Fig. 2

Overview of the DeepLOF model. DeepLOF combines a feedforward neural network and a population genetics-based likelihood function to infer the relative rate of LOF variants in a gene (\(\eta\)) with respect to the expected number under a neutral mutation model (n). The feedforward neural network transforms genomic features into a beta prior distribution of \(\eta\), which represents our belief about \(\eta\) based on genomic features. The population genetics-based likelihood function describes the probability of observing y LOF variants in a gene conditional on \(\eta\) and n, which represents our belief about \(\eta\) based on population genomic data. Finally, DeepLOF combines the prior distribution and the likelihood function to compute the posterior distribution of \(\eta\). The DeepLOF score is defined as \(1 - \mathbb {E}[\eta ]\), where \(\mathbb {E}[\eta ]\) is the mean of \(\eta\) under the posterior distribution