TransKla: A Local-Global Cross-Attention Based Transformer Approach for Prediction of Lysine Lactylation Sites

Lysine lactylation (Kla) is a novel post-translational modification that bridges metabolic flux with epigenetic signaling. Dysregulation of lactylation disrupts multiple biological pathways, driving pathological states including oncogenesis, neural hyperexcitability, and immune dysfunction. Although wet-lab experiments are considered the gold standard, they are expensive and laborious. While computational methods have contributed to alternative solutions, they often fail to capture the unique biochemical properties of lactylation or integrate local sequence patterns with global protein context. To address these challenges, we present TransKla, a novel transformer-based framework that integrates key physicochemical features (charge and hydrophobicity) and sequence embeddings into a unified representation. The model combines local 41-residue context with global protein representations via cross-attention to capture long-range dependencies. Extensive ablation studies and a comprehensive regularization strategy validate our architectural choices and prevent overfitting. TransKla results in an AUPRC of 0.891, an AUC of 0.891, and an accuracy of 0.805 on the validation test, an increase in performance of 3.4%, 2.7%, and 3.1% on test data, respectively, and significantly outperforms state-of-the-art methods on the independent human Kla dataset. Moreover, our model uses ∼4.03 million parameters, much fewer than models that leverage large language models (LLMs). TransKla stands out as a useful tool, enabling accurate lactylation predictions with minimal computational resources. The dataset and source code used in this study are freely accessible at https://github.com/abduljabbar-repo/TransKla.git.