Prediction of Lysine Acetylation Sites in Escherichia Coli Using DenseNet and Attention Mechanisms

Abstract

Lysine acetylation is an important post-translational modification that plays a critical role in regulating protein conformational stability, catalytic activity, and cellular metabolism. Owing to the cost and turnaround time of mass spectrometry–based assays, developing efficient and generalizable computational predictors is of practical significance. Using publicly available Escherichia coli protein sequences, we design a lightweight deep learning model composed of a DenseNet convolutional backbone and a CBAM attention module to identify acetylation sites. Each input peptide is encoded by a 23-channel representation that concatenates 20-amino-acid one-hot vectors with three physicochemical indicators (polarity, hydrophobicity, and charge). During training, predicted probabilities are calibrated on the validation set via temperature scaling, and a decision threshold is selected under a specificity constraint of Sp ≈ 0.900; the calibrated temperature and threshold are then applied to the test set for evaluation. On the E. coli test set, the proposed method achieves Acc = 0.645, Sn = 0.363, Sp = 0.901, Pre = 0.802, F1 = 0.528, and AUC = 0.762. These results indicate that, under a high-specificity operating regime, our model attains overall discrimination comparable to a strong baseline without relying on transfer learning, while delivering higher precision and stability at the selected operating point.