(27by) Deep Neural Networks for Predicting Single Cell Responses and Probability Landscapes

Predicting the dynamic behavior of genetic circuits is a crucial tool for engineering biology. However, the inherent stochasticity of biochemical reactions, variability between cells, and incomplete information about underlying biological processes often make such predictions challenging. Machine learning methods have immense potential for addressing these difficulties, as they are able to infer complex input-output relationships from data without a priori mechanistic knowledge. An important application of these techniques is the prediction of gene expression dynamics in a single cell given a prior history of responsiveness. To explore this, we computationally simulated single cell responses to optogenetic inputs, incorporating different types of noise and stochasticity as well as diverse genetic circuit architectures. Deep neural networks trained on these data could accurately infer single cell responses to diverse light inputs despite the presence of noise, though the stochasticity of the reactions produced small errors that grew with time. We also explored the limits of training set size and past information required for reliable predictions and showed that a cascaded genetic circuit that introduced delays increased the requirement for past information. However, our initial approach of predicting a single trajectory for a cell’s future response was ill-suited to genetic circuits with multiple steady-states. To better predict multimodal futures, we updated the neural network architecture to predict the entire distribution of future states. This new model could predict the bimodal responses of an auto-activation circuit with high precision. Together, these results highlight the power of machine learning methods for inferring complex dynamics from data and their flexibility to perform diverse prediction tasks. In the future, such methods can be adapted to predict and control a variety of natural and synthetic genetic circuits across biological systems and scales.