PFGPred: a stack ensemble classifier for the identification of fusion genes in plants

Fusion genes play crucial roles in plant biological processes but remain far less explored than their human counterparts, largely due to limited validated datasets and the absence of plant-specific prediction tools. Existing approaches often produce high false-positive rates, restricting reliable discovery. To address this gap, we developed Plant Fusion Gene Predictor (PFGPred). This ensemble machine learning framework integrates Random Forest, XGBoost, and long short-term memory (LSTM) models into a meta-classifier for accurate identification of true and false fusion genes from RNA sequencing (RNA-Seq) data. PFGPred was trained on a high-confidence dataset of fusion genes validated by both RNA-Seq and whole-genome sequencing from Arabidopsis thaliana, Oryza sativa, Triticum aestivum, and Zea mays, to predict and rank candidate fusion genes for future functional validation. It outperformed individual baseline models, achieving accuracies of 0.97 on training data and 0.77 on independent test data. When evaluated on human datasets, it achieved 0.71 accuracy at the cost of lower sensitivity, reflecting biological differences between plant and human fusion events. Comparative analyses confirmed that PFGPred reliably identifies validated fusions, demonstrating its utility as a cost-effective, plant-specific prediction tool for high-throughput fusion gene screening and functional genomics research. It is freely available as a web server at http://www.nipgr.ac.in/PFGPred.