the difference in the direction between vectors of methylation data,but also the difference in absolute distance between vectors.Besides,through adding a sine term,the function distribution is close to be a straight line,which increases the accuracy of methylation site recognition and improve the identification of CpG sites with a high degree of differential methylation.Both the simulation experiment and the real-world data experiment support the availability of the method.Second,our method is more fit the real-world methylation data compared to the published methods mainly in two aspects:(1)Methylation data does not obey any explicit distribution,but many of the proposed methods are based on the assumption of a normal distribution of methylated data.The method in this paper does not have assumptions about the distribution of data.(2)There are many outliers in the methylation data.This paper applies the skewness coefficient to model the outliers and tries to reduce its interference on the site discrimination.(2)Integrate multiple genome features to make predict tumor/control sample.The potential application of machine learning to cancer diagnosis is appealing wide attention.Determining features is a kernel step for such application.In this paper,we adopted differentially expressed genes and differentially methylated loci/regions extracted from RNA-Seq data and methylation microarray data,and then integrate these features to make machine leanring classifiers for cancer diagnosis.In the choice of classifier,we prefer to select the commonly used classifiers with good performance.Then,ROC curve is used to measure the performance of the classifiers on cancer samples and normal samples under different feature sets.The experimental results are as follows:The total AUC area of the four classifiers based on only differentially expressed gene features is 3.587.As a comparison,the total AUC area based on the integrated features(differential expressed genes and differentially methylated loci/regions)was increased by 7.4%over the former.The conclusion of the experiment is that the integration of multi-scale gene features can benefit the cancer diagnosis.this method can also be applied to other complex diseases and provide some new ideas for precision medicine. |
