Recent developments of sequence-based prediction of protein–protein interactions

The identification of protein–protein interactions (PPIs) can lead to a better understanding of cellular functions and biological processes of proteins and contribute to the design of drugs to target disease-causing PPIs. In addition, targeting host–pathogen PPIs is useful for elucidating infection mechanisms. Although several experimental methods have been used to identify PPIs, these methods can yet to draw complete PPI networks. Hence, computational techniques are increasingly required for the prediction of potential PPIs, which have never been seen experimentally. Recent high-performance sequence-based methods have contributed to the construction of PPI networks and the elucidation of pathogenetic mechanisms in specific diseases. However, the usefulness of these methods depends on the quality and quantity of training data of PPIs. In this brief review, we introduce currently available PPI databases and recent sequence-based methods for predicting PPIs. Also, we discuss key issues in this field and present future perspectives of the sequence-based PPI predictions.     

Benchmarking AlphaFold for protein complex modeling reveals accuracy determinants

High‐resolution experimental structural determination of protein–protein interactions has led to valuable mechanistic insights, yet due to the massive number of interactions and experimental limitations there is a need for computational methods that can accurately model their structures. Here we explore the use of the recently developed deep learning method, AlphaFold, to predict structures of protein complexes from sequence. With a benchmark of 152 diverse heterodimeric protein complexes, multiple implementations and parameters of AlphaFold were tested for accuracy. Remarkably, many cases (43%) had near‐native models (medium or high critical assessment of predicted interactions accuracy) generated as top‐ranked predictions by AlphaFold, greatly surpassing the performance of unbound protein–protein docking (9% success rate for near‐native top‐ranked models), however AlphaFold modeling of antibody–antigen complexes within our set was unsuccessful. We identified sequence and structural features associated with lack of AlphaFold success, and we also investigated the impact of multiple sequence alignment input. Benchmarking of a multimer‐optimized version of AlphaFold (AlphaFold‐Multimer) with a set of recently released antibody–antigen structures confirmed a low rate of success for antibody–antigen complexes (11% success), and we found that T cell receptor–antigen complexes are likewise not accurately modeled by that algorithm, showing that adaptive immune recognition poses a challenge for the current AlphaFold algorithm and model. Overall, our study demonstrates that end‐to‐end deep learning can accurately model many transient protein complexes, and highlights areas of improvement for future developments to reliably model any protein–protein interaction of interest.     

Linking structural features of protein complexes and biological function

Protein–protein interaction (PPI) establishes the central basis for complex cellular networks in a biological cell. Association of proteins with other proteins occurs at varying affinities, yet with a high degree of specificity. PPIs lead to diverse functionality such as catalysis, regulation, signaling, immunity, and inhibition, playing a crucial role in functional genomics. The molecular principle of such interactions is often elusive in nature. Therefore, a comprehensive analysis of known protein complexes from the Protein Data Bank (PDB) is essential for the characterization of structural interface features to determine structure–function relationship. Thus, we analyzed a nonredundant dataset of 278 heterodimer protein complexes, categorized into major functional classes, for distinguishing features. Interestingly, our analysis has identified five key features (interface area, interface polar residue abundance, hydrogen bonds, solvation free energy gain from interface formation, and binding energy) that are discriminatory among the functional classes using Kruskal-Wallis rank sum test. Significant correlations between these PPI interface features amongst functional categories are also documented. Salt bridges correlate with interface area in regulator-inhibitors (r = 0.75). These representative features have implications for the prediction of potential function of novel protein complexes. The results provide molecular insights for better understanding of PPIs and their relation to biological functions.     

Frequent Assembly of Chimeric Complexes in the Protein Interaction Network of an Interspecies Yeast Hybrid

Hybrids between species often show extreme phenotypes, including some that take place at the molecular level. In this study, we investigated the phenotypes of an interspecies diploid hybrid in terms of protein–protein interactions inferred from protein correlation profiling. We used two yeast species, Saccharomyces cerevisiae and Saccharomyces uvarum, which are interfertile, but yet have proteins diverged enough to be differentiated using mass spectrometry. Most of the protein–protein interactions are similar between hybrid and parents, and are consistent with the assembly of chimeric complexes, which we validated using an orthogonal approach for the prefoldin complex. We also identified instances of altered protein–protein interactions in the hybrid, for instance, in complexes related to proteostasis and in mitochondrial protein complexes. Overall, this study uncovers the likely frequent occurrence of chimeric protein complexes with few exceptions, which may result from incompatibilities or imbalances between the parental proteomes.     

Match_Motif: A rapid computational tool to assist in protein–protein interaction design

In order to generate protein assemblies with a desired function, the rational design of protein–protein binding interfaces is of significant interest. Approaches based on random mutagenesis or directed evolution may involve complex experimental selection procedures. Also, molecular modeling approaches to design entirely new proteins and interactions with partner molecules can involve large computational efforts and screening steps. In order to simplify at least the initial effort for designing a putative binding interface between two proteins the Match_Motif approach has been developed. It employs the large collection of known protein–protein complex structures to suggest interface modifications that may lead to improved binding for a desired input interaction geometry. The approach extracts interaction motifs based on the backbone structure of short (four residues) segments and the relative arrangement with respect to short segments on the partner protein. The interaction geometry is used to search through a database of such motifs in known stable bound complexes. All matches are rapidly identified (within a few seconds) and collected and can be used to guide changes in the interface that may lead to improved binding. In the output, an alternative interface structure is also proposed based on the frequency of occurrence of side chains at a given interface position in all matches and based on sterical considerations. Applications of the procedure to known complex structures and alternative arrangements are presented and discussed. The program, data files, and example applications can be downloaded from https://www.groups.ph.tum.de/t38/downloads/.     

Two distinct classes of cochaperones compete for the EEVD motif in heat shock protein 70 to tune its chaperone activities

Chaperones of the heat shock protein 70 (Hsp70) family engage in protein–protein interactions with many cochaperones. One “hotspot” for cochaperone binding is the EEVD motif, found at the extreme C terminus of cytoplasmic Hsp70s. This motif is known to bind tetratricopeptide repeat domain cochaperones, such as the E3 ubiquitin ligase CHIP. In addition, the EEVD motif also interacts with a structurally distinct domain that is present in class B J-domain proteins, such as DnaJB4. These observations suggest that CHIP and DnaJB4 might compete for binding to Hsp70’s EEVD motif; however, the molecular determinants of such competition are not clear. Using a collection of EEVD-derived peptides, including mutations and truncations, we explored which residues are critical for binding to both CHIP and DnaJB4. These results revealed that some features, such as the C-terminal carboxylate, are important for both interactions. However, CHIP and DnaJB4 also had unique preferences, especially at the isoleucine position immediately adjacent to the EEVD. Finally, we show that competition between these cochaperones is important in vitro, as DnaJB4 limits the ubiquitination activity of the Hsp70–CHIP complex, whereas CHIP suppresses the client refolding activity of the Hsp70–DnaJB4 complex. Together, these data suggest that the EEVD motif has evolved to support diverse protein–protein interactions, such that competition between cochaperones may help guide whether Hsp70-bound proteins are folded or degraded.     

Characterization of interactions within the Igα/Igβ transmembrane domains of the human B-cell receptor provides insights into receptor assembly

The B-cell receptor (BCR), a complex comprised of a membrane-associated immunoglobulin and the Igα/β heterodimer, is one of the most important immune receptors in humans and controls B-cell development, activity, selection, and death. BCR signaling plays key roles in autoimmune diseases and lymphoproliferative disorders, yet, despite the clinical significance of this protein complex, key regions (i.e., the transmembrane domains) have yet to be structurally characterized. The mechanism for BCR signaling also remains unclear and has been variously described by the mutually exclusive cross-linking and dissociation activation models. Common to these models is the significance of local plasma membrane composition, which implies that interactions between BCR transmembrane domains (TMDs) play a role in receptor functionality. Here we used an in vivo assay of TMD oligomerization called GALLEX alongside spectroscopic and computational methods to characterize the structures and interactions of human Igα and Igβ TMDs in detergent micelles and natural membranes. We observed weak self-association of the Igβ TMD and strong self-association of the Igα TMD, which scanning mutagenesis revealed was entirely stabilized by an E–X₁₀–P motif. We also demonstrated strong heterotypic interactions between the Igα and Igβ TMDs both in vitro and in vivo, which scanning mutagenesis and computational models suggest is multiconfigurational but can accommodate distinct interaction sites for self-interactions and heterotypic interactions of the Igα TMD. Taken together, these results demonstrate that the TMDs of the human BCR are sites of strong protein–protein interactions that may direct BCR assembly, endoplasmic reticulum retention, and immune signaling.     

An ultra‐high affinity protein–protein interface displaying sequence‐robustness

Protein–protein interactions are crucial in biology and play roles in for example, the immune system, signaling pathways, and enzyme regulation. Ultra‐high affinity interactions (K _d <0.1 nM) occur in these systems, however, structures and energetics behind stability of ultra‐high affinity protein–protein complexes are not well understood. Regulation of the starch debranching barley limit dextrinase (LD) and its endogenous cereal type inhibitor (LDI) exemplifies an ultra‐high affinity complex (K _d of 42 pM). In this study the LD–LDI complex is investigated to unveil how robust the ultra‐high affinity is to LDI sequence variation at the protein–protein interface and whether alternative sequences can retain the ultra‐high binding affinity. The interface of LD–LDI was engineered using computational protein redesign aiming at identifying LDI variants predicted to retain ultra‐high binding affinity. These variants present a very diverse set of mutations going beyond conservative and alanine substitutions typically used to probe interfaces. Surface plasmon resonance analysis of the LDI variants revealed that high affinity of LD–LDI requires interactions of several residues at the rim of the protein interface, unlike the classical hotspot arrangement where key residues are found at the center of the interface. Notably, substitution of interface residues in LDI, including amino acids with functional groups different from the wild‐type, could occur without loss of affinity. This demonstrates that ultra‐high binding affinity can be conferred without hotspot residues, thus making complexes more robust to mutational drift in evolution. The present mutational analysis also demonstrates how energetic coupling can emerge between residues at large distances at the interface.     

Cold spots are universal in protein–protein interactions

Proteins interact with each other through binding interfaces that differ greatly in size and physico‐chemical properties. Within the binding interface, a few residues called hot spots contribute the majority of the binding free energy and are hence irreplaceable. In contrast, cold spots are occupied by suboptimal amino acids, providing possibility for affinity enhancement through mutations. In this study, we identify cold spots due to cavities and unfavorable charge interactions in multiple protein–protein interactions (PPIs). For our cold spot analysis, we first use a small affinity database of PPIs with known structures and affinities and then expand our search to nearly 4000 homo‐ and heterodimers in the Protein Data Bank (PDB). We observe that cold spots due to cavities are present in nearly all PPIs unrelated to their binding affinity, while unfavorable charge interactions are relatively rare. We also find that most cold spots are located in the periphery of the binding interface, with high‐affinity complexes showing fewer centrally located colds spots than low‐affinity complexes. A larger number of cold spots is also found in non‐cognate interactions compared to their cognate counterparts. Furthermore, our analysis reveals that cold spots are more frequent in homo‐dimeric complexes compared to hetero‐complexes, likely due to symmetry constraints imposed on sequences of homodimers. Finally, we find that glycines, glutamates, and arginines are the most frequent amino acids appearing at cold spot positions. Our analysis emphasizes the importance of cold spot positions to protein evolution and facilitates protein engineering studies directed at enhancing binding affinity and specificity in a wide range of applications.     

Accuracy modulating mutations of the ribosomal protein S4-S5 interface do not necessarily destabilize the rps4-rps5 protein–protein interaction

During the process of translation, an aminoacyl tRNA is selected in the A site of the decoding center of the small subunit based on the correct codon–anticodon base pairing. Though selection is usually accurate, mutations in the ribosomal RNA and proteins and the presence of some antibiotics like streptomycin alter translational accuracy. Recent crystallographic structures of the ribosome suggest that cognate tRNAs induce a “closed conformation” of the small subunit that stabilizes the codon–anticodon interactions at the A site. During formation of the closed conformation, the protein interface between rpS4 and rpS5 is broken while new contacts form with rpS12. Mutations in rpS12 confer streptomycin resistance or dependence and show a hyperaccurate phenotype. Mutations reversing streptomycin dependence affect rpS4 and rpS5. The canonical rpS4 and rpS5 streptomycin independent mutations increase translational errors and were called ribosomal ambiguity mutations (ram). The mutations in these proteins are proposed to affect formation of the closed complex by breaking the rpS4-rpS5 interface, which reduces the cost of domain closure and thus increases translational errors. We used a yeast two-hybrid system to study the interactions between the small subunit ribosomal proteins rpS4 and rpS5 and to test the effect of ram mutations on the stability of the interface. We found no correlation between ram phenotype and disruption of the interface.     

In Vitro Evolution Reveals Noncationic Protein–RNA Interaction Mediated by Metal Ions

RNA–peptide/protein interactions have been of utmost importance to life since its earliest forms, reaching even before the last universal common ancestor (LUCA). However, the ancient molecular mechanisms behind this key biological interaction remain enigmatic because extant RNA–protein interactions rely heavily on positively charged and aromatic amino acids that were absent (or heavily under-represented) in the early pre-LUCA evolutionary period. Here, an RNA-binding variant of the ribosomal uL11 C-terminal domain was selected from an approximately 10¹⁰ library of partially randomized sequences, all composed of ten prebiotically plausible canonical amino acids. The selected variant binds to the cognate RNA with a similar overall affinity although it is less structured in the unbound form than the wild-type protein domain. The variant complex association and dissociation are both slower than for the wild-type, implying different mechanistic processes involved. The profile of the wild-type and mutant complex stabilities along with molecular dynamics simulations uncovers qualitative differences in the interaction modes. In the absence of positively charged and aromatic residues, the mutant uL11 domain uses ion bridging (K⁺/Mg²⁺) interactions between the RNA sugar-phosphate backbone and glutamic acid residues as an alternative source of stabilization. This study presents experimental support to provide a new perspective on how early protein–RNA interactions evolved, where the lack of aromatic/basic residues may have been compensated by acidic residues plus metal ions.     

Protein co-evolution, co-adaptation and interactions

Co-evolution has an important function in the evolution of species and it is clearly manifested in certain scenarios such as host–parasite and predator–prey interactions, symbiosis and mutualism. The extrapolation of the concepts and methodologies developed for the study of species co-evolution at the molecular level has prompted the development of a variety of computational methods able to predict protein interactions through the characteristics of co-evolution. Particularly successful have been those methods that predict interactions at the genomic level based on the detection of pairs of protein families with similar evolutionary histories (similarity of phylogenetic trees: mirrortree). Future advances in this field will require a better understanding of the molecular basis of the co-evolution of protein families. Thus, it will be important to decipher the molecular mechanisms underlying the similarity observed in phylogenetic trees of interacting proteins, distinguishing direct specific molecular interactions from other general functional constraints. In particular, it will be important to separate the effects of physical interactions within protein complexes (‘co-adaptation') from other forces that, in a less specific way, can also create general patterns of co-evolution.     

By-passing selection: direct screening for antibody–antigen interactions using protein arrays

We have developed a system to identify highly specific antibody–antigen interactions by protein array screening. This removes the need for selection using animal immunisation or in vitro techniques such as phage or ribosome display. We screened an array of 27 648 human foetal brain proteins with 12 well-expressed antibody fragments that had not previously been exposed to any antigen. Four highly specific antibody–antigen pairs were identified, including three antibodies that bind proteins of unknown function. The target proteins were expressed at a very low copy number on the array, emphasising the unbiased nature of the screen. The specificity and sensitivity of binding demonstrates that this ‘naive’ screening approach could be applied to the high throughput isolation of specific antibodies against many different targets in the human proteome.     

Uncovering New Pathogen–Host Protein–Protein Interactions by Pairwise Structure Similarity

Pathogens usually evade and manipulate host-immune pathways through pathogen–host protein–protein interactions (PPIs) to avoid being killed by the host immune system. Therefore, uncovering pathogen–host PPIs is critical for determining the mechanisms underlying pathogen infection and survival. In this study, we developed a computational method, which we named pairwise structure similarity (PSS)-PPI, to predict pathogen–host PPIs. First, a high-quality and non-redundant structure–structure interaction (SSI) template library was constructed by exhaustively exploring heteromeric protein complex structures in the PDB database. New interactions were then predicted by searching for PSS with complex structures in the SSI template library. A quantitative score named the PSS score, which integrated structure similarity and residue–residue contact-coverage information, was used to describe the overall similarity of each predicted interaction with the corresponding SSI template. Notably, PSS-PPI yielded experimentally confirmed pathogen–host PPIs of human immunodeficiency virus type 1 (HIV-1) with performance close to that of in vitro high-throughput screening approaches. Finally, a pathogen–host PPI network of human pathogen Mycobacterium tuberculosis, the causative agent of tuberculosis, was constructed using PSS-PPI and refined using filtration steps based on cellular localization information. Analysis of the resulting network indicated that secreted proteins of the STPK, ESX-1, and PE/PPE family in M. tuberculosis targeted human proteins involved in immune response and phagocytosis. M. tuberculosis also targeted host factors known to regulate HIV replication. Taken together, our findings provide insights into the survival mechanisms of M. tuberculosis in human hosts, as well as co-infection of tuberculosis and HIV. With the rapid pace of three-dimensional protein structure discovery, the SSI template library we constructed and the PSS-PPI method we devised can be used to uncover new pathogen–host PPIs in the future.