Prediction show off methylation reputation and you can peak. (A) ROC curves out of get across-genome recognition out of methylation updates prediction. Colors represent classifier trained using ability combos specified in the legend. Each ROC curve means an average untrue self-confident rate and you can correct confident price to own prediction for the held-away establishes for every anastasiadate of 10 repeated random subsamples. (B) ROC curves for different classifiers. Tone depict forecast to possess good classifier denoted throughout the legend. For every ROC bend signifies an average untrue positive rate and you may real confident price getting prediction into stored-out sets for every of the ten regular arbitrary subsamples. (C) Precision–recall curves for part-particular methylation reputation anticipate. Color depict prediction into the CpG sites inside specific genomic places given that denoted from the legend. For every accuracy–bear in mind contour means an average reliability–recall to have prediction with the stored-away set for every of the 10 constant random subsamples. (D) Two-dimensional histogram of forecast methylation account as opposed to fresh methylation accounts. x- and you may y-axes portray assayed as opposed to predict ? viewpoints, respectively. Shade represent the density of each and every matrix unit, averaged over-all forecasts having one hundred individuals. CGI, CpG isle; Gene_pos, genomic status; k-NN, k-nearest residents classifier; ROC, individual performing attribute; seq_property, succession attributes; SVM, assistance vector host; TFBS, transcription basis binding website; HM, histone modification scratching; ChromHMM, chromatin claims, once the defined from the ChromHMM app .
Cross-attempt forecast
To decide how predictive methylation users was in fact all over trials, i quantified the newest generalization error your classifier genome-greater round the individuals. Specifically, i instructed the classifier toward 10,000 internet from one private, and you will predicted methylation updates for everyone CpG web sites toward almost every other 99 somebody. The newest classifier’s abilities is extremely uniform across the people (Extra file step one: Profile S4), suggesting that individual-particular covariates – additional dimensions of cell sizes, such as for instance – don’t restriction forecast precision. The brand new classifier’s efficiency is extremely uniform when education toward lady and you will predicting CpG web site methylation condition in guys, and you will the other way around (Most document step 1: Contour S5).
To check the fresh new susceptibility in our classifier on number of CpG internet regarding the degree lay, we investigated the new forecast performance for different degree lay versions. We discovered that degree sets having greater than 1,000 CpG web sites got quite equivalent overall performance (Extra file step one: Figure S6). During these tests, i made use of a training put measurements of ten,000, so you’re able to strike an equilibrium between enough numbers of studies examples and you may computational tractability.
Cross-program prediction
To help you quantify group all over platform and you will cell-method of heterogeneity, i investigated brand new classifier’s results with the WGBS study [59,60]. Specifically, i categorized each CpG web site from inside the an effective WGBS shot based on whether one CpG webpages is assayed towards 450K number (450K webpages) or perhaps not (low 450K webpages); neighboring websites in the WGBS study try web sites which might be surrounding to the genome when they are both 450K internet. We use you to WGBS try regarding b-cells, that can match specific proportion each and every whole blood attempt; i note that the fresh 450K assortment whole blood samples will have heterogeneous phone items compared to the brand new WGBS study. Total, we see a greater proportion off hypomethylated CpG web sites towards the brand new 450K range relative to the WGBS analysis (A lot more document step 1: Figure S7) from the disproportionate signal out-of hypomethylated CpG sites within CGIs on 450K assortment.
First, we investigated cross-platform prediction, training our classifier on a 450K array sample and testing on WGBS data. We trained the classifier on 10,000 CpG sites in the 450K array samples, and then we tested on 100,000 CpG sites in WGBS data twice – once restricting the test set to 450K sites and once restricting the test set to non 450K sites. We repeated this experiment ten times. Next, we performed the same experiment but trained and tested on the WGBS data. Because the proportion of hypomethylated and hypermethylated sites was imbalanced for CpG sites not on the 450K array, we used a precision–recall curve instead of a ROC curve to measure the prediction performance . We used all 122 features and considered prediction of inverse CpG status \(<\hat>> = -(\tau – 1)\) in this experiment, to assess the quality of the predictions for the less frequent class of hypomethylated CpG sites.