Deciphering Pleiotropic Signatures of Regulatory SNPs in Zea mays L. Using Multi-Omics Data and Machine Learning Algorithms

Abstract

Maize is one of the most widely grown cereals in the world. However, to address the challenges in maize breeding arising from climatic anomalies, there is a need for developing novel strategies to harness the power of multi-omics technologies. In this regard, pleiotropy is an important genetic phenomenon that can be utilized to simultaneously enhance multiple agronomic phenotypes in maize. In addition to pleiotropy, another aspect is the consideration of the regulatory SNPs (rSNPs) that are likely to have causal effects in phenotypic development. By incorporating both aspects in our study, we performed a systematic analysis based on multi-omics data to reveal the novel pleiotropic signatures of rSNPs in a global maize population. For this purpose, we first applied Random Forests and then Markov clustering algorithms to decipher the pleiotropic signatures of rSNPs, based on which hierarchical network models are constructed to elucidate the complex interplay among transcription factors, rSNPs, and phenotypes. The results obtained in our study could help to understand the genetic programs orchestrating multiple phenotypes and thus could provide novel breeding targets for the simultaneous improvement of several agronomic traits.

Keywords: gene expression profiles; hierarchical network model; incremental feature selection; markov clustering; multi-omics; random forest; regulatory SNPs.

Figure 1

Overview of the analyses pipeline highlighting key machine learning algorithms for the identification of pleiotropic signatures of regulatory SNPs (rSNPs) to establish complex interplay of transcription factors (TFs), rSNPs and multiple phenotypes. The genotypic data (A1), consisting of 1.03 m SNP markers was filtered for MAF (<0.05) and 31,000 SNPs found within promoter regions of 37,407 maize genes were considered for association analysis with 20 quantitative agronomic traits (A2). RNA-seq (A3) dataset was utilized for the validation of pleiotropic rSNPs on the underlying gene expression. As of first step in the data analysis, rSNPs were identified (B) for their impact on the gain or loss of TFBSs, after which their association with multiple phenotypes was determined using random forest (RF) using the Boruta algorithm and incremental feature selection (IFS) technique (C). Pleiotropic signatures of rSNPs were then established by pruning weaker connections in the overall network into smaller non-overlapping fully connected clusters, using Markov clustering (MCL) algorithm (D) which provided the basis for the construction of hierarchical network models with three distinct layers modelling the complex interplay of TFs, rSNPs and multiple phenotype (B). Further, the boxplots show the impact of pleiotropic rSNPs at gene expression level as a function of gain or loss of TFBSs (E).

Figure 2

A plot to show the change of $R^2-$values versus the number of rSNPs in association with the phenotype pollen shed. The incremental feature selection (IFS) curves were drawn using the ranking of rSNPs. The $R^2-$ value reached a peak when considering the first 90 rSNPs. These rSNPs were used for the further analysis of this phenotype.

Figure 3

Number of associated rSNPs determined by the incremental feature selection (IFS) procedure for each phenotype and their overlap represented in matrix layouts using the UpSet technique [67]. Black circles in the matrix layout are related to the phenotypes that are part of the intersection. For the sake of clarity, not all intersections are displayed.

Figure 4

Hierarchical network model constructed using Cluster-7 to elucidate the complex interplay among TFs−rSNPs(genes)−Phenotypes. (A–C) show the significant changes in the gene expression values resulting from the consequences of pleiotropic rSNPs. (D) Hierarchical network model with three layers.