Fig. 4: Results of LUAD application experiments.

a Top 25 CpG sites selected by the proposed method on LUAD data. Listed are the chromosomes (CHR), coordinates (MAPINFO), and genes, as well as BH-corrected P values derived from two-sided T-test between cancer samples and normal samples and the weights obtained by the model. Different colors distinct different genes. b Classification performance of reconstructed classification models on another GEO dataset based on the top-ranked 25 CpG sites. The error bar is presented as mean value ± standard error of ten-time repetitions. The P values derived from one-sided paired-sample T-test are listed on the right side. SVM support vector machine, RF random forest. c Comparison of the top 50 CpG sites and their corresponding genes with and without contrastive scheme (denoted by w C and w/o C, respectively). d A visualization of the contrastive embeddings that are compressed into a two-dimensional plane using Principal Component Analysis (PCA)⁷¹. The data points denote paired cancer (red) and paired normal (gray) samples at the start and end stages of training. e Euclidean distances between samples with different classes in the PCA space. Violin plots represent the kernel density distribution and the box plots inside represent the median (center line), upper and lower quartiles, and 1.5× the interquartile range (whiskers). The number of sample distances for all three groups are as follows: Paired-cancer-normal: n = 29, Paired Cancer and Paired Normal: n = 841, Individual Cancer and Individual Normal: n = 1293. f Visualization of a segment of clustered five CpG sites (their IlmnID are listed on the right) on the LUAD data. The horizontal axis represents the site coordinate, and the vertical axis represents the methylation level, the BH-corrected P value derived from a two-sided T-test, and the weight given by the proposed method. Source data are provided as a Source Data file.