Predicting the functions of proteins remains a critical yet challenging task in computational biology. Advances in high-throughput sequencing, the expansion of protein databases, and the continuous development of artificial intelligence have led to the emergence of many computational methods dedicated to protein function prediction. In this study, we evaluated the performance of four state-of-the-art models - DeepGOPlus, DeepGraphGO, DeepGOZero, and DeepGOSE - using experimentally annotated proteins from the UniProt-KB/Swiss-Prot database. We also trained and tested these models on species-specific datasets from Arabidopsis thaliana and Oryza sativa to investigate their potential and applicability in plant protein studies. Our results showed that DeepGOPlus consistently achieved the best evaluation scores across all datasets. DeepGOSE and DeepGOZero performed comparably and only marginally outperformed DeepGraphGO in certain training attempts. Further analysis revealed that dataset stratification into training, validation, and testing sets introduced variations in Gene Ontology annotation specificity, which may have influenced model performance.