Discrimination between Distant Homologs and Structural Analogs: Lessons from Manually Constructed, Reliable Data Sets

Abstract

A natural way to study protein sequence, structure, and function is to put them in the context of evolution. Homologs inherit similarities from their common ancestor, while analogs converge to similar structures due to a limited number of energetically favorable ways to pack secondary structural elements. Using novel strategies, we previously assembled two reliable databases of homologs and analogs. In this study, we compare these two data sets and develop a support vector machine (SVM)-based classifier to discriminate between homologs and analogs. The classifier uses a number of well-known similarity scores. We observe that although both structure scores and sequence scores contribute to SVM performance, profile sequence scores computed based on structural alignments are the best discriminators between remote homologs and structural analogs. We apply our classifier to a representative set from the expert-constructed database, Structural Classification of Proteins (SCOP). The SVM classifier recovers 76% of the remote homologs defined as domains in the same SCOP superfamily but from different families. More importantly, we also detect and discuss interesting homologous relationships between SCOP domains from different superfamilies, folds, and even classes.

Introduction

Three-dimensional structural similarities among proteins are explained by either divergence or convergence. In divergent evolution, homologs inherit similar structures from their common ancestor. In convergent evolution, proteins from distinct evolutionary lineages arrive independently at similar structures due to a limited number of energetically favorable ways to pack secondary structural elements (SSEs),1, 2, 3 and such proteins are called analogs. Judging if two structurally similar proteins are homologous or analogous remains a difficult task. Statistically significant sequence similarity, as detected by sequence search tools such as PSI-BLAST,⁴ is generally accepted as adequate evidence for homology.5, 6 In the absence of significant sequence similarity, remote homology inference can be based on overall structural similarity, augmented by other properties such as similar arrangements of functional residues, common ligand-binding modes, shared unusual structural features, and similar domain organizations.7, 8 However, capturing these features often requires visual inspection by human experts and is more in the realm of art than science.

The Structural Classification of Proteins (SCOP) database⁹ represents a comprehensive collection of manually curated homologous superfamilies of protein domains with known structures. In the SCOP hierarchy, domains with significant sequence similarity or overwhelming structural and functional similarity (close homologs) are grouped into the same family; families with convincing structural and/or functional evidence for common ancestry are grouped into the same superfamily; superfamilies with the same overall three-dimensional structure and topology but without very strong evidence for homology are grouped into the same fold; and folds are grouped into classes based on their SSE compositions. SCOP is manually maintained by human experts, and its superfamily level is regarded as the most reliable standard for remote homologs.¹⁰

Several efforts have been made to discern the boundary between homology and analogy in an automated and quantitative way. Russell et al.¹¹ statistically analyzed structurally aligned homologous and analogous pairs and found that homologs generally retain higher sequence identity, more conserved SSEs, and solvent accessibility compared to analogs. Matsuo and Bryant¹² defined a homologous core structure representing the consensus substructure in a protein family, and used the overlap of homologous core structure to distinguish homologs and analogs. Dietmann and Holm¹³ trained a neural network to discriminate homologs and analogs based on sequence, structure, and functional similarities. All three studies used domains in the same SCOP superfamily as homologs and domains in different SCOP superfamilies as analogs in their analysis. Given the conservative nature of the SCOP hierarchy, a potential flaw of this approach is the contamination of the analog data set by homologs. Domains in different superfamilies are not necessarily analogs and may in fact be homologous when new evidence emerges.⁹ For instance, through careful analysis, Ponting and Russell¹⁴ suggested that at least five SCOP superfamilies under the β-trefoil fold were actually homologous and had descended from a common ancestor.

To avoid the aforementioned ambiguity, we approach the problem of discriminating between homologs and analogs with more clear-cut and reliable data sets. Previously, we manually constructed a homolog database (MALIDUP¹⁵) composed of duplicated domain pairs and an analog database (MALISAM¹⁶) composed of three categories of analogous pairs (a hybrid motif and a core motif, an interface motif and a core motif, and an artificial protein and a natural protein). Each pair in MALIDUP or MALISAM is carefully inspected to convincingly support homology or analogy and then manually superimposed and aligned to ensure good alignment quality. In this study, we use pairs from these two databases as reliable homologs and analogs to understand the differences as well as to develop a discriminator between homology and structural analogy.

We first characterize and compare the MALIDUP and MALISAM pairs in terms of structure, sequence, and profile scores. Combining these scores, we train support vector machines (SVMs) to discriminate between the homologs in MALIDUP and the analogs in MALISAM. Since MALIDUP and MALISAM are quite small in size and may not be representative of the total protein variety found in nature, we test the resulting SVM-based classifier on the comprehensive SCOP database. We show that although the classifier is trained on the manually constructed data sets with particular statistical properties, it can recover the majority of distant homologs classified in the same SCOP superfamily but different families. Moreover, the classifier is capable of finding more distantly related pairs between SCOP superfamilies, folds, and classes. We discuss some of these interesting pairs and argue that many of them indeed represent remote homologs.