Figure 1.

Overview of DeepGeni analysis. (a) The source microbiome profiles are augmented with realistic profiles generated from DeepBioGen and an interpretable autoencoder is trained on the augmented training data. (b) Through the encoder of the trained autoencoder, the source microbiome profiles are transformed into latent representations that are used to train classifiers predicting ICI responsiveness. (c) The informative latent variables are identified with a feature selection algorithm and their signals are propagated through weights on the decoder of the trained autoencoder to identify the informative microbial markers.