Protein language model trained to spot lactylated lysines in rice proteins

The work

Technical · from the record

The study builds PCBert-Kla, a classifier that decides whether a lysine site in a protein sequence is lactylated. It fine-tunes the ProtBert protein language model over 51-residue sequence windows, concatenates the resulting 1024-dimensional representation with a 27-dimensional vector of physicochemical descriptors computed in Biopython, and feeds the combined vector to an attention-weighted fully connected classifier. On an existing rice Kla benchmark the model reached an average accuracy of 0.9837 in five-fold cross-validation, and on the benchmark's independent external test set it reached an accuracy of 0.9497, compared with 0.9396 for the authors' ProtBert-only variant, 0.9322 for DeepKla and 0.6893 for Auto-Kla. Permutation of each physicochemical feature placed net charge highest in the resulting importance ranking, and attention maps over positive training samples showed weight concentrated on residues adjacent to each position.

How AI was used

A pretrained protein language model, Rostlab/prot_bert, supplied the sequence representation: only the first four encoder layers were retained, and those layers were fine-tuned together with the rest of the network rather than frozen, yielding a 1024-dimensional [CLS] vector per 51-residue window. Separately, Biopython routines computed a 27-dimensional physicochemical vector (molecular weight, isoelectric point, amino acid composition, three-dimensional secondary structure content, Kyte–Doolittle hydrophobicity and net charge), min–max normalised; the two vectors were concatenated into 1051 dimensions and passed to a classifier of two hidden layers (32 and 8 neurons) with a feature-weighting attention mechanism in the first hidden layer, ReLU activations, dropout of 0.1 and 0.3, and a sigmoid output. Training used the published Lv et al. benchmark unchanged, 30 epochs, batch size 4, learning rate 0.003, stochastic gradient descent, binary cross-entropy loss, five-fold cross-validation and early stopping on validation accuracy, on an NVIDIA GeForce RTX 4090. The trained model was then run over the benchmark's independent external test set, with ablation variants (no physicochemical features, no attention, no fine-tuning, ESM2 in place of ProtBert) trained under the same conditions and DeepKla and Auto-Kla run through their public web services for comparison. Interpretation used permutation of each physicochemical feature column with the resulting AUC change summed across folds, and average ProtBert attention maps over positive training samples. The model is also served through a public web interface.