Prediction of Protein B-Factors Using Multi-Class Bounded SVM

Abstract

In this paper, we propose the adoption of the bounded support vector machine (BSVM) to predict the B-factors of residues based on a number of distinctive properties of residues. Due to the ability of multi-class classification of the BSVM, we can elaborately distinguish our targets and obtain relatively higher accuracy.

Predicting the Flexibility Profile of Ribosomal RNAs

“…Although the sequence-based approach has been successfully used in several previous works to rapidly predict the B-factors for proteins without known three-dimensional structures, [19,20,22] the atom's motion in crystal lattice is affected by not only sequentially adjacent residues but also spatially vicinal atoms/groups and thus the predictive accuracy for rRNA B-factors cannot be further improved unless the information about the local structure features of P atoms are considered in the modeling. As might be anticipated, the models on the basis of a structure-based approach, as shown in Table 3, perform much better than those obtained from a sequence-based strategy, given by a considerable increase in both the fitting ability for the training set and the predictive power for the test set.…”

“…This phenomenon could be attributed to the inevitable deviation of theoretical models from real cases, which has also been observed in previous studies of protein B-factors. [19,20] The crystal structure of the ribosome made up by a longchain 16S rRNA , a tRNA segment, and a series of 30S ribosomal proteins has been successfully solved at a high-resolution of 2.5 and a moderate R value of 0.257. [52] The 16S rRNA is composed of 1540 residues (5-1544), which was used here as an independent test sample to validate the reliability of our models.…”

“…[13][14][15][16][17] Previously, numerous works have been addressed on theoretical and computational studies of protein B-factors. [18][19][20][21][22][23][24][25][26] However, the methods used for modeling and predicting Bfactor values of atoms in RNA crystals, particularly in the complicated ribosome crystals, to our knowledge, still remain unexploited. As aforementioned, the folded rRNA is actually protein-like and hence could be treated by the methods used for proteins.…”

“…Several previous studies on protein B-factors have used a window size n of 9, [22] 13, [20] and 15. [19] Considering that the rRNAs are similar to proteins and the interplay between two residues in a RNA sequence is short-range and rapidly decreases with the increase of their distance, we separately used the 9, 11, 13 and 15 as n to test modeling performance on the basis of two nucleotide parameterization methods (vide post).…”

Prediction of Water′s Mobility and Disorder in Protein Crystals Using Novel Local Hydrophobic Descriptors

“…More commonly, one adopts a univariant Gaussian function, fully characterized by the isotropic mean-square displacement < u k ·u k > s k , or the B-factor B k = 8ps k /3. [7,8] Although a number of theoretical and computational methods have been proposed to model the flexibility and motion of protein polypeptide chains, such as Gaussian network model (GNM), [9,10] translation libration screw (TLS), [11] local density theory, [12] and machine learning-based approach, [13][14][15] the strategies used for predicting B-factor [12] B-factors of atoms in packing proteins are ultimately governed by local features of a highly complex energy landscape, and later, Weiss demonstrated that up to 50 % of the total atomic displacement variation in biomolecules can be successfully predicted solely based on the atomic coordinates and just three additional parameters per structure. [16] (ii) Water dynamic behavior in protein surface and interior is dominated by the hydrophobicity of local environment.…”