Inference of Gene Networks from Single Cell Data through Quantified Inductive Logic Programming

Single cell sequencing technologies represent a unique opportunity to appreciate all the heterogeneity of gene expressions within specific biological cell types. While these data are sparse and especially noisy, it remains possible to perform multiple analysis tasks such as identifying sub cellular types and biological markers. Beyond revealing distinct sub cell populations, single cell gene expressions usually involve complex gene interactions, which may often be interpreted as an underlying gene network. In this context, logical computational approaches are particularly attractive as they provide models that are easy to interpret and verify. However, the noise is especially important in single cell sequencing data. This may appear as a limit for symbolic methods as they usually fail in addressing the statistical aspect necessary to handle efficiently such noise. In this work, we propose a computational approach based on symbolic modeling to identify gene connections from single cell RNA sequencing data. Our algorithm, LOLH, is based on Inductive Logic Programming, and intends to rapidly identify potential gene interactions by formulating discrete classification problems, which are solved through discrete optimization. By combining symbolic modeling with optimization techniques, we aim to provide an interpretable model that still fits properly on sparse and noisy data. We apply our method to the unsupervised inference of a gene correlation network from a concrete single cell dataset. We show that the output of our algorithm can be interpreted by using the data itself, and we use additional biological knowledge to validate the approach.