How Binding Affinity and Binding Specificity Map to Sequence Space

Abstract

Evidence suggests that molecular interactions may evolve differently for proteins compared to functional RNA. We suggest that for RNA interacting with other nucleic acid molecules, the selection of higher affinity interactions leads to greater specificity, whereas for proteins interacting with other proteins, there is a trade-off between selection for greater affinity vs. greater specificity. This difference arises from the nature of the molecular contacts driving intra- and inter-molecular interactions, with crucial roles for non-specific hydrophobic interactions driving affinity in proteins and for specific hydrogen bonds driving affinity and specificity in RNA. The implications of this difference are discussed.

Introduction

Genomes consist of thousands of protein-coding and RNA-expressing genes that function by interacting with other proteins, nucleic acids, and small molecules that are present in cells. These sets of interactions largely define the functions that make up the molecular phenotypes associated with a given species. For example, databases like STRING (Szklarczyk et al. 2023) and MINT (Calderone et al. 2020) detail the sets of protein–protein interactions that have either been observed biochemically or predicted in the proteomes of different species. Many proteins (both enzymes and non-enzymes) are part of metabolic and signal transduction pathways that function to produce either a yield of a particular metabolite or the generation of a transcriptional signal in the nucleus of a (eukaryotic) cell. These would be the functional interactions of a protein in a particular species, but would not reflect the totality of protein–protein interactions observed for any particular protein, as proteins not present in the same sub-cellular location at the same time will not functionally interact. A similar paradigm is true for RNAs. However, in the case of RNA-RNA interactions such contacts most frequently result in gene expression changes. The combination of all these interactions, each of which may be modified via mutation of the genome, define an organism that is under selection during evolution.

Current understanding of the evolution of genome architecture suggests that organismal effective population size is a critical modulator of selective strength (Lynch 2020). It describes the ability of a species to identify a relatively optimal sequence for any given selectable function (Liberles et al. 2011; Liberles 2023a). But the ability to find a suitably functional sequence for a given binding function also depends upon the underlying nature of the interactions themselves that are being selected (Buda et al. 2023). This is also a core part of the landscape and one that is much less well characterized in the evolutionary literature. This piece aims to offer a conceptual characterization of this, one that potentially uncovers core differences in the selectable nature of interactions of proteins and nucleic acids.

Defining Sequence Space and the Relationship to Biochemical Functions

For any of the thousands of proteins in a genome, the number of possible sequences is 20^protein length, where the protein length is itself not fixed. For an RNA expressing gene, that set of possible sequences is 4^RNA length, where again the RNA length is not itself fixed. That is a large number of possible sequences, a subset of which will be sampled throughout evolution. Of all of the possible sequences for any biological macromolecule, a tiny fraction will have binding capabilities that correspond to interactions that are positively selected. It is occupation of this tiny part of possible sequence space that then enables a biomacromolecule to have a specific suite of binding interactions (Goldstein 2008; Liberles 2023a, b; Liberles et al. 2011). The sequences that occupy this space and emerge through selection will be largely distinct from those that are not positively selected (Goldstein 2008; Liberles et al. 2011). The chemical nature of the selected sequences is what defines their properties qualitatively.

Defining Affinity and Specificity

A particular focal biomacromolecule will have an ensemble of interactions under physiological conditions at the concentration at which it is expressed (Dill and Bromberg 2010). More generally, it will have a potential affinity (binding energy) with every other sequence in sequence space, including the subset of sequences that are expressed in any given cell, as well as any small molecule metabolites present in the cellular milieu. The combination of binding energies and concentrations make most of the potential interactions for expressed molecules irrelevant, interactions that will rarely happen and not in sufficient mass numbers to have biological consequences. The relative binding energy dictates the relative fractional occupancy of two molecules about each other and the statistical average of those interactions for large numbers of individual molecules describes the concentration of different macromolecular complexes (Dill and Bromberg 2010). To add complexity to the scenario, any sequence can fold into an ensemble of structures, each with a distinct set of affinities (Siltberg-Liberles et al. 2011; Halvorsen et al 2010).

For the set of interactions of a particular sequence, mutations that alter the sequence may potentially change the binding affinities for all of the interactions (Dasmeh et al. 2024). Thus, the ensemble of potential interactions will change with mutations and is subject to selection. For any RNA or protein sequence with a particular expression level and resulting concentration, there is a set of total physical interactions with other biomolecules. However, within the total set of such interactions, there is a subset that are under positive or negative selection, and these would the set of interactions that occur under selection. Thus, we draw a distinction between physical interaction alone and evolutionary selection for that physical interaction.. In some cases, there will be a single interaction that is most strongly positively selected and the other interactions will be a byproduct of the resulting sequence and its physical properties. The distinction between selectable function and function that does not impact organismal fitness is a side issue to the arguments being made here, but is an important differentiation as interactions that are evolutionarily stable but not selected will persist without benefit to the organism solely because of the combination of mutational density in the sequences that enable the interaction. There is a now extensive literature on the importance of selection in defining function (see Doolittle et al. 2014 and references therein).

Thinking About not Binding as a Kind of Binding

Selection to bind to a particular molecule restricts the space of available sequences, as many potential sequences will not have the ability to bind with sufficiently high affinity to the target molecule under physiological concentration and other conditions. Similarly, selection to not bind a particular molecule can also restrict the number of sequences to include only those that lack the ability to interact specifically with a particular target molecule (Liberles et al. 2011). One important question then becomes if the sequences that are excluded by the selective criteria for binding at a particular affinity overlap with the sequences that are excluded by the selective criteria for not binding other partners with a particular affinity. That is, what is the relationship between affinity and specificity? Does selecting for the highest affinity molecules automatically select against binding to other molecules or does selecting for the highest affinity molecules tend to select for molecules that have high affinity for other interactions as well?

How Molecular Recognition Occurs in Proteins and in Nucleic Acids

Before addressing that question, let’s dive in to the nature of molecular recognition, of interactions that occur in proteins and in nucleic acids. In proteins, there isn’t a readily generalizable fit between interacting molecules. Hydrogen bonds in secondary structural units come from the backbone and not from amino acid side chains that are subject to mutation (Branden and Tooze 1999). Built from building blocks of amino acids, van der Waals interactions with steric fit between closely packed amino acids don’t follow a specific rule for how particular amino acids interact (Ames et al. 2016). Similarly, the interaction of charged amino acids follows the rules of Debye-Hückel theory, where the distance dependence depends upon the dielectric constant of the local environment, but does not have a regularity to orientations and distances. These interactions are stronger in more hydrophobic environments like binding interfaces where there is less charge screening. A key driving force is the hydrophobic effect, where the occlusion of water from tightly packing amino acids with relative hydrophobicity drives affinity (Ames et al. 2016). Because of the lack of regularity, any number of amino acids in the right orientation can pack together to occlude water. Aromatic residues do form very geometrically specific irregular interactions beyond the role associated with their hydrophobicity, sometimes interacting with each other, but these interactions contribute only a small fraction of the total binding affinity for protein–protein interactions in general (Anjana et al. 2012; Makwana and Mahalakshmi, 2015). The most specific interaction among aromatic residues was an edge to face interaction between Phe residues (Anjana et al. 2012). With that said, that some specific interactions (also including cation-pi interactions) occur that contribute to the binding energy specifically, these types of interactions do not drive the majority of the affinity in protein–protein interactions. Further, aromatic residues offer on average, statistically favorable interactions with ~ 40–50% of other residues according to statistical contact potentials (Miyazawa and Jernigan 1985).

For nucleic acids by contrast, every nucleobase in the chain can interact favorably with any other base, and ultimately folding is driven by pi-pi stacking of the hydrophobic bases to leave the charged edges of the planar base exposed to the solvent (Vicens and Kieft 2022). While canonical Watson–Crick base-pairing is fundamentally energetically driven by this stacking, the regularized hydrogen-bonding patterns that arise within the canonical double-stranded structure give rise to highly specific and high affinity intra- or inter-molecular interactions between two strands. However, Watson–Crick interactions are only a subset of the potential edge-edge interactions. The notion of sequence specific interactions can be extended to include the Hoogsteen edge of the base that remains exposed in the major groove of a Watson–Crick helix. This face enables regularized and reverse Hoogsteen base-pairs (T(U)-AT(U), and C + -GC base pairs in the Hoogsteen sense and A-AT(U) and G-GC base pairs in the reverse Hoogsteen sense) that may result in triple or quadruple helices, albeit with different preferences in RNA compared with DNA (Rangadurai et al. 2018; Leontis and Westhof 2001). Yet these interactions are still largely dependent on the bases involved with specific base combinations resulting stable conformations, and non-conforming bases substantially destabilizing the structure.

RNA tertiary structures are typically scaffolded by Watson–Crick paired elements, with regions not involved in Watson–Crick interactions typically forming complex motifs that incorporate non-canonical interactions to maximize stacking as described above. Adding complexity to the sequence to functional relationship for RNA is the dynamism of RNA structure. The time to fold is much shorter than the time required for synthesis resulting in rapid interchange between structures with similar folding energies. Thus, the probability of a specific base pairing pattern within the ensemble of potential stable structures is frequently invoked to understand the biological behavior of RNAs (Vicens and Kieft 2022; Ganser et al 2019).

Affinity-Specificity Trade-Offs in Proteins and Nucleic Acids

From the mechanisms giving rise to affinity, it is clear that there isn’t a general rule for specific amino acid-amino acid contacts that give rise to high affinity, specific interactions. The distribution of effects of mutations on interaction binding affinity shows a large variance, with many mutations having very small effects on a particular interaction when compared with wild type sequences (Ames et al. 2016; Brender and Zhang. 2015). The ad hoc nature of these interactions gives rise to a lack of specificity to the interactions that give rise to affinity. This is particularly true for large hydrophobic patches that give rise to affinity at binding interfaces (Desantis et al. 2022). While charges have been described as giving some specificity to protein–protein interactions as evidenced by the enrichment of charges in interfaces (Tsai et al. 1997), there are still likely to be many proteins that will have such charges, with residues of a particular charge representing 10–15% of possible amino acids at any site (this number comes from a sampling of the genetic code). The difference between matched and mismatched sequences is comparatively small, with high affinity driven by hydrophobic interactions and specificity driven by charge-charge interactions to a large extent. The highest affinity interactions will therefore have larger hydrophobic surface areas. From this, there is a natural tradeoff between affinity and specificity in protein–protein interactions. Evolving into sequences with higher affinity will more often result in sequences that are less specific in their interactions, with specificity generally harder to achieve in proteins than in nucleic acids. This is shown schematically in Fig. 1.

Fig. 1.

This schematic representation shows the different trade-offs between affinity and specificity between proteins and nucleic acids, with the affinity in one case driven by the hydrophobic surface areas, with some specificity coming from complementary charged interactions, and in the other case both affinity and specificity driven by base pairing (hydrogen bonding and coordinated stacking)

This contrasts with nucleic acids where the double-stranded structure characteristic of many biologically relevant intermolecular RNA-RNA or RNA–DNA interactions largely depends upon Watson–Crick interactions. While base stacking contributes significantly to the energetics, this contribution is dependent upon proper orientation that is steered through the hydrogen bonding interactions of base pairing (Delcourt and Blake 1991; Svozil et al. 2010). Examples of such base paired interactions include a host of regulatory interactions across all domains of life (siRNA, miRNA, sRNAs, etc.), tRNA-mRNA interactions that enable translation, interaction between the 16S rRNA and Shine-Dalgarno sequence that guide translation initiation in bacteria, and even CRISPR recognition of specific DNA or RNA sequences. The nature of Watson–Crick interactions enables both high affinity and highly specific interactions. One mark of the specificity with which such interactions act is the extent to which transcriptomes are under selection to prevent spurious regulatory RNA-RNA interactions (Umu et al. 2016; Hockenberry et al. 2018; Chen and Rajewsky 2006; Guo et al. 2008). The energetic effect of a single mismatched base in a stem structure is significant (Aboul-ela et al. 1985). The end effect of this is that a high affinity molecule is specific for a perfectly matched complement and sequences that are not perfectly matched are likely to exhibit an energetic penalty.

One caveat to note is that a cell operates at generally fixed temperature, pH, and salt/metal ion concentrations, as double-stranded structures grow longer, they gain in affinity, but in the environmental conditions will become less specific. However, there is a point that will optimize affinity and specificity for a given interaction in sequence space under cellular conditions. This has been observed in CRISPR guide RNA engineering efforts, where decreased affinity of the nucleic acid pairing ultimately increased specificity of the guided nuclease cleavage (Fu, 2014, Bisaria, 2017). From a theoretical perspective, one might expect that the length of stem loops is itself constrained by the organismal strength of selection. This is observed in many systems driven by mutation-selection-drift balance and is a well understood process with well understood effects on molecular stability (Wilson et al. 2020). In neither large nor small effective population size species would one expect to select for stem structures longer (and with greater affinity) than their selectable stability in cellular conditions. Therefore, under these conditions, one would expect RNA molecules to exist in a regime where stability and affinity are correlated, regardless of effective population size.

This argument generally holds with all macromolecular interactions that enable binding, including those that are subject to drift barriers, where the same evolutionary and biophysical dynamics interact around the added macromolecular complexity (Lynch 2020). There is an interplay between what is selected for based upon underlying thermodynamics, the strength of selection dependent upon organismal population size, and the collection of sequences available in sequence space. A similar argument as that for other macromolecular interactions,, can be made for some intrinsically disordered proteins. At least some intrinsically disordered proteins fold upon binding, where the energy of folding around the target (targets in the case of multi-specificity) is a mechanism of gaining specificity of interaction (Teilum et al. 2021). In this case, this entropic energetic cost in affinity is driving specificity. One recent study did suggest that disordered proteins can bind at high affinity and specificity, but driven by strong selection inducing strong sequence conservation (a reduction in the sampling of sequence space for key residues) (Lazar et al. 2022). Within the set of sequences, there is likely still a trade-off between affinity and specificity. Molecules with high affinity and specificity were not viewed in the context of sequence space and the density of solutions with variation in specificity and/or affinity.

An additional factor that is relevant to this discussion is macromolecular crowding in cellular milieus. The large concentration of molecules in cells inherently leads to more non-selected interactions occurring transiently and can also affect the affinity of selected interactions, both for ordered and for disordered proteins (Zosel et al. 2020). This itself will become part of the selective regime that sequences evolve in, but doesn’t alter the inherent biophysical tradeoffs between affinity and specificity in proteins and in nucleic acids.

Nucleic acids may also interact with a wide variety of molecules including proteins, metal ions, amino acids, and nucleoside analogs, in single-stranded, double-stranded or tertiary structured forms. Nucleic acid tertiary structures that interact specifically with small-molecules or proteins are termed aptamers, and while they can be generated synthetically via in vitro selection (Ellington and Szostak 1990; Tuerk and Gold 1990), they also occur naturally (Nahvi et al. 2002; Huang and Lilley 2025). It was originally argued that increased specificity in nucleic acid-small molecules interactions would result in increased affinity (Eaton 1995); however, experimental studies have found no specific relationship between affinity and specificity in large collections of synthetic aptamers (Carothers et al. 2006; Alkamis et al. 2025). While the properties of natural aptamers are distinct from those artificially selected (Kennedy et al. 2010; Meyers et al. 2004) there is no evidence to date that suggests a different paradigm to describe the relationship between affinity and specificity.

RNA–protein interactions also tend have no generalizable model of specific amino acids interacting with specific nucleotides. Even as datasets increase in size and complexity, computational prediction of RNA-RBP (RNA binding protein) interactions remains an active challenge in the field (Mizrahi et al 2025; Hentz et al. 2025) due to many factors, such as sequence accessibility and secondary or tertiary folding, that confound simple models for binding. Indeed, the relationship between primary sequence and functional conservation also remains obscure for many long noncoding RNAs (lncRNAs) as differing regions may act as isolated motifs (sequence or structural) that enable interaction with protein partners, and structural elements displaying Watson–Crick pairing scaffold these partners into close proximity (Fabbri et al. 2019). While there are a few general trends that have been noted DNA–protein interactions in the case of zinc finger transcription factor proteins (Klug 2010, see also Zhang et al. 2025), these are far from the 1:1 map of Watson–Crick base pairing that occur frequently between two nucleic acid molecules wherein specificity and affinity may correlate across an ensemble of interacting sequences.

Concluding Thoughts

Given the inherent differences in molecular recognition between proteins and nucleic acids in self-recognition, the underlying sequence space differs in how it maps to functional, selected proteins and nucleic acids. This stems from the nature of the mapping in the spaces of protein–protein interactions vs. RNA-RNA interactions. In proteins, the lack of generalizable lock and key rules for how specific amino acids interact and the ubiquity of the hydrophobic effect in driving affinity leads to high affinity sequences that are not unique in their interactions. RNA, when self-interacting via Watson–Crick base-pairing, has a 1:1 mapping due to hydrogen bonding patterns. While there are many equivalent affinity interactions, these will all be specific interactions, and will evolve via coordinated (compensatory) changes in RNA molecules. The equivalent high affinity interactions in proteins will mostly occur through replacing less specific hydrophobic residues with other hydrophobic residues. When mutations occur for selection to act upon, there will be a general trade-off in finding sequences with higher affinity and those with higher specificity in proteins. In nucleic acid self-interaction, there will be a greater correlation between mutations that increase affinity and those that increase specificity. Thus, acquiring novel RNA-RNA interactions that are both high-affinity and high specificity may be relatively facile, leading to the plethora of RNA-RNA regulatory interactions observed across all species (Dexheimer and Cochella 2020; Dutcher and Raghavan 2018). This distinction is worth noting as researchers understand the evolution of protein and RNA sequences across genomes. It also is informative for what types of molecules are more likely to evolve promiscuous interactions (proteins), for when those interactions may open up new evolutionary trajectories, during adaptation, or during the early diversification of life.

Declarations

Competing Interests

Both authors are involved as editors of the Journal of Molecular Evolution. They have no other competing interests to declare.