Predicting chemotherapy response using a variational autoencoder approach