# BiGKbhb: a bi-directional gated recurrent unit model for predicting lysine β-hydroxybutyrylation sites

## TL;DR

This paper introduces BiGKbhb, a deep learning model that accurately predicts lysine β-hydroxybutyrylation sites in proteins across multiple species.

## Contribution

The novel contribution is the development of BiGKbhb, a BiGRU-based framework that outperforms existing tools in Kbhb site prediction.

## Key findings

- BiGKbhb achieves test set accuracies of 0.824, 0.832, and 0.871 for human, mouse, and fungal datasets.
- The model demonstrates robust performance with AUC values of 0.920, 0.902, and 0.945 across species.
- Cross-species analysis confirms improved transferability and significant performance over existing methods.

## Abstract

Post-Translational Modifications (PTMs) are covalent chemical alterations that occur after protein synthesis, critically regulating protein function, localization, and interactions. β-hydroxybutyrylation (Kbhb), a metabolically derived histone modification discovered in 2016, influences gene activation and cellular metabolism. While accurate PTM site identification is essential for understanding protein regulation and disease mechanisms, experimental approaches face significant limitations, including low modification abundance, high cost, and limited proteome coverage. Kbhb remains computationally underexplored, with only three existing prediction tools exhibiting modest accuracy and limited cross-species applicability. To address this gap, we developed BiGKbhb, a deep learning framework that depends on Bidirectional Gated Recurrent Units (BiGRU). With BiGKbhb, we systematically evaluate seven protein sequence encoding strategies, and compare six deep learning architectures using datasets from human, mouse, and fungal species. Results demonstrated that BLOSUM62 evolutionary encoding combined with BiGRU architecture achieves optimal performance, with BiGKbhb consistently achieving higher accuracy than those of existing methods with test set accuracies of 0.824, 0.832, and 0.871 for human, mouse, and fungal balanced datasets, respectively, with corresponding Area Under Curve (AUC) values of 0.920, 0.902, and 0.945, while additional evaluation on imbalanced datasets confirmed model robustness under realistic conditions. Cross-species analysis revealed enhanced transferability of the general multi-species model, and statistical validation confirmed significant improvements over existing predictors (p < 0.05). These findings contribute a robust computational tool for Kbhb prediction and provide insights into sequence determinants of this important modification across evolutionarily diverse species.

The online version contains supplementary material available at 10.1186/s12864-025-12166-9.