ABSTRACT
Predicting cellular responses to perturbations is an unsolved problem in biology. Traditional approaches assume that different cell types respond similarly to perturbations. However, this assumption does not take into account the context of genome interactions in different cell types, which leads to compromised prediction quality. More recently, deep learning models used to discover gene-gene relationships can yield more accurate predictions of cellular responses. The huge difference in biological information between different cell types makes it difficult for deep learning models to encode data into a continuous low-dimensional feature space, which means that the features captured by the latent space may not be continuous. Therefore, the mapping relationship between the two conditional spaces learned by the model can only be applied where the real reference data resides, leading to the wrong mapping of the predicted target cells because they are not in the same domain as the reference data. In this paper, we propose an information-navigated variational autoencoder (INVAE), a deep neural network for cell perturbation response prediction. INVAE filters out information that is not conducive to predictive performance. For the remaining information, INVAE constructs a homogeneous space of control conditions, and finds the mapping relationship between the control condition space and the perturbation condition space. By embedding the target unit into the control space and then mapping it to the perturbation space, we can predict the perturbed state of the target unit. Comparing our proposed method with other three state-of-the-art methods on three real datasets, experimental results show that INVAE outperforms existing methods in cell state prediction after perturbation. Furthermore, we demonstrate that filtering out useless information not only improves prediction accuracy but also reveals similarities in how genes in different cell types are regulated following perturbation.
