Identification of Orphan Genes in Unbalanced Datasets Based on Ensemble Learning

Orphan genes are associated with regulatory patterns, but experimental methods for identifying orphan genes are both time-consuming and expensive. Designing an accurate and robust classification model to detect orphan and non-orphan genes in unbalanced distribution datasets poses a particularly huge challenge. Synthetic minority over-sampling algorithms (SMOTE) are selected in a preliminary step to deal with unbalanced gene datasets. To identify orphan genes in balanced and unbalanced Arabidopsis thaliana gene datasets, SMOTE algorithms were then combined with traditional and advanced ensemble classified algorithms respectively, using Support Vector Machine, Random Forest (RF), AdaBoost (adaptive boosting), GBDT (gradient boosting decision tree), and XGBoost (extreme gradient boosting). After comparing the performance of these ensemble models, SMOTE algorithms with XGBoost achieved an F1 score of 0.94 with the balanced A. thaliana gene datasets, but a lower score with the unbalanced datasets. The proposed ensemble method combines different balanced data algorithms including Borderline SMOTE (BSMOTE), Adaptive Synthetic Sampling (ADSYN), SMOTE-Tomek, and SMOTE-ENN with the XGBoost model separately. The performances of the SMOTE-ENN-XGBoost model, which combined over-sampling and under-sampling algorithms with XGBoost, achieved higher predictive accuracy than the other balanced algorithms with XGBoost models. Thus, SMOTE-ENN-XGBoost provides a theoretical basis for developing evaluation criteria for identifying orphan genes in unbalanced and biological datasets.