Improving the prediction of disulfide bonds in Eukaryotes with machine learning methods and protein subcellular localization

Abstract

piero.fariselli@unibo.it.

Disulfide Connectivity Prediction Based on Modelled Protein 3D Structural Information and Random Forest Regression

“…The second dataset, PDBCYS-R, was constructed from a recently released dataset, i.e., PDBCYS [40], for disulfide bond predictions. The original PDBCYS dataset consists of 1,797 protein sequences, the maximal pairwise sequence identity of which was reduced to 25 percent.…”

“…However, until now, predicting a complete 3D structure directly from the sequence has been far from successful and is still a challenging and open problem. In this regard, researchers have resorted to decomposing the prediction of complete 3D structure into predictions of special structure segments or characteristics, such as disordered regions [8], [9], [10], [11], transmembrane helices [12], [13], beta-sheets [14], [15], [16], residue-residue contact maps [17], [18], [19], [20], disulfide connectivity [21], [22], [23], [24], [25], [26], solvent accessibility [27], [28], and so on. The knowledge derived from the protein structure segments or characteristics may provide valuable insights into protein 3D structures and may help to understand protein functions [23].…”

“…Predicting which cysteines in a protein sequence will form disulfide connections plays a relevant role in protein structural and functional annotation. Various effective methods have been developed in previous studies, such as DISULFIND [32], Pair-wise SVM [33], GASVM [21], SS_SVR [22] and FS_SVR [23], DBCP [24], DISLOCATE [25], DMC (DISLOCATEþMIpþiCOV) [26], and so on. All these existing methods can be grouped into three categories [25]: (I) disulfide-bonding state prediction; (II) connectivity pattern prediction using the prior knowledge of the disulfide-bonding states of cysteines; and (III) methods that predict both disulfide-bonding state and connectivity patterns.…”

“…Recently, considerable attention has been paid to developing machine learning-based methods for disulfide connectivity predictions, and experimental results have shown that applying advanced machine learning algorithms is a promising route to further improve the prediction performance [21], [22], [23], [24], [25], [26], [33]. Extracting discriminative features and utilizing powerful machine learning algorithms are two crucial aspects that will significantly affect the performance of machine learning-based disulfide connectivity prediction.…”

“…Extracting discriminative features and utilizing powerful machine learning algorithms are two crucial aspects that will significantly affect the performance of machine learning-based disulfide connectivity prediction. Tremendous efforts have been made to uncover effective features and many feature sources, such as position specific scoring matrix (PSSM) [23], [34], predicted secondary structure (PSS) [34], correlated mutations (CM) [26], [35], cysteine separation distances [33], and subcellular localizations [25], were proven to be useful for disulfide connectivity prediction. As for classifiers, many advanced machine learning algorithms, e.g., support vector regression (SVR) [23], [34] and grammaticalrestrained hidden conditional random fields (GRHCRFs) [25], have been widely applied to disulfide connectivity prediction.…”