KmalPred: a deep learning framework for lysine malonylation site prediction using protein language model...

Lysine malonylation (Kmal) is a reversible post-translational modification involved in metabolic regulation and other essential biological processes across diverse species. Accurate identification of Kmal sites is important for understanding their biological functions, but experimental approaches such as mass spectrometry remain labor-intensive and time-consuming. Existing computational predictors often rely on local sequence windows and therefore may fail to fully capture long-range contextual information within proteins.

We present KmalPred, a deep learning framework for Kmal site prediction that combines ProtT5-derived residue representations with a bidirectional long short-term memory network. Unlike conventional segment-first pipelines, KmalPred first encodes the full-length protein sequence and then extracts lysine-centered windows from the resulting residue-level embeddings. This sequence-first design preserves global sequence context and avoids redundant computation over highly overlapping windows. On the independent test set, KmalPred achieved an accuracy of 0.78 and a Matthews correlation coefficient of 0.56, outperforming previously reported predictors evaluated on the same benchmark. Comparative analyses showed that the sequence-first strategy consistently performed better than the conventional segment-first strategy and that ProtT5 provided more informative representations than several other widely used protein language models. The framework also maintained stable performance under class-imbalance settings.

KmalPred provides an effective and robust approach for large-scale screening of candidate Kmal sites. More broadly, the results highlight the value of full-sequence protein language model representations for residue-level post-translational modification prediction and suggest that this framework can be readily extended to other post-translational modification site prediction tasks.