Introduction

Protein–protein interactions (PPIs) play fundamentally important roles in cellular functions and biological processes, and structural understanding of the PPIs is important for the elucidation of those functions (Jones and Thornton 1995, 1996). In 2001, the Critical Assessment of PRediction of Interactions (CAPRI, 2022) began as a community-wide experiment designed to assess methods for predicting PPIs based on the estimation of PPIs for previously solved structures of protein complexes. The latest experiment (Round 54) was conducted during May–August in 2022 (CAPRI Round 54, 2022). A recent report of a CAPRI experiment (Lensink et al. 2020) indicated that the increase of structures of protein complexes deposited in the Protein Data Bank (PDB) enables their use as structural complex templates for predicting the structures of other protein complexes, particularly in homo protein–protein docking. That finding implies that classical docking may no longer be strictly necessary for the prediction of homo PPIs. Nevertheless, these methods have not yet been successfully applied to the prediction of hetero PPIs, including antibody–antigen interactions, and this area of research has room for improvement. In this review, we place our focus on computational docking-based approaches and AI-based approaches, introducing a wide array of methods, including our own, for PPI prediction.

Protein–protein docking

Traditional protein–protein docking methods have been of central importance for sampling the conformational space of protein complexes (Smith and Sternberg 2002). In the last 10 years, sophisticated high-precision docking methods such as HADDOCK (van Zundert et al. 2015), ClusPro (Desta et al. 2022a, b), MegaFold (Liu et al. 2022), and HelixFold (Wang et al. 2022), it was only Uni-Fold and Uni-Fold-symmetry (Li et al. 2022b) that could succeed not only in monomer prediction but also in multimer prediction upon using the original trained parameters or protocols. No third-party group has yet published a large benchmark result. Actually, pLM-based predictors such as OmegaFold (Wu et al. 2022), ESMFold (Lin et al. 2022), and IgFold (Ruffolo et al. 2022) are anticipated as the next breakthroughs in structure prediction because they omit the construction of MSA, which is crucially important for performance, and which is the most time-consuming part of AF2. In spite of these expectations, no MSA-free method has achieved performance equal to that of AF2, with no explicit implementation of multimer modeling with the exception that IgFold was built specifically for antibody modeling.

Although no explicit implementation of multimer treatment in OmegaFold has been described, it is relatively straightforward to use pseudo-multiple sequence inputs as an AF2-linker which we designate as OmegaFold-linker. Figure 1 presents multimer modeling results in terms of the DockQ score (Basu and Wallner 2016) using methods developed post AF2 such as OmegaFold-linker, AF2-Multimer (v2.1.1) with parameters released on Nov. 2021, and AF2-Multimer (v.2.2.0) with parameters released on Mar. 2022. In the construction of Fig. 1, AF2 was run with “max_template_date = 2020–10–01”. The targets are taken from those of CASP14 (not included in the training set of AF2 and AF2-multimer, but two of these structures (PDB IDs: 6N64 and 6YA2) of them are released before 2020–10–01). They are all dimers, and the target IDs are T1032 (PDB ID: 6N64), T1038 (PDB ID: 6YA2), T1054 (PDB ID: 6V4V), T1078 (PDB ID: 7CWP), H1045 (PDB ID: 6XOD), and H1065 (PDB ID: 7M5F). For OmegaFold-linker, due to the memory-limitation, T1032 with more than 500 residues was omitted. For Omega-linker and AF2-linker, the input sequences were connected via 21 length poly-GLY-GLY-SER linker according to the previous research (Evans et al. 2021). For comparison, DockQ scores of docking structures calculated using ZDOCK (Pierce et al. 2014) are also shown. As the initial monomer structures of ZDOCK, the same models predicted by AF2-Multimer (v.2.2.0) were used. Basically, the prediction difficulty depends on targets. For instance, all methods except ZDOCK show 0.77 or more for H1065, while all methods show DockQ < 0.04 for T1054. These trends, except for T1038 (6YA2) whose structure was released before 2020–10–01, are generally similar to those observed in CASP14. Although, this may imply that the prediction difficulty of PPI depends on the availability of “good” templates, the results accord with those reported from benchmark research (Evans et al. 2021; Bryant et al. 2022a): AF2-Multimer shows the highest performance, and AF2-based methods using parameters trained for monomers are still comparable in some cases.

Fig. 1
figure 1

Distributions of DockQ scores as boxplots for different modeling methods: ZDOCK with AF2 models, OmegaFold-linker, AF2-linker, AF2-Multimer (v.2.1.1), and AF2-Multimer (v2.2.0). Six targets (T1032 (shown in black dot), T1038 (blue), T1054 (green), T1078 (red), H1045 (orange), and H1065 (pink)) are taken from CASP14 (see text). For the case of OmegaFold-linker, T1032 was omitted because its amino acid length is beyond the limitation of OmegaFold

Full size image

Antibody–antigen interaction

The prediction of PPI between antibody and antigen proteins is not an easy task because of the flexibility of the antibody’s hypervariable loops, particularly the complementary determining loop 3 in the heavy chain (CDR-H3 loop). Four high-precision docking software suites ClusPro (Brenke et al. 2012), LightDock (Jiménez-García et al. 2018), ZDOCK (Pierce et al. 2014), and HADDOCK (van Zundert et al. 2015) were used to examine which structural information contributes to the accuracy of the model building of an antibody–antigen complex, such as the structural information of CDR loops, paratope (antigen-binding residues on an antibody CDR), antigen surface, and epitope (antibody-binding residues on an antigen) (Ambrosetti et al. 2020). The findings showed for all docking methods that the overall performance decreased without epitope information. They were improved by consideration of the low-resolution epitope information. Accurate modeling of the structure of the long CDR-H3 loop remains challenging. However, the flexible refinement of HADDOCK led to the improvement of the prediction accuracy of CDR-H3 loop conformations if the epitope information was available, even though the resolution of the information is low (Ambrosetti et al. 2020).

The AI-based prediction of antibody–antigen interfaces has been developed as PECAN (Pittala and Bailey-Kellogg 2020). Here, antibody and antigen structures are presented as the respective graphs. They are input to the neural network that consists of graph convolution, attention, and fully connected layers. The network discriminates antibody and antigen residues between interface and non-interface residues. The predictive accuracies of PECAN achieved when using the datasets provided with epitope (Krawczyk et al. 2014), and paratope prediction methods (Daberdaku and Ferrari 2019) were found to have higher precision and recall rates than these providers’ methods. It is noteworthy that the providers’ methods also predict epitope and paratope regions with high accuracies. Epitope prediction, using the program EpiPred, predicts epitopes based on geometric fitting and knowledge-based asymmetric antibody–antigen scoring (Krawczyk et al. 2014). The paratope prediction method uses a SVM classifier to distinguish interface surface patches from non-interface ones based on 3D Zernike descriptors that represent global and local protein surface shapes and physicochemical properties on the surfaces (Daberdaku and Ferrari 2019). Again, note that even recent AI-based approaches, such as AF2-Multimer, suffer from an inability to predict accurate antibody-antigen complex structures (as mentioned above).

To obtain information about antibody-specific epitopes, some improvements in the accuracy of docking and affinity predictions must be achieved. For this purpose, a large and non-redundant benchmark set for antibody–antigen docking and affinity prediction was constructed. It includes camelid nanobodies, therapeutic monoclonal antibodies, and broadly neutralizing antibodies that target viral glycoproteins (Guest et al. 2021).

Another possible approach is the search of similar regions on proteins to known antibody-binding epitopes. This approach is based on the fact that some antibodies can cross-reactively recognize different antigen proteins having similar surface regions in the structures and properties (Vieths et al. 2006; Negi and Braun 2017). Information about cross-reactivity might provide information about the repurposing of antibody drugs. We are striving to develop a database of known and putative epitopes on proteins (PoSSuMAg 2022), which has a similar scheme to that of the PoSSuM database (PoSSuM 2021), which is helpful to detect putative pockets that are similar to known ligand-binding sites on protein structures (Tabei et al. 2010; Ito et al. 2012b, a, 2015). Our new database presents information about putative epitopes that are similar to known epitopes on the antigen proteins in complex with antibodies, including antibody drugs. Information about known and putative epitopes on SARS-CoV-2 proteins will also be available. Although the current version includes information of putative epitopes on non-redundant protein structures, we expect to increase the data through future studies.

Concluding remarks

Here we have introduced and discussed a diverse range of computational methods, from docking-based to AI-based approaches, for PPI prediction. Along with the increase in the amount of information related to protein sequences and structures, AI-based approaches, especially those based on evolutionary information and templates, are expected to become more powerful and useful. Yet, prediction of hetero PPIs leaves room for improvement. Particularly, antibody–antigen interaction predictions remain very limited in terms of their accuracy, although many sophisticated prediction methods have been developed (Ambrosetti et al. 2020).