Performance of ZDOCK and IRAD in CAPRI rounds 39-45

Abstract

We report docking performance on the six targets of CAPRI rounds 39–45 that involved heteromeric protein-protein interactions and had the solved structures released since the rounds were held. Our general strategy involved protein-protein docking using ZDOCK, reranking using IRAD, and structural refinement using Rosetta. In addition, we made extensive use of experimental data to guide our docking runs. All the experimental information at the amino-acid level proved correct. However, for two targets, we also used protein-complex structures as templates for modeling interfaces. These resulted in incorrect predictions, presumably due to the low sequence identity between the targets and templates. Albeit a small number of targets, the performance described here compared somewhat less favorably with our previous CAPRI reports, which may be due to the CAPRI targets being increasingly challenging.

Introduction

The interactions between proteins are essential for many fundamental cellular processes, playing important roles in the immune system, signaling pathways, and enzyme inhibition. High-throughput proteomics studies have shown that most proteins interact with other proteins.¹ Solving protein-protein complex structures experimentally, however, is laborious and not always successful. As an alternative, a large number of computational approaches were developed to predict complex structures using unbound structures as input, and a wide variety of techniques were used for these approaches: Fast Fourier Transform (FFT)², geometric hashing³, Monte Carlo simulation⁴, template-based modeling^5,6, and swarm optimization⁷. Increasingly, these methods can incorporate auxiliary information to guide the prediction, with a recent focus on coevolution data.⁸

Since 2001, the Critical Assessment of PRedicted Interactions (CAPRI) experiment has stimulated the development of methods for the prediction of protein-protein interactions,⁹ and we participated since the earliest rounds using our series of ZDOCK algorithms^10,11 and docking post-processing programs.^12–16 In its early years, CAPRI focussed on the prediction of protein-protein complex structures from its unbound components, but more recently has expanded to the prediction of mutation energies, protein design, protein-oligosaccharide binding, and protein-peptide binding. In addition, integration with CASP¹⁷ provided a wealth of challenging targets that often required modeling of both the tertiary and quaternary structures.

Here we report our results for the rounds that were held from 2016 to 2018 and were covered by the 7^th CAPRI evaluation meeting. As the focus of our lab is primarily on protein-protein interaction, we did not make predictions for other types of targets. Also the CASP-CAPRI rounds 37 and 46 are excluded as they are evaluated separately,¹⁸ and several targets can not be assessed because the solved structures have not been released or described yet. The resulting six targets all involve heteromeric interactions, but for two, homomeric interactions also needed to be predicted as components of higher-order multimers.

Methods

Our base strategy starts with the ZDOCK program for rigid-body docking,^10,11 with a 6° angular sampling that generates 54,000 predictions. ZDOCK uses FFT techniques to make the exhaustive search computationally feasible. As a starting point for the docking, we used unbound X-ray or nuclear magnetic resonance (NMR) structures when available, homology models using Modeller¹⁹ if we could identify good templates, or threading-based models using I-TASSER²⁰ when templates could not be found. Because the scoring function in ZDOCK is constrained by the FFT formalism, we reranked the prediction using the more sophisticated and accurate scoring function IRAD²¹. In addition, we pruned the results to remove highly similar predictions.^15,22,23 The pruning algorithm also identified regions of high prediction-density, which we used to guide our selection of the ten final models to submit.

When we could identify interface information from the literature, we used it to either guide the ZDOCK runs or to filter the list of predictions. The details on the information we applied are given in the respective sections in Results and Discussion. From the retained predictions (with or without using experimental information), we pooled the best predictions based on ZDOCK score, IRAD score²¹, and prediction density to make a shortlist of structures for manual inspection.

Finally, the ten manually selected models were refined using Rosetta’s fixed backbone repacking (fixbb) module to remove possible steric clashes.²⁴ After the most recent rounds of CAPRI, however, we realized that this approach did not always bring the number of clashes within the range acceptable to CAPRI, due to backbone-backbone clashes or tight packing between the receptor and the ligand. Therefore, we developed a new approach that iterates Rosetta’s fixbb procedure with global separation of the component proteins, summarized as follows. We first use fixbb to repack the protein side-chains. Next, we averaged the clash vectors, which are the vectors defined by the atom positions of each clashing pair, to produce a global clash vector. This vector was then used to separate the ligand and receptor, moving 1/4th of its magnitude. We repeated cycles of Rosetta fixbb repacking and global separation until the complex was under a specified clash threshold.

Results and Discussion

Target 122: Interleukin-23 and its cognate receptor IL-23R

This target involves Interleukin-23 (IL-23), a cytokine in the IL-12 family, and its cognate receptor IL-23R. We were provided with the structure of IL-23, which is a heterodimer of the IL-23 alpha subunit (IL-23A) and the IL-12 beta subunit (IL-12B), but we needed to model the structure of IL-23R. Thus, we performed homology modeling using the beta-receptor of another cytokine IL-6, IL-6Rβ (PDB ID 1P9M, chain A) as the template. The crystal structure of the IL-23:IL23R complex has since been released (PDB ID 5MZV)²⁵ and shows that our modeled IL-23R structure was reasonably accurate, despite the low sequence identity (~25%) between IL-23R and IL-6Rβ.

The crystal structure we used to homology model IL-23R also contained IL-6, whose structure superimposed well with that of IL-23A, although their sequence identity was low (~20%). Furthermore, mutation experiments on the murine IL-23 suggested that the tryptophan residue at position 156 (W156) of IL-23A was part of its binding interface with IL-23R.²⁶ Structural alignment showed that the IL-6Rβ residue corresponding to IL-23A’s W156 was not in the IL6:IL-6Rβ interface. We nonetheless decided to use the IL6:IL-6Rβ interface as the template for some of our predictions. To address the conflicting mutation data, we generated three sets of predictions: (1) modeled using the IL6:IL-6Rβ complex as the template; (2) docked with W156 in the interface as a constraint; and (3) docked without constraints.

The binding interface in the released crystal structure of the IL-23:IL-23R complex (PDB ID 5MZV) did not agree with the IL6:IL-6Rβ interface, which we used for template-based modeling of the complex. The template suggested binding through the center and C-terminal domains of IL-23R, whereas the IL-23:IL-23R crystal structure shows that binding occurs through IL-23R’s N-terminal domain and that IL-23A’s W156 is indeed at the binding interface with IL-23R (Figure 1). Despite the correct constraint of placing W156 at the interface, our ab initio docking runs showed a strong but erroneous preference to bind through the C-terminal domain of IL-23R, and we did not make correct predictions.

Figure 1:

Crystal structure of the complex of IL-23 (a heterodimer of IL-12B and IL-23A) with IL-23R in green, blue, and magenta, respectively. In cyan and light green, we show the modeled orientations of IL-23 using the IL6:IL-6Rβ complex structure as the template (PDB ID 1P9M), which clearly does not agree with the crystal structure of the IL-23:IL-23R complex. Residue W156 of IL-23A (in orange sticks) is indeed at the interface in the IL-23:IL-23R structure, but not at the interface in the modeled structure based on the IL6:IL-6Rβ complex.

Target 124: PorM with nanobody nb130 crystallization chaperone

PorM is one of the several porins that compose the bacterial Type IX secretion system (T9SS) and proved difficult to crystallize during a structural characterization of T9SS.²⁷ In an effort to identify a crystallization chaperone, llama antibodies were raised against PorM. Nanobody nb130 resulted from the effort—It could bind and stabilize PorM so that crystal structures could be obtained.

For this target, we needed to predict the complex of the PorM-Cter dimer—the homodimer of the C-terminal portion of PorM’s periplasmic domain—with the chaperone nanobody nb130. The unbound structure of nb130 was available (PDB ID 5FWO), but neither the PorM-Cter monomer structure nor any of the interfaces in the complex were known, hence we needed to predict both the PorM-Cter dimer and its interaction with nb130. We could not identify a suitable template to perform homology modeling of PorM-Cter, so we modeled its structure using the template-based fragment assembly algorithm I-TASSER²⁰. The template employed by I-TASSER (PDB ID 3MTR) happened to originate from a homodimer of neural cell adhesion molecule (NCAM), and we decided to use this template to model the PorM-Cter dimeric interface despite the low sequence identity. We subsequently docked the PorM-Cter dimer model with the nb130 unbound crystal structure, where we only allowed interaction of nb130 through its complementarity determining regions (CDRs).

The structure of the PorM-Cter-nb130 complex has been released (PDB ID 6EY6)²⁷ and shows that our model for PorM-Cter was incorrect, including its homodimeric interface. The released structure also shows that nb130 interacts with both chains of the PorM-Cter dimer. An accurate model of the PorM-Cter dimer was essential for a successful prediction of its interaction with nb130, and we did not correctly predict the interaction with nb130 either. However, the complex structure did confirm that our assumption of blocking the non-CDR regions of nb130 was correct.

Target 125: Lectin-like ligand transcript with a C-type lectin-like receptor

Target 125 comprises NKR-P1 (a C-type lectin-like receptor on human natural-killer cells) in complex with extracellular domains of LLT1 (lectin-like ligand transcript 1). According to CAPRI, one homodimer of LLT1 associates with two homodimers of NKR-P1.

The homodimer structure of LLT1 was available (PDB ID 4QKI), but for NKR-P1, we performed homology modeling (using Modeller¹⁹) using the mouse NKR-P1A structure as the template, which had a 45% sequence identity (PDB ID 3T3A). Under the assumption that the LLT1:NKR-P1 complex was C₂ symmetric, we first docked the LLT1 homodimer and the NKR-P1 homodimer, and then placed the second NKR-P1 using symmetry operations. If the two resulting NKR-P1 homodimers clashed, we discarded such predictions.

The LLT1:NKR-P1 complex structure has been released (PDB ID 5MGT), but the asymmetric unit does not represent the biologically relevant structure that we needed to predict, and the paper describing this complex structure is not yet published: In the asymmetric unit, one NKR-P1 homodimer associated with an LLT1 homodimer and another NKR-P1 homodimer (thus A2-B2-B2, where A stands for LLT1 and B stands for NKR-P1), whereas the information provided by CAPRI indicated association of an LLT1 homodimer with two NKR-P1 homodimers (thus B2-A2-B2). Therefore we cannot be certain about all the correct heteromeric interfaces and compare those with our predictions. However, our model of the NKR-P1 homodimer did not overlap well with the homodimers in the released structure, which presumably prevented us from making correct predictions of any of the heteromeric interfaces.

Targets 131 and 132: HopQ isotypes I and II bound to CEACAM1

The gastric pathogen Helicobacter pylori is implicated in a range of gastroduodenal diseases including stomach cancers. Binding of Helicobacter pylori’s adhesin protein HopQ to the human cell adhesion protein CEACAM is implicated in the delivery of the oncoprotein CagA into the host gastric epithelial cells.²⁸ Two targets of this round are related: the Helicobacter pylori cell adhesin proteins HopQ type I (target 131) and HopQ type II (target 132), bound to human cell adhesion protein CEACAM1. Unbound structures were available for both components of target 131 (PDB IDs 5DZL and 5LP2), but HopQ type II of target 132 needed to be homology modeled (template PDB ID 5LP2, the HopQ type I structure, with a sequence identity of 56%).

A recent report showed that mutation of CEACAM residues Y34 or I91 in a related CEACAM-HopQ complex affected binding,²⁹ thus we filtered the ab initio docking results to retain predictions with these residues at the interface. The complex structures have been published since the round was held (PDB IDs 6GBG and 6GBH for target 131 and target 132, respectively)³⁰. They show that the binding modes of the two complexes are virtually identical and that Y34 and I91 are indeed at the interface.

For target 131 we obtained a near-native prediction. However, it was rejected in the CAPRI assessment due to a large number of steric clashes, otherwise, it would have been of acceptable quality (f_nat was within the cutoff of a medium accuracy prediction, but I-RMSD and L-RMSD were not). We want to note that in order to avoid any further rejections due to steric clashes, we have revised our strategy of removing clashes since this round was held (see Methods). Our new strategy applied to the rejected prediction resulted in a clash count well below the cutoff, although the more aggressive clash removal also caused the number of native contacts to drop from 21 to 17. Despite this drop, f_nat and the other CAPRI criteria remained within the cutoffs of an acceptable prediction.

For target 132 we did not obtain a near-native prediction. This may be due to the docking algorithms being sensitive to the quality of the input structures, in particular, those obtained through homology modeling,^13–16,31 which was required for target 132. In retrospect, a more promising approach for the unbound/homology target 132 would have been to assume that the CEACAM1:HopQ type II complex adopts a binding mode similar to that of the CEACAM1:HopQ type I complex and generate predictions for the former by superposing the components onto the docking predictions of the latter. We tested this hypothesis after the results of this round were released, and indeed we obtained a model that is acceptable according to the CAPRI criteria. In Figure 2 we show the improvement resulting from this alternative approach, with which the best prediction was much closer to the crystal structure of the complex than the best prediction obtained using our original approach.

Figure 2:

Crystal structure of HopQ type II bound to the human cell-adhesion protein CEACAM1 (PDB ID 6GBH, green/cyan), and the best prediction from both the docking approach (orange) and from modeling using the predictions of target 131 as templates (magenta). The receptors of the three complex structures are superposed on each other, but for clarity, only the receptor in the crystal structure is shown (green).

Target 133: Colicin E2 and its cognate immunity protein, redesigned

Colicins are a class of antibiotics released by E. coli under stress conditions to kill competing strains. To avoid self (or same strain) destruction, the bacteria that produce colicin also produce an immunity protein (Im), which is an antagonist specific to its toxin. The complexes of colicin with Im proteins display a broad range of affinities, with dissociation constants ranging from a millimol to a femtomol per liter. Target 133 is complex between engineered forms of the endonuclease domain of colicin E2 and its cognate immunity protein Im2 (PDB ID 3U43), designed as part of a network of specific and promiscuous homologous protein-protein interactions.³²

A designed CAPRI target like this allowed us to ‘reverse engineer’ the design process, which we then incorporated in our prediction. We followed similar strategies in prior CAPRI rounds.³¹ Sequence alignment of the wild-type and designed proteins identified the mutated residues, which were heavily centered on the wild-type interface (Figure 3). We reasoned that the mutated residues either disrupted binding of the wild type complex or facilitated binding in the engineered complex. With the mutated residues centered on the wild-type interface, this implied that at least one of the engineered proteins had an interface similar to its wild type. We docked homology models of the two proteins generated by I-TASSER and filtered the predictions to ensure that at least several mutated residues were at the interface.

Figure 3:

Colicin E2 and its immunity protein Im2 (PDB ID 3U43), with the residues mutated in the designed complex colored red and magenta, for Im2 and Colicin E2, respectively. Each docking prediction was required to have at least several mutated residues in the interface, and the termini (colored in orange) were blocked from the interface in the docking calculation.

The released complex structure (PDB ID 6ERE)³² shows that our strategy was correct and three of our predictions were deemed acceptable accuracy. In fact, the interfaces of both proteins of the wild-type complex overlap nearly identically with the interfaces of the designed complex, and in retrospect, we might have achieved medium or high accuracy predictions had we homology modeled the complex using the wild type structure as the template. However, we did not anticipate such a close match and exclusively used our rigid-body docking approach.

Conclusions

Overall, our performance on these CAPRI rounds compared unfavorably with past rounds^13–16,31 or on benchmarks³³, but it must be noted that CAPRI targets are selected to be challenging.³⁴ This appears particularly the case for the current set, as evidenced by the performance of other CAPRI participants. Specifically for us, our performance was hampered by targets that involved higher-order multimers that required multiple cycles of interface prediction; for both targets 124 and 125, we made wrong predictions of a homodimeric subunit, which then prevented us from making correct predictions of the higher-order heteromeric interactions.

As in our previous CAPRI report,³¹ we made use of experimental information to guide our docking for all targets. This included mutation experiments (targets 122, 131, 132) and blocking non-CDR regions of a nanobody (target 124). Less conventionally, we also guided our prediction by reverse-engineering a designed complex (target 133). Although we did not make correct predictions for all these targets, all the experimental data we used were correct. On the other hand, for two targets we used complex templates to predict one of the interfaces of the complex (targets 124 and 125), and the information proved wrong. This may be due to the low sequence identity between the targets and templates, but it illustrates the importance of integrated approaches to account for erroneous or noisy experimental data.