Fig. 2: Flexynesis supports single-task modeling for regression (A), classification (B), and survival (C).

For all three tasks, both a fully-connected-network and a supervised variational auto-encoder was trained and best-performing model’s results were presented. A Performance evaluation of Flexynesis on drug response prediction of a model trained on 1051 cell lines from CCLE (using RNA and CNV profiles) and evaluated on 1075 cell lines from GDSC2 for the drugs Lapatinib (Pearson correlation test, r = 0.6, p = 7.750175e-42) and Selumetinib (Pearson correlation test, r = 0.61, p = 3.873949e-50). The x-axis depicts observed drug response values (AAC-recomputed as in Pharmacogx package)⁶⁰ and the y-axis depicts the predicted drug response values for the test samples. B Evaluation of Flexynesis on microsatellite instability (MSI) status prediction using gene expression and/or promoter methylation data from seven different TCGA cohorts (gastrointestinal and gynocological cancers) with microsatellite instability (MSI) annotations: TCGA-COAD (Colon Adenocarcinoma), TCGA-ESCA (Esophageal Carcinoma), TCGA-PAAD (Pancreatic Adenocarcinoma), TCGA-READ (Rectum Adenocarcinoma), TCGA-STAD (Stomach Adenocarcinoma), TCGA-UCEC (Uterine Corpus Endometrial Carcinoma), TCGA-UCS (Uterine Carcinosarcoma). The models were trained on 70% of the samples (N = 1133) with MSI status annotations and evaluated on the remaining 30% of the samples (N = 283). The tSNE (t-distributed Stochastic Neighbor Embedding) plot represents the sample embeddings colored by MSI status and the ROC curve represents the best performing deep learning model based on both gene expression and methylation data. C Evaluation of Flexynesis on a survival modeling task on a merged cohort of LGG (Lower Grade Glioma) and GBM (Glioblastoma Multiforme) (using mutations and copy-number-alteration profiles). The model is trained on 557 samples and evaluated on 239 test samples. The tSNE plot depicts the sample embeddings obtained from the model encoder for the test samples colored by the predicted Cox proportional hazard risk scores stratified into “high-risk” and “low-risk” based on the median risk score. The Kaplan-Meier-Plot represents the survival stratification of the test samples based on this risk stratification (Logrank Test, p = 9.94475168880626e − 10). Source data are provided as a Source Data file.