Predicting the function of proteins is an imperative task in bioinformatics, qualifying depth wise analysis of speck mechanisms underlying life and supporting breakthroughs in drug discovery, research of disease and refined biology. Classical methods for protein function prediction, for example, BLAST (Altschul et al. in J Mol Biol 215:403–410, 1990) and template-based methods, depend profoundly on sequence homology (Pearson, W.R.: An introduction to sequence similarity (“homology”) searching. Curr. Protoc. Bioinforma. Chapter 3, 3.1.1–3.1.8 (2013). https://doi.org/10.1002/0471250953.bi0301s4 ) and structural motifs. One of the drawbacks of the classical methods is dealing with unknown proteins and sequences that have a low similarity score to annotated terms, limiting their effectiveness in the verity of predicting functions across manifold datasets. The deep-learning technique has appeared as a potent approach having various models designed to learn compound links directly from diverse protein datasets. This survey provides a comprehensive assessment of different protein function prediction methods, revolving around classical homology-based techniques to modern deep learning models. Results obtained from the evaluation signify that deep-learning models surpass classical methods in peculiar tasks. The hybrid techniques that incorporate sequence, structure and protein–protein interactions to provide balanced solutions are promising candidates for large-scale functional annotation. This survey emphasizes the necessity for continuous advancements in computational efficiency, model interpretability and integration of various data sources to fully analyse the potency of protein function prediction in biological research and practical applications.