Global mapping of the protein structure space and application in structure-based inference of protein function

Abstract

We have constructed a map of the ''protein structure space'' by using the pairwise structural similarity scores calculated for all nonredundant protein structures determined experimentally. As expected, proteins with similar structures clustered together in the map and the overall distribution of structural classes of this map followed closely that of the map of the ''protein fold space'' we have reported previously. Consequently, proteins sharing similar molecular functions also were found to colocalize in the protein structure space map, pointing toward a previously undescribed scheme for structure-based functional inference for remote homologues based on the proximity in the map of the protein structure space. We found that this scheme consistently outperformed other predictions made by using either the raw scores or normalized Z-scores of pairwise DALI structure alignment. global map of protein universe ͉ multivariate analysis ͉ protein function prediction ͉ protein structure universe T he molecular functions of a protein can be inferred from either its sequence or structure information. Sequence-based function inference methods annotate molecular function of a protein from its sequence homologues. Most genome-wide functional annotations are carried out with this scheme, by using sequence alignment tools such as BLAST (1), or motif͞profile-based search tools such as PROSITE (2, 3) and PFAM (4, 5). However, when two functionally similar proteins do not share detectable sequence homology, molecular function cannot be inferred based solely on sequence information. Low sequence homology results either from an early branching point at the protein evolution (also known as remote homologues) or a convergent evolution. Many studies were focused on the detection of remote homologues (6-8). In general, methods using statistical models extracted from multiply aligned sequences perform better than pairwise sequence comparison methods (9). However, even these improved methods fail to recognize remote homologues with sequence identity Ͻ25-30%, which is estimated to be Ͼ25% of all sequenced proteins.Structure-based function inference, however, depends less on sequence information. During protein evolution, homology on sequence level is far less preserved compared with homology on structure level. Because proteins fold into specific structures to perform their molecular functions, structure-based functional inference is able to characterize remote homologous relationships of proteins that are impossible to detect by using sequences. By using different random sampling methods and similarity measuring functions, a large number of structural alignment algorithms have been developed to measure similarity of a pair of protein structures. Among these algorithms, DALI (10), SSAP (11), CE (12), and VAST (13) have been widely used, and their performances have been assessed [see Koehl (14) for a review].The issue of predicting the function of remote homologues has become more prominent recently: the Structural Genomics initiative (15...

Automatic classification of protein structures using low-dimensional structure space mappings

“…Based on Table 5 and Table 6, it appears this model has little impact on classification of remotely homologous structures of shared function. It has been argued previously that the down-weighting of alignments between large structures impairs the ability of Dali Z-scores to capture functional relationships implied by local similarity between such structures [12]. However, we find only that these Z-scores offer little improvement.…”

“…The observation that MDS may confer an advantage over pairwise alignment distances for annotation inference was made previously in the context of Gene Ontology [12] and was recently corroborated by us in experiments in automatic classification of remotely homologous relationships from CATH and SCOP [13]. In that work we examined the influence of certain critical meta-parameters on construction of MPSS using MDS, including choice of the underlying protein alignment method and particular implementation of MDS.…”

A galaxy of folds

“…[25][26][27] Some studies also considered structure-based function inference 28 and the fold usage of organisms. 27 All these studies used structural similarity to connect different folds, a property that primarily reflects analogy.…”

Solving Nuclear Norm Regularized and Semidefinite Matrix Least Squares Problems with Linear Equality Constraints

“…We also conducted numerical experiments on a much larger protein data set to evaluate the performance of our algorithm. We used the PDB SELECT 25 data set [13], a representative subset of the Protein Data Bank database, which contains 1898 protein chains. The numerical results for the PDB SELECT 25 data set are reported in Table 2.…”