Inferring DNA Methylation from RNA-Seq in Renal Papillary Carcinoma Using Paired Multi Omics Data - a GenAI model

DNA methylation regulates gene expression and is pivotal in cancer biology but its profiling is expensive and less common that RNA seuqencing. Here, we present an ensemble machine learning approach to predict methylation beta values from matched RNA-seq on the paired transcriptome–methylome data of the The Cancer Genome Atlas Kidney Renal Papillary Cell Carcinoma cohort (TCGA-KIRP). Our pipeline employs a patient-wise 80/20 train/test split, pre-filters RNA features iteratively based on high variance and model-informative fit, imputes missing values, scales inputs, and then trains two different complementary predictors; (1) a fully-connected auto encoder (MethylNetAutoencoder) that reconstructs overall methylation patterns from the RNA features, and (2) a fast, tree-based multi-output regressor (XGBoost where available, or RandomForest) that has been trained on PCA compressed methylation observations. The final predictions from each patient are the mean of the two model outputs. On the held-out test set, the final ensemble made predictions exhibiting R² = 0.386, mean absolute error (MAE) = 0.185 and a Pearson correlation R = 0.622 to true observations of the methylation beta values. Visual diagnostics show global hexbin plots, and per-patient distributions of correlation. Overall, these results demonstrate that RNA expression bears enough signal to estimate DNA methylation patterns in papillary renal cell carcinoma at a moderate level of accuracy.IntroductionDNA methylation is among the most widely studied epigenetic changes, with important effects on chromatin architecture, gene expression, and tumor biology [1,5]. Aberrant methylation patterns have been associated with cancer initiation, cancer progression, and cancer prognosis, suggesting it is useful class of biomarkers for clinical diagnosis and treatment decision [2,3]. However, methylation assays are still technically challenging and relatively expensive compared to RNA sequencing, which is now common place in large scale projects such as The Cancer Genome Atlas (TCGA) [4]. This disparity has motivated a growing interest in computational approaches that predict DNA methylation patterns based on transcriptomic data. In the present study, we describe a learning framework that is trained on paired RNA–methylation data from TCGA-KIRP, with the goal of performing RNA-only prediction of methylation profiles. We implement ensemble modeling strategies to evaluate prediction performance across CpG sites and patient-level correlations. Our findings show that transcriptomic data can capture meaningful methylation signals in papillary renal cell carcinoma, and that RNA-only prediction models can expand the use of transcriptomic datasets in cancer research.MethodsThis model was trained using paired RNA-Sequencing and Methylation of same patients to later enable methylation only from RNA sequencing. The code was generated with assitance from Github-CopilotData SourceRNA sequencing (FPKM) and DNA methylation (Illumina 450K) data for kidney renal papillary cell carcinoma (TCGA-KIRP) were obtained from The Cancer Genome Atlas (TCGA) portal. Only patients with both RNA-seq and methylation data available were included.Data ProcessingMethylation beta values were filtered to include high-quality CpG sites.RNA-seq expression data were log-transformed and normalized.Data were aligned by patient identifiers to generate paired RNA–methylation matrices .Model TrainingData were split into 80% training and 20% independent test sets (patient-wise split).Two base models were trained:A fully connected deep neural network (DNN).A gradient boosting regression model (XGBoost).An ensemble predictor was constructed by averaging predictions from the two models.Evaluation MetricsModel performance was evaluated on the held-out 20% test set using:R² scoreMean Absolute Error (MAE)Pearson correlation coefficient (R)ResultsEnsemble Model PerformanceOn the independent test set, the ensemble model achieved:R² score : 0.386MAE : 0.185Pearson correlation (R) : 0.622Spearman correlation : 0.615Example PredictionsA subset of predictions compared to ground truth values is shown inTable 1 .