Unsupervised Gene Regulatory Network Inference using K-Nearest-Neighbor Based Mutual Information

Abstract

Biological cells exhibit a variety of responses to internal cues and different external perturbations. This is possible due to elaborate gene regulatory networks (GRNs). The structure of GRN cannot be drawn by reading the genetic code, instead, a series of knockdown experiments on a downstream process or cell state was used to infer the GRN structure. The advancement in large scale gene expression data collection, enables inference of large networks. Basic networks were built by clustering co-expressed genes using a statistical correlation metric, like covariance. Information theory is a widely used correlation metric, as it can detect any correlation (linear and non-linear) between any number of variables (n-dimensional). Using information theory for higher dimensions and continuous data (i.e. normalizes expression data) is sensitive to data size, correlation strength and underlying distributions. To estimate it, few techniques are used: fixed width binning, adaptive partitioning, kernel density estimator and k-nearest neighbor (KNN). Unfortunately, most techniques require user defined parameters and “tweaking” which lead to many different implementations. The era of single cell transcriptome only worsens the problem as data is noisier, and while KNN was shown to be accurate and robust option for any user defined parameter, it was hardly used due to its computational costs. Here, we present a novel GRN inference method that solves two issues in the field: A) Implementation of unsupervised and accurate three-way MI estimator based on KNN, B) Reconstructing the network - we use all two- and three-way MI quantities (total of seven) to generate a digital signature (or “barcode”) to most 3-node network topologies, instead of ranking the interactions according to a single MI quantity with an arbitrary threshold to solve conflicts.