Fig. 5: Pan-cancer neural network model predicts patient outcome based on epifactor gene expression patterns.

a A Cox-nnet model⁴⁰ was used as a framework for predicting patient outcomes. The patient cohort was randomly split (80:20) into training and test sets. The model was trained on input features consisting of the expression values of the 720 epifactor genes, and the age and sex of the patients in the training set. The model consisted of an “input layer” that accepts the input features and is fully connected to a “hidden layer.” The output of the nodes of the “hidden layer” was fed to a “cox-regression layer.” The final output of the model was the log hazard ratios of the patients (prognostic index, PI). To evaluate the performance of the model, the test set patients were divided into high PI and low PI groups based on the median PI of the patients. The clinical outcomes between these two groups were compared using the log-rank Mantel–Cox test (Kaplan–Meier method). Created using Biorender. b–d Kaplan–Meier plots evaluating the performance of the model. b Results when the model was trained and tested on patients from the 5-cancer group (ACC, KIRC, LGG, LIHC, and LUAD). High PI n = 71, Low PI n = 70. c Results when the model was trained and tested on four cancer types (ACC, LGG, LIHC, and LUAD). High PI n = 55, Low PI n = 56. d Results for a model trained and tested on only KIRC. High PI n = 17, Low PI n = 17. e Prognostic status of the top 20 input features (left panel) ranked on the basis of their importance in the Cox-nnet machine learning (ML) model for the five cancer types is shown. A heatmap indicating which of the top 20 features from the left panel are also top NMF genes across the five cancer types is shown on the right. Only the features that are a top NMF gene for at least one cancer type are shown. f Same as (e), but for the four cancer type model. KIRC was not included in the Cox-nnet model, but is included in these heatmaps for comparison. Supporting information for this figure can be found in Supplementary Data 9.