Training PBertKla on an Integrated Multi-Source Dataset with a Machine-Learning Layer for Lysine Lactylation Site...

Area of Science:

• Biochemistry
• Computational Biology
• Genomics

Background:

• Lysine lactylation (Kla) is a critical post-translational modification impacting cellular processes like metabolism and reprogramming.
• Understanding Kla site regulation is vital for disease research.

Purpose of the Study:

• To develop and validate an accurate computational model for predicting lysine lactylation (Kla) sites.
• To create and release the largest curated Kla dataset for community use.

Main Methods:

• Trained an enhanced ProteinBERT-based predictor (PBertKla + ML) using an integrated multi-source dataset.
• Assembled and quality-controlled a large Kla dataset (Multi) from nine sources.
• Validated model generalization using leakage-controlled splits (protein-level, homology, leave-one-study-out).

Main Results:

• PBertKla + ML achieved an AUROC of 0.9126 on the integrated dataset, comparable to state-of-the-art tools.
• The model demonstrated strong generalization, with AUROC ≈0.90 under stringent leakage-controlled conditions.
• Ablation and SHAP analyses confirmed the primary contribution of ProteinBERT features, with a modest gain from the ML layer.

Conclusions:

• The PBertKla + ML model is a highly accurate and generalizable tool for Kla site prediction.
• The released Multi dataset serves as a valuable community resource for Kla research.
• The study provides insights into the predictive signal for Kla sites and their potential structural distribution.