Encoding the space of protein-protein binding interfaces by artificial intelligence.

The physical interactions between proteins are largely determined by the structural properties at their binding interfaces. It was found that the binding interfaces in distinctive protein complexes are highly similar. The structural properties underlying different binding interfaces could be further captured by artificial intelligence. In order to test this hypothesis, we broke protein-protein binding interfaces into pairs of interacting fragments. We employed a generative model to encode these interface fragment pairs in a low-dimensional latent space. After training, new conformations of interface fragment pairs were generated. We found that, by only using a small number of interface fragment pairs that were generated by artificial intelligence, we were able to guide the assembly of protein complexes into their native conformations. These results demonstrate that the conformational space of fragment pairs at protein-protein binding interfaces is highly degenerate. Features in this degenerate space can be well characterized by artificial intelligence. In summary, our machine learning method will be potentially useful to search for and predict the conformations of unknown protein-protein interactions.

Computational simulations of bispecific T cell engagers by a multiscale model.

The use of bispecific antibodies as T cell engagers can bypass the normal T cell receptor-major histocompatibility class interaction, redirect the cytotoxic activity of T cells, and lead to highly efficient tumor cell killing. However, this immunotherapy also causes significant on-target off-tumor toxicologic effects, especially when it is used to treat solid tumors. To avoid these adverse events, it is necessary to understand the fundamental mechanisms involved in the physical process of T cell engagement. We developed a multiscale computational framework to reach this goal. The framework combines simulations on the intercellular and multicellular levels. On the intercellular level, we simulated the spatial-temporal dynamics of three-body interactions among bispecific antibodies, CD3 and tumor-associated antigens (TAAs). The derived number of intercellular bonds formed between CD3 and TAAs was further transferred to the multicellular simulations as the input parameter of adhesive density between cells. Through the simulations under various molecular and cellular conditions, we were able to gain new insights into how to adopt the most appropriate strategy to maximize the drug efficacy and avoid the off-target effect. For instance, we discovered that the low antibody-binding affinity resulted in the formation of large clusters at the cell-cell interface, which could be important to control the downstream signaling pathways. We also tested different molecular architectures of the bispecific antibody and suggested the existence of an optimal length in regulating the T cell engagement. Overall, the current multiscale simulations serve as a proof-of-concept study to help in the future design of new biological therapeutics.

Prediction of interactions between cell surface proteins by machine learning.

Cells detect changes in their external environments or communicate with each other through proteins on their surfaces. These cell surface proteins form a complicated network of interactions in order to fulfill their functions. The interactions between cell surface proteins are highly dynamic and, thus, challenging to detect using traditional experimental techniques. Here, we tackle this challenge using a computational framework. The primary focus of the framework is to develop new tools to identify interactions between domains in the immunoglobulin (Ig) fold, which is the most abundant domain family in cell surface proteins. These interactions could be formed between ligands and receptors from different cells or between proteins on the same cell surface. In practice, we collected all structural data on Ig domain interactions and transformed them into an interface fragment pair library. A high-dimensional profile can then be constructed from the library for a given pair of query protein sequences. Multiple machine learning models were used to read this profile so that the probability of interaction between the query proteins could be predicted. We tested our models on an experimentally derived dataset that contains 564 cell surface proteins in humans. The cross-validation results show that we can achieve higher than 70% accuracy in identifying the PPIs within this dataset. We then applied this method to a group of 46 cell surface proteins in Caenorhabditis elegans. We screened every possible interaction between these proteins. Many interactions recognized by our machine learning classifiers have been experimentally confirmed in the literature. In conclusion, our computational platform serves as a useful tool to help identify potential new interactions between cell surface proteins in addition to current state-of-the-art experimental techniques. The tool is freely accessible for use by the scientific community. Moreover, the general framework of the machine learning classification can also be extended to study the interactions of proteins in other domain superfamilies.

Machine-learning-based Structural Analysis of Interactions between Antibodies and Antigens.

Computational analysis of paratope-epitope interactions between antibodies and their corresponding antigens can facilitate our understanding of the molecular mechanism underlying humoral immunity and boost the design of new therapeutics for many diseases. The recent breakthrough in artificial intelligence has made it possible to predict protein-protein interactions and model their structures. Unfortunately, detecting antigen-binding sites associated with a specific antibody is still a challenging problem. To tackle this challenge, we implemented a deep learning model to characterize interaction patterns between antibodies and their corresponding antigens. With high accuracy, our model can distinguish between antibody-antigen complexes and other types of protein-protein complexes. More intriguingly, we can identify antigens from other common protein binding regions with an accuracy of higher than 70% even if we only have the epitope information. This indicates that antigens have distinct features on their surface that antibodies can recognize. Additionally, our model was unable to predict the partnerships between antibodies and their particular antigens. This result suggests that one antigen may be targeted by more than one antibody and that antibodies may bind to previously unidentified proteins. Taken together, our results support the precision of antibody-antigen interactions while also suggesting positive future progress in the prediction of specific pairing.

Protein-lipid interactions drive presynaptic assembly upstream of cell adhesion molecules.

Textbook models of synaptogenesis position cell adhesion molecules such as neurexin as initiators of synapse assembly. Here we discover a mechanism for presynaptic assembly that occurs prior to neurexin recruitment, while supporting a role for neurexin in synapse maintenance. We find that the cytosolic active zone scaffold SYD-1 interacts with membrane phospholipids to promote active zone protein clustering at the plasma membrane, and subsequently recruits neurexin to stabilize those clusters. Employing molecular dynamics simulations to model intrinsic interactions between SYD-1 and lipid bilayers followed by in vivo tests of these predictions, we find that PIP2-interacting residues in SYD-1's C2 and PDZ domains are redundantly necessary for proper active zone assembly. Finally, we propose that the uncharacterized yet evolutionarily conserved short γ isoform of neurexin represents a minimal neurexin sequence that can stabilize previously assembled presynaptic clusters, potentially a core function of this critical protein.

Prediction of Interactions between Cell Surface Proteins by Machine Learning.

Cells detect changes of external environments or communicate with each other through proteins on their surfaces. These cell surface proteins form a complicated network of interactions in order to fulfill their functions. The interactions between cell surface proteins are highly dynamic and thus challenging to detect using traditional experimental techniques. Here we tackle this challenge by a computational framework. The primary focus of the framework is to develop new tools to identify interactions between domains in immunoglobulin (Ig) fold, which is the most abundant domain family in cell surface proteins. These interactions could be formed between ligands and receptors from different cells, or between proteins on the same cell surface. In practice, we collected all structural data of Ig domain interactions and transformed them into an interface fragment pair library. A high dimensional profile can be then constructed from the library for a given pair of query protein sequences. Multiple machine learning models were used to read this profile, so that the probability of interaction between the query proteins can be predicted. We tested our models to an experimentally derived dataset which contains 564 cell surface proteins in human. The cross-validation results show that we can achieve higher than 70% accuracy in identifying the PPIs within this dataset. We then applied this method to a group of 46 cell surface proteins in C elegans. We screened every possible interaction between these proteins. Many interactions recognized by our machine learning classifiers have been experimentally confirmed in the literatures. In conclusion, our computational platform serves a useful tool to help identifying potential new interactions between cell surface proteins in addition to current state-of-the-art experimental techniques. The tool is freely accessible for use by the scientific community. Moreover, the general framework of the machine learning classification can also be extended to study interactions of proteins in other domain superfamilies.

A computational study for understanding the impact of p120-catenin on the cis-dimerization of cadherin.

A prototype of cross-membrane signal transduction is that extracellular binding of cell surface receptors to their ligands induces intracellular signaling cascades. However, much less is known about the process in the opposite direction, called inside-out signaling. Recent studies show that it plays a more important role in regulating the functions of many cell surface receptors than we used to think. In particular, in cadherin-mediated cell adhesion, recent experiments indicate that intracellular binding of the scaffold protein p120-catenin can promote extracellular clustering of cadherin and alter its adhesive function. The underlying mechanism, however, is not well understood. To explore possible mechanisms, we designed a new multiscale simulation procedure. Using all-atom molecular dynamics simulations, we found that the conformational dynamics of the cadherin extracellular region can be altered by the intracellular binding of p120-catenin. More intriguingly, by integrating all-atom simulation results into coarse-grained random sampling, we showed that the altered conformational dynamics of cadherin caused by the binding of p120-catenin can increase the probability of lateral interactions between cadherins on the cell surface. These results suggest that p120-catenin could allosterically regulate the cis-dimerization of cadherin through two mechanisms. First, p120-catenin controls the extracellular conformational dynamics of cadherin. Second, p120-catenin oligomerization can further promote cadherin clustering. Our study, therefore, suggests a mechanistic foundation for the inside-out signaling in cadherin-mediated cell adhesion, while the computational framework can be generally applied to other cross-membrane signal transduction systems.

Understanding the General Principles of T Cell Engagement by Multiscale Computational Simulations.

The use of bispecific antibodies as T cell engagers can bypass the normal TCR-MHC interaction, redirect the cytotoxic activity of T-cells, and lead to highly efficient tumor cell killing. However, this immunotherapy also causes significant on-target off-tumor toxicologic effects, especially when they were used to treat solid tumors. In order to avoid these adverse events, it is necessary to understand the fundamental mechanisms during the physical process of T cell engagement. We developed a multiscale computational framework to reach this goal. The framework combines simulations on the intercellular and multicellular levels. On the intercellular level, we simulated the spatial-temporal dynamics of three-body interactions among bispecific antibodies, CD3 and TAA. The derived number of intercellular bonds formed between CD3 and TAA were further transferred into the multicellular simulations as the input parameter of adhesive density between cells. Through the simulations under various molecular and cellular conditions, we were able to gain new insights of how to adopt the most appropriate strategy to maximize the drug efficacy and avoid the off-target effect. For instance, we discovered that the low antibody binding affinity resulted in the formation of large clusters at the cell-cell interface, which could be important to control the downstream signaling pathways. We also tested different molecular architectures of the bispecific antibody and suggested the existence of an optimal length in regulating the T cell engagement. Overall, the current multiscale simulations serve as a prove-of-concept study to help the future design of new biological therapeutics. SIGNIFICANCE: T-cell engagers are a class of anti-cancer drugs that can directly kill tumor cells by bringing T cells next to them. However, current treatments using T-cell engagers can cause serious side-effects. In order to reduce these effects, it is necessary to understand how T cells and tumor cells interact together through the connection of T-cell engagers. Unfortunately, this process is not well studied due to the limitations in current experimental techniques. We developed computational models on two different scales to simulate the physical process of T cell engagement. Our simulation results provide new insights into the general properties of T cell engagers. The new simulation methods can therefore serve as a useful tool to design novel antibodies for cancer immunotherapy.

Computational analyses of the interactome between TNF and TNFR superfamilies.

Proteins in the tumor necrosis factor (TNF) superfamily (TNFSF) regulate diverse cellular processes by interacting with their receptors in the TNF receptor (TNFR) superfamily (TNFRSF). Ligands and receptors in these two superfamilies form a complicated network of interactions, in which the same ligand can bind to different receptors and the same receptor can be shared by different ligands. In order to study these interactions on a systematic level, a TNFSF-TNFRSF interactome was constructed in this study by searching the database which consists of both experimentally measured and computationally predicted protein-protein interactions (PPIs). The interactome contains a total number of 194 interactions between 18 TNFSF ligands and 29 TNFRSF receptors in human. We modeled the structure for each ligand-receptor interaction in the network. Their binding affinities were further computationally estimated based on modeled structures. Our computational outputs, which are all publicly accessible, serve as a valuable addition to the currently limited experimental resources to study TNF-mediated cell signaling.

How does the same ligand activate signaling of different receptors in TNFR superfamily: a computational study.

TNFα is a highly pleiotropic cytokine inducing inflammatory signaling pathways. It is initially presented on plasma membrane of cells (mTNFα), and also exists in a soluble variant (sTNFα) after cleavage. The ligand is shared by two structurally similar receptors, TNFR1 and TNFR2. Interestingly, while sTNFα preferentially stimulates TNFR1, TNFR2 signaling can only be activated by mTNFα. How can two similar receptors respond to the same ligand in such a different way? We employed computational simulations in multiple scales to address this question. We found that both mTNFα and sTNFα can trigger the clustering of TNFR1. The size of clusters induced by sTNFα is constantly larger than the clusters induced by mTNFα. The systems of TNFR2, on the other hand, show very different behaviors. Only when the interactions between TNFR2 are very weak, mTNFα can trigger the receptors to form very large clusters. Given the same weak binding affinity, only small oligomers were obtained in the system of sTNFα. Considering that TNF-mediated signaling is modulated by the ligand-induced clustering of receptors on cell surface, our study provided the mechanistic foundation to the phenomenon that different isoforms of the ligand can lead to highly distinctive signaling patterns for its receptors.

Dissecting the general mechanisms of protein cage self-assembly by coarse-grained simulations.

The development of artificial protein cages has recently gained massive attention due to their promising application prospect as novel delivery vehicles for therapeutics. These nanoparticles are formed through a process called self-assembly, in which individual subunits spontaneously arrange into highly ordered patterns via non-covalent but specific interactions. Therefore, the first step toward the design of novel engineered protein cages is to understand the general mechanisms of their self-assembling dynamics. Here we have developed a new computational method to tackle this problem. Our method is based on a coarse-grained model and a diffusion-reaction simulation algorithm. Using a tetrahedral cage as test model, we showed that self-assembly of protein cage requires of a seeding process in which specific configurations of kinetic intermediate states are identified. We further found that there is a critical concentration to trigger self-assembly of protein cages. This critical concentration allows that cages can only be successfully assembled under a persistently high concentration. Additionally, phase diagram of self-assembly has been constructed by systematically testing the model across a wide range of binding parameters. Finally, our simulations demonstrated the importance of protein's structural flexibility in regulating the dynamics of cage assembly. In summary, this study throws lights on the general principles underlying self-assembly of large cage-like protein complexes and thus provides insights to design new nanomaterials.

Understanding the functional role of membrane confinements in TNF-mediated signaling by multiscale simulations.

The interaction between TNFα and TNFR1 is essential in maintaining tissue development and immune responses. While TNFR1 is a cell surface receptor, TNFα exists in both soluble and membrane-bound forms. Interestingly, it was found that the activation of TNFR1-mediated signaling pathways is preferentially through the soluble form of TNFα, which can also induce the clustering of TNFR1 on plasma membrane of living cells. We developed a multiscale simulation framework to compare receptor clustering induced by soluble and membrane-bound ligands. Comparing with the freely diffusive soluble ligands, we hypothesize that the conformational dynamics of membrane-bound ligands are restricted, which affects the clustering of ligand-receptor complexes at cell-cell interfaces. Our simulation revealed that only small clusters can form if TNFα is bound on cell surface. In contrast, the clustering triggered by soluble TNFα is more dynamic, and the size of clusters is statistically larger. We therefore demonstrated the impact of membrane-bound ligand on dynamics of receptor clustering. Moreover, considering that larger TNFα-TNFR1 clusters is more likely to provide spatial platform for downstream signaling pathway, our studies offer new mechanistic insights about why the activation of TNFR1-mediated signaling pathways is not preferred by membrane-bound form of TNFα.

EXCESP: A Structure-Based Online Database for Extracellular Interactome of Cell Surface Proteins in Humans.

The interactions between ectodomains of cell surface proteins are vital players in many important cellular processes, such as regulating immune responses, coordinating cell differentiation, and shaping neural plasticity. However, while the construction of a large-scale protein interactome has been greatly facilitated by the development of high-throughput experimental techniques, little progress has been made to support the discovery of extracellular interactome for cell surface proteins. Harnessed by the recent advances in computational modeling of protein-protein interactions, here we present a structure-based online database for the extracellular interactome of cell surface proteins in humans, called EXCESP. The database contains both experimentally determined and computationally predicted interactions among all type-I transmembrane proteins in humans. All structural models for these interactions and their binding affinities were further computationally modeled. Moreover, information such as expression levels of each protein in different cell types and its relation to various signaling pathways from other online resources has also been integrated into the database. In summary, the database serves as a valuable addition to the existing online resources for the study of cell surface proteins. It can contribute to the understanding of the functions of cell surface proteins in the era of systems biology.

A structural-based machine learning method to classify binding affinities between TCR and peptide-MHC complexes.

The activation of T cells is triggered by the interactions of T cell receptors (TCRs) with their epitopes, which are peptides presented by major histocompatibility complex (MHC) on the surfaces of antigen presenting cells (APC). While each TCR can only recognize a specific subset from a large repertoire of peptide-MHC (pMHC) complexes, it is very often that peptides in this subset share little sequence similarity. This is known as the specificity and cross-reactivity of T cells, respectively. The binding affinities between different types of TCRs and pMHC are the major driving force to shape this specificity and cross-reactivity in T cell recognition. The binding affinities, furthermore, are determined by the sequence and structural properties at the interfaces between TCRs and pMHC. Fortunately, a wealth of data on binding and structures of TCR-pMHC interactions becomes publicly accessible in online resources, which offers us the opportunity to develop a random forest classifier for predicting the binding affinities between TCR and pMHC based on the structure of their complexes. Specifically, the structure and sequence of a given complex were projected onto a high-dimensional feature space as the input of the classifier, which was then trained by a large-scale benchmark dataset. Based on the cross-validation results, we found that our machine learning model can predict if the binding affinity of a given TCR-pMHC complex is stronger or weaker than a predefined threshold with an overall accuracy approximately around 75 %. The significance of our prediction was estimated by statistical analysis. Moreover, more than 60 % of binding affinities in the ATLAS database can be successfully classified into groups within the range of 2 kcal/mol. Additionally, we show that TCR-pMHC complexes with strong binding affinity prefer hydrophobic interactions between amino acids with large aromatic rings instead of electrostatic interactions. Our results therefore provide insights to design engineered TCRs which enhance the specificity for their targeted epitopes. Taken together, this method can serve as a useful addition to a suite of existing approaches which study binding between TCR and pMHC.

Coarse-grained simulations of phase separation driven by DNA and its sensor protein cGAS.

The enzyme cGAS functions as a sensor that recognizes the cytosolic DNA from foreign pathogen. The activation of the protein triggers the transcription of inflammatory genes, leading into the establishment of an antipathogen state. An interesting new discovery is that the detection of DNA by cGAS induced the formation of liquid-like droplets. However how cells regulate the formation of these droplets is still not fully understood. In order to unravel the molecular mechanism beneath the DNA-mediated phase separation of cGAS, we developed a polymer-based coarse-grained model which takes into accounts the basic structural organization in DNA and cGAS, as well as the binding properties between these biomolecules. This model was further integrated into a hybrid simulation algorithm. With this computational method, a multi-step kinetic process of aggregation between cGAS and DNA was observed. Moreover, we systematically tested the model under different concentrations and binding parameters. Our simulation results show that phase separation requires both cGAS dimerization and protein-DNA interactions, whereas polymers can be kinetically trapped in small aggregates under strong binding affinities. Additionally, we demonstrated that supramolecular assembly can be facilitated by increasing the number of functional modules in protein or DNA polymers, suggesting that multivalency and intrinsic disordered regions play positive roles in regulating phase separation. This is consistent to previous experimental evidences. Taken together, this is, to the best of our knowledge, the first computational model to study condensation of cGAS-DNA complexes. While the method can reach the timescale beyond the capability of atomic-level MD simulations, it still includes information about spatial arrangement of functional modules in biopolymers that is missing in the mean-field theory. Our work thereby adds a useful dimension to a suite of existing experimental and computational techniques to study the dynamics of phase separation in biological systems.

Computational Simulation of Holin S105 in Membrane Bilayer and Its Dimerization Through a Helix-Turn-Helix Motif.

During the final step of the bacteriophage infection cycle, the cytoplasmic membrane of host cells is disrupted by small membrane proteins called holins. The function of holins in cell lysis is carried out by forming a highly ordered structure called lethal lesion, in which the accumulation of holins in the cytoplasmic membrane leads to the sudden opening of a hole in the middle of this oligomer. Previous studies showed that dimerization of holins is a necessary step to induce their higher order assembly. However, the molecular mechanism underlying the holin-mediated lesion formation is not well understood. In order to elucidate the functions of holin, we first computationally constructed a structural model for our testing system: the holin S105 from bacteriophage lambda. All atom molecular dynamic simulations were further applied to refine its structure and study its dynamics as well as interaction in lipid bilayer. Additional simulations on association between two holins provide supportive evidence to the argument that the C-terminal region of holin plays a critical role in regulating the dimerization. In detail, we found that the adhesion of specific nonpolar residues in transmembrane domain 3 (TMD3) in a polar environment serves as the driven force of dimerization. Our study therefore brings insights to the design of binding interfaces between holins, which can be potentially used to modulate the dynamics of lesion formation.

Methane Clathrate Formation is Catalyzed and Kinetically Inhibited by the Same Molecule: Two Facets of Methanol.

Here, we perform molecular dynamics simulations to provide atomic-level insights into the dual roles of methanol in enhancing and delaying the rate of methane clathrate hydrate nucleation. Consistent with experiments, we find that methanol slows clathrate hydrate nucleation above 250 K but promotes clathrate formation at temperatures below 250 K. We show that this behavior can be rationalized by the unusual temperature dependence of the methane-methanol interaction in an aqueous solution, which emerges due to the hydrophobic effect. In addition to its antifreeze properties at temperatures above 250 K, methanol competes with water to interact with methane prior to the formation of clathrate nuclei. Below 250 K, methanol encourages water to occupy the space between methane molecules favoring clathrate formation and it may additionally promote water mobility.

A multiscale study on the mechanisms of spatial organization in ligand-receptor interactions on cell surfaces.

The binding of cell surface receptors with extracellular ligands triggers distinctive signaling pathways, leading into the corresponding phenotypic variation of cells. It has been found that in many systems, these ligand-receptor complexes can further oligomerize into higher-order structures. This ligand-induced oligomerization of receptors on cell surfaces plays an important role in regulating the functions of cell signaling. The underlying mechanism, however, is not well understood. One typical example is proteins that belong to the tumor necrosis factor receptor (TNFR) superfamily. Using a generic multiscale simulation platform that spans from atomic to subcellular levels, we compared the detailed physical process of ligand-receptor oligomerization for two specific members in the TNFR superfamily: the complex formed between ligand TNFα and receptor TNFR1 versus the complex formed between ligand TNFß and receptor TNFR2. Interestingly, although these two systems share high similarity on the tertiary and quaternary structural levels, our results indicate that their oligomers are formed with very different dynamic properties and spatial patterns. We demonstrated that the changes of receptor's conformational fluctuations due to the membrane confinements are closely related to such difference. Consistent to previous experiments, our simulations also showed that TNFR can preassemble into dimers prior to ligand binding, while the introduction of TNF ligands induced higher-order oligomerization due to a multivalent effect. This study, therefore, provides the molecular basis to TNFR oligomerization and reveals new insights to TNFR-mediated signal transduction. Moreover, our multiscale simulation framework serves as a prototype that paves the way to study higher-order assembly of cell surface receptors in many other bio-systems.

Mechanistic dissection of spatial organization in NF-κB signaling pathways by hybrid simulations.

The nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB) is one of the most important transcription factors involved in the regulation of inflammatory signaling pathways. Inappropriate activation of these pathways has been linked to autoimmunity and cancers. Emerging experimental evidences have been showing the existence of elaborate spatial organizations for various molecular components in the pathways. One example is the scaffold protein tumor necrosis factor receptor associated factor (TRAF). While most TRAF proteins form trimeric quaternary structure through their coiled-coil regions, the N-terminal region of some members in the family can further be dimerized. This dimerization of TRAF trimers can drive them into higher-order clusters as a response to receptor stimulation, which functions as a spatial platform to mediate the downstream poly-ubiquitination. However, the molecular mechanism underlying the TRAF protein clustering and its functional impacts are not well-understood. In this article, we developed a hybrid simulation method to tackle this problem. The assembly of TRAF-based signaling platform at the membrane-proximal region is modeled with spatial resolution, while the dynamics of downstream signaling network, including the negative feedbacks through various signaling inhibitors, is simulated as stochastic chemical reactions. These two algorithms are further synchronized under a multiscale simulation framework. Using this computational model, we illustrated that the formation of TRAF signaling platform can trigger an oscillatory NF-κB response. We further demonstrated that the temporal patterns of downstream signal oscillations are closely regulated by the spatial factors of TRAF clustering, such as the geometry and energy of dimerization between TRAF trimers. In general, our study sheds light on the basic mechanism of NF-κB signaling pathway and highlights the functional importance of spatial regulation within the pathway. The simulation framework also showcases its potential of application to other signaling pathways in cells.

Computational Assessment of Protein-protein Binding Affinity by Reversely Engineering the Energetics in Protein Complexes.

The cellular functions of proteins are maintained by forming diverse complexes. The stability of these complexes is quantified by the measurement of binding affinity, and mutations that alter the binding affinity can cause various diseases such as cancer and diabetes. As a result, accurate estimation of the binding stability and the effects of mutations on changes of binding affinity is a crucial step to understanding the biological functions of proteins and their dysfunctional consequences. It has been hypothesized that the stability of a protein complex is dependent not only on the residues at its binding interface by pairwise interactions but also on all other remaining residues that do not appear at the binding interface. Here, we computationally reconstruct the binding affinity by decomposing it into the contributions of interfacial residues and other non-interfacial residues in a protein complex. We further assume that the contributions of both interfacial and non-interfacial residues to the binding affinity depend on their local structural environments such as solvent-accessible surfaces and secondary structural types. The weights of all corresponding parameters are optimized by Monte-Carlo simulations. After cross-validation against a large-scale dataset, we show that the model not only shows a strong correlation between the absolute values of the experimental and calculated binding affinities, but can also be an effective approach to predict the relative changes of binding affinity from mutations. Moreover, we have found that the optimized weights of many parameters can capture the first-principle chemical and physical features of molecular recognition, therefore reversely engineering the energetics of protein complexes. These results suggest that our method can serve as a useful addition to current computational approaches for predicting binding affinity and understanding the molecular mechanism of protein-protein interactions.