Using Residue Interaction Networks (RINs) to Understand Protein Function and Evolution and to Engineer New Proteins

Abstract

Residue interaction networks (RINs) provide graph-based representations of interaction networks within proteins, providing important insight into the factors driving protein structure, function, and stability relationships. There exists a wide range of tools with which to perform RIN analysis, taking into account different types of interactions, input (crystal structures, simulation trajectories, single proteins or comparative analysis across proteins), as well as formats, including standalone software, webservers, and a web API. In particular, the ability to perform comparative RIN analysis across protein families using “meta-RINs” provides a valuable tool with which to dissect protein evolution. This, in turn, highlights hotspots to avoid (or target) for in vitro evolutionary studies, providing a powerful framework that can be exploited to engineer new proteins.

Graphical Abstract

Introduction

Enzymes remain the ultimate “green” catalysts, with continued interest in exploiting them as biological catalysts for chemical synthesis in industry [1]. Because of this, there is significant interest in understanding protein evolution from an engineering perspective - learning from nature how to control enzyme activity and selectivity [2]. While enzyme functional evolution is complex, there is mounting evidence that modulating conformational dynamics through evolutionary conformational selection is an important component of this process [3], a hypothesis put forward by James and Tawfik over two decades ago [4]. Appreciation of the importance of conformational dynamics to enzyme evolution has allowed its exploitation in protein engineering with ever-increasing success stories in the literature [3]. However, actually achieving this in practice is non-trivial, and would greatly benefit from predictive computational tools that can be used to guide protein engineering efforts.

In this context, non-covalent interactions play an important role in driving protein conformational dynamics, and if it is possible to identify such interactions across protein scaffolds, as well as evaluating the importance of each individual interaction, they can identify hotspots to target in engineering efforts focused on altering protein dynamics in a controlled fashion. Further, understanding which interactions are evolutionarily conserved across protein families can conversely help identify regions of the protein to avoid in engineering efforts (Figure 1). Being able to identify and predictively exploit such interactions is, however, non-trivial, in particular as experimental approaches typically lack the high-throughput capacities and resolution to study such interactions comprehensively.

Figure 1.

A visual summary of application examples of Residue Interaction Networks (RINs). These analyses include mutational hotspot identification [12], identification of allosteric signal networks [13], protein engineering [14], protein-ligand interaction analysis [15], and more broad analysis of dynamic data [16].

There is a long history of using computational sequence-based approaches to understand protein-residue co-evolution: EVCouplings [5] and EVcomplex2 [6], GREMLIN [7], Direct Coupling Analysis [8], and Protein Sparse InverseCOVariance [9] are all examples of this. Complementary to this, residue interaction networks (RINs) [10] provide a valuable tool with which to understand the co-evolution of protein interaction networks across protein families. Similarly, the Evoformer module of Alphafold2 outputs residue-residue interaction graphs [11]. When such approaches are coupled with tools to predict allosterically important residues and/or interactions within a protein, this provides a powerful framework for using biomolecular simulations and analysis to manipulate protein conformational ensembles in order to create designer proteins.

Herein, we will discuss the concept of RINs, and illustrate the promise of this approach as a tool for understanding protein evolution and engineering, through showcasing recent successes of RIN-based approaches in rationalizing and mapping protein folding, stability, function and evolution. We note that although applications of such interaction network analysis to understanding protein evolution are more recent, RIN methods represent a vast subgroup of an even greater field of graph analysis applied in biological and chemical sciences. For example, network models have been used to understand the evolutionary rates in protein interaction networks [17], the evolution of promiscuity and specificity in such networks [18], and even the evolution of resilience in evolutionarily diverse protein interactomes [19], to name just a very few examples. While discussing graph and network models more broadly is beyond the scope of this review, recent developments in the area of network biology can be found in more comprehensive reviews of the field, such as, for example, most recently ref. [20].

A feature of such analysis is that it is large scale, examining interactions across protein networks, rather than the interaction networks within an individual protein (or family of proteins). The modeling principles are, however, transferable also to individual proteins (or families of proteins), and this is valuable in light of the fact that not only do individual enzymes have conserved functional networks of interactions [21], but these are clearly perturbed by mutations during evolution [22]. In this context, residue interaction networks (RINs) provide a powerful analytical tool with which to characterize such interaction networks and assign importance to them. This then allows for the expansion of the original concept of RINs, as a representational tool for protein interactions, to a “meta-RIN” that allows for the analysis of such interactions across families of proteins, providing insights that can both rationalize protein evolutionary trajectories and inform protein engineering efforts.

Residue Interaction Networks (RINs)

RINs (also referred to as protein contact networks (PCN)) are graph-based representations of protein structures where residues are represented by nodes and the non-covalent interactions between these residues – by graph edges. While the exact formulation of the edges may vary from method to method, these interactions commonly include hydrogen bonds, salt bridges, hydrophobic interactions, and other physicochemical contacts that contribute to the protein’s stability, folding, and function. Once the network for protein structure is constructed, graph analysis techniques can be applied to extract physical knowledge about the system through calculation of parameters like clustering coefficients, closeness, betweenness, edge betweenness, average shortest path, spectral clustering or participation coefficients [23]. These and many other mathematical parameters have been applied to gather insight about the structural and functional roles of residues within the protein, such as their catalytic or allosteric role and mutation effects [24].

The majority of the widely used RIN packages define the edges of the graph in terms of the geometric descriptors such as relative distances [10,25–27]. Wide application of packages and web-based platforms such as RING [10], PDBe-Arpeggio [25], RINMaker [26] and more have shown that these networks are able to capture protein-wide interaction and provide insight into proteins structure-function relationships, drug design and allosteric mechanisms [13,28] and even the underlying factors governing disease severity in mutated proteins [29].

Importantly, the efficiency of these tools allows for scaled application of these methods from single crystal structures to full datasets of protein-ligand complexes. In a recent study, Carlucci and coworkers used Arpeggio RIN tool to identify weak hydrogen bond (wHB) interactions, specifically those involving C5−H groups in the context of protein−ligand interactions of 1,2,3-triazole drug design scaffold [15]. This analysis highlighted that in the 220 PDB database of 123T-protein cocrystal structures weak hydrogen bonds involving C5−H groups were present in nearly half of the data set examined providing insights into the interaction capabilities of the 123T-triazole scaffold for future drug design efforts.

Frequently, RIN analysis is performed to analyze NMR and MD-based ensembles. A set of unweighted interaction graphs are constructed from each analyzed frame which are then used to form a dynamic interaction graph with the edges weighted by the persistence of interaction in the set of dynamic data. RING 3.0 is an example of this approach: the software creates a probabilistic network where the edges are weighted between 0 and 1 based on frequency of connectivity between two nodes and nodes retain other statistical metrics like mutual information and Shannon distances [10]. This technique combined with graph-based analysis can highlight useful properties relevant to mutation effects and allosteric communication pathways. One of the most common graph-extracted quantities used for this analysis is residue-based betweenness centrality. Betweenness counts the number of the shortest paths between any two nodes in a network pass through a node given the total number of shortest paths in the network. Based on this per-residue quantity a measure of the overall importance of the given residue is assigned to the node [28]. This analysis has been recently highlighted in the application of RING 3.0 to the study of SARS-CoV-2 Spike Omicron Receptor Binding Domain with hACE2. In this work, the authors identified key residues that facilitate information flow within the protein structure used dynamic RIN technique to identify non-additive epistatic effects of mutations that impact binding interactions of the complex and to highlight allosteric communication pathways in the system [12].

The previously mentioned centrality metrics from RIN of each frame in the time series can also be used to construct the dynamic system representation directly [27]. Franke et al. used residue closeness centrality from an independently-computed RINs along the MD trajectory combination with an autoencoder-based dimensionality reduction (EncoderMap) to construct interpretable 2D interaction landscape [30]. This analysis helps to retain physical interpretability of identified states and can provide additional information during clustering analysis of protein conformations, allosteric analysis and construction of Markov State Models.

RINs can be an informative tool in the space of evolutionary analysis and protein engineering. While great depth of information in protein evolution has been captured by methods that rely on sequence analysis [5,9], structural information provides a deep insight into dynamics and functional mechanisms of the proteins [31]. Tools like Elastic Network Models [32] and Community-Hopping Models [33] combine evolutionary information like coevolutionary coupling measured by proximity mutual information with network theory to identify key functional residues and allosteric pathways involved in protein mechanisms [34,35]. While the previously mentioned techniques focus on network description of a specific protein supplemented with evolutionary interaction, tools like Key Interaction Networks (KIN, [36]) utilize multiple sequence alignment (MSA) and RIN graph analysis for a protein group such as protein family to identify common interaction patterns, identify outliers and understand trends that may lie along evolutionary timeline. Specifically, KIN examines families of networks (i.e. a meta-RIN) rather than individual networks within a protein. In this context, KIN has been used to characterize a family of class A β-lactamases and identify properties that aid in the evolution of ligand binding specificity in these systems [36,37].

A big advantage of RIN representations is the fact that they provide important information about the topological properties of proteins [26], without the need for detailed and complex 3D representations of proteins, thus significantly reducing computational cost while providing valuable insight into the system. As a result of this, RIN-based approaches are being implemented widely into a range of analysis tools, typically as a means of analyzing molecular dynamics trajectories (for a likely inexhaustive list of these, see Table 1). Some examples of such network analysis based models include NAPS (Network Analysis of Protein Structures) [38], RING 3.0 [10]] / 4.0 [39]], and AQcalc [40], which are available as webservers, PyInteraph [41], PyInKnife [42]] and RING-PyMOL [43]] which interface with PyMOL [44]], and gRINN (get Residue Interaction eNergies and Networks) [16], RIP-MD [45]], Arpeggio (and PDBe-Arpeggio) [25]], Rinmaker [26] and KIN (Key Interaction Networks) [36]], which are available as (typically open source) packages (note that Rinmaker is also available as a webserver or a web API service [26]). The most popular of these tools for dedicated RIN analysis, as measured by citation count as of August 2024, are the RING, PDBe-Arpeggio, NAPS, NetworkView, and the MD/MODE-Task methodologies, all of which are being broadly used by the community. Taken together, these tools have been used to significantly expand our understanding of protein function and evolution, as described below.

Table 1.

A non-exhaustive list of examples of current computational approaches for network analysis of biomolecules.

Method	Method Category	Ligand Handling	Notes
Multiple Categories
RING [39]	Standalone application, web server	Can process any ligand known from the Protein Data Bank.	A web server is available in addition to the application. An API is available for this web server, allowing for calculations on remote clusters.
PDBe-Arpeggio [25]	Standalone application, web server	All atoms are typed using SMARTS, all ligands treated the same as other atoms.	A web server is available in addition to the application. This is a newer version of Arpeggio, which is also accessible.
RINMaker [26]	Standalone application, web server	This method is focused exclusively on proteins.	A web server is available in addition to the application.
MD-Task [46] / MODE-TASK [47]	Python package, web server	Nodes exclusively represent amino acids, ligands are not represented.	MD-Task is a general analysis tool for biomolecular MD simulations using graph theory and dynamic cross-correlation. MODE-TASK is related, and specifically used for large-scale protein motion; it also has a PyMol plugin available. A web server is available for both tools.
Standalone Applications
gRINN [16]	Standalone application	Requires valid ligand parameters compatible with protein force field.	Performs network analysis on the strength of interactions, not the presence. Primarily compatible with GROMACS/NAMD.
RIP-MD [45]	Standalone application	All non-protein atoms are deleted before processing.	Optional VMD integration is available. A web server existed but is currently inaccessible.
WISP [48]	Standalone application	Only amino acids are condensed to nodes in the network.	Optional VMD integration is available, otherwise there is no GUI.
PyInteraph2 and PyInKnife2 [49]	Standalone application	This method is focused exclusively on proteins.	A pair of programs for the creation of protein structure networks; PyInteraph creates the network, PyInKnife estimates convergence and selects cutoffs.
Programming Language Packages
DyNetAn [50]	Python package	The user must define all ligands present.	A jupyter lab notebook-based method to calculate dynamic networks.
Allopath [51,52]	Python package	Ligands can be included as interactor nodes, and must be individually defined.	Network analysis to identify allosteric pathways.
dCNA [53]	R and CPPTRAJ packages	Ligands are not explicitly supported.	Uses CPPTRAJ and Bio3D to perform difference contact network analysis.
Bio3D [54]	R package	Ligands are not explicitly supported.	An R package with a variety of analytical methods, including network analysis.
Correlationplus [55]	Python package	Input is coarse-grained by amino acid, ligands are not represented.	Uses dynamical pairwise correlations to investigate allosteric communication.
MDiGest [56]	Python package	Nodes exclusively represent amino acids, ligands are not represented.	A comprehensive Python package for network analysis of biomolecules.
Key Interaction Finder (KIF) [57] and Key Interaction Networks (KIN) [37]	Python packages	Ligands are explicitly supported	KIF is a package designed to identify key interactions in a given conformational change, while KIN analyzes conserved networks of these interactions across a family of proteins.
Web servers
Network Analysis of Protein Structures (NAPS) [58]	Web server	Only amino acids and nucleic acids are supported; ligands are not represented.	This web server is used to perform network analysis of proteins and nucleic acids, including visualization of the results.
AQCalc [40]	Web server	Amino acids, nucleic acids, and lipids and select ions are supported; custom ligands are not supported.	This web server is used specifically to calculate networks of anion-quadrupole (AQ) interactions; these are often neglected in other network analysis.
SPM Web Tool [59]	Web server	Each node is centered on the alpha carbon atom of an amino acid; ligands are not represented.	This web server is used to perform Shortest Path Map (SPM) analysis to identify residue pairs with large contributions to communication pathways.
Application Plugins
NetworkView [60]	VMD plugin	Nodes represent either an amino acid or a nucleic acid; ligands are not represented.	NetworkView is designed to perform network analysis of biomolecules and show visual representations of this analysis using the VMD program.
RING-PyMol [43]	PyMol plugin	Can process any ligand known from the Protein Data Bank.	RING-PyMol is an implementation of the RING method as a PyMol plugin, and has identical functionality.

Applications of RINs to Understanding Protein Function, Dynamics, Evolution and Engineering

Network-based models have played an important role in expanding our understanding of protein function and evolution, insights which can in turn be exploited for the evolutionary-based design of new proteins. For example, dynamic residue networks (DRIN) have provided insight into the differential dynamics of different redox states of human protein disulfide isomerase [61]. Amino acid residue interaction networks have provided insight into the mechanisms of enzyme repair by the AAA+ chaperone rubisco activase [62]. Analysis of correlated dynamic networks has provided insight into enzyme activity-stability relationships [14,63] (and networks models have been valuable for dissecting protein stability more broadly [64]). Network proximity analysis has provided insight into long-range allosteric regulation of conformational dynamics in germline PTEN hamartoma tumor syndrome mutations associated with either autism spectrum disorder or cancer [65]. Finally, the RING server has been used to dissect the terpene pheromone and defense evolution of stink bugs and hemipteran insects based on analysis of crystallographic data [66], to name just a few examples.

Such models can also be applied to dissecting enzyme catalysis, and to exploit this for engineering (for instance to determine key sequences during directed evolution [67]). Enzyme catalysis relies on selective transition state stabilization, and RIN analysis can play a pivotal role in understanding the mechanistic details of this stabilization. As an example, network analysis was used to understand how the evolution of dynamical networks enhances catalysis in a computationally designed Kemp eliminase, through tightening the transition state ensemble and altering the activation heat-capacity and temperature dependence of the reaction [68]. This work also showed, using extensive molecular dynamics simulations, how prior directed evolution experiments [69] enhanced enzyme activity by closing solvent-exposed loops and enhancing active site packing [68].

Residue interaction networks have also been effectively used to understand the allosteric impact of mutational effects on DNA and RNA polymerases, including the effect of distal mutations on the exonuclease activity of DNA polymerase III [70], of a single, distal proofreading mutation (F60S) in SARS-CoV-2 variants [71], as well as for identifying and understanding functionally and allosterically important regions of the SARS-CoV-2 nucleocapsid and spike proteins (including identifying cryptic binding pockets) [72,73], as well as the impact of stabilizing mutations on the interactions between different SARS-CoV-2 variants and the ACE receptor [74,75].

From an engineering perspective, understanding is critical to being able to design. In this context, node-weighted amino acid networks (NAANs) have been suggested as a means to identify protein functional residues [76], and EvoRater extracts features identified from network analysis of protein structures in order to predict residue-level evolutionary rates in proteins [77]. Network models can also be used to predict the propagation of mutational effects across entire protein scaffolds [78], which is valuable in the context of incorporating distal mutations into the protein engineering workflow (a problem that is otherwise significantly challenging) [79]. Finally, tools such as coRIN (Comparator of Residue Interaction Networks) [80] and KIN (Key Interaction Networks) [36] can be exploited to compare protein interaction networks across pairs or families of proteins, to understand their evolutionary conservation across systems.

While clearly evolutionary conserved interactions are regions one likely does not want to disrupt in protein engineering efforts (in order to avoid significantly compromising function or stability, for instance), identifying which interactions are important provides an important guide to regions to avoid in engineering efforts. These can then be coupled with a wide range of tools that can characterize protein allostery and identify allosterically important residues (Key Interactions Finder, KIF [57], as one example of many) to predict potential mutational hotspots that are less likely to disrupt functionally important residue interaction networks.

Overview and Conclusions

Network analysis models are becoming an increasingly important tool in biology in order to characterize protein-protein interactions, protein allostery, protein ligand interactions, and protein stability, among a few examples. There exist a wide range of approaches to perform network analysis more broadly. In this context, residue interaction networks (RINs) are a valuable tool that can dissect functionally important non-covalent interaction networks in both individual proteins and families of proteins, and provide a valuable tool for the engineering of new protein functions. Such approaches, which are rapidly gaining in poluarity, can be applied to static structures, simulation trajectories, and there are methods available as standalone software, webservers, and even a web API (Table 1). As the diversity and ease of use of computational tools for RIN analysis continues to expand, RINs are likely to emerge as an indispensable tool in protein engineering efforts.

Figure 2.

Illustration of Key Interaction Networks [36] analysis applied to the study of the class A β-lactamases, as a model system. The highlighted functionality is used to (A) identify strongly conserved interaction networks among the family of related proteins, (B) identify new interactions that characterize changes in protein function from generalist to specialist along a phylogenetic tree produced from Ancestral Sequence Reconstruction, and (C) evaluate the conservation degree of interactions between the catalytically relevant residues among the family of modern enzymes.