Extended Data Fig. 4: Neural Network prediction of IFN-γ and Il17-producing phenotypes.
From: Gut CD4+ T cell phenotypes are a continuum molded by microbes, not by TH archetypes
a, A Keras neural network was trained to use as input the expression of 500 most variable genes in Teff single-cell RNAseq data to predict Ifng or Il17a expression in each cell. Loss as a function of training epochs plotted here. Note the overfitting beyond 10 epochs (representative of >50 independent training runs with random 80/20 training/test). b, Accuracy of DNN-predicted cytokine expression by individual Teff cells, relative to their actual expression in the test scRNAseq data (non-expressing cells were not included as input, since there is uncertainty as to their real nature given drop-out frequencies in scRNAseq data). Numbers shown represent the range observed in 10 independent training runs (with different training/test sets). c, Contribution of each transcript to the prediction of Il17a or Ifng expression, as score in the Integrated Gradients, comparing the model learned in two independent runs. A positive score indicates influence on predicting Il17a expression, a negative score influence in predicting Ifng expression.
