Ferebee, T. H. | bioRxiv Science

bioRxiv2023

Fishing for a reelGene: evaluating gene models with evolution and machine learning

Assembled genomes and their associated annotations have transformed our study of gene function. However, each new assembly generates new gene models. Inconsistencies between annotations likely arise from biological and technical causes, including pseudogene misclassification, transposon activity, and intron retention from sequencing of unspliced transcripts. To evaluate gene model predictions, we developed reelGene, a pipeline of machine learning models focused on (1) transcription boundaries, (2) mRNA integrity, and (3) protein structure. The first two models leverage sequence characteristics and evolutionary conservation across related taxa to learn the grammar of conserved transcription boundaries and mRNA sequences, while the third uses conserved evolutionary grammar of protein sequences to predict whether a gene can produce a protein. Evaluating 1.8 million gene models in maize, reelGene found that 28% were incorrectly annotated or nonfunctional. By leveraging a large cohort of related species and through learning the conserved grammar of proteins, reelGene provides a tool for both evaluating gene model accuracy and genome biology.

genomics↗

bioRxiv2023

Exploring the utility of regulatory network-based machine learning for gene expression prediction in maize

Genomic selection and gene editing in crops could be enhanced by multi-species, mechanistic models predicting effects of changes in gene regulation. Current expression abundance prediction models require extensive computational resources, hard-to-measure species-specific training data, and often fail to incorporate data from multiple species. We hypothesize that gene expression prediction models that harness the regulatory network structure of Arabidopsis thaliana transcription factor-target gene interactions will improve on the present maize models. To this end, we collect 147 Oryza sativa and 99 Sorghum bicolor gene expression assays and assign them to maize family-based orthologous groups. Using three popular graph-based machine learning frameworks, including a shallow graph convolutional autoencoder, a deep graph convolutional autoencoder, and the inductive GraphSage strategy, we encode an Arabidopsis thaliana integrated gene regulatory network (iGRN) structure and TF gene expression values to predict gene expression both within and between species. We then evaluate the network methods against a partial least-squares baseline. We find that the baseline gives the best predictions within species, with Spearman correlations averaging between 0.74 and 0.78. The graph autoencoder methods were more variable with correlations between -0.1 and 0.65. In particular, the GraphSage and deep autoencoders performed the worst, and the shallow autoencoders performed the best. In the most challenging prediction context, where predictions were in new species and on genes that were not seen, we found that the shallow graph autoencoder framework averaged around 0.65. Unlike initial thoughts about preserved network structure improving gene expression predictions, this study shows that within-species predictions only need simple models, such as partial least squares, to capture expression variations. In cross-species predictions, the best model is often a more complex strategy utilizing regulatory network structure and other studies expressions.

bioinformatics↗