Linear grouping of predictor instances to infer gene networks

Abstract

One of the most important problems in the context of systems biology is to infer gene regulatory networks from gene expression data, since most of the control of cellular processes are performed by the multivariate activity of genes by means of their transcribed mRNA expression. Although many methods have been proposed to deal with this problem in the last two decades, gene network inference from gene expression data is still considered an open problem, mostly because the huge dimensionality (thousands of genes) and the very limited number of data samples typically available, even considering the fact that the network is sparse (limited number of input genes per target gene). In this work, we propose two variants of a previously published method which alleviates the curse of dimensionality by grouping predictor gene configurations in their respective linear combination values. Such values are assigned to equivalence classes. In this way, the number of instances of predictor values (equivalence classes) grows as a linear function of the dimensionality (number of predictors) instead of increasing as an exponential function when considering the original configurations. Both proposed method and its variants follow the probabilistic gene networks approach, applying local feature selection to achieve a satisfactory predictor gene set for each target gene. Although the results obtained from the aforementioned previous work were very promising, it applies the grouping unconditionally, even in cases where the number of samples is enough to estimate the conditional probabilities of the target given the predictors, which leads to unnecessary information loss. Results from simulated and real data indicate that, despite suffering from information loss, the inference with linear grouping tends to provide networks with better topological similarities than those obtained without grouping in cases where the number of samples is quite limited and the inference involves a larger number of predictors per gene. Besides, the variants proposed here displayed better results in cases where part of the parameters could be properly estimated without grouping, thus achieving better balance between information loss and estimation power gain.