Predicting interacting hotspots for nanobodies' binding using triplets of residues

Abstract

Protein–protein interactions (PPI) are fundamental to cellular signaling, forming robust networks that govern critical biological processes such as immune response, cell growth, and signal transduction. Nanobody‐based therapies have emerged as a key strategy for modulating PPIs, offering exceptional potential due to their high specificity, stability, and ability to access challenging epitopes on PPI interfaces inside cells. The rational design of nanobodies relies mainly on understanding and predicting their binding regions, particularly the residues that contribute the most to the binding energy (binding hotspots). Existing computational methods do not fully provide a scalable solution for hotspot identification in nanobody design, leaving a critical gap in the rational design of these therapeutics. Here, we present a scalable and structure‐aware algorithm for hotspot prediction in nanobody design. The algorithm queries a curated database of triplets of interacting residues obtained from ~20,000 non‐redundant PDB structures. We showed that these triplets contain structural and energetic information, being able to assess the stability effect of residue variations in protein structures, Pearson R = 0.63 (MSE = 1.58 kcal/mol). More important than effects on stability is the ability of the algorithm to predict binding hotspots of protein–protein generic complexes and more specifically in complexes containing nanobodies. HotspotPred reached an accuracy of 0.73 for hotspot residue identification in a protein interaction dataset of 1160 Alanine mutants and correctly identified in 63.4% of the cases we predicted at least 2 residues on the binding surface.

1. INTRODUCTION

Protein–protein interactions (PPIs) are central to cellular function, orchestrating essential processes such as cell growth, signal transduction, enzymatic regulation, and immune signaling (Komarova et al., 2017). Disruptions in these networks can lead to severe diseases, making precise modulation of PPIs essential for therapeutic interventions. PPIs rely on non‐covalent forces, including hydrogen bonds, van der Waals interactions, electrostatic forces, and hydrophobic effects (Adhav & Saikrishnan, 2023; Rost, ¹⁹⁹⁹). However, not all residues on their surface contribute equally to the interaction. Certain key residues, known as binding hotspots (HS), disproportionately stabilize the interaction, contributing greatly to the binding energy (Chen et al., 2024). Thus, identifying those residues and how their disruption leads to disease is critical for designing PPI interactions and docking.

Nanobody‐based therapeutics have emerged as powerful tools for targeting PPIs (Adhav & Saikrishnan, 2023; Jin et al., ²⁰²³). Their small size, high specificity, and ability to access challenging epitopes and function inside a cell make them ideal for disrupting disease‐driving interactions. The rational design of nanobodies with optimal affinity requires identifying their binding HS on their target, as well as incorporating negative design (Farasat et al., 2017). Non‐determinant residues (ND) do not contribute strongly to the interaction affinity but have a role in specificity.

Identifying HS and ND residues in a crystal structure of a protein complex can be done using Alanine scanning mutagenesis, where individual residues are mutated to Alanine and the binding energy changes are measured. Deep Alanine mutagenesis provides valuable insights into hotspot identification but is usually resource‐intensive and time‐consuming (Bromberg & Rost, 2008), computational approaches were developed. These methods generally fall into two main categories (Ovek et al., 2022; Rosário‐Ferreira et al., ²⁰²²).

(1) Energy‐based approaches, such as FoldX (Delgado et al., 2019; Schymkowitz et al., ²⁰⁰²) and Rosetta (Rohl et al., 2004), predict binding energy difference upon mutation using force fields or empirical scoring functions.

(2) Machine learning‐based methods that rely on sequence like SPOTONE (Preto & Moreira, 2020) and structure‐based methods such as Hotpoint (Tuncbag et al., 2010) KFC2 (Ovek et al., 2022; Rosário‐Ferreira et al., ²⁰²²; Tuncbag et al., ²⁰¹⁰; Zhu & Mitchell, ²⁰¹¹) and PredHS (Deng et al., 2014) trained on related features such as conservation rate, solvent‐accessible surface area, or deep mutagenesis scanning fitness data (Chen et al., 2024; Fu et al., ²⁰²³).

Current computational methods face key limitations: energy‐based approaches rely heavily on having a 3D structure model of the complex, while machine learning models struggle with small training datasets due to the scarcity of experimental HS data. Moreover, most existing models are trained on general PPIs rather than nanobody‐specific complexes, failing to capture their unique structural constraints (Chen et al., 2024).

The main challenge is to pinpoint specific regions on a protein's surface, such as those on a nanobody, where HS or ND residues could bind. Identifying these critical sites and the corresponding residues enables the strategic docking of target proteins and the engineering of mutations to create new interaction sites tailored for specific amino acids.

To validate our approach, we first assessed whether the triplet‐based interactions used in our method encapsulate energetic information by using them to predict ΔΔG values for point mutations from a curated dataset of 2484 mutants (Delgado et al., 2025). Then we tested our tool for HS prediction over 1160 Alanine mutants from the SKEMPI database (Jankauskaite et al., 2019). Finally, we analyzed a database of 123 non‐redundant nanobody complex structures (Zavrtanik & Hadži, 2019), to predict the interacting HS and ND residues of the protein ligand which will allow positioning and docking of a target protein on a nanobody. Our approach provides a powerful, scalable solution for rational nanobody engineering, facilitating the development of next‐generation therapeutics.

2. RESULTS

2.1. TriX database building

The TriX database (TriXDB) was originally obtained from 5095 high‐resolution (<1.5 Å) crystal structures, to which a 95% sequence similarity filter was applied selecting one representative structure per group. Structures were later virtually digested into triplets of residues (Cianferoni et al., 2025). Here we expanded the TriXDB database and refined the selection criteria by introducing other structural constraints than just the Cα and Cβ distances previously used to identify and query the triplets. We built a new TriXDB_20K from 19,958 high‐resolution (≤2.5 Å) crystal structures, containing the biological units, from the RCSB ( https://www.rcsb.org). To ensure broad structural coverage while reducing redundancy, we performed a two‐step curation process. First, protein crystal structures with a resolution equal to or less than 2.5 Å were selected to ensure high structural quality for downstream geometric analysis. Next, to reduce redundancy and avoid the overrepresentation of closely related proteins, we clustered the filtered structures based on a 30% sequence identity cluster provided by the RCSB website and consistent with the threshold at which structural homology is preserved (Rost, 1999). These clusters were obtained from the RCSB, (Bank, 2025) which groups protein chains based on pairwise global alignments.

For each resulting cluster, we selected a single representative structure (the one with the best resolution). This ensured that our dataset captures non‐redundant and diverse structural features from the widest structural landscape while avoiding overrepresentation of closely related proteins. The workflow for constructing TriXDB_20K, including structure selection, virtual digestion, and triplet extraction, is outlined in Figure 1.

FIGURE 1.

Overview of the TriX Database (TriXDB_20K) construction pipeline. Representative high‐resolution crystal structures (≤2.5 Å) were selected based on a 30% sequence similarity cluster to ensure structural diversity. The selected structures are fragmented into triplets of interacting residues and triangle features are extracted: Sequence of the triangle, name of the molecules each residue belongs to, Cα–Cα and Cβ–Cβ distances between the different residues, Dihedral angles (Cβ–Cα–Cα′–Cβ′) and backbone torsional angles (φ, ψ).

After selecting representative structures, we applied a digestion algorithm to fragment each protein into triplets of interacting residues, preserving spatial and biochemical context. As in the original TriXDB methodology, a triplet was defined as three residues in which at least one pair of side chain atoms coming from different residues is distant within 4.5 Å (Threshold used for contact distance), and where the maximum Cα–Cα distance allowed is 18. After our virtual digestion, we identified 176,143,505 unique triplets. For each selected triplet, we recorded the following geometrical features: (1) distances between the α‐carbons and β‐carbons of the three residues, (2) backbone torsional angles phi (φ) and psi (ψ) for each residue, and (3) dihedrals formed between the α and β carbons (Cβ–Cα–Cα′–Cβ′) of interacting residues. These features provide a comprehensive geometric fingerprint of each triplet, facilitating accurate superimposition on two anchoring residues of the chosen protein. Triplets were also labeled based on the number of contacts between the implied residues. A contact was defined as any atom from one residue found within 4.5 Å of another residue's atom. We assigned a contact count from 1 to 3, with 1 indicating a single contact between two residues of the triplet; 2 indicating two contacts where at least one of the triplet residues contacts two others; 3 indicating all three residues contact the two other residues of the triplet. Triplets with a count equal to 0 (no residues in contact) were excluded from the dataset.

2.2. Assessing stability information in triplet‐based interactions

In order to use triplets to identify possible hot spot residues, we first needed to demonstrate that they contain energy information. To do so, we extracted all relevant triplets for each chosen amino acid substitution, using a three‐step query process (Figure 2).

FIGURE 2.

Assessing stability information in triplet‐based interactions. (a)–(c) Triplet extraction process for hotspot prediction. (a)Extracting wild type triplets. (b)Extracting mutant triplets. (c)Background triplets. The adjacent table demonstrates an example of triplet‐based terms computation. The minimum ratio for the wild type is coming from the triplet with the sequence Arginine‐Alanine‐Methionine and for the mutation is coming from triplet Tryptophan‐Leucine‐Valine. (d)Regression analysis of predicted ΔΔG and the experimental for the validation set. (e)distribution of the important ratio of the features used for the random forest. The table shows the calculation of avgRatioWt, avgRatioMt, minRatioWt, and minRatioMt; minRatioWt is obtained by dividing the number of triangles that involved the WT residue with two interacting residues by the number of times you find triangles involving any amino acids interacting with the same two residues (9 divided by 16 as the minimum ratio). You repeat the same for the three triangles involving the mutated residue and calculate avgRatioWt. Then we do the same for the mutant residue to calculate the other two parameters.

(1) Wild‐type triplets: The first query identifies all triplets containing the wild‐type residue at a mutation site, later retrieving compatible triplets from the TriXDB_20K. The search implies allowing structural dubiety for Cα–Cα and Cβ–Cβ distances and also with the option of considering additional geometric descriptors such as dihedral angles (Cβ–Cα–Cα′–Cβ′) and backbone torsional angles (φ, ψ) (Figure 2a). (2) Mutant triplets: This query follows the same process as above, except that the wild‐type residue identity is substituted with the mutant one, while preserving the surrounding structural environment constraints (Figure 2b). (3) Background term: The final query to the database is meant to establish a background reference and is obtained by masking the identity of the residue at the mutation site while preserving the surrounding structural neighborhood. The number of retrieved triplets containing any possible amino acid at the mutation position in this step is the background term (Figure 2c).

After collecting this information from the database, we compute two types of frequency‐based ratios per triplet: one ratio equal to the frequency of wild type residue triplets with respect to the background (all possible amino acids at the chosen position) and another ratio equal to the frequency of mutant triplets with respect to the wild type background. After this step, we end up with a set of ratios for all the triangles that have the mutated position at one of its vertices (see “Methods,” Supplementary Equations (1) and (2)) (see table in Figure 2). The different terms derived from the ratios (avgRatioWt, avgRatioMt, minRatioWt and minRatioMt) (see supplementary file) quantify how frequently the mutant and wild type residues appear in interacting triplets (see legend of Figure 2).

To see if we could predict the change in stability upon mutation using triplets, we used a dataset of 2484 experimentally determined ΔΔG values (Delgado et al., 2025). The dataset was split into 80% training and 20% test subsets (Supplementary Table 1) and we trained a random forest regression model on the first one. The model incorporated terms derived from the triplet ratios, as defined above (see “Methods”), as well as biophysical properties of wild‐type and mutant residues (Supplementary Table 2).

The predicted ΔΔG values from our model showed a good agreement with experimental values for the training dataset (Supplementary Table 1 and Supplementary Figure 1) as well as for the validation dataset (Pearson correlation of 0.63 and MSE value of 1.58 kcal/mol) (Figure 2d).

To assess the relative contribution of each input feature to the prediction of ΔΔG values, we examined the feature importance scores derived from the Random Forest model. Among the triplet‐based features, the minimum and the average ratio features emerged as the most important features to predict the changes in energy (Figure 2e).

Based on these findings, only the minimum and average ratio terms were included in Figure 2, as they provide the clearest insights into the structural constraints learned by the model. The maximum ratio was excluded from the figure to avoid visual redundancy and focus on understanding the most impactful features.

2.3. HS assessment using the SKEMPI database

To further evaluate the accuracy of our hotspot prediction algorithm, we validated it against the SKEMPI database (Jankauskaite et al., 2019), which contains experimentally determined ΔΔG_binding values for different mutations at protein–protein interfaces. To test our approach, we selected all mutations to Alanine involving COR (at the center of the interface) and RIM (border of the interface) residues (Levy, 2010). Our goal was to determine how well our predictions align with experimentally defined binding HS residues (based on changes in experimental binding energy). Triplets coming from PDB structures in this dataset were excluded from the TriXDB_20K search step.

To predict whether a given residue in our target protein is a HS, we select the list of possible anchoring residues in the interacting protein (see “Methods”). Then pairs of residues are made combining the anchors' list; their identities, along with the structural features defined above and in Methods, are used to query TriXDB_20K, retrieving the most relevant triplets for the search: the floating residues. After doing multiple queries all along the interface pairs of the input structure, we end up with clouds of floating residues. Residues whose α‐carbons are found at a distance less than 3.6 Å (Li & Zhang, 2009) and would present steric clashes with the input structure are removed, and the remaining ones get clustered in clouds using a spatial clustering algorithm (see “Methods”). After the cloud cleaning (minimum three residues per cluster), the structure of the target is reintroduced, and we search for any residue in the cloud of floating residues (see “Methods”) that: (1) shares the same amino acid identity as the mutated residue in SKEMPI, (2) is located within 3.5 Å (Cα–Cα distance) or less of the mutated residue, and (3) is contacting the two residues on the triplet that match the anchors. The Cα–Cα distance threshold was selected to be less than the average distance between two Cα of two consecutive residues on a peptide (3.8 Å) (Chakraborty et al., 2013) and to prevent confusion regarding which residue on the target protein we are trying to predict. We opted for a large dubiety for the Cα–Cα distance because our triangle approach could bring an interacting residue where the tip of the side chain interacts in a similar manner as the real residue in the target structure, but due to torsional differences for the rest of the side chain, the Cα could be distant from the real one. If these criteria are met, our algorithm marks the residue as a hotspot.

We then considered, as reported in the literature, that an experimental HS residue is the one where its mutation to Alanine causes a loss of more than 2 kcal/mol (Chen et al., 2024). This choice of the threshold is grounded in the biophysical interpretation of ΔΔG: a 2 kcal/mol change typically corresponds to a ~100‐fold loss in binding affinity, indicating the disruption of key stabilizing interactions such as a hydrogen bond, salt bridge, or hydrophobic contact (Morrow & Zhang, 2012). Such energetic losses have been consistently used to identify functionally critical residues at protein–protein interfaces.

In this aspect, it is important to mention that the SKEMPI ΔΔG_binding energies are the result of true changes in binding energies plus the possible destabilizing effect of a mutation that could cause conformational changes.

We then compare the predicted HS to the experimentally verified ones in SKEMPI, quantifying our method's performance through a confusion matrix (Figure 3a). To further assess the predictive capability of our algorithm, we computed the true positive rate (TPR) as a function of the experimental ΔΔG_binding threshold (Figure 3b). The observed increasing trend in TPR suggests that our method performs particularly well in identifying functionally critical residues, those with the highest binding energy changes. This confirms that our approach is effective at capturing residues that play a major role in binding interactions. We also determined the recall, the accuracy, and the specificity as functions of the Cα–Cα distance threshold in the SKEMPI database (see Supplementary Figure 4). We found that increasing the Cα–Cα distance threshold from 1.0 to 3.5 Å improves recall from about 16%–25%, indicating that allowing broader spatial clustering enhances the method's ability to identify experimentally validated HS. Both specificity and accuracy remain relatively high but decrease slightly (from 0.8 to 0.73) with the distance threshold, suggesting that widening the distance for clustering leads to more false positives.

FIGURE 3.

Analysis of HS prediction on SKEMPI database. (a) Confusion matrix comparing predicted and experimentally determined hotspots from the SKEMPI database. True hotspots (HS) were defined based on a ΔΔG_binding threshold of 2 kcal/mol. (b) True positive rate (TPR) as a function of the ΔΔG_binding threshold. (c) Analysis of false positives (FP): Binding versus interaction effects for false positives. Each point represents a residue predicted as a hotspot by our algorithm but did not meet the experimental ΔΔG_binding threshold. We plotted FoldX‐calculated ΔΔG_interaction values for these residues. (d) Analysis of false negatives (FN): The distribution of FoldX ΔΔG_stability values for the FN cases that have a ΔΔG_binding > = 2 kcal/mol but were not identified by our algorithm. Distributions of ΔΔG_stability were plotted based on the sign: The mutations with a positive ΔΔG_stability are associated with a destabilizing effect when mutated to Alanine and the ones with negative ΔΔG_stability are associated with a stabilizing effect when mutated to Alanine.

2.4. Analysis of false positive (FP) cases

False positives correspond to residues that were identified by our algorithm but do not meet the experimental ΔΔG_binding threshold considered in the literature of 2 kcal/mol (Bogan & Thorn, 1998). Thus, they are not false positives in a strict sense and could correspond to either experimental errors (see Supplementary Figure 7 for mutation Valine 34 to Alanine in the PDB 1emv) or the strictness of the 2 kcal/mol cutoff (see Supplementary Figure 6 for mutation Tyrosine 42 in the PDB 1a22). To understand these cases, we examined their ΔΔG_interaction values computed using FoldX (Figure 3c). The results indicate that the majority of the FP predicted residues contribute significantly to the interaction energy, thus having functional importance in the binding interface (ND residues) despite not being classified as HS residues under the experimental energy criteria. This implies that the FP residues are relevant for molecular interactions but are not picked up because of the experimental energy threshold used.

2.5. Analysis of false negative (FN) cases

FNs represent residues that are experimentally classified as HS residues but were not detected by our algorithm. This could be happening for two reasons that are not mutually exclusive. The first one could be due to incomplete data in the PDB at the time we created our database. The second could be the result of a mutation causing a large destabilization and therefore not contributing directly to protein interactions as discussed above. A large ΔΔG_binding change does not necessarily mean a residue solely contributes to interaction energy; instead, it may play a crucial role in maintaining structural stability, indirectly influencing binding interactions. This suggests that the residues may have been involved in stabilizing interactions (hydrogen bonds, salt bridges, van der Waals contacts, bonds with metals) that are lost upon Alanine substitution, which results in significant changes in binding energy via destabilization for the mutated subunit of the complex.

To see the relative contribution of both reasons, we analyzed the change in stability of the protein upon mutation in the absence of its partner (ΔΔG_stability) using FoldX. The distribution plot (Figure 3d) compares the FoldX ΔΔG_stability distributions for false‐negative HS that were not initially identified but are experimentally validated as binding‐relevant. The distributions are plotted based on the sign of FoldX ΔΔG_stability for the 177 FN cases and are divided in two: mutations that result in stabilization (negative ΔΔG_stability) and destabilizing ones (positive ΔΔG_stability). Although not majorly, many of the FNs fall into the destabilizing category, suggesting that their role extends beyond binding affinity. Thus, these residues will not be detected by our algorithm that focuses on interaction‐driven HS rather than those that primarily contribute to structural integrity. The extreme cases where FoldX ΔΔG_Stability >4 kcal/mol (Figure 3d) revealed 2 metal‐chelating residues (Histidine and Aspartate), an occluded Glycine that, when mutated to Alanine, introduced large Van der Waals clashes and a fully buried tryptophan. Obviously, these residues are going to be considered as experimental HS, but their function is mostly structural in one of the subunits of the complex.

Thus, a significant number of FN corresponds to mutations that destabilize the subunit of the complex where the residue is located and thus affecting binding energy, plus those not found probably due to lack of cases in our database.

To benchmark our method against other state‐of‐the‐art hotspot prediction tools, we selected the BID18 dataset used in Wang et al. (2014) to compare most of the tools of the state of the art as a common benchmark. BID18 contains 125 interface residues, with 38 HS and 87 non‐HS taken from 18 protein complexes.

In our performance benchmarking using the BID18 dataset, we systematically varied the distance threshold for identifying HS residues and then compared the resulting metrics across multiple thresholds (see Supplementary Figure 5). We found that at 1 Å, the prediction performance reached an optimal balance between a high precision and a favorable MCC value. Then we evaluated performance across standard metrics (accuracy, recall, precision, specificity, sensitivity, and MCC) as shown in Supplementary Table 3, comparing it against established hotspot prediction methods. The results illustrated by Supplementary Table 3 placed HotspotPred among the top‐performing methods in terms of balanced precision and sensitivity, also achieving the highest MCC value.

2.6. HS prediction in nanobody binders

Once we have shown that there is information in the triplets regarding stability and binding, we proceeded to predict key binding residues in nanobodies using the structural constraints embedded in the structure. To validate our HS prediction algorithm, we used a dataset of 123 non‐redundant nanobody (Nb)‐target complexes (Zavrtanik & Hadži, 2019). All the 123 complexes used in this analysis were excluded from the digestion of our triplet database.

The method follows a structured pipeline illustrated in Figure 4 to identify important amino acids on a target surface. Once we have retrieved all triangles that could be superimposed on the two selected anchoring residues from a nanobody (see “Methods”), we refine the obtained cloud of floating residues (non‐anchors) using the same process defined in the validation on SKEMPI. Then we go through the structure of the nanobody target in complex with the nanobody and select, for each target residue, the list of floating ones that are found in a distance of 3.5 Å (Cα–Cα). If we manage to find in the selected cloud a residue that has the same identity as the target residue in question, then we mark it as HS.

FIGURE 4.

Stepwise execution of the hotspot prediction algorithm. Illustration of the computational pipeline for predicting binding HS. The process begins with the selection of anchoring residues on the input nanobody based on structural constraints. As indicated above, the retrieved triplets are structurally aligned to the input nanobody structure by superimposing the α‐carbons of the anchoring residues with their corresponding residues in the triplet (left). After the superimposition process, the third residue remains unconstrained structurally and is labeled as floating. After performing multiple queries all along the interface of the protein, a cloud of floating residues accumulates, representing plausible hotspot candidates (right).

We compared our predicted HS residues to variations in interaction energy (ΔΔG values) obtained through the FoldX AlaScan command that mutates every residue to Alanine. The true labels were assigned based on the FoldX‐calculated energetic impact of Alanine mutations on binding energy, where residues with a ΔΔG value above 2 kcal/mol (Chen et al., 2024) were considered HS.

We determined the recall, the accuracy, and the specificity as functions of the Cα–Cα distance threshold in the Nanobody‐target dataset (Figure 5a). As in the SKEMPI database (see Supplementary Figure 4) we found that increasing the Cα–Cα distance threshold from 1.0 to 3.5 Å improves recall from about 0.42%–0.55%. Both specificity and accuracy remain relatively high but decrease slightly with the distance threshold. The algorithm achieves high accuracy in detecting non‐hotspot residues (true negatives: 57% of the total number) (see Supplementary Figure 8A), demonstrating that it effectively excludes non‐functional residues from predictions. We identified 42 residues, while 34 HS predicted by FoldX were missed, suggesting that certain energetic HS may not be captured due to lack of data in the TriXDB_20K. Regarding the 385 false positives (see Supplementary Figure 8B) when we plot their FoldX ΔΔG_interaction energies, we find, similarly to what we observed in the analysis of the SKEMPI database, that almost 23.64% of them are predicted to increase ΔΔG_binding by more than 0.8 kcal/mol (error margin in FoldX predictions (Delgado et al., 2025)) indicating that some FP residues may not be HS residues but ND. While these ND residues may not contribute strongly to the binding energy in their optimal configuration, their presence could play a critical role in the specificity of the binding. This means that even residues with moderate or low ΔΔG_interaction can still be functionally significant, as they help maintain specificity by discouraging unfavorable conformations of the binding interface. In nanobody‐target interactions, these residues may act as structural support elements that determine the orientation of critical binding residues, ensuring that the interaction remains strong.

FIGURE 5.

Combined hotspot prediction validation. (a) Performance metrics versus distance threshold (same residue identity). This panel shows the variation of recall, accuracy, and specificity for hotspot prediction as a function of the Cα–Cα distance threshold (1.0–3.5 Å) using exact residue identity. (b) True positive rate versus ΔΔG threshold. This plot shows how the TPR (sensitivity) changes as the ΔΔG threshold for labeling a residue as a hotspot increases. Higher thresholds focus on stronger interaction contributions, thereby increasing the fraction of hotspots correctly identified (true positives), albeit with fewer total positives. (c) Performance metrics versus distance threshold (same polarity group). For this analysis, a true positive is found when at least one residue that belongs to the same group as the target residue is found within a 3.5 Å distance cutoff and a ΔΔG_interaction threshold is set to 2 kcal/mol. (d) Hotspot distribution across targets. A bar plot representing how many residues are predicted per PDB structure.

To further assess the performance of our predictions, we evaluated the TPR as a function of the ΔΔG threshold (Figure 5b). For 2 kcal/mol, this number is around 55%. As the ΔΔG threshold increases, indicating stronger energetic contributions to binding, the TPR rises accordingly, reaching a peak close to 100% at high ΔΔG values. This positive correlation suggests that our method successfully prioritizes residues contributing significantly to interaction stability. The observed increase in TPR with ΔΔG reinforces that our model is able to capture highly destabilizing mutations, making it a reliable tool for identifying functionally important binding HS in Nb‐target interactions.

Regarding the FNs, it could be as discussed above that our database is not complete, and we could be missing some triangles. To see if this is the case, we grouped the amino acids by chemical properties and repeated the analysis of the different metrics done in Figure 5a. In this case, if the predicted residue should be a tyrosine, for example, but we find a phenylalanine, we will consider the prediction correct (Figure 5c). By doing this, we decrease significantly the number of FNs while increasing that of true positives. In fact, the recall value reaches 0.73 for a 3.5 distance; thus, supporting the hypothesis that some of the FNs are due to missing information in our database. Regarding the number of false positives, we also see an increase, which affects both the accuracy and the specificity, but the relative proportion that FoldX predicts to contribute to binding energy remains almost the same 21.98% of the total count FP.

To further assess the robustness of our approach, we examined the distribution of predicted hotspot residues per nanobody structure (Figure 5d). The results show that in most cases, at least two residues were successfully identified per target (see Figure 6 for an example). These results are valuable because, in computational docking, two correctly identified residues on a target structure can provide valuable constraints for refining docking solutions.

FIGURE 6.

Stepwise prediction of HS on a nanobody target structure (RCSB code: 5BOP). (a)Is the input structure of the nanobody in “Cyan.” (b) After selecting the anchors and starting the search for triplets, a cloud of floating residues forms around the nanobody's binding interface and undergoes a spatial clustering of 2A and the residues that cause steric clashes with the input structure are removed. (c) The structure of the original binder or target is reintroduced in “magenta.” (d) The residues that are found in a distance of 3.5A from any residue in the target (Cα–Cα distance) are selected in “red.” In panel (e) and (f) we are selecting four cases of overlap between α‐carbons coming from floating residues on Cα of residues Glutamate 155, Aspartate 152, Phenylalanine 46 and Histidine 39 from the target. We colored in “green” the residues from the target that were predicted as true positive HS and in “blue” the residues set as false positives; the ones that our algorithm predicts as HS but they do not pass the threshold of FoldX ΔΔG_interaction and that could be ND residues.

3. DISCUSSION

Our method presents a novel approach to predict positions on the surface of a protein where HS residues could dock, and this could help dock a target protein with a specific focus on HS for nanobody binding. It can also be used to identify possible mutations of residues involved in the interaction of two proteins.

While we demonstrated success in identifying HS residues, several considerations need to be presented.

3.1. Limitations of experimental ΔΔG interpretation

A fundamental challenge in the validation of computational methods is the interpretation of changes in experimental binding energy values upon mutation (Geng et al., 2016). The values of ΔΔG do not often distinguish between the contribution to binding itself or to the structural stability of the residue being mutated. Thus, a change in the free energy of a complex upon mutation may result from an alteration in binding and/or structural stability. This limitation could lead to a misclassification of residues as false positives when a mutation significantly destabilizes the corresponding protein, as we witnessed in our analysis.

3.2. Limitations of the triplet‐motif

Another inherent limitation of our triplet‐based approach is the dependence on the three‐residue configuration. While this motif enables the detection of multi‐residue interaction patterns, it may fail to identify functionally important interactions that arise from simple pairwise contact, particularly when only two residues are responsible for the contribution to affinity, either by intermolecular interaction or by complex stabilization. This structural constraint could lead to missed predictions of true HS that fall outside of a triplet‐compatible geometric context.

3.3. ΔΔG threshold

Another explanation of the misclassification of some residues as false positives is the commonly used threshold of 2 kcal/mol to define a HS residue. This threshold is arbitrary and may overlook residues that introduce significant changes in ΔΔG_binding and could play crucial roles in binding specificity. These residues that we have defined as ND residues participate in the binding and could be essential to prevent non‐specific interactions and off‐target binding.

3.4. Sequence similarity filtering and data representation

To construct our TriXDB_20K, we applied a clustering step based on the percentage of sequence identity, discarding sequences sharing more than 30% similarity. This approach is good at avoiding redundancy; however, it results in reduced datasets and a limited representation of HS interactions. This is supported by the fact that when grouping the amino acids by chemical and size properties in the case of the prediction of HS in the nanobody dataset, we observe a significant improvement in the predictions. To address this limitation, future work could apply similarity filtering at the triplet level instead, allowing the inclusion of more structures with higher sequence identity, thus increasing the number of different triplets while omitting redundancy.

Something we have not tested here but could be extremely useful is to look not only at the position of the Cα of the predicted binding residue, but also at the interaction of the side chain with the anchoring residues. Correct positioning of the side chain will mean that we just need to explore different rotamers to explore different positioning of the Cα and Cβ and use the generated cloud to search for that residue in the target protein.

4. CONCLUSIONS

We introduced a novel computational approach for predicting possible docking of hotspot residues, which could be used for engineering PPIs and could be applied to antibody design. Unlike traditional methods, our tool retrieves triplet‐based structural motifs to infer HS locations.

Our method successfully predicts binding HS residues in known PPI datasets. Given that the definition of HS is not very well defined, but according to energetic thresholds that are commonly used (2.0 kcal/mol, citation), they must be composed in many cases of more than one pairwise interaction. This makes the triple residue fragment approach able to capture these multiresidue interactions.

Our method, which prioritizes direct interaction‐based hotspot detection, naturally omits stabilizing residues, which could lead to FNs when compared with experimental energies. Conversely, the presence of false positives suggests that some residues with significant interaction contributions may be functionally relevant even if they do not exceed HS ΔΔG_binding threshold.

Future improvements could integrate stability considerations to refine predictions and distinguish between residues essential for direct binding and those that exert their effects through global structural stabilization.

5. MATERIALS AND METHODS

5.1. FoldX RepairPDB

All the structures used for the validation were energetically minimized with FoldX (Delgado et al., 2019) RepairPDB.

5.2. FoldX AlaScan

All the amino acids on the target structure were mutated to Alanine using the FoldX AlaScan command executed after the RepairPDB command. Then, we used the Analyze complex command to obtain the interaction energy between the two molecules.

5.3. HotSpotPred

HotspotPred is a portable command‐line tool included in the ModelX toolsuite, written in C++, that employs SQL queries to retrieve triplets that are saved in a relational database delivered with a template SQL dump called TriXDB_20K.

5.4. Random forest

The initial dataset of amino acid substitutions was randomly split into training (80%) and testing (20%) subsets (Supplementary Table 1). The model evaluation was performed using a 10‐fold cross validation on the training set to ensure generalization and robustness of the used features. The model used a fixed number of 100 trees in the forest and a minimum of five samples per leaf node. This regression included two sets of features (Supplementary Table 2). (1) Amino acid type features: The molecular mass, volume, and number of atoms of both mutant and wild type residues. These features capture how the change in the amino acid's size and composition could influence local packing and have an energetic effect on a complex's integrity. (2) Features extracted from the triplet frequencies: The average of the ratios involving all triplets with a common position (avgRatioWt and avgRatioMt) which captures the distribution of geometric relationships across the retrieved triplets for both the wild type residue and the mutant one; the minimum ratio of all triplets retrieved for a single substitution (minRatioWt and minRatioMt) (see Figure 2) which indicates the most geometrically constrained interactions. Additionally, we included the maximum ratio of all triplets retrieved for a single substitution, which describes the least constrained or most common interactions.

These statistical terms were included in the training to capture the structural characteristics of the environment surrounding the substitution and also to evaluate its effects on the stability of the structures. No direct computation of ΔΔG from the triplets' ratios was performed. Instead, ΔΔG values were predicted relying on the model's understanding of the complex relationships between the input features to infer the stability changes upon substitution.

Each decision tree in the forest out of the 100 ones is trained on a random subset of the data and features, then outputs a predicted value of ΔΔG for a given substitution. The final prediction is obtained by averaging all the outputs across all trees. The mathematical formulas of all the ratios used are provided in Supplementary Note 1.

5.5. HS prediction algorithm

1. Selecting anchoring residues

Anchoring residues are chosen to define the local binding environment and constrain the search space for structurally relevant interactions. Selecting optimal anchors ensures the retrieval of meaningful triplet interactions while avoiding structurally irrelevant residues. The user passes the list of the residues that could be involved in the binding interface on the input structure surface and that serve as anchoring points. In the case where the list is not passed (in our validation set) as an argument, the algorithm does an automated search on the input structure for any residues with at least one backbone atom distant by 7.5 Å or less from any backbone atom on a residue from the target. The choice of these residues is critical for the query process since it serves as structural references to constrain the search space.

• 2Querying TriXDB_20K for structurally similar triplets

For each chosen pair of anchoring residues, corresponding triplets are retrieved from the TriXDB_20K. Each retrieved triplet contains two residues that match the sequence of the original anchors. Structural flexibility is allowed within defined tolerances to account for natural variations in protein conformation. These tolerances include (1) Cα–Cα and Cβ–Cβ distances: In both of our validations we chose dubieties Cα–Cα = 0.5 and Cβ–Cβ = 0.4; (2) dihedral angles between interacting residues equal to 25° (allowing flexibility in backbone orientation); (3) backbone torsional angles (φ, ψ) with a dubiety of 25°.

Each retrieved triplet leaves the third residue (the one not corresponding to any of the anchors) unconstrained.

• 3Superimposition and floating residue cloud formation

The retrieved triplets are structurally aligned to the input nanobody structure by superimposing the α‐carbons of the anchoring residues with their corresponding residues in the triplet (see Figure 4). After the superimposition process, the third residue remains unconstrained structurally and is labeled as floating. After performing multiple queries all along the interface of the protein, a cloud of floating residues accumulates representing plausible hotspot candidates. The cloud then undergoes a spatial clustering process and all the residues resulting in backbone steric clashes with the input structure are eliminated. Then we extract frequency terms from the resulting clusters to assess the retrieved residues.

• 4Hotspot prediction validation and ranking

The original target structure is reintroduced, and the floating residues are analyzed in a user‐defined distance threshold. If we find the WT residue in a cloud at a distance less than 3.5 Å from the original residue of the interacting structure, we define this as a HS.

5.6. ClusterPdbs

To refine the floating residue clouds in the predicted structures, we applied a distance‐based clustering approach using the ClusterPdbs command within ModelX. This function processes clouds of floating residues resulting from our prediction by computing pairwise Cα–Cα distances between them. The clustering process begins by identifying the residue with the smallest average Cα–Cα distance to all others, designating it as the first cluster centroid residue. Additional amino acids are assigned to this cluster if their distance falls below a user‐defined threshold (in our case 2.0 Å) (see Supplementary Figures 2 and 3). This process is iteratively repeated until all floating residues are assigned to a cluster. After clustering, the resulting clusters are sorted by size, and only the most significant clusters (containing more than three elements) are retained for further analysis.

AUTHOR CONTRIBUTIONS

Rahma Hamdani: Writing – original draft; conceptualization; investigation; methodology; validation; software; data curation; visualization; writing – review and editing; formal analysis. Damiano Cianferoni: Conceptualization; methodology; writing – review and editing. Raul Reche: Software. Javier Delgado: Conceptualization; investigation; writing – original draft; writing – review and editing; methodology; software; data curation; supervision; project administration. Luis Serrano: Conceptualization; investigation; funding acquisition; writing – original draft; writing – review and editing; methodology; software; formal analysis; project administration; supervision; resources; data curation.

FUNDING INFORMATION

This publication is part of the grant PRE2022‐101389, funded by MICIU/AEI /10.13039/501100011033 and by the ESF+, and of the project PID2021‐122341NB‐I00, funded by MCIN/ AEI / 10.13039/501100011033 / FEDER, UE). This project has also received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement No 101020135).

Supporting information

Supplementary Figure 1: A scatter plot of the predicted ΔΔG and the experimental one from the training test. Pearson correlation of 0.86 and an MSE of 0.77.

Supplementary Figure 2: Box plot of the different F1 scores computed varying the clustering distance parameter ran on the nanobody dataset. We tried running the spatial clustering command using 4 different cluster distances that reflect the distance of the center of each cluster to the rest of the elements.

Supplementary figure 3: Comparison between two clustering distances using a heatmap of the true positive rates variations. (A)True positive rates variation with different interaction energy threshold for a clustering distance of 1.5A. (B)True positive rates variation with different interaction energy threshold for a clustering distance of 2A.

Supplementary Figure 4: Plotting the recall, the accuracy and the specificity as functions of the Cα–Cα distance threshold in the SKEMPI database.

Supplementary figure 5: Performance Metrics Across Varying Distance Thresholds for the BID18 dataset analysis.

Supplementary figure 6: Structural context of a false positive hotspot prediction (PDB ID: 1a22). Tyrosine 42 was predicted as a hotspot by our method due to its direct π‐stacking and hydrogen bonding interactions across the interface. However, it does not meet the experimental ΔΔG threshold for classification as a hotspot (ΔΔG = 0.2 kcal/mol).

Supplementary figure 7: Structural context of a false negative hotspot prediction (PDB ID: 1emv). Valine 34 is an experimentally validated hotspot contributing to a hydrophobic cluster that stabilizes the interface. Although our method did not predict it as a hotspot, the experimental ΔΔG is significant (2.58 kcal/mol).

Supplementary figure 8: Analysis of HS prediction on Nanobody‐target dataset database. (A) Confusion matrix comparing HotspotPred determined hotspots and FoldX predicted ones. True hotspots (HS) were defined based on a FoldX ΔΔG_interaction threshold of 2 kcal/mol. (B) A bar plot showing the distribution of ΔΔG_interaction values for the false positive subset.

Supplementary Table 1: Table illustrating the point mutations used for the validation, the associated features, the experimental values and our predicted ones.

Supplementary Table 2: Table illustrating all the features extracted from the different point mutations including the triplet‐based ratios and the amino acid related terms.

Supplementary Table 3: Comparison of different hot spot prediction methods in the independent test set BID18.

Hamdani R, Cianferoni D, Reche R, Delgado J, Serrano L. Predicting interacting hotspots for nanobodies' binding using triplets of residues. Protein Science. 2025;34(8):e70220. 10.1002/pro.70220

DATA AVAILABILITY STATEMENT

The data that supports the findings of this study are available in the supplementary material of this article.