An Evaluation of Machine Learning Approaches for the Prediction of Essential Genes in Eukaryotes Using Protein Sequence-Derived Features

Abstract

The availability of whole-genome sequences and associated multi-omics data sets, combined with advances in gene knockout and knockdown methods, has enabled large-scale annotation and exploration of gene and protein functions in eukaryotes. Knowing which genes are essential for the survival of eukaryotic organisms is paramount for an understanding of the basic mechanisms of life, and could assist in identifying intervention targets in eukaryotic pathogens and cancer. Here, we studied essential gene orthologs among selected species of eukaryotes, and then employed a systematic machine-learning approach, using protein sequence-derived features and selection procedures, to investigate essential gene predictions within and among species. We showed that the numbers of essential gene orthologs comprise small fractions when compared with the total number of orthologs among the eukaryotic species studied. In addition, we demonstrated that machine-learning models trained with subsets of essentiality-related data performed better than random guessing of gene essentiality for a particular species. Consistent with our gene ortholog analysis, the predictions of essential genes among multiple (including distantly-related) species is possible, yet challenging, suggesting that most essential genes are unique to a species. The present work provides a foundation for the expansion of genome-wide essentiality investigations in eukaryotes using machine learning approaches.

High-quality nuclear genome for Sarcoptes scabiei—A critical resource for a neglected parasite

“…Combined with the bioinformatic prediction and prioritisation of essential genes from functional information (e.g., lethality) available for other metazoan organisms, particularly D . melanogaster , using machine learning approaches [ 55 ], RNAi-based screening of S . scabiei stages provides a powerful functional genomics tool to validate prioritised targets.…”

Combined use of feature engineering and machine-learning to predict essential genes in Drosophila melanogaster

“…Features were selected by random subsampling from 10 to 90% of data representing ‘ essential ’ or ‘ non-essential ’’ genes (in 10% stepwise increments) based on a consensus between elasticNet (alpha = 0.5) and ensemble Sparse Partial Least Squares (SPLS) methods using ‘glmnet’ and ‘enspls’ in R, respectively ( 26 ). The individual feature values were then normalized by subtracting the mean and dividing by the standard deviation calculated for each feature column.…”

“…The individual feature values were then normalized by subtracting the mean and dividing by the standard deviation calculated for each feature column. Normalized features were used to train each of six ML-models (GBM (Gradient Boosting Machine), GLM (Generalized Linear Model), NN (Neural Network—perceptron), Random Forest (RF), SVM (Support-Vector Machine) ( 26 ) and XGB (eXtreme Gradient Boosting—xgbTree) in the ‘caret’ R-package. During the training process, we employed parameter-tuning and 5-fold cross-validation, ultimately selecting the models with highest ROC-AUC.…”

“…A range of genomic features, such as gene size, evolutionary rate, phyletic retention, transcription level, connectivity in protein–protein interaction (PPI) networks, cellular or subcellular localization, and/or sequence-derived features ( 24–27 ) have been linked to essentiality in eukaryotes. For D. melanogaster , studies have sought to infer essential genes computationally using features based on gene homology/orthology/ontology, PPI/co-expression networks, sequence-derived (nucleotide/protein), or combinations thereof ( 24 , 26–27 ). However, there are limitations in genome-wide predictions, particularly using PPI data and/or sequence-based features.…”

“…In addition, such datasets are limited or unavailable for non-model organisms. Additionally, although sequence-derived features alone have proven useful for essential gene predictions, their performance is still suboptimal, even if combined with PPI network features ( 26 , 27 ).…”