Strategies for high-throughput comparative modeling: applications to leverage analysis in structural genomics and protein family organization

Structure-Based Approaches for Protein-Protein Interaction Prediction Using Machine Learning and Deep Learning

Protein-Protein Interaction (PPI) prediction plays a pivotal role in understanding cellular processes and uncovering molecular mechanisms underlying health and disease. Structure-based PPI prediction has emerged as a robust alternative to sequence-based methods, offering greater biological accuracy by integrating three-dimensional spatial and biochemical features. This work summarizes the recent advances in computational approaches leveraging protein structure information for PPI prediction, focusing on machine learning (ML) and deep learning (DL) techniques. These methods not only improve predictive accuracy but also provide insights into functional sites, such as binding and catalytic residues. However, challenges such as limited high-resolution structural data and the need for effective negative sampling persist. Through the integration of experimental and computational tools, structure-based prediction paves the way for comprehensive proteomic network analysis, holding promise for advancements in drug discovery, biomarker identification, and personalized medicine. Future directions include enhancing scalability and dataset reliability to expand these approaches across diverse proteomes.

Recent advances in predicting protein-protein interactions with the aid of artificial intelligence algorithms

Protein-protein interactions (PPIs) are essential in the regulation of biological functions and cell events, therefore understanding PPIs have become a key issue to understanding the molecular mechanism and investigating the design of drugs. Here we highlight the major developments in computational methods developed for predicting PPIs by using types of artificial intelligence algorithms. The first part introduces the source of experimental PPI data. The second part is devoted to the PPI prediction methods based on sequential information. The third part covers representative methods using structural information as the input feature. The last part is methods designed by combining different types of features. For each part, the state-of-the-art computational PPI prediction methods are reviewed in an inclusive view. Finally, we discuss the flaws existing in this area and future directions of next-generation algorithms.

Predicting peptide-mediated interactions on a genome-wide scale

We describe a method to predict protein-protein interactions (PPIs) formed between structured domains and short peptide motifs. We take an integrative approach based on consensus patterns of known motifs in databases, structures of domain-motif complexes from the PDB and various sources of non-structural evidence. We combine this set of clues using a Bayesian classifier that reports the likelihood of an interaction and obtain significantly improved prediction performance when compared to individual sources of evidence and to previously reported algorithms. Our Bayesian approach was integrated into PrePPI, a structure-based PPI prediction method that, so far, has been limited to interactions formed between two structured domains. Around 80,000 new domain-motif mediated interactions were predicted, thus enhancing PrePPI’s coverage of the human protein interactome.

Complexes formed between a structured domain on one protein and an unstructured peptide on another are ubiquitous. However, they are often quite difficult to detect experimentally. The development of computational approaches to predict domain-motif interactions is therefore an important goal. We report a method to predict domain-motif interactions using a Bayesian approach to integrate evidence from a variety of sources, including three-dimensional structural and non-structural information. The method was applied to the entire human proteome and showed significant improvement over existing methods. The method was incorporated into PrePPI, a computational pipeline for the prediction of protein-protein interactions that relies heavily on structural information. Approximately 80,000 new interactions were detected. The new PrePPI database provides easy access to about 400,000 human protein-protein interactions and should thus constitute a valuable resource in a variety of biological applications including the characterization of molecular interaction networks and, more generally, in the study of interactions mediated by proteins in families that may not be extensively studied experimentally.

Global approaches for the elucidation of phosphoinositide-binding proteins

Phosphoinositide lipids (PIPns) control numerous critical biological pathways, typically through the regulation of protein function driven by non-covalent protein-lipid binding interactions. Despite the importance of these systems, the unraveling of the full scope of protein-PIPn interactions has represented a significant challenge due to the massive complexity associated with these events, including the large number of diverse proteins that bind to these lipids, variations in the mechanisms by which proteins bind to lipids, and the presence of multiple distinct PIPn isomers. As a result of this complexity, global methods in which numerous proteins that bind PIPns can be identified and characterized simultaneously from complex samples, which have been enabled by key technological advancements, have become popular as an efficient means for tackling this challenge. This review article provides an overview of advancements in large-scale methods for profiling protein-PIPn binding, including experimental methods, such as affinity enrichment, microarray analysis and activity-based protein profiling, as well as computational methods, and combined computational/experimental efforts.

PrePPI: a structure-informed database of protein-protein interactions

PrePPI (http://bhapp.c2b2.columbia.edu/PrePPI) is a database that combines predicted and experimentally determined protein-protein interactions (PPIs) using a Bayesian framework. Predicted interactions are assigned probabilities of being correct, which are derived from calculated likelihood ratios (LRs) by combining structural, functional, evolutionary and expression information, with the most important contribution coming from structure. Experimentally determined interactions are compiled from a set of public databases that manually collect PPIs from the literature and are also assigned LRs. A final probability is then assigned to every interaction by combining the LRs for both predicted and experimentally determined interactions. The current version of PrePPI contains ∼2 million PPIs that have a probability more than ∼0.1 of which ∼60 000 PPIs for yeast and ∼370 000 PPIs for human are considered high confidence (probability > 0.5). The PrePPI database constitutes an integrated resource that enables users to examine aggregate information on PPIs, including both known and potentially novel interactions, and that provides structural models for many of the PPIs.

Structure-based prediction of protein-protein interactions on a genome-wide scale

The genome-wide identification of pairs of interacting proteins is an important step in the elucidation of cell regulatory mechanisms^1,2. Much of our current knowledge derives from high-throughput techniques such as yeast two hybrid and affinity purification³, as well as from manual curation of experiments on individual systems⁴. A variety of computational approaches based, for example, on sequence homology, gene co-expression, and phylogenetic profiles have also been developed for the genome-wide inference of protein-protein interactions (PPIs)^5,6. Yet, comparative studies suggest that the development of accurate and complete repertoires of PPIs is still in its early stages^7–9. Here we show that three-dimensional structural information can be used to predict PPIs with an accuracy and coverage that are superior to predictions based on non-structural evidence. Moreover, an algorithm, PrePPI, that combines structural information with other functional clues is comparable in accuracy to high-throughput experiments, yielding over 30,000 high confidence interactions for yeast and over 300,000 for human. Experimental tests of a number of predictions demonstrate the ability of the PrePPI algorithm to identify unexpected PPIs of significant biological interest. The surprising effectiveness of three-dimensional structural information can be attributed to the use of homology models combined with the exploitation of both close and remote geometric relationships between proteins.

The Protein Structure Initiative: achievements and visions for the future

The Protein Structure Initiative (PSI) was established in 2000 by the National Institutes of General Medical Sciences with the long-term goal of providing 3D (three-dimensional) structural information for most proteins in nature. As advances in genomic sequencing, bioinformatics, homology modelling, and methods for rapid determination of 3D structures of proteins by X-ray crystallography and nuclear magnetic resonance (NMR) converged, it was proposed that our understanding of the biology of protein structure and evolution could be greatly enabled by ‘genomic-scale’ protein structure determination. Over the past 12 years, the PSI has evolved from a testing bed for new methods of sample and structure production to a core component of a wide range of biology programs.

Using structure to explore the sequence alignment space of remote homologs

Protein structure modeling by homology requires an accurate sequence alignment between the query protein and its structural template. However, sequence alignment methods based on dynamic programming (DP) are typically unable to generate accurate alignments for remote sequence homologs, thus limiting the applicability of modeling methods. A central problem is that the alignment that is “optimal” in terms of the DP score does not necessarily correspond to the alignment that produces the most accurate structural model. That is, the correct alignment based on structural superposition will generally have a lower score than the optimal alignment obtained from sequence. Variations of the DP algorithm have been developed that generate alternative alignments that are “suboptimal” in terms of the DP score, but these still encounter difficulties in detecting the correct structural alignment. We present here a new alternative sequence alignment method that relies heavily on the structure of the template. By initially aligning the query sequence to individual fragments in secondary structure elements and combining high-scoring fragments that pass basic tests for “modelability”, we can generate accurate alignments within a small ensemble. Our results suggest that the set of sequences that can currently be modeled by homology can be greatly extended.

It has been suggested that, for nearly every protein sequence, there is already a protein with a similar structure in current protein structure databases. However, with poor or undetectable sequence relationships, it is expected that accurate alignments and models cannot be generated. Here we show that this is not the case, and that whenever structural relationship exists, there are usually local sequence relationships that can be used to generate an accurate alignment, no matter what the global sequence identity. However, this requires an alternative to the traditional dynamic programming algorithm and the consideration of a small ensemble of alignments. We present an algorithm, S4, and demonstrate that it is capable of generating accurate alignments in nearly all cases where a structural relationship exists between two proteins. Our results thus constitute an important advance in the full exploitation of the information in structural databases. That is, the expectation of an accurate alignment suggests that a meaningful model can be generated for nearly every sequence for which a suitable template exists.

Genome-wide structural analysis reveals novel membrane binding properties of AP180 N-terminal homology (ANTH) domains

An increasing number of cytosolic proteins are shown to interact with membrane lipids during diverse cellular processes, but computational prediction of these proteins and their membrane binding behaviors remains challenging. Here, we introduce a new combinatorial computation protocol for systematic and robust functional prediction of membrane-binding proteins through high throughput homology modeling and in-depth calculation of biophysical properties. The approach was applied to the genomic scale identification of the AP180 N-terminal homology (ANTH) domain, one of the modular lipid binding domains, and prediction of their membrane binding properties. Our analysis yielded comprehensive coverage of the ANTH domain family and allowed classification and functional annotation of proteins based on the differences in local structural and biophysical features. Our analysis also identified a group of plant ANTH domains with unique structural features that may confer novel functionalities. Experimental characterization of a representative member of this subfamily confirmed its unique membrane binding mechanism and unprecedented membrane deforming activity. Collectively, these studies suggest that our new computational approach can be applied to genome-wide functional prediction of other lipid binding domains.

Internal organization of large protein families: relationship between the sequence, structure, and function-based clustering

The protein universe can be organized in families that group proteins sharing common ancestry. Such families display variable levels of structural and functional divergence, from homogenous families, where all members have the same function and very similar structure, to very divergent families, where large variations in function and structure are observed. For practical purposes of structure and function prediction, it would be beneficial to identify sub-groups of proteins with highly similar structures (iso-structural) and/or functions (iso-functional) within divergent protein families. We compared three algorithms in their ability to cluster large protein families and discuss whether any of these methods could reliably identify such iso-structural or iso-functional groups. We show that clustering using profile-sequence and profile-profile comparison methods closely reproduces clusters based on similarities between 3D structures or clusters of proteins with similar biological functions. In contrast, the still commonly used sequence-based methods with fixed thresholds result in vast overestimates of structural and functional diversity in protein families. As a result, these methods also overestimate the number of protein structures that have to be determined to fully characterize structural space of such families. The fact that one can build reliable models based on apparently distantly related templates is crucial for extracting maximal amount of information from new sequencing projects.

Toward prediction of functional protein pockets using blind docking and pocket search algorithms

Location of functional binding pockets of bioactive ligands on protein molecules is essential in structural genomics and drug design projects. If the experimental determination of ligand-protein complex structures is complicated, blind docking (BD) and pocket search (PS) calculations can help in the prediction of atomic resolution binding mode and the location of the pocket of a ligand on the entire protein surface. Whereas the number of successful predictions by these methods is increasing even for the complicated cases of exosites or allosteric binding sites, their reliability has not been fully established. For a critical assessment of reliability, we use a set of ligand-protein complexes, which were found to be problematic in previous studies. The robustness of BD and PS methods is addressed in terms of success of the selection of truly functional pockets from among the many putative ones identified on the surfaces of ligand-bound and ligand-free (holo and apo) protein forms. Issues related to BD such as effect of hydration, existence of multiple pockets, and competition of subsidiary ligands are considered. Practical cases of PS are discussed, categorized and strategies are recommended for handling the different situations. PS can be used in conjunction with BD, as we find that a consensus approach combining the techniques improves predictive power.

Identification and structural characterization of FYVE domain-containing proteins of Arabidopsis thaliana

BACKGROUND

FYVE domains have emerged as membrane-targeting domains highly specific for phosphatidylinositol 3-phosphate (PtdIns(3)P). They are predominantly found in proteins involved in various trafficking pathways. Although FYVE domains may function as individual modules, dimers or in partnership with other proteins, structurally, all FYVE domains share a fold comprising two small characteristic double-stranded beta-sheets, and a C-terminal alpha-helix, which houses eight conserved Zn2+ ion-binding cysteines. To date, the structural, biochemical, and biophysical mechanisms for subcellular targeting of FYVE domains for proteins from various model organisms have been worked out but plant FYVE domains remain noticeably under-investigated.

RESULTS

We carried out an extensive examination of all Arabidopsis FYVE domains, including their identification, classification, molecular modeling and biophysical characterization using computational approaches. Our classification of fifteen Arabidopsis FYVE proteins at the outset reveals unique domain architectures for FYVE containing proteins, which are not paralleled in other organisms. Detailed sequence analysis and biophysical characterization of the structural models are used to predict membrane interaction mechanisms previously described for other FYVE domains and their subtle variations as well as novel mechanisms that seem to be specific to plants.

CONCLUSIONS

Our study contributes to the understanding of the molecular basis of FYVE-based membrane targeting in plants on a genomic scale. The results show that FYVE domain containing proteins in plants have evolved to incorporate significant differences from those in other organisms implying that they play a unique role in plant signaling pathways and/or play similar/parallel roles in signaling to other organisms but use different protein players/signaling mechanisms.

High-throughput computational structure-based characterization of protein families: START domains and implications for structural genomics

SkyLine, a high-throughput homology modeling pipeline tool, detects and models true sequence homologs to a given protein structure. Structures and models are stored in SkyBase with links to computational function annotation, as calculated by MarkUs. The SkyLine/SkyBase/MarkUs technology represents a novel structure-based approach that is more objective and versatile than other protein classification resources. This structure-centric strategy provides a multi-dimensional organization and coverage of protein space at the levels of family, function, and genome. The concept of "modelability", the ability to model sequences on related structures, provides a reliable criterion for membership in a protein family ("leverage") and underlies the unique success of this approach. The overall procedure is illustrated by its application to START domains, which comprise a Biomedical Theme for the Northeast Structural Genomics Consortium as part of the Protein Structure Initiative. START domains are typically involved in the non-vesicular transport of lipids. While 19 experimentally determined structures are available, the family, whose evolutionary hierarchy is not well determined, is highly sequence diverse, and the ligand-binding potential of many family members is unknown. The SkyLine/SkyBase/MarkUs approach provides significant insights and predicts: (1) many more family members (approximately 4,000) than any other resource; (2) the function for a large number of unannotated proteins; (3) instances of START domains in genomes from which they were thought to be absent; and (4) the existence of two types of novel proteins, those containing dual START domain and those containing N-terminal START domains.

Structure and function of Pseudomonas aeruginosa protein PA1324 (21-170)

Pseudomonas aeruginosa is the prototypical biofilm-forming gram-negative opportunistic human pathogen. P. aeruginosa is causatively associated with nosocomial infections and with cystic fibrosis. Antibiotic resistance in some strains adds to the inherent difficulties that result from biofilm formation when treating P. aeruginosa infections. Transcriptional profiling studies suggest widespread changes in the proteome during quorum sensing and biofilm development. Many of the proteins found to be upregulated during these processes are poorly characterized from a functional standpoint. Here, we report the solution NMR structure of PA1324, a protein of unknown function identified in these studies, and provide a putative biological functional assignment based on the observed prealbumin-like fold and FAST-NMR ligand screening studies. PA1324 is postulated to be involved in the binding and transport of sugars or polysaccharides associated with the peptidoglycan matrix during biofilm formation.

Unique opportunities for NMR methods in structural genomics

This Perspective, arising from a workshop held in July 2008 in Buffalo NY, provides an overview of the role NMR has played in the United States Protein Structure Initiative (PSI), and a vision of how NMR will contribute to the forthcoming PSI-Biology program. NMR has contributed in key ways to structure production by the PSI, and new methods have been developed which are impacting the broader protein NMR community.

The Protein Model Portal

Structural Genomics has been successful in determining the structures of many unique proteins in a high throughput manner. Still, the number of known protein sequences is much larger than the number of experimentally solved protein structures. Homology (or comparative) modeling methods make use of experimental protein structures to build models for evolutionary related proteins. Thereby, experimental structure determination efforts and homology modeling complement each other in the exploration of the protein structure space. One of the challenges in using model information effectively has been to access all models available for a specific protein in heterogeneous formats at different sites using various incompatible accession code systems. Often, structure models for hundreds of proteins can be derived from a given experimentally determined structure, using a variety of established methods. This has been done by all of the PSI centers, and by various independent modeling groups. The goal of the Protein Model Portal (PMP) is to provide a single portal which gives access to the various models that can be leveraged from PSI targets and other experimental protein structures. A single interface allows all existing pre-computed models across these various sites to be queried simultaneously, and provides links to interactive services for template selection, target-template alignment, model building, and quality assessment. The current release of the portal consists of 7.6 million model structures provided by different partner resources (CSMP, JCSG, MCSG, NESG, NYSGXRC, JCMM, ModBase, SWISS-MODEL Repository). The PMP is available at http://www.proteinmodelportal.org and from the PSI Structural Genomics Knowledgebase.

Functional genomics and structural biology in the definition of gene function

By mid-2007, the three-dimensional (3D) structures of some 45,000 proteins have been solved, over a period where the linear structures of millions of genes have been defined. Technical challenges associated with X-ray crystallography are being overcome and high-throughput methods both for crystallization of proteins and for solving their 3D structures are under development. The question arises as to how structural biology can be integrated with and adds value to functional genomics programs. Structural biology will assist in the definition of gene function through the identification of the likely function of the protein products of genes. The 3D information allows protein sequences predicted from DNA sequences to be classified into broad groups, according to the overall 'fold', or 3D shape, of the protein. Structural information can be used to predict the preferred substrate of a protein, and thereby greatly enhance the accurate annotation of the corresponding gene. Furthermore, it will enable the effects of amino acid substitutions in enzymes to be better understood with respect to enzyme function and could thereby provide insights into natural variation in genes. If the molecular basis of transcription factor-DNA interactions were defined through precise 3D knowledge of the protein-DNA binding site, it would be possible to predict the effects of base substitutions within the motif on the specificity and/or kinetics of binding. In this chapter, we present specific examples of how structural biology can provide valuable information for functional genomics programs.