Deep learning-based contact prediction in C-I-TASSER; violin plots for portions of residues predicted by TMHMM2.0 to be within transmembrane helices for JCVI-syn3.0 proteins that are annotated versus unannotated by C-I-TASSER/ COFACTOR with C-score > 0.5 for specific GO terms in the MF, BP, and CC aspects; structural alignment of the predicted structure of MMSYN1_0877 with the nine closest structural homologues identified in the Protein Data Bank; substrate binding domains for ECF systems targeting riboflavin; and a random PPI network for syn3.0, where 2483 of all 95,703 protein pairs are randomly selected as the positive PPI pairs (PDF)

protein methylation plays a vital role in cell processing. Many novel methods try to predict methylation sites from protein sequence by sequence information or predicted structural information, but none of them use protein tertiary structure information in prediction. In particular, most of them do not build models for predicting methylation types (mono-, di-, tri-methylation). To address these problems, we propose a novel method, Met-predictor, to predict methylation sites and methylation types using a support vector machine-based network. Met-predictor combines a variety of sequence-based features that are derived from protein sequences with structure model-based features, which are geometric information extracted from predicted protein tertiary structure models, and are firstly used in methylation prediction. Met-predictor was tested on two independent test sets, where the addition of structure model-based features improved AUC from 0.611 and 0.520 to 0.655 and 0.566 for lysine and from 0.723 and 0.640 to 0.734 and 0.643 for arginine. When compared with other state-of-the-art methods, Met-predictor had 13.1% (3.9%) and 8.5% (16.4%) higher accuracy than the best of other methods for methyllysine and methylarginine prediction on the independent test set I (II). Furthermore, Met-predictor also attains excellent performance for predicting methylation types.