Automatic Explanation of Protein-Protein Binding Mechanism: A Preliminary Study

Abstract

Understanding the biochemical mechanisms that drive protein-protein interactions is a challenging task, traditionally requiring mutation studies and expert interpretation of protein structures. A method that can generate mechanistic explanations from the biochemical properties and contributions of interactions would enhance our ability to study protein-protein interactions. In this study, we present a novel approach to interpreting mechanistic insights from machine learning methods; we manually annotated a dataset of 1225 mutation experiments with mechanistic insights focused on electrostatic, hydrogen bonding, steric and hydrophobic interactions. To show a preliminary process for evaluating mechanism prediction models, we extracted SHAP features that are representative of protein binding mechanisms from a Gradient-Boosting Tree (GBT) model trained to predict binding affinity. We found that the SHAP values generally agreed with the annotated mechanisms from our dataset, especially when looking at electrostatic and steric features. We also found that hydrophobicity consistently played a dominant role and hydrogen bonds consistently played a secondary role, challenging conventional assumptions about the role of these interactions.

1. Introduction

Interacting proteins play a central role in organizing the functions of the cell. They form complexes that regulate the cell cycle [9], present antigens for the adaptive immune system [4], influence the plasticity of the nervous system [20], and many other important tasks. In order to reverse engineer these interactions for the study of healthy systems, or to create interventions for unhealthy ones, it is essential to discover the biochemical mechanisms that underpin selective binding. To this end, the traditional approach has been to perform mutations of amino acids at the interface and infer the role they play in binding. This process is expensive and time consuming, because there are many amino acids involved and multiple chemical interactions to consider.

To hasten this approach, categories of software are used to help select amino acids for mutation. Structure prediction (e.g. [1,2]) and docking algorithms (e.g. [13,10]) predict the structure of proteins in complex in order to illustrate which amino acids are present at the interface and what interactions they form. When using these tools, researchers visually examine the predicted complex and imagine how the biochemical mechanisms interact, building insight into the how the binding mechanism might work. Molecular simulation tools (e.g. [5,7]) form a second category of software that offer insights into how protein structures change in order to recognize each other. Careful scrutiny of simulated motions offers insight into how specific amino acids interact with each other and a second distinct perspective on the possible roles they play in binding.

These are powerful methods, but they rely substantially on the capacity of the human user to integrate their outputs into models of interacting biochemical mechanisms. When there are many variations, or many mechanisms involved, the number of possibilities can easily exceed human attention. To examine this problem, this paper explores the possibility of creating structural analyses that infer the mechanistic role of amino acids directly. Our approach is to interpret the features of a machine learning method that was originally designed to predict binding affinity between two proteins. Rather than examining the accuracy of the binding affinity predictions, which we do elsewhere [18], we asked whether or not accurate explanations could be generated for the biochemical contributions of specific amino acids.

Evaluating this question requires a ground truth dataset that describes the biochemical role of individual amino acids and maps it to a discrete list of mechanism. This could simply be a list of tags that indicate, for example, that amino acid X plays an electrostatic role in binding, or that amino acid Y forms hydrogen bonds to support binding. While the diversity of structural biology will always exhibit novel mechanisms for molecular function, these basic mechanisms can form the building blocks larger interactions. This information is common in the structural literature, but it exists as freeform text and is not organized as a ground truth dataset that can be leveraged for explanations.

For this reason, one major contribution of this paper is to begin creating that dataset. Our dataset was constructed from a subset of the SKEMPI 2.0 database [11], which connects the structures of experimentally determined protein complexes with the binding affinity of different mutations of amino acids in the complex [11]. We manually examined the literature relating to these structures, identifying specific quotes that document the mechanisms in which many amino acids are involved. In this work, we tagged mechanisms that are documented to have [steric], [electrostatic], [hydropathic], [salt bridge] or [hydrogen bonding] contributions to binding affinity, and we denote them with square brackets. This dataset, while already sizable, remains incomplete, as we continue to add additional literature and additional mechanisms. Currently, we have annotated 1225 experiments from the SKEMPI dataset and aim to complete 2639 annotations in total.

Our explanations are generated using SHAP (SHapley Additive exPlanations) values [19], which offer a way to explain the importance of different feature groups used by a Gradient-Boosting Tree (GBT) algorithm. A number of GBT-based algorithms are compatible with this purpose, including topNetTree[21], GeoPPI[17], MechPPI[18], and MuToN[16]. In this work, we use MechPPI, and we generate SHAP values from steric, electrostatic, hydropathic, and hydrogen bonding feature groups computed for a mutant. We treat high absolute SHAP values for a given feature group (e.g. the electrostatic group of features) as being an indicator that the mutated amino acid influences affinity through that group’s mechanism (e.g. an electrostatic influence). Measuring the correlation between SHAP values and our literature annotations, we observed numerous correlations, indicating that this strategy for explaining binding mechanisms can be effective, based on our well defined ground truth data.

Our method is the first technique, to our knowledge, for automatically generating explanations of the biochemical role of specific amino acids in protein-protein interfaces. In our results, we observed varying degrees of accuracy in identifying mechanisms of different types, and correlations between some mechanisms that have similar physical foundations. Two examples from our detailed literature search illustrate these relationships. Overall, these findings suggest that structure based features can indeed be used to generate explanations about the roles of amino acids.

2. Methods

2.1. Manual annotation process

The labeling of SKEMPI mutation mechanisms was carried out through a manual, multi-step process to ensure high-quality and reliable mechanism information. We first verified that the mutant in SKEMPI was accurately referenced in the literature. The authors then reviewed the papers that were referenced by SKEMPI and the PDB [3], seeking mechanistic information about the amino acid in question. Generally, the amino acid that was mutated because it’s biochemical role was being established experimentally. The authors then annotated the role of the amino acid with tags representing biochemical interactions, which in this work are [hbond], [electrostatic], [salt bridge], [steric], or [hydrophobic]. These labels for experimentally documented interactions are indicated using square brackets (e.g, [hbond], [electrostatic]) throughout this paper, so that they cannot be easily confused with mechanistic predictions with similar names. Note that we have distinct annotations for [salt bridge] and [electrostatic] mechanisms despite their electrostatic similarity; we define them separately due to the notable differences in effect between short range salt bridge interactions and general long range ionic interactions [15]. For each entry, the final mechanism annotation includes: PDB structure, mutated amino acid, mechanism tag, relevant quotes taken directly from the referenced literature, as well as the paper reference for traceability. While this process is extremely time consuming, it ensures high quality annotations with well defined mechanistic information.

2.2. SHapley Additive exPlanations (SHAP values)

Prediction models often use numerous features, each contributing towards the final output prediction. However, the degree of contribution from each feature can be difficult to decipher from the final prediction. To quantify these contributions, we used SHapley Additive exPlanations (SHAP), a method that assigns an importance value to each feature based on its contribution on the model’s prediction. These values were extracted from a pre-trained regression model which predicts binding affinity of protein complexes [18]. To determine the importance of a feature j, we calculate the marginal contribution - the difference in the model’s prediction when the feature is included versus when it is excluded.

The marginal contribution of feature j depends on the subset of features to which j is added (subsets of size 1, 2, 3, …, up to N). SHAP values are calculated by determining all possible combinations of feature subsets, computing the total marginal contribution of feature j for each subset, and finding the average of these contributions. The average contribution is weighted based on the number of subset permutations, ensuring equal distribution of feature contributions across all possible combinations. This results in a SHAP value that represents the average contribution of a feature across all possible combinations. The SHAP value calculation can be formally expressed as:

$${\varphi }_i=\sum _{S\subseteq N\backslash \left\{i\right\}}\frac{\left|S\right|!(N-|S|-1)!}{N!}\left[f_x(S\cup \{i\left\}\right)-f_x\left(S\right)\right]$$ (1)

Here, ${\varphi }_i$ is the SHAP value for feature i,S is the subset of features from all features N except for features i,N is the total number of features, for each feature i, we need to consider all possible subsets S of features that do not include i $(S\subseteq N\backslash \{i\})$, $\frac{\left|S\right|!(N-|S|-1)!}{N!}$ is the weighting factor counting the number of permutations of the subset S, $f_x\left(S\right)$ is the prediction when only the features in subset S are present, $f_x(S\cup \{i\left\}\right)$ is the prediction when feature i is added to subset S, marginal contribution of feature i is $f_x(S\cup \{i\left\}\right)-f_x\left(S\right)$, it represents the change in the prediction when feature i is included in the set compared to when it is excluded. So the SHAP value ${\varphi }_i$ for feature i is the sum of its weighted marginal contributions across all subsets.

After obtaining SHAP values of all features, we can use SHAP values to explore the most influential mechanisms from a mutation on the binding affinity of protein complexes. The larger the absolute SHAP value a feature has, the greater the contribution that feature has towards the final prediction.

2.3. Interactions represented by SHAP Values

In this work, we divide the features of MechPPI into four groups: electrostatic complementarity, hydrogen bonding, steric interactions, and hydrophobic interactions. One SHAP value was computed for each group. Below, we describe some specific details also covered by SHAP values within each interaction and specify the number of SHAP features assigned to each group.

For our electrostatic feature group, our GBT method uses VASP-E to measures electrostatic complementarity by using Constructive Solid Geometry (CSG) operations to generate solid representations of electrostatic fields [8]. This approach generates regions of electrostatic complementarity by computing the over-lapping volume between multiple positive and negative isopotentials from interacting molecules. This method is particularly useful in finding bond interactions like salt bridges which innately involve oppositely charged residues.

For our hydrogen bonding feature group, we identify the gain and loss of hydrogen bonds at the binding site and at the binding interface. We count the number of backbone-backbone, backbone-sidechain, sidechain-sidechain, and sidechain-backbone hydrogen bonds. Finally, we use a volumetric analysis that represents the range of bond angles and bond distances accessible to a hydrogen bond [6].

We also have a steric feature group that uses a solid volumetric representation of the interface molecular surface in a protein protein interface. It also incorporates the volume of the mutated residue and distance of mutated residue to the binding partner. The SHAP value that represents this group helps capture the contribution of steric hindrance and Van der Waals complementarity towards protein-protein interactions.

Hydrophobic interactions are known to provide substantial stability for protein binding [12]. In our hydrophobic feature group, we describe hydropathy environments using counts of neighboring hydrophobic, hydrophilic, and neutral residues to the mutant residue.

2.4. Evaluating SHAP feature contribution to Mechanisms

We developed a systematic process for selecting the top-contributing mechanism group to a protein-protein interaction based on SHAP values, and use these to compare contributions of each mechanism group to the recorded mechanistic interaction observed in SKEMPI proteins.

Aggregating SHAP Features

The 97 SHAP features i were divided into four categories based on how each SHAP feature is representative of the physicochemical interaction mechanisms discussed above, which we call mechanism groups. The SHAP values associated to each SHAP feature are then aggregated by mechanism group by taking the absolute sum A; this is done separately for each mutation. This ensures that both positive and negative contributions are captured equally, reflecting the magnitude of impact of all SHAP values for each mechanism group.

$$A_{mut}=\sum _{i\in X}\left|{\varphi }_i\right|,\text{for}X\in \{E,H,S,P\}$$ (2)

Counting Top Mechanism Groups

For each mutation, we calculated an aggregate SHAP value A_X for each major mechanism group. We used these aggregate SHAP values to produce two mechanism prediction strategies. The first strategy finds the mechanism group that has the highest aggregate SHAP value; we predict that the mechanism of the mutated amino acid is that of the mechanism group that has the highest absolute SHAP value and we call this the “top 1” mechanism $G_{top1}$. For example, if, for a given amino acid, the SHAP values associated with electrostatic complementarity has the highest aggregate SHAP value, then the “top 1” prediction strategy predicts that the amino acid in question influences binding via an electrostatic complementarity mechanism.

$$G_{top1}=\mathrm{a}\mathrm{r}\mathrm{g}\underset{X\in \{E,H,S,P\}}{\mathrm{m}\mathrm{a}\mathrm{x}}{A}_{mut}$$ (3)

Our second strategy makes the same prediction as “top 1” and also finds the mechanism group with the second highest SHAP value; the mechanism group of both the highest and second highest SHAP value we call the “top 2” mechanisms. With the “top 2” mechanisms, we predict that the primary and secondary mechanism of the mutated amino acid is that of the “top 2” mechanism groups. In this scenario, it might predict, for example, that the amino acid acts via an electrostatic mechanism as the top 1 primary mechanism and hydrogen bonding as the top 2 secondary mechanism.

In our results, we illustrate how often these predictions coincide with tagged annotations from SKEMPI. The five tagged annotations are [electrostatic], [hydrogen bond], [steric], [hydrophobic], and [salt bridge]. We specifically evaluate salt bridge annotations because of its high degree of strength compared to long-range electrostatic interactions, its innate electrostatic properties, as well as its co-occurrence with hydrogen bonds and electrostatic interactions. As a result of these steps, we obtain a count of how many times each major mechanism group had the top 1 or top 2 SHAP value across the subset of mutations labelled with the five interaction mechanisms mentioned. The count is divided by the total number of mutations for each mechanism subset and is presented in the results.

3. Results

3.1. Annotating mechanisms from SKEMPI

The SKEMPI dataset maps mutations in protein-protein interactions (PPIs) to changes in binding free energy, relative to a wild type complex. While SKEMPI provides extensive mutation data, it lacks detailed mechanistic information on the biochemical interaction that occurs. To bridge this gap, we undertook a manual annotation process to supplement the dataset with mechanistic insights.

First, we identified 1225 distinct mutation experiments from 121 proteins described in the SKEMPI dataset. For each amino acid mentioned in SKEMPI, we manually searched through the literature references by SKEMPI and the Protein Data Bank, identifying any specific interactions related to the amino acid that could explain how the mutant affects the protein-protein interaction.

From the 1225 mutation experiments, we removed several proteins where mechanistic information are difficult to extract. 374 mutations had no discussion of mechanistic information about the amino acid. The majority of these mutations came from alanine-scanning experiments which did not consider the structural role of individual amino acids. Another 39 un-annotated mutations were from T-cell receptors (TCR) interacting with a peptide in the major histocompatibility complex (pMHC). We found that most papers focused their discussion on the general TCR/pMHC/ligand complex rather than individual amino acids and interactions, and so we removed all TCR/pMHC mutations. When looking at just the five mechanisms tagged in this work, our manual annotation effort resulted in a total of 819 amino acids with documented mechanisms. The composition of these annotations is shown in Table 1.

Table 1.

Count of occurrences for each biochemical mechanism.

Mechanism	hbond	hydrophobic	salt bridge	electrostatic	steric
Count	470	139	101	79	30

Note that this dataset of biochemical mechanisms is still a work in progress. Our subset consists of a sequence non-redundant selection of the SKEMPI dataset, reducing biases from highly similar sequences and ensuring diversity in the biochemical interactions studied, but quite a few proteins have not yet been annotated. Nonetheless, the number of annotations is not small, so we feel that this initial study could provide some insight into the value of producing the whole dataset.

Altogether, we annotated mechanisms for 819 mutants and collected the SHAP values for those same mutants to produce explanations of their binding mechanisms.

3.2. Comparing SHAP features with mechanism annotations

Here, we evaluate how well the SHAP values from our dataset predict annotated binding mechanisms. The SHAP values were aggregated into four major mechanism groups: electrostatic, hydrogen bond, steric, and hydrophobic interactions. Next, we performed both the top 1 or top 2 prediction strategy for each mutation in our SKEMPI subset. These selections were used to count the proportion of the top contributing mechanism and evaluated if mechanism annotations of the SKEMPI data agreed with the top contributing mechanism.

Two sets of results are presented in Fig. 2. The top 1 predictions are shown in the bottom segment of each bar (dark colors), while the top 2 predictions are shown in a long bar behind them (light colors). Since top 2 predictions strictly contain all predictions made by the top 1 strategy, the top 2 bar must always be at least as tall as the top 1 bar. Each bar is presented as a proportion (left axis) of the dataset.

Fig. 2.

The proportion of top 2 counts for each mechanism tag, showing the combined contribution of both dominant and secondary mechanisms. Dark color bars can be interpreted as the proportion of annotated mutations which have the mechanism group as the top 1 SHAP value. Light color bars are the proportion of annotated mutations which have the mechanism group as the top 2 SHAP value.

Top 1 Mechanism Group Analysis

The bottom segments (dark colors) in Fig. 2 show the proportion of how frequently a mechanism group has the highest SHAP value G_top1, and therefore plays the most dominant role in prediction for each annotated mechanism tag. For example, when observing [hbond] labelled mutations (dark blue) in the “Electrostatic” mechanism group (leftmost column grouping), we see that for mutations tagged as involving hydrogen bonding, around 27% of these mutations have electrostatic as the dominant mechanism group based on SHAP value. In comparison, around 58% of mutations tagged as involving [salt bridges] (dark green) have electrostatic as the dominant mechanism group. This is consistent with the involvement of electrostatic interactions in salt bridge formation, as opposed to hydrogen bonds which involve only partial charges.

In general, we observe substantial agreement between SHAP mechanism groups and our mechanism annotations. For [salt bridge] and [electrostatic] tagged mutations, the electrostatic mechanism group is most frequently the dominant mechanism, consistent with expectations. Similarly, [steric] and [hydrophobic] tagged mutations are well-aligned with the steric and hydrophobic mechanism groups. Surprisingly, the hydrogen bond mechanism group appears less frequently as the top 1 feature for [hbond] tagged mutations, where the electrostatic mechanism was more prevalent than expected.

Top 2 SHAP Feature Analysis

The lighter colored columns in Fig. 2 show the proportion of major mechanism group contributions when the top 2 mechanism groups G_top2 within each category are selected for each mutation. This figure provides insight into the secondary contributions of mechanism groups, complementing the top 1 analysis. In general, the top 2 mechanisms show slightly better alignment with our expectation for the relationship between SHAP values and annotated mechanism tags than for top 1. For example, hydrogen bonds are more frequently a top 2 contributing mechanism for [hbond] tagged mutations, as opposed to the top 1 where hydrogen bonds were significantly lower in proportion than electrostatic for [hbond] tagged mutations; similarly, the SHAP value for the steric mechanism group are better represented for [steric] tags.

The electrostatic mechanism group continues to play a significant role for mutations labeled as involving salt bridges, with more than 70% of these mutations having electrostatics among the top 2 mechanism groups. This reinforces the earlier observation of electrostatics in salt bridge formation. Interestingly, hydrogen bonding and “hydroneutral” mechanism contributions had the greatest increase across most tags in the top 2 analysis, suggesting a more subtle role for hydrogen bonding and neutral hydrophilicity that may not appear when only considering the top 1 mechanism.

In general, the inclusion of the top 2 analysis highlights the importance of considering multiple mechanisms in understanding complex mutations, as interactions that were previously less visible (e.g., steric and hydrogen bonding) now show greater influence.

Despite several trends aligning with known biochemical principles, some results in Fig. 2 do not match expectations. Notably, the hydrophilic mechanism group had almost no frequency as top 1 or top 2; this is in stark contrast to the relatively large proportion for hydrophobic and neutral mechanism groups across all tags. Additionally, SHAP values for hydrogen bonding remain lower than neutral hydrophilicity for [hbond] labels, suggesting that hydrogen bond SHAP features may not be the main contributing force in predicting [hbond] tagged mutations.

Examples of top SHAP features supported by other studies

Here we present two examples that illustrate the explanations generated by our method. The first example is derived from the complex of RAP and C-RAF1 (pdb: 1C1Y) [22], where ARG67 was mutated to Ala. The mutation R67A is shown in Fig. 3 where the mutation of the original Arg residue creates a reduction in electrostatic isopotential and therefore a reduction in electrostatic complementarity. This shows that the electrostatic mechanism is the dominant mechanism underlying the mutation R67A. This is consistent with the specific top SHAP feature found for this sample. The prediction is verified in the literature from Zhang, et al [22], who state: “Similar to the RalGDS complexes, the polar and positively charged amino acid side-chains of the effector contribute most to the interaction, namely Arg59, Arg67, Lys84 and Arg89…”.

Fig. 3.

Examples of two protein complexes 1C1Y and 3BK3, their Top 1 mechanisms underlying mutation of 1C1Y-RB67A and 3BK3-IC21A. Both complexes are illustrated like an open book, where the interfacial regions are opened from the center to face the reader. The regions of the binding interface that are nearest to the mutated amino acid are circled. A red/blue scale of surface charge is shown from −5 kt/e to +5kt/e. For the 1C1Y-RB67A mutation, electrostatic complementarity reduction is the predicted mechanism of the RB67A mutation (according to the top 1 SHAP feature). This complementarity reduction is apparent in the reduction of blue surface area on chain B in the mutant. For 3BK3-1C21A, reduction of hydrophobic contact area induced by mutation is predicted to be the most influential mechanism (according to the top 1 SHAP feature), and is visible in the reduction of purple surface area on chain C in the mutant. A Kyte-Doolitle hydrophobicity scale [?] is provided from purple (hydrophobic) to yellow (hydrophillic). These two predicted mechanisms are consistent with earlier structural studies [?,?].

The second example is based on the complex of the BMP-2 Receptor and the N-Terminal Von Willebrand factor type C domain of crossveinless 2 (CV-2) (pdb: 3BK3). This complex is involved in Bone Morphogenetic Protein (BMP) signaling pathways [14]. We show an example of mutation of Ile21 to Ala in Fig. 3. The top SHAP feature we generated for this mutation is hydrophobic which is expected when mutating from a large hydrophobic residue like Ile to a smaller, less hydrophobic Alanine as well as a reduction in hydrophobic contact area.Therefore, this indicates hydrophobic mechanism is the most important mechanism underlying mutation I21A, which is also consistent with our predicted top SHAP feature. This observation is verified in the literature by Kiel, et al [14], who write “…large hydrophobic side chains (mostly tryptophane) at a position corresponding to Ile21…Upon mutation of Ile21 or Ile27 to Ala, about 1.3 kcal/mol of binding free energy is lost.” Together, these examples illustrate how the predictions made with our tool and can be used to produce mechanistic explanations that are consistent with quotes structural publications.

4. Discussion

We present our work which uses explainable products of a trained prediction model to suggest a biochemical mechanism for amino acids involved in protein-protein interactions. Specifically, we evaluated the alignment between SHAP values from a GBT model trained to predict binding affinity and ground truth mechanistic annotations from experimental data. This preliminary work introduces a process for integrating fine-grained explainable model outcomes directly with functional real-world lab reports, allowing us to evaluate specific mechanisms while bypassing the need for exhaustive prediction processes typically employed to approximate real-world data.

In order to develop a ground truth mechanistic dataset, we used a novel annotation process to associate experimental mutation data from the SKEMPI 2.0 dataset with relevant mechanism tags. On SKEMPI dataset annotation, we determined the mechanism of 1225 amino acid mutation experiments through intensive manual literature review. Our goal in annotating mechanisms directly is to use these annotations as labels or features in downstream prediction tasks. In this study, we evaluated the contribution of four different interaction mechanisms -electrostatic, hydrogen bonding, steric, and 4 categories of hydrophobic interactions - by comparing SHAP values generated from a GBT model with the annotated mechanism tags.

To allow comparison, we grouped our SHAP features into 7 “mechanism groups” - electrostatic, hydrogen bonding, steric, and 4 different hydrophobicity groups - and evaluated which of the mechanism groups have the top 1 and top 2 total SHAP value among the four. Across the board, we found substantial agreement between SHAP mechanism groups and our annotated mechanisms for both top 1 and top 2 mechanisms. We notice, however, that hydrophobic and neutral mechanisms are much more prominent for many different tagged mutations, like [steric], [hydrophobic] and [hbond] tagged mutations. This suggests that hydrophobicity may play a more important role in stabilizing these interactions than previously recognized. Electrostatic interactions, as expected, were dominant for [electrostatics] tagged mutations and particularly dominant in [salt bridge] tagged mutations, confirming their crucial role in salt bridge formation. Unexpectedly, hydrogen bond SHAP features have low contribution across all tags in Top 1, especially for [hbond] tagged mutations where we expected for hydrogen bond mechanism to be the highest contributing SHAP value. Similarly, steric contributions were lower than anticipated across for all tags except [hydrophobic] tag, indicating that their influence on binding affinity may be hidden by higher contributing forces like hydrophobicity or is less effectively captured in our model.

In the top 2 mechanism analysis, the contribution of secondary interactions became clearer. The SHAP value for hydrogen bonding mechanism group, for example, showed a more significant role in [hbond] tagged mutations; similarly, steric and hydroneutral interactions were better represented. This highlights the importance of considering multiple features when evaluating complex mutations, as interactions that are hidden by the top contributing mechanism can still play an essential and complementary role. Overall, the inclusion of both top 1 and top 2 analyses generally agreed with the mechanism tags while providing more insights to how primary and secondary interactions cooperate in protein-protein interactions.

Despite the higher contribution of hydrogen bond and steric to [hbond] and [steric] tagged mutations respectively when examining top 2 mechanisms, these two mechanism groups still differ from classical expectations. Models, especially in energy function models and in traditional protein structure models, have assumed that hydrogen bonds and steric effects are significant to protein-protein interactions, however our results challenge these traditional assumptions. Overall, our results suggest that SHAP values can effectively capture both dominant and subtle mechanistic influences on binding affinity. In addition, we found that our manual SKEMPI annotations may work well with protein interaction models that utilize fine-grained mechanistic features or with explainable values like SHAP.

These differences also indicate the presence of additional structural nuances in this approach to producing biochemical explanations for binding mechanisms. Where explanations based on features disagree with annotations, it is important to understand how such disagreements can occur, and whether it results from a matter of biochemical representations, a lack of detailed annotations, or both.

One way to better capture interaction mechanisms is to review and completely annotate all interactions in the SKEMPI 2.0 dataset. This would result in a curated mechanism annotation dataset of over 2600 mutation experiments. We established a procedure for labeling and tagging large interaction datasets for mechanisms; the resulting dataset would be useful in many downstream tasks that are impossible today due to a lack of datasets providing biochemical interaction insights. For example, this dataset could be augmented for language-based tasks with recent advancements in LLM-based systems. Altogether, the practice of explaining how protein structures function, based on structural data working in tandem with annotations derived from the literature, represents a rich opportunity to advance the field.

Fig. 1.

Overview of the Preliminary Study and SHAP Value Generation Process. A) SHAP Value Generation Workflow: SHAP is applied alongside a binding affinity prediction model. The prediction model computes changes in binding affinity, while SHAP provides biochemical explanations underlying these changes for the SKEMPI data. B) Preliminary StudyWorkflow: Building on SHAP values obtained from a previous publication, we aggregated all feature SHAP values into four mechanism types and identified the top 1 and top 2 mechanisms for each sample. We then conducted a comprehensive literature review of the SKEMPI dataset and manually curated a PPI mutation mechanism annotation dataset. Finally, we compared the mechanisms derived from the model-generated explanations (top SHAP features) with the annotated mechanism labels from literature sources to assess the consistency and accuracy of the model’s biochemical interpretations.

Data

We have made the mechanism annotation data subset of SKEMPI available here.