Target highlights from the first post-PSI CASP experiment (CASP12, May–August 2016)

Abstract

The functional and biological significance of the selected CASP12 targets are described by the authors of the structures. The crystallographers discuss the most interesting structural features of the target proteins and assess whether these features were correctly reproduced in the predictions submitted to the CASP12 experiment.

Introduction

Integrity of the CASP experiment rests on the blind prediction principle requesting models to be built on proteins of unknown structures. To get a supply of modeling targets, the CASP organization relies on the help of the experimental structural biology community. In the latest seven experiments (2002–2014), the vast majority (>80%) of CASP targets came from structural genomics centers participating in the Protein Structure Initiative (PSI) program. With the end of the PSI in 2015, CASP faced a challenging task of replenishing the target supply normally provided by the PSI Centers. Dealing with this problem required diversification of target sources and going beyond the existing network of the recurring CASP target providers. Soliciting for targets, the organizers directly approached a wider set of structure determination groups, and also worked out a better protocol for obtaining and analyzing information about the structures placed on hold with the PDB. These efforts bore fruits, and 82 targets were secured for the CASP12 experiment. This number is quite impressive (considering that targets were collected in a short 3-month span of time) and is only somewhat smaller than the number of targets in a typical PSI-era CASP experiment (cf. 100 targets in the most recent CASP11 experiment). It is also worth mentioning that CASP12 targets came from 33 different protein crystallography groups stationed in 17 countries worldwide. Because of this variety, CASP12 targets exhibited wide diversity of sizes (from 75 to 670 residues), difficulties (from high accuracy modeling targets to new folds), quaternary structure composition (from single-domain targets to hetero-complexes), organisms (from rare extremophilic archaea from the depths of the Red Sea to Homo sapiens), and protein types (from globular to viral and membrane). Such diversity is vital for comprehensive testing of prediction methods. CASP organizers, who are co-authors of this paper, want to thank every experimentalist who contributed to CASP12 and thereby helped promote the development of more effective protein structure prediction methods. The list of all crystallographers who contributed targets for the CASP12 experiment is provided in Table 1 of the Supplementary material.

This manuscript is the fourth in a series of CASP target highlight papers^1–3. The chapters of the paper reflect the views of the contributing authors on twelve CASP12 targets: 1) the flagellar cap protein from Pseudomonas aeruginosa – T0886; 2) bacteriophage AP205 coat protein – T0859; 3) toxin-immunity protein complex from the contact-dependent growth inhibition system of Cupriavidus taiwanensis – T0884/T0885; 4) sorbitol dehydrogenase from Bradyrhizobium japonicum – T0889; 5) C-terminal domain of human gasdermin-B – T0948; 6) receptor-binding domain of the whitewater arroyo virus glycoprotein – T0877; 7) glycoside hydrolase family 141 founding member BT1002 – T0912; 8) a DNA-binding protein from Aedes aegypti – T0890; 9) snake adenovirus-1 LH3 hexon-interlacing protein – T0909; 10) an ice-binding protein from Antarctica – T0883; 11) a domain of UDP-glucose glycoprotein glucosyltransferase from Chaetomium thermophilum – T0892; and 12) a cohesin from Ruminococcus flavefaciens scaffoldin protein complexed with a dockerin – T0921/T0922. The results of the comprehensive numerical evaluation of CASP12 models are available at the Prediction Center website (http://www.predictioncenter.org). The detailed assessment of the models by the assessors is provided elsewhere in this issue.

1. FliD, the flagellar cap protein from Pseudomonas aeruginosa PAO1 (CASP: T0886, Ts886, PDB: 5FHY) – provided by Sandra Postel and Eric J. Sundberg

Bacterial flagella are long helical cell appendages that are important for bacterial motility and pathogenicity ⁴. These extracellular hollow filaments are formed by thousands of copies of FliC (flagellin) molecules and connected via a hook to the flagellar rotary motor anchored in the bacterial membrane ⁵. The motor drives the propeller-like motion of the filament, which confers swimming motility to the bacteria ⁶. An important structural and functional component of bacterial flagella is the flagellar capping protein, FliD, that is located at the distal end of the flagellar filament ⁷. Unfolded FliC molecules are translocated from the cell cytoplasm through the hollow filament pore to the tip of the growing flagellum where FliD regulates flagellar assembly by chaperoning and sorting FliC proteins. An absence of FliD leads to improperly constructed filaments and, consequently, impaired bacterial motility and infectivity ⁸. In the most commonly studied organism for flagella, Salmonella serovar Typhimurium, FliD is known to form a homopentameric complex on the tip of the flagellum, as shown in a low-resolution cryo-EM structure ^7,9,10. Until recently, these data provided the only available structural insight to FliD. Our crystal structure of a large fragment of FliD, FliD_78–405, from Pseudomonas aeruginosa PAO1 was the first high-resolution structure of any FliD from any bacterium, providing novel details concerning FliD function ¹¹.

In our crystal structure ¹¹, the Pseudomonas FliD_78-405 monomer exhibits an L-shaped structure (Figure 1A), which can be divided into two globular domains and a helical region. Domain D3 is a loop insertion into domain D2, and both domains have structural similarity to other flagellar proteins. Residues 309 to 405 of FliD_78-405 are highly flexible as revealed by hydrogen/deuterium exchange (HDX) and we were also unable to model those residues in our structure. Full-length Pseudomonas FliD_1-474 encodes predicted N- (residues 1 to 77) and C-terminal (residue 406 to 474) coiled coil domains that prohibited crystallization in our hands.

Figure 1. (A) Crystal structure of the Pseudomonas FliD_78-405 monomer subunit.

in which the domain D3 (CASP domain D2, green), domain D2 (CASP domain D1, blue) and the helical region (red), which belongs to domain D1 (not evaluated in CASP), are indicated. (B) Side view (top panel) and top view (bottom panel) showing cartoon representations of the hexameric FliD_78-405 crystal structure. Each monomer subunit is colored distinctly. (C) SAXS-generated molecular envelope of the monomeric FliD_1-474 with the CASP prediction model T0886TS036_1 (cyan). (D) Superposition of CASP prediction models T0886TS247_1_D1 (orange) and T0886TS247_1_D2 (orange) with D2 (CASP domain D1, blue) and D3 (CASP domain D2, green) of the FliD_78-405 monomer crystal structure. (E) Superposition of CASP prediction model T0886TS247_1 (orange) with the FliD_78-405 monomer crystal structure (domain coloring as in Panel A). (F) Superposition of CASP prediction model T0886TS247_1 (orange) with the E. coli FliD_43-416 crystal structure 5H5V (magenta). (G) Superposition of CASP prediction models T0886TS247_1 (orange), T0886TS011_1 (cyan), T0886TS064_1_1 (light blue), T0886TS411_1 (yellow) with the FliD_78-405 monomer crystal structure (domain coloring as in Panel A). (H) Superposition of CASP prediction models T0886TS247_1-D2 (orange), T0886TS064_1_1-D2 (light blue), T0886TS011_1-D2 (cyan), T0886TS411_1-D2 (yellow), T0886TS456_1-D2 (dark grey), T0886TS173_1_1-D2 (red) with D3 of the FliD_78-405 monomer crystal structure (green).

In contrast to the Salmonella FliD, which forms a pentamer, Pseudomonas FliD adopts a hexameric oligomeric state in the crystal structure (Figure 1B), as well as in solution and functions as a hexamer in vivo ¹¹. The number of protofilaments that comprise the flagellar filament upon which FliD oligomers reside varies among bacteria ¹², suggesting that FliD stoichiometries also vary between bacteria, which is supported by our results. More recently, the crystal structure of FliD from E. coli, which includes all residues except the N- and C-terminal coiled coils showed that this FliD protein also forms a hexamer ¹³.

Pseudomonas FliD was included in CASP12 as a regular target T0866 and small-angle X-ray scattering (SAXS)-assisted target Ts886. SAXS data of the monomeric full-length protein, FliD_1-474, for which no crystal structure yet exists, were collected and the data provided to the modelers to aid the structure prediction process of the shorter construct that we had crystallized. All the SAXS-assisted target models exhibit low similarity to the FliD crystal structure as shown in an overlay of the best model Ts886TS036_1 with our crystal structure in Figure 1C, but do fit well into the SAXS envelope (Figure 1C).

The models obtained during the regular prediction round without using the SAXS envelopes to assist model-building vary greatly. The highest ranked model T0886TS247_1 closely resembles the crystal structure of Pseudomonas FliD_78-405 on the individual domain level (Figure 1D). However, the connection between domain D2 (CASP domain D1) and domain D3 (CASP domain D2) diverges resulting in a relative positioning of these two domains that is different than in the crystal structure (Figure 1E). The low resolution of the SAXS molecular envelope of FliD_1-474 is potentially compatible with multiple, various domain arrangements and may have made it difficult to predict the exact positioning of the individual domains (Figure 1C). Residues 309 to 405 of FliD_78-405, which we could not model in the crystal structure due to poor or missing electron density, were in general modeled as helical bundles in T0886TS247_1. A superposition with the recently solved crystal structure of E. coli FliD_43-416 (PDB 5H5V ¹³), which covers a larger fragment of FliD, shows the correct prediction of helical bundles in those regions (Figure 1F). However, the bundles are placed in a different orientation relative to the D2 and D3 domains, and do show a differences in the placement of individual helices. These discrepancies between the model and the experimental structure may be due to the high flexibility in the linker region and in the helical regions that we detected by HDX¹¹.

Compared to T0886TS247_1, all of the other models exhibit substantially less similarity to the FliD_78-405 crystal structure (Figure 1G). Models of domain D3 (CASP domain D2) alone, however, exhibited greater likenesses to the crystal structure, with secondary structural elements generally predicted properly (Figure 1H). This might be related to the lower flexibility (as shown by HDX) of domain D3 in comparison to the rest of the FliD molecule. Overall, FliD seemed to be a difficult target to model, despite the SAXS data provided, and only domain D3 appeared to yield models by multiple modeling groups that resembled the actual crystal structure very well.

2. Structure of bacteriophage AP205 coat protein (CASP: T0859; PDB: 5FS4, 5JZR, 5LQP) - provided by Kaspars Tars, Roman I. Koning and Guido Pintacuda

ssRNA phages like MS2, Qβ and AP205 infect various gram-negative bacteria and are among the simplest known viruses used for decades as models to study various problems in molecular biology. Lately, ssRNA phages and their components have found several applications, notably in vaccine development ¹⁴. Capsid of ssRNA phages contains 178 copies of coat protein (CP) and a single copy of maturation protein, responsible for attachment of phage particles to bacterial receptor ¹⁵. When produced in bacteria, recombinant CP of ssRNA phages spontaneously assembles in virus-like particles (VLPs), containing 180 copies of CP. Due to strong interactions between two adjacent CP monomers, VLPs can be regarded as built from 90 CP dimers.

In general, VLPs are empty, non-infectious shells of viruses, devoid of genomic nucleic acid, but morphologically similar to the corresponding viruses. VLPs have several applications, the best known of which is vaccine development. For example, VLPs of Hepatitis B virus have been used as successful vaccines for a few decades¹⁶. VLPs can be used not only as vaccines against the disease caused by the virus of VLP origin, but also as scaffolds to induce strong immune response against virtually any antigen¹⁷. In this case, multiple copies of the antigen of interest should be attached to the surface of VLP. The immune system recognizes patterns of regularly repeating antigens on VLP surface as a potential threat to the organism, inducing highly elevated titres of antibodies and stronger T-cell responses compared to free antigen ¹⁸. To avoid pre-existing immune responses, pathogens that do not target humans are preferable as carriers of antigens. For this purpose, VLPs of ssRNA phages like MS2, Qβ and AP205 have been widely used ¹⁴.

For creation of vaccine candidate, the antigen of choice can be efficiently attached to VLPs by genetic fusion of CP and antigen genes. Since antigens must be presented on the surface of VLPs, the knowledge of the exact three-dimensional structure of VLP provides useful information about suitable sites of insertion of antigens in coat protein sequences. Due to folding problems, large insertions are often tolerated only at either N- or C-termini of CP, but this is possible only if the terminal end of CP is well exposed on the VLP surface. However, in VLPs of ssRNA phages studied so far, like MS2 ¹⁹, Qβ ²⁰, GA ²¹, PP7 ²², PRR1 ²³ and Cb5 ²⁴ both terminal ends are poorly exposed on the surface. Additionally, three N- and three C- terminal ends of neighbouring CP dimers on the VLP surface are clustered together, resulting in steric clashes among any N- or C-terminal insertions. Instead, a so-called AB loop is well exposed and well separated from AB loops of neighbouring CP subunits, but only relatively short amino acid sequences can be inserted in it without compromising the VLP stability. In contrast, AP205 VLPs have been known before to tolerate significantly longer insertions at both C- and N- termini²⁵, but the structural reason for this remained unknown. Since we failed to obtain high resolution crystals of recombinant AP205 VLPs, we constructed and crystallized an assembly-deficient AP205 CP mutant, capable to form dimers, but not VLPs. The obtained crystal structure was further fitted into a medium resolution cryo-EM map of native recombinant AP205 VLPs. Additionally, a solid-state NMR structure of AP205 coat protein was obtained from labelled AP205 VLPs. The obtained results revealed that compared to related ssRNA phages, the structure of AP205 CP is circularly permuted ²⁶, meaning that about 20 N-terminal residues including the first β-strand are found at the C-terminal part instead. This feature is made possible due to the close proximity of N- and C-terminal parts of two monomers within the dimer (Figure2AB). The result is that in AP205 VLPs both N- and C- termini are found in the same position as AB loops in other phages (Figure 2CD). This provides a structural basis for construction of vaccine candidates using AP205 VLPs.

Figure 2. Structural features of bacteriophage AP205 coat protein.

Coat protein in AP205 and related phages, such as MS2, builds very stable dimers. Two monomers are shown in different shades of grey (panels A and B). Notice the close proximity of N- (blue) and C- (red) termini in dimers. 90 dimers further assemble into VLPs (panels C and D). In MS2, AB loop (green) is the most exposed structure on the surface of VLPs. Compared to MS2, in AP205 the first β-strand (yellow) is shifted to the C-terminus, although it remains in the same position in 3D. As a result, in AP205, C-and N- termini are the most exposed features on VLPs. In panel (E), crystal structure of AP205 monomer (green) is superimposed with the modeled structure (blue and red). The overall fold of model is approximately correct, except that it lacks C-terminal β-strand. Residues 1-39 (blue) are correctly placed in respect to the sequence, corresponding to the first four β-strands. For the rest of model (red) residues are placed incorrectly according to the sequence and out-of-register errors occur. Notice also that position of N-terminus is relatively well predicted, while C-terminus is in a very different position.

Out of 499 models submitted on CASP12 target T0859, only one had a reasonably accurate overall structure (Figure 2E, red and blue). Model T0859TS001, made by researchers at Francis Crick institute, included almost all of the actual secondary structure elements apart from the C-terminal β-strand, which is unique for AP205, compared to other similar phages. About one third of the protein, comprising approximately 40 N-terminal residues was placed fairly accurately in respect to sequence, as compared to the crystal structure. This means that researchers correctly deduced that the first β-strand is missing in AP205. After residue 40, progressively increasing out-of-register errors occur in the model. At the C-terminal part the register shift is about 20 residues. Due to this shift, the C-terminal residues are modeled as α-helix although in crystal structure they form the extra (C-terminal) β-strand, not observed in similar phages. Therefore, the C-terminal part is not modeled correctly and does not suggest the placement of C-termini on the surface of VLP, close to AB loops in related phages. Even though the overall precision of the model is somewhat limited, the model correctly suggests that N-terminal part is indeed well-exposed on the surface of VLP and occupies the position of AB loops in related phages. If experimental data had not been available, the model T0859TS001 would have provided significant biologically relevant information for construction of VLP based vaccines.

3. Structure of the toxin-immunity protein complex from the contact-dependent growth inhibition system of Cupriavidus taiwanensis (CASP: T0884/T0885, PDB: 5T87) – provided by Karolina Michalska, Christopher S. Hayes, Celia W. Goulding and Andrzej Joachimiak

Contact-dependent growth inhibition (CDI) is an important mechanism of inter-cellular competition found in Gram-negative bacteria. Bacteria utilizing the CDI system (CDI⁺) use diverse CdiB-CdiA two-partner secretion systems to deliver protein toxins directly into neighboring bacteria ^27,28. CdiB is an outer membrane transport protein exporting the CdiA effector onto the cell surface. CdiA recognizes specific receptors on susceptible bacteria and translocates its C-terminal toxin domain (CdiA-CT) into the target cell ^29–31. The variable CdiA-CT toxin region is usually demarcated by a conserved peptide motif, such as the VENN sequence found in enterobacterial CdiAs ³². Different CdiA-CTs can be fused to heterologous CdiA proteins at the VENN motif to generate novel chimeric effectors ^28,32,33. CdiA proteins carry a variety of toxin domains, most commonly exhibiting nuclease or pore-forming activities ^32–35. To protect against self-inhibition, CDI⁺ bacteria produce CdiI immunity proteins, which bind and neutralize cognate CdiA-CT toxins.

We have selected the CdiA-CT/CdiI complex from Cupriavidus taiwanensis LMG 19424 for structural analysis. PSI-BLAST searches for CdiA-CT homologs recover several predicted S-type pyocins from Pseudomonas species and MafB toxins from Neisseria species ³⁶ (50–70% sequence identity). Other hits include CdiA-CT domains from Rhizobium leguminosarum and Achromobacter strains, and Rhs peptide-repeat proteins from Streptomyces species. All of these homologs are predicted to mediate inter-bacterial competition ^37,38, though none have been validated experimentally. An HHpred-based search identified the C-terminal domain of 16S rRNA-cleaving colicin E3 ^39,40 as a possible structural homolog having 9% sequence identity to CdiA-CT. The CdiI immunity protein is less conserved than CdiA-CT, with homologs sharing ~30–40% sequence identity. An HHpred analysis recovered proteins with α-helical hairpin repeats, with the armadillo-like γ-COP coatomer (13% sequence identity with CdiI) being the closest match.

The 2.40 Å resolution crystal structure of the CdiA-CT/CdiI complex (Figure 3A) shows that the toxin putative catalytic domain (75 residues) consists of a central four-stranded antiparallel β-sheet, sandwiched by two N- and C-terminal α-helices and one 3₁₀ helix. The immunity protein (116 residues) is composed of three consecutive α-hairpins creating an armadillo-like structure. The N-terminal β-strand of CdiI protrudes from the helical body to complement the CdiA-CT β-sheet, potentially influencing toxin conformation. This arrangement also suggests that the N-terminal segment of CdiI is likely disordered in the free CdiI. A Dali server search for CdiA-CT structural homologs identified only low-similarity matches: inorganic triphosphatase (Z-score 3.7, RMSD 3.3 Å, PDB:3TYP) (Figure 3B) and WW domain of human transcription elongation regulator 1 (Z-score 3.5, RMSD 2.9 Å, PDB:2DK7). More distant hits include E. coli ParE toxin (Z-score 3.0, RMSD 2.4 Å, PDB:3KXE) (Figure 3C), which belongs to the barnase/EndoU/colicin E5-D/RelE (BECR) family (PMID:22731697). Although structurally related, these toxins display different activities: ParE family poison DNA gyrase ⁴¹, RelE is a ribosome-dependent mRNase ⁴², and colicins D/E5 cleave the anticodon loops of specific tRNAs ⁴³. Therefore, the exact biochemical function of CdiA-CT cannot be predicted easily and may include RNase or DNase activity. The CdiI fold is well-represented in the PDB and is a popular scaffold for designer proteins. The closest match corresponds to human deoxyhypusine hydroxylase (Z-score 12.3, RMSD 2.0 Å, PDB:4D4Z), followed by protein phosphatase 2 (Z-score 12.3, RMSD 2.5 Å, PDB:2IE3) and other proteins with virtually no sequence similarity to CdiI. Though many of the homologs engage in protein-protein interactions, none are annotated as an immunity protein.

Figure 3. The CdiA-CT/CdiI^Ctai complex.

(A) Experimental structure with the most conserved residues and their interactions shown in stick representation. The CdiA-CT toxin domain is shown in teal and the CdiI immunity protein in pink. Hydrogen bonds are depicted as red broken lines. Superposition of CdiA-CT with (B) the closest PDB homolog, inorganic triphosphatase (coral, PDB:3TYP), (C) with ParE toxin from E. coli (yellow, PDB:3KXE) and (D) with model T0884TS183_1 (purple) and refined model TR884TS118_1 (blue). The strand β1 from CdiI is shown for reference. (E) Superposition of CdiI with model T0885TS005_2 (cyan) and refined model TR885TS247_1 (blue).

Antitoxin proteins often bind over nuclease toxin active sites to prevent substrate access. Typically, nuclease toxins are highly electropositive and the cognate immunity proteins carry complementary acidic residues to promote electrostatic interactions. CdiA-CT contains several basic residues, including conserved His212, His214 and Arg216 (Figure 3A), which may be key catalytic residues. CdiI is more electrostatically neutral than previously characterized immunity proteins. It directly interacts with the toxin’s putative active site using the conserved His72, Arg75 and Asp108 residues, which form a hydrogen bond, stacking interaction and salt-bridge, respectively. As mentioned above, β1 of CdiI complements the toxin fold.

For the CASP12 competition, CdiA-CT and CdiI were modeled as monomers and as a hetero-complex.

For CdiA-CT (T0884), the best model (out of 185 total monomeric predictions) was generated by QUARK (T0884TS183_1), which uses ab initio algorithms with no global template information. This model scored 66 GDT_TS points, 10 points higher than the next model, T0884TS236_1 generated by MULTICOM-construct. The highest-scoring regular prediction model T0884TS183_1 was subsequently released for refinement, where it was further improved to GDT_TS of 76 by the PKUSZ_Wu_group (TR884TS118_1). Model T0884TS183_1-D1 closely resembles the crystal structure, though helix α1 is misoriented and the β3-β4 hairpin is distorted (Figure 3D). However, we note that toxin helix α1 is constrained by the immunity protein in the CdiA-CT/CdiI complex. Therefore, it is possible that the free toxin domain adopts the conformation predicted by the computational model. Toxin residues that interact with the immunity protein are generally located in proper positions, though a more accurate spatial prediction of β4 would bring the conserved His212 and His214 to better agreement with the crystal structure.

CdiI (T0885) is a more straightforward structure prediction target with fewer discrepancies among the 190 predicted models. The best five models for this target were generated by the BAKER-ROSETTAserver group, with the top model T0885TS005_2 scoring 88 (out of 100) GDT_TS points (Figure 3E). The next model in the accuracy ranking was generated by the MULTICOM-novel group scoring 15 GDT_TS points below the best. As we found with CdiA-CT, the major misalignments were observed for peripheral elements (β1 and the C-terminus of helix α6) involved in protein-protein interactions. Similarly to the CdiA-CT, the best server model for CdiI, T0885TS005_2, was released for the refinement (without the 11 N-term residues trimmed by the assessors) and was further improved to 95 GDT_TS points (TR885TS247_1).

These examples show that computational prediction can yield models with correct folds, and when combined with sequence conservation analysis, can inform rational mutagenesis and biochemical analyses.

Even though the monomeric subunits of the CdiA-CT/CdiI (T0884/T0885) hetero-complex were predicted quite well, the full complex was modeled poorly. Although some of the multimeric models reached reasonable global accuracy scores (e.g., LDDT of 0.73 for TS239_1), the visual inspection showed that all models left the putative active site of toxin fully exposed and failed to properly predict the correct protein-protein interface. Accuracy of interface contacts in the submitted predictions is rather poor, with the highest recall of 23.4% achieved in the prediction TS203_3, where subunit molecules partly overlap. Thus, for the CdiA-CT/CdiI complex, in silico approaches did not provide useful information to confidently determine complex organization important for understanding function and catalysis.

4. Sorbitol dehydrogenase (BjSDH) from Bradyrhizobium japonicum (CASP: T0889; PDB: 5JO9) - provided by Leila Lo Leggio, Folmer Fredslund and Gert-Wieland Kohring

Rare sugars are defined as monosaccharides and their derivatives which are rare in nature, and these sugars have attracted interest for potential medical and food applications ⁴⁴. Consequently, enzymes able to produce and interconvert rare sugars have also attracted attention. One such enzyme is the Zn-independent short chain dehydrogenase from Bradyrhizobium japonicum (BjSDH) which uses NAD⁺/NADH as a non-covalently bound cofactor. We initiated structural studies of BjSDH (CASP ID T0889) as part of a collaborative EU project devoted to the development of an electro-enzymatic flow-cell device for the production of rare sugars ⁴⁵. BjSDH was selected for structure determination due to some favorable properties. First of all, while BjSDH preferentially catalyses the oxidation of D-glucitol (a synonym for D-sorbitol) to D-fructose, it can also catalyse the oxidation of L-glucitol to the rare sugar L-sorbose with enzymatic cofactor regeneration and high D-sorbose yield ⁴⁶ (Figure 4A). Sorbitol dehydrogenases are additionally of particular interest in biosensor technology, since D-sorbitol is a marker for onset of diabetes as well as a food ingredient ⁴⁷. Furthermore, it is a thermostable enzyme with T_m of 62 °C ⁴⁶, which is a desirable property for potential industrial use and biosensor technology, as thermostability often correlates with general stability.

Figure 4.

(A) Products of reaction catalyzed by BjSDH with D-glucitol and L-glucitol as substrates; (B) Structure based sequence alignment of region around loop 193-203 covering the active site of BjSDH. Sequences of GatDH, RsSDH and top 5 DALI hits searching with the BjSDH structure are shown; (C) BjSDH structure shown as cartoon (gold) and symmetry related molecule packing against is (grey). Ligands in the structure are shown as sticks, while loop 193-203 in top 5 models from CASP12 are shown as lines; (D) Continuous β-sheet between two monomers in BjSDH crystal structure, and same region in the RsSDH crystal structure.

Structure determination ⁴⁸ was not straightforward due to limited resolution, which was estimated to be at 2.9Å according to CC_1/2 of about 50% in the outer resolution shell ⁴⁹, but closer to 3.2Å with more conventional evaluation of resolution limit at I/σ(I) around 2. The Molecular Replacement model (PDB code 4NBU ⁵⁰) had only 29 % sequence identity to the target after structure-based alignment. As all short chain dehydrogenases, BjSDH adopts a Rossman fold ⁵¹ and has a catalytic tetrad (Asn112, Ser140, Tyr153 and Lys157). BjSDH was co-crystallized with NAD⁺ and D-glucitol. D-glucitol could be modeled in the electron density map and phosphate is clearly bound, mimicking part of the cofactor, however a full co-factor molecule could not be modeled. This is probably due to the presence of 1.4 M NaH₂PO₄/K₂HPO₄ in the crystallization conditions, competing with the cofactor. Although there is only one molecule in the asymmetric unit, the enzyme forms a tetramer in the crystal structure due to crystallographic symmetry, and this is also assumed to be the predominant form in solution ⁴⁸.

All the closest structural relatives identified with DALI after structure determination (reported in Fredslund et al ⁴⁸), have only around 30% sequence identity, and while most are dehydrogenases, none are denoted as sorbitol dehydrogenases. When compared to the DALI results, the most structurally diverse part of the structure is a helix-turn-helix motif or “lid” loop, residues 189–205 in BjSDH, partly responsible for ligand binding. This loop is different in length, sequence and conformation (Figure 4B), compared to enzymes with relatively similar specificity like R. sphaeroides sorbitol dehydrogenase RsSDH ⁵². The analysis of the DALI results also confirmed that the catalytic tetrad is highly conserved structurally in BjSDH compared to similar dehydrogenases. All the top DALI hits also form tetramers with similar symmetry.

To see if structural features of target T0889 were correctly predicted in CASP12 models, we analyzed the top 5 monomeric models (based on the GDT_TS score) and the top oligomeric model (based on the recall score for interface contacts).

The monomeric models were based solely or in part on the structure of clavulanic acid dehydrogenase from Streptomyces clavuligerus ⁵³ (PDB entry 2JAH or 2JAP), which was also the top DALI hit. Unsurprisingly, the models predict correctly the positioning of the catalytic tetrad and overall predict the structure of BjSDH in a satisfactory manner. However, the helix-turn-helix loop is different in the 5 top scoring models as compared to the crystal structure and the model used for molecular replacement. Since the resolution of the crystal structure is limited, and this loop in particular was difficult to trace in the electron density, there might be errors in the crystallographic model, but the conformation of the loop from several CASP12 models is definitely incompatible with crystal packing (Figure 4C) and cannot accurately represent the conformation it assumes in the crystal. On the other hand, crystal packing could have affected the conformation and furthermore, the loop is involved in ligand binding, which would not be taken into account explicitly by the modeling programs and could also affect its conformation.

One of the most important features of BjSDH was its thermostability⁴⁶, as the knowledge of its structural determinants may help stabilize related enzymes by protein engineering. In particular, we compared the structure to the sorbitol dehydrogenase RsSDH, for which the melting temperature by CD spectroscopy was also measured and found to be much lower than for BjSDH under similar conditions (T_m of 47 °C vs 62°C). One of the striking features in BjSDH is a much higher proline/glycine ratio compared to RsSDH, a feature which is obvious from the sequence and does not require knowledge of the 3D structure. An additional feature which is likely to affect stability becomes obvious only through analysis of the quaternary structure. As previously mentioned BjSDH is a tetramer in the structure and in solution, as are many members of the short chain dehydrogenase family, and probably also RsSDH⁵². In BjSDH, two monomers of the tetramer have a large interaction surface via a continuous β-sheet formed between the two monomers, while this is not the case in RsSDH, indicating a less stable tetramer in the latter (Figure 4D). As the top CASP12 models for BjSDH were all based on the clavulanic acid dehydrogenase structure, which is also a tetramer and includes the continuous β-sheet between subunits, the top monomeric models are all compatible with an intersubunit β-sheet formation.

Among the oligomeric models, model TS188_4 from the chuo-u group was the best as judged by the interface contact recall (http://predictioncenter.org/casp12/multimer_results.cgi?target=T0889o). The model represents the same homo-tetrameric assembly as the target structure T0889 (BjSDH) and the PDB structure 2JAH, which was used as a template. The tetramer interfaces are modeled reasonably well, with 72% of the native interface contacts being correctly reproduced, while the constituting monomers lack some details, which may affect the analysis of the protein stability. It should be noted, though, that the top model does not have much added value compared to the 2JAH template, as their superposition yields a Cα RMSD of only 0.7 Å.

In conclusion, the top CASP12 models reproduce correctly some but not all biologically and biotechnologically interesting features of SDH, specifically they cannot predict the lid loop conformation, which is part of the substrate binding pocket, or subtle details of the interactions in the tetramer.

5. Crystal Structure of the C-terminal Domain of Human Gasdermin-B (CASP: T0948; PDB: 5TJ4, 5TJ2, 5TIB) - provided by Kinlin L. Chao and Osnat Herzberg

Biological Significance of Gasdermin-B

The human genome encodes four gasdermins (GSDMA-D) that are expressed in epithelial cells of the gastrointestinal tract and skin, regulating the maintenance of the epithelial cell barrier, cell proliferation, differentiation and programmed cell-death processes ^54,55. Based on the different protein levels in cancers, human GSDMA, GSDMC and GSDMD are considered tumor suppressors and GSDMB (CASP12 target T0948), a tumor promoter. GSDMB amplification and GSDMB overexpression lead to poor response to HER2-targeted therapy in HER2-positive breast cancer ⁵⁶. The N-terminal domain of gasdermins possesses membrane-binding activity, whereas the C-terminal domain autoregulates the lipid binding function. Multiple genome-wide association studies (GWAS) revealed a correlation between single nucleotide polymorphisms (SNPs) in the protein coding and transcriptional regulatory regions of the neighboring GSDMA, GSDMB and ORDML3 genes with susceptibility to asthma ⁵⁷, type 1 diabetes ^58,59, Crohn’s disease, ulcerative colitis ^59,60 and rheumatoid arteritis ^59,61. Pal and Moult identified 2 GSDMB SNPs (dbSNP:rs2305479 and dbSNP:rs2305480) in linkage disequilibrium with a marker of disease risk ⁵⁹. They correspond to a Gly299 → Arg299 change (rs230549), and a Pro306 → Ser306 change (rs2305480) in the C-terminal domain of GSDMB (GSDMB_C) (numbering scheme according to Uniprot isoform Q8TAX9-1). Analyses of the 1000 Genomes Project Consortium data ⁶² showed co-occurrence of the 2 SNPs (Gly299:Pro306 or Arg299:Ser306) with ~50% occurrence of each combination in the general population (Pal and Moult, unpublished). Unlike monogenic diseases which are caused by high penetrance SNPs in single genes, complex-trait diseases are associated with multiple low penetrance SNPs in multiple genes. Most of the SNPs present in a genome are actually not disease causative. However, because of linkage disequilibrium within the genome SNPs the challenge for the large-scale genome sequencing is to reveal the disease causative SNPs. The structural studies of GSDMB_C were undertaken to provide insights into possible mechanisms that the SNPs may contribute to disease risk ⁶³.

Key features of Gasdermin-B C-terminal domain

GSDMB amino acid sequence is homologous to the sequence of Gsdma3, the mouse homolog of GSDMA. The structure of Gsdma3 (PDB 5B5R) revealed 2 domains connected by a long flexible linker. The N-terminal lipid-binding domain folds into an α+β structure, and the C-terminal inhibitory domain adopts an α-helical fold comprising 8 helices ⁶⁴. The 7-helix bundle topology of GSDMB_C (α5-α11 in our paper ⁶³ describing the crystal structure, PDB 5TJ4, 5TJ2, 5TIB ) is the same as that of Gsdma3, except that it lacks a Gsdma3 subdomain comprising an α-helix and a 3-stranded β-sheet between the last two α-helices (Fig 5A–C).

Figure 5.

(A) Structure-based sequence alignment of the GSDMB (T0948 comprises GSDMB’s C-terminal domain) and mouse Gsdma3 C-terminal domains with secondary structure elements shown above or below the respective sequences. Identical and conservatively replaced residues are colored in red and blue. The alignment was performed using the programs Clustal Omega ¹¹⁸ and ESPript 3 (espript.ibcp.fr/Espript/). (B) Ribbon diagram of the GSDMB_C fold (PDB 5TIB). The α7–α8 GSDMB loop containing the polymorphism residues is colored in red. (C) Superposition of the experimental GSDMB_C structure (colored yellow) and the corresponding Gsdma3 domain that served as a modeling template (blue, 5B5R), (D) Superposition of the experimental GSDMB_C structure (colored yellow) and the best GTD_TS CASP12 scored model of group 251 (green). (E) Superposition of the polymorphism loop of the experimental structure (colored gray with α′ highlighted in orange) with the corresponding loop assessed as the closest (Group 330) based on the position specific criterion (colored cyan with α′ highlighted in magenta).

We determined three crystal structures of the GSDMB_C containing (1) the Arg299:Ser306 pair corresponding to individuals with increased disease risk, (2) the Gly299:Pro306 present in healthy individuals, and (3) the Gly299:Ser306 combination, one from each allele. The second possible combination, Arg299:Pro306, did not yield well diffracting crystals ⁶³. The SNP residues at positions 299 and 306 are located on a loop connecting the α7 and α8 helices of GSDMB (Figure 5AB). Three GSDMB_C structures provide 16 independently determined molecules in their asymmetric units: 6 with Ser at position 306 and 10 molecules with Pro at that position. All 16 versions of this loop contain a 5-residue α-helix (α′, Pro309-Ser313) (Figure 5AB). However, the loops with Ser306 adopt an additional well-ordered 4-residue helical turn (Met303-Ser306) between the α7 and α′ helices (Figure 5B). By contrast, the loops with a Pro306 do not form this helical turn and each loop version assumes different backbone conformations ⁶³. In addition, a Gly299→ Arg299 alters the charge distribution on the protein surface. Examination of the structures shows that, unlike a more flexible Ser306 side chain, Pro306 cannot be accommodated at the end of the helical turn because its side chain would clash with main chain carbonyl atoms of the preceding residues. One or both of these changes may contribute to the susceptibility of individuals to develop diseases by possibly modulating the selectivity and binding affinity of its N-terminal domain to lipids or the association with partner proteins, for example HSP90β or fatty acid synthase ⁶⁵.

CASP12 predictions for the functionally important regions of GSDMB_C

The 166-residue GSDMB_C CASP12 target sequence (T0948) contained the Arg299:Ser306 pair found in individuals with increased disease risk (PDB 5TIB). The publication of the full-length Gsdma3 structure shortly prior to the CASP12 prediction deadline provided a homologous template for T0948 (PDB 5B5R ⁶⁴). T0948 and the 198-residue Gsdma3 C-terminal domain share 34.5% sequence identity, and superpositioning yields a RMSD of 2.3 Å for 113 shared Cα positions (Figure 5C). However, a 33 amino acid Gsdma3 subdomain between α10 and the last helix (Gsdma3 α12 or GSDMB α11) corresponds to a disordered loop in GSDMB that is too short to form an analogous subdomain (Met366–Tyr382) ⁶³, and therefore could not be predicted. This Gsdma3 region is functionally important because it interacts with a segment on the N-terminal domain that is involved in membrane disruption ⁶⁴.

A total of 422 predictions for T0948 were deposited in CASP12, and 150 of them had GDT_TS scores > 70. The Gsdma3-based models for T0948 were quite accurate for the well-aligned core 7-helix bundle region, but not for the functionally important polymorphism loop. The superposed structures of GSDMB_C and the highest GDT_TS scored model, from group 251 (myprotein-me server, Skwark and colleagues) illustrate the similarity within the core 7-helix bundle (Figure 5D). However, the predictions for the polymorphism loop conformation (i.e. residues Arg299–Val322 of GSDMB corresponding to Arg54-Val77 in T0948) were poor, presumably because the GSDMB loop is 8 residues longer than that of Gsdma3 and lacks significant sequence homology ⁶³ (Figure 5A). Encouragingly, many top models (although not TS251_1-D1, Figure 5D) predicted the α′ helix (Pro309-Ser313) in the polymorphism loop. However, its length was overestimated and its orientation was wrong in all cases. Examination of the CASP analyses tables including position-specific alignment shows that large differences exist even for the polymorphism loop closest to the crystal structure (e.g., group 330, Laufer_seed, Perez and colleagues - Figure 5E). No group reproduced in their prediction the 4-residue helical turn preceding Ser306, a key structural difference that distinguishes the GSDMB produced by Crohn’s, ulcerative colitis, and asthma patients from that of healthy individuals. Thus, the GSDMB example shows that prediction of the conformations of large loops that deviate substantially from their template structures has not yet achieved the level of accuracy required for drawing conclusions about structure-function relationships.

6. Receptor-binding domain of the Whitewater Arroyo Virus glycoprotein: studying pathogenicity from a structural point of view (CASP: T0877; PDB: 5NSJ) – provided by Amir Shimon and Ron Diskin

Some enveloped RNA viruses from the Arenaviridae family attach to Transferrin Receptor 1 (TfR1) and use it as a cellular receptor for cell entry. For binding to TfR1, they utilize the receptor-binding domain (GP1) that is part of their class-I trimeric spike complex. Several arenaviruses can infect humans and cause acute disease due to their ability to bind the human-TfR1 (hTfR1) in addition to TfR1 from rodents and bats that naturally serve as hosts for these viruses.

Since both pathogenic and non-pathogenic arenaviruses use similar rodent-TfR1 receptors but only the pathogenic viruses can utilize hTfR1, we wanted to understand what the structural barriers are that prevent non-pathogenic viruses from doing so. This information is important if we want to understand the molecular mechanisms that may allow non-pathogenic viruses to emerge into the human population as novel pathogens. To compare non-pathogenic and pathogenic arenaviruses, we crystallized the GP1 domain from the non-pathogenic Whitewater Arroyo virus (WWAV) ^66,67 and compared its structure with the GP1 from the pathogenic Machupo arenavirus determined in complex with hTfR1 by the Harrison group⁶⁸.

This structural information allowed us to analyze a putative interaction of WWAV-GP1 with hTfR1 (Figure 6A). We found several structural features that preclude hTfR1 usage ⁶⁹, including electrostatic incompatibility between WWAV-GP1 and hTfR1 (Figure 6B). Interestingly, similar incompatibilities equally affect the pathogenic viruses. These pathogenic viruses can nevertheless use hTfR1 due to more elaborated sets of weak interactions throughout their binding sites that allow them to energetically overcome the structural incompatibilities ⁶⁹. Thus, viruses within this family make different interactions with TfR1, giving rise to a range of affinities toward TfR1, which ultimately determine their potential to utilize hTfR1 despite the structural barriers ⁶⁹.

Figure 6. The structure of WWAV-GP1 compared to the top three models.

(A): Ribbon diagrams of the WWAV-GP1 colored in rainbow and shown in a putative complex with hTfR1 (surface representation) (PDB ID: 3KAS). (B): A potential charge-repulsion between two negatively charged groups on WWAV and hTfR1 that was identified using this analysis. (C): Comparison of the top three models from ‘MULTICOM-construct’, ‘MULTICOM-novel’, and ‘GOAL’ (designated S236, S345, and S220, respectively) with WWAV-GP1. (D): A close-up view comparing the loops of WWAV-GP1 that interact with hTfR1 to the top model. Structures were rendered using PyMOL (www.pymol.org).

This study required an accurate structure of WWAV-GP1. Sequence conservation of viral glycoproteins like the GP1 domains from TfR1-tropic viruses is generally very low, due to rapid evolution under strong immunological pressure (i.e. 24 % identity between the GP1s of Machupo and Whitewater Arroyo viruses). Thus, a modeling approach may not fully reveal the fine details that are needed for such an analysis. In CASP12, the GP1 domain from WWAV was designated as a target for automated servers (T0877). Most of the predictors were able to provide models that faithfully represent the overall structure of this domain with GDT_TS > 50. We compared the top three models to the crystal structure of WWAV-GP1 (Figure 6C). ‘MULTICOM-construct’, ‘MULTICOM-novel’, and ‘GOAL’ achieved the best overall ranking with GDT_TS of 67.8, 68.7, and 70.3, respectively. The central β-sheet and the α-helices were modeled correctly along the primary structure but slightly deviate from their real positions in space. Interestingly, a disulfide bond that WWAV has but is not shared by GP1 domains for which structural information was previously available, was not modeled although the cysteine residues were placed in their correct orientations. Since this bond influences the local geometry of a near-by loop, the modelers were unable to accurately model its conformation. In general, the conformations of the loops from the various predictors cluster together, but deviate from the real structure of WWAV-GP1. Considering the goal of our study, this is a major drawback since some of the important contacts that GP1 makes with TfR1 are mediated through these loops (Figure 6D). Thus, modeling loops is a challenging task and since loops are often involved in protein-protein interactions, bona fide structural information would be preferred for the type of analysis that we have performed.

7. Structure features and biological significance of a new glycoside hydrolase family 141 founding member BT1002 (CASP: T0912; PDB: 5MPQ) - provided by Didier Ndeh, Arnaud Baslé and Harry J. Gilbert

Rhamnogalacturonan II (RG-II) is a primary cell wall pectin of plants present in fruits, vegetables, wine and chocolate. It is the most complex carbohydrate known and despite its remarkable structural complexity, it is highly conserved across the plant kingdom ^70,71. RG-II is a complex 10 kDa acidic polysaccharide ^70,72. To elucidate how the human gut microbiota (HGM) has evolved to utilise complex glycans we investigated the RG-II degradome of the prominent gut microbe Bacteroides thetaiotaomicron. The organism is capable of metabolising RG-II in in-vitro growth experiments, and combined transcriptomic and biochemical data revealed that at least 23 enzymes induced in culture conditions with RG-I as the sole carbon source are directly involved in its metabolism ^72,73. The organism is capable of cleaving 20 out of the 21 unique glycosidic linkages in RG-II and biochemical evidence suggests that the CASP12 target T0912 (BT1002) is one of 7 novel enzymes recruited by B. thetaiotaomicron to achieve this purpose⁷².

BT1002 is a novel α-L-fucosidase and founding member of the new glycoside hydrolase family 141 (GH141) ⁷⁴. BT1002 targets the complex tetrasaccharide structure mXFRA found in RG-II. The importance of BT1002 in RG-II metabolism is exemplified by the fact that genetic mutants lacking the enzyme are unable to metabolise mXFRA during in-vitro growth on RG-II, leading to accumulation of mXFRA in the growth medium. This implies that the enzyme is unique and indispensable for the breakdown of its target in RG-II.

We solved the BT1002 phase problem using selenomethionine single-wavelength anomalous diffraction. The crystallized construct diffracted to a resolution of 2 Å. It comprises 624 amino acids of which 605 were modeled (PDB ID 5MQP). BT1002 contains 12 α-helices and 50 β-strands forming 6 sheets. The catalytic domain is made of the C-terminal and N-terminal ends of the protein (residues 19-113 and 300-618 respectively), which fold into a β-helix. An extended loop of the catalytic domain comprising residues 323 to 370 mediates contacts between the β-helix and the β-sandwich domains (D1 and D2) made of residues 114 to 299. Domain D3 is flanked by two α-helices (Figure 7, panel A). While efforts to identify specific active site interactions between BT1002 and its tetrasaccharide target are ongoing, we identified two aspartates (Asp523 and Asp564) as potential catalytic residues through site directed mutagenesis⁷². The residues are 6.1 Å apart in a pocket suggesting an acid-base assisted double displacement mechanism. The closest structural homolog we found using a DALI search with the catalytic domain was a GH-120 β-xylosidase (PDB code 3VSU) with a root mean square deviation of 2.7 Å. While the active site pockets are conserved their primary sequence (20% identity), their catalytic centers and their specificities are very different.

Figure 7. (A): Cartoon representation of BT1002.

(5MPQ, chain A) aligned with T0912TS349_1 in pymol (sequence alignment followed by structural superposition with Cα atoms only). Residues are colored by a RMSD gradient (dark blue is a good alignment and red are higher deviations). Residues not used are colored grey. The domain are labelled D1 to D3. (B): Binding pocket surface representation. The predicted model (T0912TS303_1) surface is represented in solid dark grey and the PDB model surface in yellow mesh. The putative catalytic residues in the predicted model are colored magenta and red in the PDB model.

The BT1002 protein was included in CASP as target T0912 and was evaluated in the full-length and domain-based modes (domain D1: residues 24-113 and 299-622; D2: 114-154 and 258-299; D3: 155-257). Out of the 456 models submitted on the target, 175 models scored 40 GDT_TS or higher. Considering large size of the target and its multi-domain composition, such prediction can be considered as successful. The best top ranked model (i.e. the best model among models assigned as #1 by each of the groups) was submitted by the wfMESHI-TIGRESS group (T0912TS303_1, GDT_TS=48.2). To illustrate how well different regions of the protein are predicted, we aligned the BT1002 crystal structure with a mid-range model (T0912TS349_1, HHPred1, GDT_TS=40.8). The result is presented in Figure 7 (panel A) where colder colors indicate a close match and hotter colors a higher RMSD (residues in grey were not used). The backbone of the catalytic domain D1 was very well predicted with the 11 parallel β-strand stacks of the β-helix correctly identified (194 models scored above GDT_TS=50 with the best model’s GDT_TS=66.4). This is not surprising as such a domain is well described with multiple examples in the PDB data bank. Side chain positioning is more distant to the crystal protein structure. For instance, the catalytic residues Asp564 and Asp523 are separated by about 9 Å in the best D1 model rather than 6.1 Å in the crystal structure. The domain D2 was also correctly modeled overall (85 models scored above GDT_TS=50 with the best model’s GDT_TS=77.7). The third domain was poorly predicted, with the best model scoring only GDT_TS=42.0. Nevertheless, this model (T0912TS247_1-D3) correctly predicted the β-strands and the β-sandwich, though with a register error. As a consequence, the flanking α-helices were missed. The overall fold prediction accuracy is essential for this target. Indeed, the binding pocket important for ligand recognition and binding, is not only constituted by the surface of the catalytic domain D1 and its extended loop but also the surface of domain D3. Therefore we had to consider only the full target predictions. Figure 7 (panel B) shows an overlay of the best predicted model (T0912TS303_1) and the experimental model (5MQP). The PDB model surface represented as a yellow mesh is clearly smaller than the predicted model surface in dark grey. Additionally, the putative catalytic residues are more distant in the predicted model (magenta surface) than in the PDB model (red mesh).

In summary, the BT1002 structure prediction results are very encouraging but show the challenges facing the community in order to elucidate complex biological functions.

8. A cryptic DNA-binding protein from Aedes aegypti (CASP: T0890; PDB: N/A) - provided by Reinhard Albrecht and Marcus D. Hartmann

During their development, pupating insects (holometabola) may accumulate uracil in the DNA of larval tissues. The protein UDE has been implicated in the development of holometabola in the late larval stages as a uracil-DNA degrading factor. At the time of its experimental identification in Drosophila larval extracts, homologs were only found in holometabola⁷⁵. Its sequence revealed a domain organization with a tandem sequence repeat in the N-terminal half, and several conserved motifs in the C-terminal half of the protein. In some holometabola, only one copy of the N-terminal tandem repeat is found, and it was shown for UDE from Drosophila melanogaster (DmUDE), that the first copy of the tandem repeat may be functionally dispensable⁷⁶. Now, however, with more genomes sequenced, sequence searches result in a more diverse picture, including UDE proteins with a more complex domain arrangement in holometabola and homologs in plant-pathogenic fungi.

With its developmental implications and narrow phylogenetic distribution, UDE posed an attractive target for the development of insecticides specific to holometabola, or fungicides specific to certain plant pathogens. Initially, UDE caught our attention as we just had identified a novel uracil-binding mode in the protein cereblon, which we thought could be linked to the recognition of uracil in DNA, and which can be mimicked by the binding of the drug thalidomide ^77,78. Inspired by the topicality of the Zika virus at that time, we decided to tackle the UDE protein from the yellow fever mosquito Aedes aegypti (AaUDE; AAEL003864), a major virus vector.

AaUDE is a canonical UDE protein with the N-terminal tandem repeat and a length of 306 residues; In vitro, it showed DNA binding properties similar to DmUDE. While full-length AaUDE withstood crystallization attempts, a recombinant protein corresponding to a proteolytic fragment encompassing residues 87-277, thus omitting the first copy of the tandem repeat and the potentially flexible C-terminal end, yielded well-diffracting crystals. The structure, which we solved via SAD phasing using a platinum derivative (CASP target T0890), shows an all-helical two-domain protein. The N-terminal domain corresponds to the second copy of the tandem repeat and forms a three-helix bundle, while the C-terminal half is folded into a compact domain consisting of six helices; the interfacial surface area between the two domains amounts to about 500 Ų (Figure 8A).

Figure 8. The crystal structure of AaUDE(87-277) in comparison to the best DALI matches and CASP predictions.

(A) The full crystal structure in cartoon representation. (B) The crystal structure (red) superimposed with the best DALI matches for the N-terminal (PDB: 3UN9; DALI Z-score 7.5) and the C-terminal domain (PDB: 3TD7; DALI Z-score 10.1). (C) The two best CASP predictions for the N-terminal domain (D1), models T0890TS236_1 (MULTICOM-construct) and T0890TS486_1 (TASSER), yielded a GDT_TS of 68.0 and 67.7 for D1 and of 30.0 and 31.8 for the whole structure. (D) The best CASP predictions for the C-terminal domain (D2). T0890TS250_1 (Seok-server) yielded a GDT_TS of 74.8 for D2 and 44.7 for the whole structure. T0890TS119_1 represents the three almost identical models T0890TS119_1 (HHPred0), T0890TS349_1 (HHPred1) and T0890TS313_1 (HHGG), which yielded a GDT_TS of 69.8, 69.8 and 70.5 for D2 and of 40.8, 40.8 and 41.0 for the whole structure. T0890TS464_1 (tsspred2) yielded a GDT_TS of 59.2 for D2 and 33.4 for the whole structure.

A DALI search with the full structure returned many hits for the N-terminal domain, but only one hit for the C-terminal domain. For the N-terminal domain, the hits yielded Z-scores of up to 7.5. It had previously been predicted to be a three-helix bundle and had been implicated in DNA binding⁷⁶. This notion is supported by our crystal structure, as this domain presents extended stretches of positively charged residues along its helices. The highest-scoring DALI hit was, however, the single hit for the C-terminal domain. With a Z-score of 10.1 it matches a non-conserved additional C-terminal domain of the mimivirus sulfhydryl oxidase R596, which had previously been described as an ORFan domain of novel fold, and which is functionally not understood⁷⁹ (Figure 8B).

For the CASP predictors, AaUDE posed a tough but not intractable target. There were many good predictions for the simpler N-terminal domain (T0890-D1), and a few good predictions for the C-terminal domain (T0890-D2). Curiously, none of the groups could predict both domains. The five best overall models, ranging between a GDT_TS of 44.7 and 33.4 (submitted by the Seok-server, HHGG, HHPred1, HHPred0 and tsspred2) owe their accuracy to the correctly identified similarity of the C-terminal domain to the aforementioned mimivirus ORFan domain. They fail, however, to reasonably predict the N-terminal domain. The overall models from rank six on mostly contain fair-to-good predictions of the N-terminal but not the C-terminal domain, as they miss the link to the mimivirus protein. The best-matching predictions for the individual domains are depicted in Figure 8C and 8D. Despite the good predictions for the individual domains, the inter-domain interface and thus the relevant biological assembly could not be predicted.

9. The snake adenovirus 1 LH3 hexon-interlacing protein (CASP: T0909; PDB: 5G5N and 5G5O) – provided by Thanh H. Nguyen, Abhimanyu K. Singh, and Mark J van Raaij

Adenoviruses are non-enveloped double-stranded DNA viruses with a diameter of around 100 nm ⁸⁰. At the vertices of the icosahedral adenovirus particles, a pentameric penton base protein is located, while the faces are covered with trimeric hexon proteins. Fiber proteins protrude from the penton bases and are responsible for primary host cell recognition ⁸¹. Internalization of human adenoviruses is known to be mediated by the penton base protein interacting with cell surface integrins, but some other adenoviruses lack known integrin-binding motifs in their penton base sequence. Five genera of adenoviruses are known, one of which is the Atadenovirus genus. Atadenoviruses infect birds, snakes, lizards, ruminants or possums. The LH3 gene is a genus-specific atadenovirus gene found at the left end of the genome. The LH3 gene product is believed to be involved in stabilization of the viral capsid ^82,83. The LH3 protein forms trimeric protrusions on the faces of the atadenovirus particle ⁸³. In total, four LH3 trimers are present on each of the faces, and 80 in the entire atadenovirus particle.

The Snake Atadenovirus 1 LH3 protein (CASP target T0909) was expressed in E. coli, crystallized, the structure was solved using SAD from a mercury derivative crystal and refined using native data of a different crystal form at 2.0 Å resolution⁸⁴. Evidence of proteolysis was observed and is consistent with the first 25 residues missing from the experimentally determined structure (Figure 9). The structure revealed a compact, knob-like trimer of right-handed β-helices, as predicted by the BetaWrap server ⁸⁵. The missing part was evident when fitting the structure into an 11 Å cryo-EM map of SnAdV-1 ⁸⁴.

Figure 9. Crystal structure of SnAdV-1 LH3 in comparison with the best CASP12 model.

Superposition of one of the best predicted regular (monomeric) models (T0909TS303_1, magenta) onto a monomer (left; side view) and the trimer (middle; top view, C-termini closest to the reader) of the experimentally determined structure (cyan). On the right, one of the best predicted trimeric models (T0909TS247_1o, orange) is shown viewed from the bottom, N-termini closest to the reader. Chain termini are indicated where possible and a loop that is disordered in two monomers of the trimer in the crystal structure is highlighted by asterisks.

Each LH3 monomer contains eleven β-helical rungs stacked on top of each other. Each β-helical rung consists of three β-strands that form long parallel β-sheets with their counterparts from the other rungs. The β-sheets are named PB1, PB2 and PB3, following the nomenclature proposed by Mayans et al ⁸⁶. Turns between β-strands are named T1 (between PB1 and PB2), T2 (between PB2 and PB3), and T3 (between PB3 and PB1). PB1 connects to PB2 mainly by short β-turns, at the trimer interface, while PB2 connects to PB3 and PB3 to PB1 by longer loops.

Amino acid ladders are observed in the structure of the LH3 protein, as is common for β-helical structures ^86,87. Asparagine-, isoleucine- and phenylalanine- ladders are found in the core of each monomer, stabilizing the basic β-helical architecture of the monomer. The asparagine ladder (residues 193, 214, 248 and 291) is located right at the T1 turn, while the isoleucine (residues 68, 98, 134, 167, 311, 357) and phenylalanine (residues 103, 139, 172, 195) ladders are found in the PB1 and PB2 sheets, respectively. A ladder containing isoleucines and a leucine (Ile84, Ile147, Ile179 and Leu125) is present in the PB3 sheet. It is possible that the hydrogen bonds in the asparagine ladder help avoid out-of-register interactions when the β-helix folds.

A structural homology search using the DALI server⁸⁸ showed the best matches for tailspikes from Bacillus phage phi29 ⁸⁹, Shigella phage Sf6 ⁹⁰ and Salmonella phage P22 ⁹¹. Structure superposition between SnAdV-1 LH3 and Sf6 TSP with its ligands revealed a strikingly similar β-helix topology, despite the low sequence identity (13%). It should be noted that the Shigella phage SF6 tailspike has endorhamnosidase activity. At the binding site, loops from T2 and T3 turns were found to be involved in the interaction with the lipopolysaccharide substrate. Superposition of the two structures do not show conservation of the loop conformations, however, it is possible to form a potential ligand binding groove in the structure of SnAdV-1 LH3 either between two subunits or on the surface of a single monomer (like in the phage P22 tailspike ⁹¹). Evidence for non-conserved binding sites among bacteriophage tailspike proteins was discussed previously ⁹². The structural similarity with bacteriophage tailspikes and its location on the viral cell surface suggested the LH3 protein may be involved in binding a (carbohydrate) ligand. However, we have not been able to demonstrate this or a role for the LH3 protein in host interaction.

Structural superposition of the crystal structure and the best CASP12 models showed they share a similar β-helical fold. The β-helix motif was predicted correctly. The best model, with a DALI z-score of 30.4, suggested a structure comprising three anti-parallel β-sheets PB1, PB2 and PB3 connected by β-turns T1, T2 and T3, as observed in the experimentally determined structure. The length and orientation of β-strands are represented quite accurately, although there are some mismatches. Surface loop conformations are, as expected, predicted much less reliably. Structural superposition of the other CASP12 models also showed that the main β-helix is generally predicted accurately, but loop conformations are different. Most of the β-strands in the models have correct length and location, which is impressive given the low sequence identity (less than 15%) of the SnAdV-1 LH3 protein to known structures. The N-terminal α-helix is identified and, for the most part, the asparagine and hydrophobic amino acid ladders are predicted correctly. It is noteworthy that the N-terminal, virus-facing part of the protein, appears to be somewhat better predicted than the C-terminal, virus-distal part.

It should be kept in mind that SnAdV-1 LH3 protein is a homo-trimer. The standard predictions did not use this given feature. However, some of the predictions that took the homo-trimeric state into account correctly predicted the trimerization interface and reproduced almost 40% of the native interface contacts. This, in turn, might have assisted us in solving the structure by molecular replacement without having to resort to a heavy atom derivative (searching for independent monomers is also possible, but more difficult than searching for correctly assembled trimers). The availability of a SAXS envelope might also have helped to derive an accurate trimeric model computationally, even without prior knowledge of the oligomeric state (see SAXS paper, this issue).

10. Crystal structure of an ice binding protein from an Antarctic Biological Consortium (CASP:T0883; PDB:6EIO) – provided by Valentina Nardone, Marco Mangiagalli and Marco Nardini)

Organisms exposed to permanent subzero temperatures or seasonal temperature dropping are protected from freezing damage by producing Ice Binding Proteins (IBPs) which adsorb to the ice surface and stop ice crystal growth in a non-colligative manner⁹³. A measurable effect of ice binding is that IBPs decrease the water freezing temperature, thereby creating a thermal hysteresis (TH) gap between the melting and the freezing temperature⁹⁴. TH has been explained by the fact that IBP induces a micro curvature on the ice surface. In this way, ice growth is restricted in between the adsorbed IBP and the curved surface. This makes the association of other water molecules thermodynamically unfavorable, causing the decrease of water freezing temperature. The second activity of IBPs is the ice recrystallization inhibition (IRI), which prevents the growth of large ice crystals at the expenses of smaller ones. Growth of these large crystals causes dehydration and cellular damage ⁹⁵. Because of these properties, in recent years the potential application of IBPs has been recognized in several different fields in which materials and substances have to be preserved from freezing, including food processing, cryopreservation, cryosurgery, fishery and agricultural industries, and anti-icing materials development^93,96.

IBPs have been isolated in different species, including fishes, insects, plants, algae, fungi, yeasts and bacteria. Proteins from different sources share the ability to bind ice crystals, but they can exhibit very diverse 3D structures, including small globular proteins, single α-helices, four helix bundles, polyproline type II helix bundles and β-solenoids. This structural diversity suggests that ice binding activity arose independently multiple times in evolution ⁹³.

As a result, it is very difficult to determine the structural features important for ice binding. Structural studies may provide useful information on the ice-binding sites and on their mechanism of action. For instance, structural comparison of IBPs with different folds may highlight common general features, such as the presence of single/multiple flat surfaces and their hydrophobic/hydrophilic residue distribution, in order to grant an efficient ice binding. Furthermore, many IBPs contain threonine-rich repeats, such as Thr-X-Thr or Thr-X-Asx, usually located on the protein surface. The comparison of position/conformation of these repeats in structurally diverse IBPs, coupled with site-directed mutagenesis studies, could help recognize their role in ice binding.

We focused our attention on EfcIBP, a bacterial IBP identified by metagenomic analysis of the Antarctic ciliate Euplotes focardii and the associated bacterial consortium. Tested for its effects on ice, recombinant EfcIBP shows atypical combination of TH and IRI activities not reported in other bacterial IBPs. Its TH activity was only 0.53 °C at 50 μM, but it had one of the highest IRI activities described to date, with an effective concentration in the nanomolar range. As a result, EfcIBP effectively protected purified proteins and bacterial cells from ice damages. Furthermore, the presence in the EfcIBP sequence of a secretion signal seems to indicate that EfcIBP might be either concentrated around cells or anchored at the cell surface, permitting the entire consortium to thrive/survive at challenging temperatures ⁹⁷. To shed light on the antifreeze properties of EfcIBP at the molecular level it is crucial to elucidate its ice-binding mechanism through a combination of structural and molecular biology studies. Therefore, we solved the EfcIBP structure by means of X-ray crystallography.

EfcIBP crystals diffracted to atomic resolution (up to 0.84 Å), and the EfcIBP structure was solved by molecular replacement with the crystal structure of the IBP from the antarctic bacteria Colwellia sp. (PDB-code 3WP9; DALI Z-score of 32.3, residue identity of 38%) as a search model ⁹⁸. The overall structure of EfcIBP consists of a right-handed β-helix with a triangular cross-section formed by three faces made by parallel β-sheets, and by an additional single 5-turn α-helix, aligned along the axis of the β-helix. The first face of the β-helix (9 β-strands) is screened from the solvent region by the long α-helix and by the N-terminal region. This protein surface is, therefore, not suited for the interaction with ice crystals. The second face (8 β-strands) is flat and regular, while the third (8 β-strands) is only partly flat, with two β-strands which markedly diverge towards the exterior of the protein body. The latter two faces are fully exposed to the solvent region and, therefore, potentially suited for the interaction with ice crystals. Interestingly, both faces host multiple threonine-rich repeats, a feature not found so far in IBPs with fold similar to EfcIBP.

Overall, the CASP12 results on target T0883 indicate that right-handed β-helix can be predicted extremely well. All β-strands of the three faces of the EfcIBP structure are correctly positioned as well as the 5-turn α-helix, aligned along the β-helix axis. It should be noted, however, that the β-strand located immediately after the α-helix is correctly placed within the β-helix fold in the model but is shifted by two residues, such that the preceding loop is two residues longer and the following loop two residues shorter than in the experimental structure.

The top ten ranked models (CASP GDT_TS score >89.0) are characterized by an RMSD of ~1.4 Å for the core of the protein (181 Cα pairs over 207 residues). The structure of the first 9 N-terminal residues is not predicted correctly partly because this region is shorter in the homologous proteins used as templates, partly because its conformation might be selected by crystal contacts and, therefore, difficult to predict. The CASP12 models contain a deletion, correctly identified at the top of the right-handed β-helix, where a small cap subdomain of about 12 residues is present in homologous proteins. In this region, however, the Gly-Pro-Pro sequence at the closure of the deletion does not superimpose well with the corresponding EfcIBP crystal structure.

Finally, it is worth noting that the overall quality of the CASP12 prediction does not seem to improve significantly when multiple protein templates are used for modeling instead of a single template. This is probably due to the high structural conservation and rigidity of the β-helix scaffold which tolerates insertion/deletion of several residues without any significant perturbation of the core structure and which is reproduced similarly in all protein templates.

11. The TRXL1 domain of Chaetomium thermophilum UGGT (CASP: T0892; PDB: 5MU1, 5MZO, 5N2J and 5NV4) – provided by Pietro Roversi, Alessandro T. Caputo, Johan C. Hill and Nicole Zitzmann

One of the last unsolved mysteries of the eukaryotic endoplasmic reticulum glycoprotein folding quality control (ERQC) machinery is its single checkpoint enzyme, the ER UDP-glucose glycoprotein glucosyltransferase (UGGT). Once monoglucosylated by this enzyme, glycoproteins are retained in the ER bound to the lectins calnexin and/or calreticulin and the associated chaperones and foldases that assist their folding⁹⁹. The mechanism by which UGGT recognizes and glucosylates a large variety of misfolded glycoprotein substrates remains unknown.

The N-terminal ~1200 residues of UGGT harbor the enzyme’s misfold sensing activity ^100,101. The lack of any obvious sequence homology of this portion of UGGT with proteins of known fold led to the creation of a UGGT-specific protein fold family (Pfam family PF06427) which gathers all known eukaryotic UGGT N-terminal sequences. The most recent secondary structure and domain boundary predictions for UGGT detected three thioredoxin-like (TRXL) domains in this region^102,103. The canonical TRXL fold (Pfam family PF13848) comprises a thioredoxin fold (a four-stranded β sheet sandwiched between three α-helices, TRX=βαβ–αββα Pfam family PF00085, red in Figure 10), modified by the insertion of a 4-helix subdomain (TRXL=βαβ–αααα–αββα blue in Figure 10)^104,105.

Figure 10. The TRXL1 domain of CtUGGT.

(A) In blue, the CtUGGT TRXL1 N-terminal α-helical subdomain (residues 43-110). In red, the TRXL1 thioredoxin subdomain (residues 111-216). The disulphide bridge C138-C150 is represented as spheres. (B) The structure of the closest structural homologue to CtUGGT TRXL1, Staphylococcus aureus DsbA, with the α-helical insertion subdomain (residues 63-129) in blue and the thioredoxin subdomain (residues 14-62 and 130-177) in red. In (A) and (B) N- and C-termini are denoted by the letters “N” and “C”, respectively. (C) The superposition of the top ten CASP12 T0892 models, overlayed on the CtUGGT TRXL1 crystal structure in the region of the N-terminal helical subdomain and the first helix of the thioredoxin subdomain. The CtUGGT TRXL1 crystal structure is colored and represented as in panel A. The top ten CASP12 T0892 models are in ribbon representation and colored as follows: T0892TS011_1:green; T0892TS011_2: cyan; T0892TS017_1: magenta; T0892TS017_2: yellow; T0892TS017_5: grey; T0892TS411_2; T0892TS017_3: salmon pink; T0892TS079_5: violet; T0892TS479_3: steel blue; T0892TS320_4: orange. A black star marks the hinge between the helical subdomain and the thioredoxin subdomain. A dotted circle marks the first helix in the thioredoxin subdomain. (D) The superposition of the top two CASP12 T0892 models (T0892TS011_1 and T0892TS011_2, in green and cyan respectively, in ribbon representation), overlayed on the CtUGGT TRXL1 crystal structure in the region of the C-terminal thioredoxin subdomain, without its first α-helix. The CtUGGT TRXL1 crystal structure is colored and represented as in panel A. The wrongly predicted first two strands of the thioredoxin subdomain are circled, and an asterisk marks the incorrectly predicted α-helix for the stretch of residues 151-164 of CtUGGT TRXL1.

To aid our understanding of UGGT structure and function, we determined four distinct crystal structures of Chaetomium thermophilum UGGT, aka CtUGGT ¹⁰⁶. An unexpected structural feature of the UGGT molecule is the unusual subdomain structure of the first thioredoxin-like domain (TRXL1), encoded by residues 43-216 in CtUGGT. The published sequence–based secondary structure predictions in this region was rather accurate, with most helices and sheets correctly predicted from sequence – but the UGGT TRXL1 domain boundaries were not well predicted ^104,105.

Indeed, the UGGT TRXL1 domain folds with sequential pairing of a helical subdomain with a thioredoxin subdomain (blue and red in Figure 10), while all other known TRXL domains present a helical subdomain as an insertion within the thioredoxin subdomain (see for example in Figure 10B the closest structural homologue of CtUGGT TRXL1, Staphylococcus aureus DsbA, PDB ID 3BD2). The CtUGGT crystal structures also reveal that the CtUGGT TRXL1 domain harbors a disulfide bridge between Cys138 and Cys150 (represented as spheres in Figure 10A).

We submitted the CtUGGT TRXL1 sequence to CASP12 (target T0892) in order to test prediction methods for their ability to model i) its non-canonical subdomain structure, in which an N-terminal α–helical subdomain is followed by a C-terminal thioredoxin subdomain and ii) the presence of a disulfide bridge between CtUGGT TRXL1 C138 and C150.

We compare here the top 10 CASP12 T0892 models (as ranked by the GDT_TS score on the CASP12 results server) to the coordinates of the TRXL1 domain in the 2.8 Å CtUGGT crystal structure (PDB ID 5NV4), residues 43-216. The overall RMSD_Cα across the ensemble of the top ten T0892 models is 10.7 Å over 174 Cαs¹⁰⁷. All these CASP12 T0892 models predict an N-terminal 4-helix subdomain followed by a C-terminal subdomain which resembles to various degrees a TRX fold. None of the top T0892 CASP12 models predicts the CtUGGT TRXL1 C138-C150 disulfide bond.

If one restricts the analysis to the CtUGGT TRXL1 N-terminal, helical subdomain (residues 43-110) and the first α-helix (residues 111-126) of the C-terminal, thioredoxin subdomain, the top ten T0892 models align rather well with each other and with the crystal structure. The overall RMSD_Cα for the ten structures over these 84 Cαs is 1.7 Å. The major differences between the CASP12 T0892 models in the 43-126 portion arise at the hinge (CtUGGT residues 108-111, denoted by a black star in Figure 10C) between the helical subdomain and the first α-helix of the thioredoxin subdomain. The two top-ranked CASP12 models (T0892TS011_1 and T0892TS011_2, green and cyan in Figures 10C#x02013;D) show a different hinge region from the rest. As a result of these differences, in the same top-ranking two models, the relative angle between the N-terminal helical subdomain and the first helix of the thioredoxin subdomain also differs from the crystal structure and the rest of the T0892 CASP12 ensemble of models. The CtUGGT 111-126 α-helix is marked by a dotted circle in Figure 10C.

In the C-terminal thioredoxin subdomain (residues 111-216), the top ten CASP12 T0892 models align poorly with each other and with the crystal structure of the target. The overall RMSD_Cα for the ten models over these 84 Cαs is 9.5 Ź⁰⁷. Only the two top ranking CASP12 T0892 models (T0892TS011_1 and T0892TS011_2, green and cyan in Figures 10C–D) correctly contain a 4-stranded β-sheet at the center of the TRXL1 thioredoxin subdomain. Even restricting attention to these two models only, across residues 127-216 the RMSD_Cα between the models and the crystal structure is still as high as 6.5 Å over 90 Cαs¹⁰⁷ (see Figure 10D). In particular, the first two β-strands of the thioredoxin subdomain β-sheet in the models do not superimpose well on the same β-strands in the crystal structure (circled in Figure 10D). Moreover, in both models, the stretch of sequence 151-164 – which immediately follows those strands - is wrongly predicted to fold as an α-helix (marked by an asterisk in Figure 10D) which is not present in the crystal structure.

Overall, none of the models predict the CtUGGT TRXL1 C138-C150 disulfide bond, and the 128-181 region between the first TRX helix and the third TRX strand is not well defined in any of the models. On the other hand, the best CASP12 T0892 models are successful in predicting the structure of the N-terminal 4-helix subdomain, and the two top-scoring ones also manage to correctly predict that the domain is a linear fusion of an N-terminal 4-helix subdomain and a C-terminal subdomain of TRX fold. In summary, as far as this target was concerned, the CASP12 predictors did well, but did not put us out of our job just yet.

12. Structural characterization of the third cohesin from Ruminococcus flavefaciens scaffoldin protein, ScaB (RfCohScaB3) complexed with a group 1a dockerin (RfDoc1a) (CASP: T0921/T0922; PDB: 5AOZ (RfCohScaB3), 5M2O (RfCohScaB3/Doc1a complex) – provided by Pedro Bule, Ana Luisa Carvalho, Carlos M.G.A. Fontes and Shabir Najmudin

The plant cell wall represents a major untapped global source of carbon and energy. Herbivores, in particular ruminants, are able to utilize this energy source thanks to the presence of cellulolytic bacteria in their gastrointestinal tract. Ruminococcus flavefaciens, a Gram-positive Firmicute, is a major symbiont in the rumen. R. flavefaciens possesses a highly intricate multi-enzyme complex, termed the cellulosome, which comprises a range of cellulases and hemicellulases that degrade the structural polysaccharides in a highly efficient and concerted way. The assembly of cellulosomes occurs via highly ordered protein–protein interactions between cohesins (Cohs), which are located in multi-modular macromolecular scaffolds (scaffoldins), and dockerin molecules (Docs), which are found in the enzymes or on the scaffoldins themselves ^108,109. Strain FD-1 of R. flavefaciens produces one of the most intricate and potentially versatile cellulosomes described to date. The genome of R. flavefaciens FD-1 encodes 223 dockerin-bearing proteins, which are predominantly enzymes displaying catalytic activity that modifies carbohydrates ¹¹⁰. In this highly elaborate cellulosome, scaffoldin B (ScaB) acts as the backbone to which other components attach. ScaB comprises 9 cohesins of 2 distinct types. Cohesins 1 to 4 are similar to the two cohesins of a second, smaller scaffoldin ScaA, whose dockerin binds to ScaB cohesins 5 to 9 through a different protein-protein specificity. ScaB contains a C-terminal dockerin that binds to the cohesin of cell-surface ScaE providing a mechanism to anchor the entire complex to the bacterial cell. A distinct scaffoldin, ScaC, acts as an adaptor that binds predominantly hemicellulases while connecting to the first type ScaB cohesins, thus serving to increase the repertoire of proteins that can be integrated into the complex. In Clostridium species studied so far, enzyme-borne Docs interact with their cognate Cohs through a dual-binding mode ¹⁰⁹. Internal dockerin symmetry allows them to bind to cohesins in either of two orientations resulting in two different Coh-Doc conformations that are related by ~180° rotation. This dual-binding mode results from the characteristic internal symmetry of the Doc primary sequence and is believed to add flexibility to the cellulosomal macromolecular organization. Based on primary sequence similarity, R. flavefaciens dockerins are classified in different groups. Recent studies have shown that groups 3 and 6 R. flavefaciens Docs display a single-binding mode for their target Cohs, i.e. binding occurs in one orientation only ¹¹¹. Intriguingly, Group 1 Docs also do not seem to possess the internal sequence symmetry required to support the dual-binding mode. Thus, modeling studies are required to predict the correct binding mode between various types of Coh-Doc complexes and to predict which amino acid residues act as molecular specificity determinants.

X-ray crystal structures of the third R. flavefaciens cohesin from ScaB (RfCohScaB3) and group 1 Doc (RfDoc1a) in complex with RfCohScaB3 (Figure 11A) were recently solved, and characterized by comprehensive biochemical analyses ¹¹². RfCohScaB3 forms an elongated nine-stranded β-sandwich in a classical jelly-roll topology. The overall RfCohScaB3 structure is similar to other enzyme-borne Doc-binding Cohs (RMSD of less than 3.0 Å between at least 130 Cα atom pairs) despite the very low sequence similarity (4–12%). The major structural differences are in the Doc-binding interface formed by β-strands 8, 3, 6, 5. In turn, the overall tertiary structure of RfDoc1a is very similar to other enzyme-borne Docs (RMSD of less than 2.0 Å between at least 60 Cα atoms; sequence identity 20–32%). The structure contains two Ca²⁺ ions coordinated by several amino-acid residues, similar to the canonical EF-hand loop motif described in all other Docs ¹⁰⁹. The whole of helix-1 makes predominantly hydrophobic interactions with the Coh, while helix-3 interacts mainly through its C-terminus. Ile-39 and Val-43 on helix-1 of the RfDoc1a and Ala-38 and Leu-79 on the binding platform of RfCohScaB3 were shown to be the key specificity determinants.

Figure 11. Structure of the RfCohScaB3-Doc1a complex.

(A) Structure of RfCohScaB3-Doc1a complex with the dockerin in red and the cohesin in blue. The dockerin N- and C- terminus and the α-helices are labeled, and a transparent gray molecular surface of the cohesin is shown. (B) Superposition of CASP12 prediction models T0921TS220_2_D1 (light blue) and T0921TS166_1_D1 (light green) with RfCohScaB3 crystal structure (black). (C) Superposition of CASP12 prediction models T0922TS005_3_D1 (light blue) and T0922TS077_4_D1 (light green) with the RfDoc1a crystal structure (black). Ca²⁺ ions are depicted as green spheres.

How do the modeling studies of CASP12 compare with the experimental structural studies of RfCohScaB3 (T0921) and RfDoc1a (T0922) and the complex between them? Predictions for both, the RfCohScaB3 and RfDoc1a were very successful, with 147 models for the former and 143 for the latter (out of 186 total for each of the subunits) having GDT_TS scores greater than 50. The top model for each target and a slightly poorer model scoring ~10 GDT_TS below the top model were chosen for comparative purposes. For RfCohScaB3, these were models T0921TS220 from the GOAL group (GDT_TS of 70.7) and T0921TS452 from the Zhou-Sparks-X group (GDT_TS of 60.7). Superpositions of these models using SSM onto the X-ray structure gave RMSD of 2.1 Å for 127 Cα atoms and 2.4 Å for 120 Cα atoms, respectively (Figure 11B). Though the core structure matches really well, there are major differences in the β6-7 and β8-9 loops and in the β8 strand on the dockerin binding interface. Ala 38 is generally in the correct position, but there is considerable variation in the Leu 79 position. For RfDoc1a, we chose T0922TS005 from the BAKER-ROSETTAserver group (the top scorer with GDT_TS of 83.8) and T0922TS077 from the Falcon_Topo group (GDT_TS of 73.7). Superpositions of these models using SSM onto the X-ray structure gave RMSD of 1.4 Å for 69 Cα atoms and 1.6 Å for 63 Cα atoms, respectively (Figure 11C). Generally, the α-helices 1 and 3 are well modeled and consequently so are the key specificity residues, like Ile 39 and Val 43, with differences mainly in the loop regions and N- and C-termini, which are not involved in Coh recognition. However, the modeling of the RfCohScaB3/Doc1a heterocomplex was less successful, with only three models out of 325 (TS203_4 from the Seok group, TS188_1 from the Chuo_U group and TS208_3 from the SVMQA group) correctly modeling half or more of the intermolecular surface contacts compared to the crystal structure. One reason for this could be incorrect modeling of the loops in the binding surface of the cohesins. In these three predicted complexes the cohesins have similar or less prominent loops between β-strands 6 & 7, and 8 & 9 compared to the crystal structure (cf. Figure 11B), thus avoiding steric clashes when complexing with the cognate dockerin models in the single-binding mode.

In summary, the monomeric subunits of the RfCohScaB3/Doc1a complex (T0921/T0922) were modeled very successfully despite relatively low sequence similarity to available homologues, while the whole complex was not. A more advanced approach is needed to predict whether the cohesin-dockerin interaction can operate through a single-binding mode, where only one binding orientation is possible (mainly through helix-1 or through helix-3 of the dockerin) or through a dual binding mode, where the binding can be in one of two orientations (by either helix-1 or helix-3 to the cohesin binding surface).

Discussion

The paper provides insights into structural and functional details of twelve selected CASP12 targets and analyzes to what extent the most interesting features of the targets are reproduced in the predictions from the standpoint of the authors of the structures. Since specific features of the targets are difficult for CASP assessors to address on a large scale, the authors’ insights represent a critical piece of information for both understanding the utility of models and developing protein structure prediction methodologies and assessment strategies.

The examples presented in the article highlight a series of reoccurring themes that challenge current modeling methods and that therefore deserve attention from method developers.

• Oligomers. The structural integrity and biological function of proteins often depends on their quaternary structure and the ability to form specific macromolecular complexes. However, protein oligomerization is not always taken into account in modeling. To address this issue, CASP introduced a separate ‘Assembly modeling’ category in CASP12 ¹¹³, and will continue to encourage modelers to develop methods for predicting hetero- and homo-oligomeric structures. When modelers do predict oligomers, they are more successful in modeling the subunits than full complexes (e.g., T0884/T0885, T0921/T0922). That is not surprising as prediction of complexes oftentimes involves more than just direct docking of the initial subunit models. One of the complications is conformational changes of protein fragments upon complex formation, as in the T0884/T0885 case. In that complex, the N-terminal segment of the immunity protein T0885 is disordered in the predicted free form, and possibly undergoes disorder-to-order transition upon binding to the toxin putative catalytic domain T0884. This transition is important for the physiological function of the complex. In general, advanced modeling techniques capable of accounting for such scenarios are needed. The authors of homo-multimers (e.g., T0909, T0889) suggest that building multimeric models is beneficial for functional annotation of the proteins and that using information about the oligomeric state can help generate better monomeric models. Sometimes higher-order structures are not only desirable, but necessary to maintain the stability of a protein, as exemplified by some of the CASP12 viral protein targets (e.g., T0880, T0909).

• Multi-domain proteins. The majority of proteins exist as multi-domain entities ¹¹⁴, and a sizeable portion of targets in each CASP (1/3 in CASP12) is multi-domain proteins. Constituent protein domains can either have independent functions or contribute to the function of a multi-domain protein in cooperation with other domains. In the latter case, the relative orientation of domains may be important for protein activity. For example, surfaces of two structural domains of target T0912 interact to form a pocket responsible for ligand recognition and binding. Therefore, accurate prediction of the full target was necessary in this case. CASP evaluates multi-domain targets as both per-domain and whole-structure models, however only the domain-based results are usually accounted in the assessors’ reports. A more comprehensive evaluation of multi-domain targets may require additional analysis of the biological relevance of inter-domain architecture, and a separate approach for assessment.

• Loops. The specific structure of individual loops is often a key for the understanding protein function. Unfortunately, prediction of loop conformations in general has not achieved the level of accuracy required to confidently establish their role in interactions with small molecules or partner proteins (e.g., T0877, T0889, T0948). The problem is more pronounced for long loops that deviate substantially from their homologues. Taking into account the importance of the problem, future assessors might consider a more careful scrutiny of loop modeling accuracy. This might include assessment of the (a) accuracy of the loop main chain in isolation, (b) relationship of the loop to rest of the structure and (c) errors in protein-ligand interactions. The local-structure evaluation measures (e.g., CAD ¹¹⁵, LDDT ¹¹⁶ or SphereGrinder ¹¹⁷) and interface accuracy measures (Interface Contact Score and Interface Patch Distance ¹¹³), which were recently introduced in CASP, can be used for this purpose. It also important to evaluate whether a loop conformation is robustly determined experimentally, and not influenced by the crystal environment.

• Conserved residues. When faced with a specific biological system, different sources of information should be checked to yield more accurate models. For example, sequence analysis of the protein family may highlight conserved surface residues potentially involved in complex formation. For example, in CDI systems it is known that immunity proteins typically block access to the active site of the toxins. For the CdiA-CT/CdiI complex (T0884/T0885 target), the three highly conserved residues of the CdiI immunity protein (H73, R76 and D109) interact directly with three highly conserved and presumably catalytic residues of CdiA-CT toxin (H181, H183 and R185) identifying part of the surface in contact. Including this information in predictions would strongly constrain possible solutions.

• Disulfides. It is well known that closely spaced cysteines tend to form disulfide bridges in extracellular proteins. However, these were not properly modeled in at least two CASP12 targets (T0877, T0892). Since disulfide bonds play an important role in the stability of some proteins, their proper modeling seems to be an easy and obvious way of improving models.

• Alignment. In spite of enormous progress, correct sequence alignment remains a challenge in structure modeling and improved methods are likely to enhance modeling accuracy (T0859, T0883). For example, for target T0859, an alignment register shift resulted in an incorrect secondary structure assignment, which in turn hindered surface exposure of functionally important residues.

• Purification tags and signal peptides. A number of CASP sequences included purification tags or signal peptides. If not identified and removed before the modeling, these structural extensions of protein domains might complicate modeling routine. Even though it is usually easy to identify the tags and there are several programs to predict the presence of signal peptide sequences, many structure prediction methods still do not make use of them and attempt to build models of these regions (e.g., T0886, T0922).

• Low resolution data. Data from low resolution structure determination experiments are expected to help build atomic-resolution models of proteins. However, the data-assisted component of CASP12 showed that utilizing SAXS or cross-linking data had only marginal effect on the atomic-level structure modeling (T0886, T0909). This outcome shows that either the additional information is too coarse-grained to assist current methods or that the computational community has not been able to fully utilize the potential hidden in the data.

We hope that these general conclusions will guide future CASP assessments and encourage methods developers to address the issues.

Abbreviations

CASP

community wide experiment on the Critical Assessment of Techniques for Protein Structure Prediction

VLP

virus-like particle

TfR1

Transferrin Receptor 1

WWAV

Whitewater Arroyo Virus

GP1

glycoprotein 1

RG-II

Rhamnogalacturonan-II

HGM

Human gut microbiota

GH

Glycoside hydrolases (GH)

IBP

ice binding protein

TH

thermal hysteresis

IRI

ice recrystallization inhibition