Inferring Conditional Probability Distributions of Noisy Gene Expression from Limited Observations by Deep Learning

Biological systems are inherently noisy so that two genetically identical cells in the exact same environment will sometimes behave in dramatically different ways. This imposes a big challenge in building traditional supervised machine learning models that can only predict determined phenotypic variables or categories per specific input condition. Furthermore, biological noise has been proven to play a crucial role in gene regulation mechanisms. The prediction of the average value of a given phenotype is not always sufficient to fully characterize a given biological system. In this study, we develop a deep learning algorithm that can predict the conditional probability distribution of a phenotype of interest with a small number of observations per input condition. We show that the deep neural network automatically generates the probability distributions based on 10 or less noisy measurements for each input condition, with no prior knowledge or assumption of the probability distributions.