Search

Abstract Knowledge of B cell epitopes is critical to vaccine design, diagnostics, and therapeutics. As experimental validation for epitopes is time-consuming and costly, many in silico tools have been developed to computationally predict the B cell epitopes. While most methods show poor performance, deep learning methods in recent years have shown promising results. We developed a method called EpiGraph that outperformed previous methods, including those that showed a significant improvement in performance in recent years. Our model’s performance can be attributed to the following factors: (1) a combination of structure and sequence feature embeddings obtained from pretrained ESM-IF1 and ESM-2 models could capture the structural and evolutionary features of B cell epitopes, (2) a graph attention network could learn the spatial proximity of B cell epitopes with high graph homophily, and (3) residual connections in the model framework mitigate the over-smoothing problem in the graph neural network. Our model achieved the highest performance on an independent benchmark dataset. The results were also consistent on a different dataset. The datasets and source codes are available at https://github.com/sj584/EpiGraph . A user-friendly web server is freely available at http://epigraph.kaist.ac.kr .

Abstract Minimum inhibitory concentration (MIC) denotes the in vitro benchmark indicating the quantity of antibiotic required to inhibit proliferation of specific bacterial strains. Determining MIC values corresponding to the infecting bacterial strain is paramount for tailoring appropriate antibiotic therapy. In the interim between specimen collection and laboratory-derived MIC outcomes, clinicians frequently resort to empirical therapy informed by retrospective analyses. Here introduces two deep learning approaches, a Convolutional Neural Network (CNN)-based model and an Enformer-based model, integrating genomic data of Klebsiella Pneumoniae and molecular structural data of 20 antibiotics to anticipate the MIC value of the bacterium for each antibiotic under consideration. These models demonstrate enhanced raw accuracy over the existing state-of-the-art model, which rely exclusively on genomic data. The CNN-based model achieves a notable 20% increase in raw accuracy while further mirroring the 1-tier accuracy of the state-of-the-art model. Although the Enformer-based model does not quite reach the performance levels of the CNN-based model, it offers an advantage by eliminating the need for arbitrary data processing steps. This streamlining of the data processing pipeline facilitates fast updates and improves the model interpretability. It is expected that these deep learning paradigms can significantly inform and bolster clinician decision-making during the empirical treatment phase.

Abstract Identifying the binding between the target proteins and molecules is essential in drug discovery. The multi-task learning method has been introduced to facilitate knowledge sharing among tasks when the amount of information for each task is small. However, multi-task learning sometimes worsens the overall performance or generates a trade-off between individual task’s performance. In this study, we propose a general multi-task learning scheme that not only increases the average performance but also minimizes individual performance degradation, through group selection and knowledge distillation. The groups are selected on the basis of chemical similarity between ligand sets of targets, and the similar targets in the same groups are trained together. During training, we apply knowledge distillation with teacher annealing. The multi-task learning models are guided by the predictions of the single-task learning models. This method results in higher average performance than that from single-task learning and classic multi-task learning. Further analysis reveals that multi-task learning is particularly effective for low performance tasks, and knowledge distillation helps the model avoid the degradation in individual task performance in multi-task learning.

In the last few decades, adverse reactions to pharmaceuticals have been evaluated using 2D in vitro models and animal models. However, with increasing computational power, and as the key drivers of cellular behavior have been identified, in silico models have emerged. These models are time-efficient and cost-effective, but the prediction of adverse reactions to unknown drugs using these models requires relevant experimental input. Accordingly, the physiome concept has emerged to bridge experimental datasets with in silico models. The brain physiome describes the systemic interactions of its components, which are organized into a multilevel hierarchy. Because of the limitations in obtaining experimental data corresponding to each physiome component from 2D in vitro models and animal models, 3D in vitro brain models, including brain organoids and brain-on-a-chip, have been developed. In this review, we present the concept of the brain physiome and its hierarchical organization, including cell- and tissue-level organizations. We also summarize recently developed 3D in vitro brain models and link them with the elements of the brain physiome as a guideline for dataset collection. The connection between in vitro 3D brain models and in silico modeling will lead to the establishment of cost-effective and time-efficient in silico models for the prediction of the safety of unknown drugs.

Abstract The development of an accurate and reliable variant effect prediction tool is important for research in human genetic diseases. A large number of predictors have been developed towards this goal, yet many of these predictors suffer from the problem of data circularity. Here we present MTBAN (Mutation effect predictor using the Temporal convolutional network and the Born-Again Networks), a method for predicting the deleteriousness of variants. We apply a form of knowledge distillation technique known as the Born-Again Networks (BAN) to a previously developed deep autoregressive generative model, mutationTCN, to achieve an improved performance in variant effect prediction. As the model is fully unsupervised and trained only on the evolutionarily related sequences of a protein, it does not suffer from the problem of data circularity which is common across supervised predictors. When evaluated on a test dataset consisting of deleterious and benign human protein variants, MTBAN shows an outstanding predictive ability compared to other well-known variant effect predictors. We also offer a user-friendly web server to predict variant effects using MTBAN, freely accessible at http://mtban.kaist.ac.kr . To our knowledge, MTBAN is the first variant effect prediction tool based on a deep generative model that provides a user-friendly web server for the prediction of deleteriousness of variants.

Abstract We developed a computational method named Molecule Optimization by Reinforcement Learning and Docking (MORLD) that automatically generates and optimizes lead compounds by combining reinforcement learning and docking to develop predicted novel inhibitors. This model requires only a target protein structure and directly modifies ligand structures to obtain higher predicted binding affinity for the target protein without any other training data. Using MORLD, we were able to generate potential novel inhibitors against discoidin domain receptor 1 kinase (DDR1) in less than 2 days on a moderate computer. We also demonstrated MORLD’s ability to generate predicted novel agonists for the D4 dopamine receptor (D4DR) from scratch without virtual screening on an ultra large compound library. The free web server is available at http://morld.kaist.ac.kr .

Prediction of protein-ligand interactions is a critical step during the initial phase of drug discovery. We propose a novel deep-learning-based prediction model based on a graph convolutional neural network, named GraphBAR, for protein-ligand binding affinity. Graph convolutional neural networks reduce the computational time and resources that are normally required by the traditional convolutional neural network models. In this technique, the structure of a protein-ligand complex is represented as a graph of multiple adjacency matrices whose entries are affected by distances, and a feature matrix that describes the molecular properties of the atoms. We evaluated the predictive power of GraphBAR for protein-ligand binding affinities by using PDBbind datasets and proved the efficiency of the graph convolution. Given the computational efficiency of graph convolutional neural networks, we also performed data augmentation to improve the model performance. We found that data augmentation with docking simulation data could improve the prediction accuracy although the improvement seems not to be significant. The high prediction performance and speed of GraphBAR suggest that such networks can serve as valuable tools in drug discovery.

Incorporation of unique barcodes into fission yeast gene deletion collections has enabled the identification of gene functions by growth fitness analysis. For fine tuning, it is important to examine barcode sequences, because mutations arise during strain construction. Out of 8,708 barcodes (4,354 strains) covering 88.5% of all 4,919 open reading frames, 7,734 barcodes (88.8%) were validated as high-fidelity to be inserted at the correct positions by Sanger sequencing. Sequence examination of the 7,734 high-fidelity barcodes revealed that 1,039 barcodes (13.4%) deviated from the original design. In total, 1,284 mutations (mutation rate of 16.6%) exist within the 1,039 mutated barcodes, which is comparable to budding yeast (18%). When the type of mutation was considered, substitutions accounted for 845 mutations (10.9%), deletions accounted for 319 mutations (4.1%), and insertions accounted for 121 mutations (1.6%). Peculiarly, the frequency of substitutions (67.6%) was unexpectedly higher than in budding yeast (~28%) and well above the predicted error of Sanger sequencing (~2%), which might have arisen during the solid-phase oligonucleotide synthesis and PCR amplification of the barcodes during strain construction. When the mutation rate was analyzed by position within 20-mer barcodes using the 1,284 mutations from the 7,734 sequenced barcodes, there was no significant difference between up-tags and down-tags at a given position. The mutation frequency at a given position was similar at most positions, ranging from 0.4% (32/7,734) to 1.1% (82/7,734), except at position 1, which was highest (3.1%), as in budding yeast. Together, well-defined barcode sequences, combined with the next-generation sequencing platform, promise to make the fission yeast gene deletion library a powerful tool for understanding gene function.

Abstract Kinases are major targets of anti-cancer therapies owing to their importance in signaling processes that regulate cell growth and proliferation. However, drug resistance has emerged as a major obstacle to cancer therapy. Resistance to drugs has various underlying mechanisms, including the acquisition of mutations at drug binding sites and the resulting reduction in drug binding affinity. Therefore, the identification of mutations that are relevant to drug resistance may be useful to overcome this issue. We hypothesized that these mutations can be identified by combining recent advances in computational methods for protein structure modeling and ligand docking simulation. Hence, we developed a web-based tool named the Kinase Resistance Docking System (KRDS) that enables the assessment of the effects of mutations on kinase-ligand interactions. KRDS receives a list of mutations in kinases, generates structural models of the mutants, performs docking simulations, and reports the results to users. The changes in docking scores and docking conformations can be analyzed to infer the effects of mutations on drug binding and drug resistance. We expect our tool to improve our understanding of drug binding mechanisms and facilitate the development of effective new drugs to overcome resistance related to kinases; it may be particularly useful for biomedical researchers who are not familiar with computational environments. Our tool is available at http://bcbl.kaist.ac.kr/KRDS/.

Abstract Background The identification of target molecules is important for understanding the mechanism of “target deconvolution” in phenotypic screening and “polypharmacology” of drugs. Because conventional methods of identifying targets require time and cost, in-silico target identification has been considered an alternative solution. One of the well-known in-silico methods of identifying targets involves structure activity relationships (SARs). SARs have advantages such as low computational cost and high feasibility; however, the data dependency in the SAR approach causes imbalance of active data and ambiguity of inactive data throughout targets. Results We developed a ligand-based virtual screening model comprising 1121 target SAR models built using a random forest algorithm. The performance of each target model was tested by employing the ROC curve and the mean score using an internal five-fold cross validation. Moreover, recall rates for top-k targets were calculated to assess the performance of target ranking. A benchmark model using an optimized sampling method and parameters was examined via external validation set. The result shows recall rates of 67.6% and 73.9% for top-11 (1% of the total targets) and top-33, respectively. We provide a website for users to search the top-k targets for query ligands available publicly at http://rfqsar.kaist.ac.kr . Conclusions The target models that we built can be used for both predicting the activity of ligands toward each target and ranking candidate targets for a query ligand using a unified scoring scheme. The scores are additionally fitted to the probability so that users can estimate how likely a ligand–target interaction is active. The user interface of our web site is user friendly and intuitive, offering useful information and cross references.

Abstract Background Computational scanning of peptide candidates that bind to a specific major histocompatibility complex (MHC) can speed up the peptide-based vaccine development process and therefore various methods are being actively developed. Recently, machine-learning-based methods have generated successful results by training large amounts of experimental data. However, many machine learning-based methods are generally less sensitive in recognizing locally-clustered interactions, which can synergistically stabilize peptide binding. Deep convolutional neural network (DCNN) is a deep learning method inspired by visual recognition process of animal brain and it is known to be able to capture meaningful local patterns from 2D images. Once the peptide-MHC interactions can be encoded into image-like array(ILA) data, DCNN can be employed to build a predictive model for peptide-MHC binding prediction. In this study, we demonstrated that DCNN is able to not only reliably predict peptide-MHC binding, but also sensitively detect locally-clustered interactions. Results Nonapeptide-HLA-A and -B binding data were encoded into ILA data. A DCNN, as a pan-specific prediction model, was trained on the ILA data. The DCNN showed higher performance than other prediction tools for the latest benchmark datasets, which consist of 43 datasets for 15 HLA-A alleles and 25 datasets for 10 HLA-B alleles. In particular, the DCNN outperformed other tools for alleles belonging to the HLA-A3 supertype. The F1 scores of the DCNN were 0.86, 0.94, and 0.67 for HLA-A*31:01, HLA-A*03:01, and HLA-A*68:01 alleles, respectively, which were significantly higher than those of other tools. We found that the DCNN was able to recognize locally-clustered interactions that could synergistically stabilize peptide binding. We developed ConvMHC, a web server to provide user-friendly web interfaces for peptide-MHC class I binding predictions using the DCNN. ConvMHC web server can be accessible via http://jumong.kaist.ac.kr:8080/convmhc . Conclusions We developed a novel method for peptide-HLA-I binding predictions using DCNN trained on ILA data that encode peptide binding data and demonstrated the reliable performance of the DCNN in nonapeptide binding predictions through the independent evaluation on the latest IEDB benchmark datasets. Our approaches can be applied to characterize locally-clustered patterns in molecular interactions, such as protein/DNA, protein/RNA, and drug/protein interactions.

Bar-code (tag) microarrays of yeast gene-deletion collections facilitate the systematic identification of genes required for growth in any condition of interest. Anti-sense strands of amplified bar-codes hybridize with ~10,000 (5,000 each for up- and down-tags) different kinds of sense-strand probes on an array. In this study, we optimized the hybridization processes of an array for fission yeast. Compared to the first version of the array (11 µm, 100K) consisting of three sectors with probe pairs (perfect match and mismatch), the second version (11 µm, 48K) could represent ~10,000 up-/down-tags in quadruplicate along with 1,508 negative controls in quadruplicate and a single set of 1,000 unique negative controls at random dispersed positions without mismatch pairs. For PCR, the optimal annealing temperature (maximizing yield and minimizing extra bands) was 58℃ for both tags. Intriguingly, up-tags required 3× higher amounts of blocking oligonucleotides than down-tags. A 1:1 mix ratio between up- and down-tags was satisfactory. A lower temperature (25℃) was optimal for cultivation instead of a normal temperature (30℃) because of extra temperature-sensitive mutants in a subset of the deletion library. Activation of frozen pooled cells for >1 day showed better resolution of intensity than no activation. A tag intensity analysis showed that tag(s) of 4,316 of the 4,526 strains tested were represented at least once; 3,706 strains were represented by both tags, 4,072 strains by up-tags only, and 3,950 strains by down-tags only. The results indicate that this microarray will be a powerful analytical platform for elucidating currently unknown gene functions.

Summary: The homeodomain is found in hundreds of transcription factors that play roles in fate determination via cell-autonomous regulation of gene expression. However, some homeodomain-containing proteins (HPs) are thought to be secreted and penetrate neighboring cells to affect the recipient cell fate. To determine whether this is a general characteristic of HPs, we carried out a large-scale validation for intercellular transfer of HPs. Our screening reveals that intercellular transfer is a general feature of HPs, but it occurs in a cell-context-sensitive manner. We also found the secretion is not solely a function of the homeodomain, but it is supported by external motifs containing hydrophobic residues. Thus, mutations of hydrophobic residues of HPs abrogate secretion and consequently interfere with HP function in recipient cells. Collectively, our study proposes that HP transfer is an intercellular communication method that couples the functions of interacting cells. : Lee et al. evaluate capabilities of homeodomain proteins (HPs) for transfer between cells. They find that intercellular transfer is a general but cell-context-sensitive property of HP. Intercellular HP transfer can be an unconventional way for the cells to communicate with neighboring cells that associate structurally and functionally. Keywords: homeodomain, cell-penetrating protein, protein secretion, intercellular protein transfer

ABSTRACT Butanol production by Clostridium acetobutylicum is accompanied by coproduction of acetone and ethanol, which reduces the yield of butanol and increases the production cost. Here, we report development of several clostridial aldehyde/alcohol dehydrogenase (AAD) variants showing increased butanol selectivity by a series of design and analysis procedures, including random mutagenesis, substrate specificity feature analysis, and structure-based butanol selectivity design. The butanol/ethanol ratios (B/E ratios) were dramatically increased to 17.47 and 15.91 g butanol/g ethanol for AADF716L and AADN655H, respectively, which are 5.8-fold and 5.3-fold higher than the ratios obtained with the wild-type AAD. The much-increased B/E ratio obtained was due to the dramatic reduction in ethanol production (0.59 ± 0.01 g/liter) that resulted from engineering the substrate binding chamber and the active site of AAD. This protein design strategy can be applied generally for engineering enzymes to alter substrate selectivity. IMPORTANCE Renewable biofuel represents one of the answers to solving the energy crisis and climate change problems. Butanol produced naturally by clostridia has superior liquid fuel characteristics and thus has the potential to replace gasoline. Due to the lack of efficient genetic manipulation tools, however, clostridial strain improvement has been slower than improvement of other microorganisms. Furthermore, fermentation coproducing various by-products requires costly downstream processing for butanol purification. Here, we report the results of enzyme engineering of aldehyde/alcohol dehydrogenase (AAD) to increase butanol selectivity. A metabolically engineered Clostridium acetobutylicum strain expressing the engineered aldehyde/alcohol dehydrogenase gene was capable of producing butanol at a high level of selectivity.

RNA-binding proteins (RBPs) are important in gene expression regulations by post-transcriptional control of RNAs and immune system development and its function. Due to the help of sequencing technology, numerous RNA sequences are newly discovered without knowing their binding partner RBPs. Therefore, demands for accurate prediction method for RBP binding sites are increasing. There are many attempts for RBP binding site predictions using various machine-learning techniques combined with various RNA features. In this work, we present a new deep convolution neural network model trained on CLIP-seq datasets using multi-sized filters and multi-modal features to predict the binding property of RBPs. With this model, we integrated sequence and structure information to extract sequence motifs, structure motifs, and combined motifs at the same time. The RBP binding site prediction on RBP-24 dataset was compared with two multi-modal methods, GraphProt and Deepnet-rbp, using area under curve (AUC) of receiver-operating characteristics (ROC). Our method (average AUC = 0.920) outperformed 20 RBPs with GraphProt (average AUC = 0.888) and 15 RBP with Deepnet-rbp (average AUC = 0.902). The improvement was achieved by using multi-sized convolution filters, where average relative error reduction was 17%. By introducing new RNA structure representation, structure probability matrix, average relative error was reduced by 3% when compared to one-hot encoded secondary structure representation. Interestingly, structure probability matrix was more effective on ALKBH5, where relative error reduction was 30%. We developed new sequence motif enrichment method, which we stated as response enrichment method. We successfully enriched sequence motif for 12 RBPs, which had high resemblance with other literature evidences, RBPgroup and CISBP-RNA. Finally by analyzing these results altogether, we found intricate interplay between sequence motif and structure motif, which agreed with other researches.

The immunoglobulin (Ig)-like domain containing receptor 1 (ILDR1) gene encodes angulin-2/ILDR1, a recently discovered tight junction protein, which forms tricellular tight junction (tTJ) structures with tricellulin and lipolysis-stimulated lipoprotein receptor (LSR) at tricellular contacts (TCs) in the inner ear. Previously reported recessive mutations within ILDR1 have been shown to cause severe to profound nonsyndromic sensorineural hearing loss (SNHL), DFNB42. Whole-exome sequencing of a Korean multiplex family segregating partial deafness identified a novel homozygous ILDR1 variant (p.P69H) within the Ig-like domain. To address the pathogenicity of p.P69H, the angulin-2/ILDR1 p.P69H variant protein, along with the previously reported pathogenic ILDR1 mutations, was expressed in angulin-1/LSR knockdown epithelial cells. Interestingly, partial mislocalization of the p.P69H variant protein and tricellulin at TCs was observed, in contrast to a severe mislocalization and complete failure of tricellulin recruitment of the other reported ILDR1 mutations. Additionally, three-dimensional protein modeling revealed that angulin-2/ILDR1 contributed to tTJ by forming a homo-trimer structure through its Ig-like domain, and the p.P69H variant was predicted to disturb homo-trimer formation. In this study, we propose a possible role of angulin-2/ILDR1 in tTJ formation in the inner ear and a wider audiologic phenotypic spectrum of DFNB42 caused by mutations within ILDR1.

Chromosome conformation capture (3C)-based techniques have recently been used to uncover the mystic genomic architecture in the nucleus. These techniques yield indirect data on the distances between genomic loci in the form of contact frequencies that must be normalized to remove various errors. This normalization process determines the quality of data analysis. In this study, we describe two systematic errors that result from the heterogeneous local density of restriction sites and different local chromatin states, methods to identify and remove those artifacts, and three previously described sources of systematic errors in 3C-based data: fragment length, mappability, and local DNA composition. To explain the effect of systematic errors on the results, we used three different published data sets to show the dependence of the results on restriction enzymes and experimental methods. Comparison of the results from different restriction enzymes shows a higher correlation after removing systematic errors. In contrast, using different methods with the same restriction enzymes shows a lower correlation after removing systematic errors. Notably, the improved correlation of the latter case caused by systematic errors indicates that a higher correlation between results does not ensure the validity of the normalization methods. Finally, we suggest a method to analyze random error and provide guidance for the maximum reproducibility of contact frequency maps.

Repeat proteins have recently attracted much attention as alternative scaffolds to immunoglobulin antibodies due to their unique structural and biophysical features. In particular, repeat proteins show high stability against temperature and chaotic agents. Despite many studies, structural features for the stability of repeat proteins remain poorly understood. Here we present an interesting result from in silico analyses pursuing the factors which affect the stability of repeat proteins. Previously developed repebody structure based on variable lymphocytes receptors (VLRs) which consists of leucine-rich repeat (LRR) modules was used as initial structure for the present study. We constructed extra six repebody structures with varying numbers of repeat modules and those structures were used for molecular dynamics simulations. For the structures, the intramolecular interactions including backbone H-bonds, van der Waals energy, and hydrophobicity were investigated and then the radius of gyration, solvent-accessible surface area, ratio of secondary structure, and hydration free energy were also calculated to find out the relationship between the number of LRR modules and stability of the protein. Our results show that the intramolecular interactions lead to more compact structure and smaller surface area of the repebodies, which are critical for the stability of repeat proteins. The other features were also well compatible with the experimental results. Based on our observations, the repebody-5 was proposed as the best structure from the all repebodies in structure optimization process. The present study successfully demonstrated that our computer-based molecular modeling approach can significantly contribute to the experiment-based protein engineering challenge.

Chromatin regulation underlies a variety of DNA metabolism processes, including transcription, recombination, repair, and replication. To perform a quantitative genetic analysis of chromatin accessibility, we obtained open chromatin profiles across 96 genetically different yeast strains by FAIRE (formaldehyde-assisted isolation of regulatory elements) assay followed by sequencing. While 5∼10% of open chromatin region (OCRs) were significantly affected by variations in their underlying DNA sequences, subtelomeric areas as well as gene-rich and gene-poor regions displayed high levels of sequence-independent variation. We performed quantitative trait loci (QTL) mapping using the FAIRE signal for each OCR as a quantitative trait. While individual OCRs were associated with a handful of specific genetic markers, gene expression levels were associated with many regulatory loci. We found multi-target trans-loci responsible for a very large number of OCRs, which seemed to reflect the widespread influence of certain chromatin regulators. Such regulatory hotspots were enriched for known regulatory functions, such as recombinational DNA repair, telomere replication, and general transcription control. The OCRs associated with these multi-target trans-loci coincided with recombination hotspots, telomeres, and gene-rich regions according to the function of the associated regulators. Our findings provide a global quantitative picture of the genetic architecture of chromatin regulation.

Repeat proteins are increasingly attracting much attention as alternative scaffolds to immunoglobulin antibodies due to their unique structural features. Nonetheless, engineering interaction interface and understanding molecular basis for affinity maturation of repeat proteins still remain a challenge. Here, we present a structure-based rational design of a repeat protein with high binding affinity for a target protein. As a model repeat protein, a Toll-like receptor4 (TLR4) decoy receptor composed of leucine-rich repeat (LRR) modules was used, and its interaction interface was rationally engineered to increase the binding affinity for myeloid differentiation protein 2 (MD2). Based on the complex crystal structure of the decoy receptor with MD2, we first designed single amino acid substitutions in the decoy receptor, and obtained three variants showing a binding affinity (K(D)) one-order of magnitude higher than the wild-type decoy receptor. The interacting modes and contributions of individual residues were elucidated by analyzing the crystal structures of the single variants. To further increase the binding affinity, single positive mutations were combined, and two double mutants were shown to have about 3000- and 565-fold higher binding affinities than the wild-type decoy receptor. Molecular dynamics simulations and energetic analysis indicate that an additive effect by two mutations occurring at nearby modules was the major contributor to the remarkable increase in the binding affinities.

Protein-protein interactions play an essential role in cellular processes. Certain proteins form stable complexes with their partner proteins, whereas others function by forming transient complexes. The conventional protein-protein interaction model describes an interaction between two proteins under the assumption that a protein binds to its partner protein through a single binding site. In this study, we improved the conventional interaction model by developing a Multiple-Site (MS) model in which a protein binds to its partner protein through closely located multiple binding sites on a surface of the partner protein by transiently docking at each binding site with individual binding free energies. To test this model, we used the protein-protein interaction mediated by Src homology 3 (SH3) domains. SH3 domains recognize their partners via a weak, transient interaction and are therefore promiscuous in nature. Because the MS model requires large amounts of data compared with the conventional interaction model, we used experimental data from the positionally addressable syntheses of peptides on cellulose membranes (SPOT-synthesis) technique. From the analysis of the experimental data, individual binding free energies for each binding site of peptides were extracted. A comparison of the individual binding free energies from the analysis with those from atomistic force fields gave a correlation coefficient of 0.66. Furthermore, application of the MS model to 10 SH3 domains lowers the prediction error by up to 9% compared with the conventional interaction model. This improvement in prediction originates from a more realistic description of complex formation than the conventional interaction model. The results suggested that, in many cases, SH3 domains increased the protein complex population through multiple binding sites of their partner proteins. Our study indicates that the consideration of general complex formation is important for the accurate description of protein complex formation, and especially for those of weak or transient protein complexes.