Discrete analysis of camelid variable domains: sequences, structures, and in-silico structure prediction

Datasets

V_HH (nanobody)sequences were obtained from NCBI Genbank (Benson et al., 2005) (https://www.ncbi.nlm.nih.gov/genbank/) and Uniprot (Boutet et al., 2016) (http://www.uniprot.org/). V_HH structures were taken from the Protein Databank (PDB, https://www.rcsb.org/pdb/home/home.do) (Berman et al., 2000). This dataset was put together by March, 2018. Sequences were aligned using Clustal Omega v1.2.1 (http://www.clustal.org) (Sievers et al., 2011) and using Jalview v.2.10 (http://www.jalview.org) (Waterhouse et al., 2009), and analysed using WebLogo (Schneider & Stephens, 1990; Crooks et al., 2004). CDRs were analysed using PyIgClassify classification (http://dunbrack2.fccc.edu/pyigclassify/) (Adolf-Bryfogle et al., 2015) (described in greater detail below).

Comparative modelling of V_HH

Modeller.9v.16 was used in the study to model V_HH. Briefly, an alignment file in the prescribed format of the query with the desired template(s) is generated. From this alignment, the algorithm derives spatial restraints or probability density functions for each residue to be modelled. 3D model(s) of the query protein are obtained with the optimisation of the input restraints. The best structural model is selected with best DOPE score (Shen & Sali, 2006; Melo & Sali, 2007).

In the study case, two different scenarios were used: (i) 4 templates were used to propose a multi-template approach and (ii) individual templates were used, leading to five distinct cases. In each case a total of 100 models were generated and were ranked using DOPE score.

Assessment of structural similarity and disulphide bridge conformations of V_HH domains

The most popular method to evaluate similarity is the distanc-based similarity measure Root Mean Squared Deviation (RMSD); it is calculated by the averaging of distances between n-pairs of equivalent atoms. It can be applied to a whole protein or a subset of specified equivalent atoms from two proteins. It must be noted that RMSD value takes only into account equivalent residue, i.e., it is not appropriate when the segments are of different lengths. RMSD was computed with ProFit (http://www.bioinf.org.uk/programs/profit/) that is based on the McLachlan algorithm (McLachlan & IUCr, 1982) and iPBA (http://www.dsimb.inserm.fr/dsimb_tools/ipba/) (Gelly et al., 2011); this last also provides Global Distance Test Total Score (GDT-TS) values (McLachlan & IUCr, 1982; Zemla, 2003).

A recent classification of disulphide bridges (Schmidt, Ho & Hogg, 2006) includes 20 kinds of disulphide bonds based on the signs of the five dihedral angles between the Cα atoms of any two cysteines involved and the Dihedral strain energy. For example, the sign pattern ‘- - - - - -’ indicates that all the dihedral angles χ1, χ2, χ3, χ ‘1, χ ‘2 between the two C α atoms Cys and Cys’ have negative values.

Local conformational analysis

Secondary structures were assigned with the most widely used algorithm, namely DSSP (CMBI version 2000) with default parameters (Kabsch et Sander 83). Protein Blocks (PBs) were also used. PBs are a structural alphabet composed of 16 local prototypes (Joseph et al., 2010) five residues in length. PBs give a reasonable approximation of all local protein 3D structures (de Brevern, Etchebest & Hazout, 2000) and are very efficient in protein superimpositions (Joseph, Srinivasan & de Brevern, 2012) and MD (Molecular Dynamics) analyses (de Brevern et al., 2005). They are labelled from a to p. PBs m and d can be roughly described as prototypes for α-helix and central β-strand, respectively. PBs a to c primarily represent β-strand N-caps and PBs e and f represent β-strand C-caps; PBs g to j are specific to coils; PBs k and l to α-helix N-caps, while PBs n to p to α-helix C-caps. PB (de Brevern, 2005) assignment was carried out using our PBxplore tool (available at GitHub) (Barnoud et al., 2017).

The equivalent number of PBs (N_eq) is a statistical measurement similar to an entropy index. It represents the average number of PBs for a residue at a given position which is calculated as follows (de Brevern, Etchebest & Hazout, 2000): (1)${\displaystyle N_{eq}=exp\left(-\sum _{x=1}^{16}{f}_{x}In{f}_x\right)}$ Where, f_x is the probability of PB x. A N_eq value of 1 indicates that only one type of PB is observed, while a value of 16 is equivalent to a random distribution.

To detect a change in PBs profile, a ΔPB value was calculated. It corresponds to the absolute sum of the differences for each PB between the probabilities of a PB x being present in the first and the second structures (x goes from PB a to PB p). ΔPB is calculated as follows: (2)${\displaystyle \Delta PB=\sum _{x=1}^{16}\left|\left(f_x^1-f_x^4\right)\right|.}$ Where, f¹_x and f²_x are the percentages of occurrence of a PB x in the analysed structures. A value of 0 indicates perfect PB identity, while a score of 2 indicates a total difference.

The change in PB entropy at given position between two different sets of structures analysed is denoted using ΔN_eq which is calculated as the following: (3)${\displaystyle \Delta N_{eq}=|N_{eq1}-N_{eq2}|.}$

PyIgClassify database

CDRs support most of the interactions with the epitope. Different classifications have been proposed to analyse them (Chothia & Lesk, 1987; Martin & Thornton, 1996; Al-Lazikani, Lesk & Chothia, 1997; Shirai, Kidera & Nakamura, 1999). The most recent classification of CDRs was proposed by North and co-workers (North, Lehmann & Dunbrack, 2011). The AHo numbering scheme (Honegger & Plückthun, 2001) was used by the authors to enumerate the residues in variable domains. The clusters are regularly updated and are made available through the PyIgClassify database (http://dunbrack2.fccc.edu/PyIgClassify/) (Adolf-Bryfogle et al., 2015). All the CDRs except CDRH3 are included in the classification. The classified CDR loop conformations are based on the (a) type of CDR, (b) length of the CDR loop, and (c) affinity-propagation clustering method (Frey & Dueck, 2007), using a dihedral angle distance function was used to group CDRs. The clusters are grouped into three categories, (1) Type I, i.e., one cluster CDR-lengths, (2) Type II, i.e., predictable CDR-lengths, and (3) Type III, i.e., unpredictable CDR-lengths. According to the authors, CDR H1 (i.e., CDR1 for V_HH) falls into the Type II category, where the CDR-length combination has multiple possible structures. CDR H2 (i.e., CDR2 for V_HH) falls under Type I, the one-cluster CDR-length. Since Protein Blocks are descriptors of local backbone conformations, they have also been used to analyse the conformations of dense clusters, which include V_HH.

Sequence datasets

Uniprot contains a limited number of V_HH annotated sequences (only 9), while 245 are retrieved from Genbank. In the latter, the majority of V_HH domains (233) are from Camelids. In the PDB, 140 V_HH structures were available at the time of dataset generation. Sequences from PDB and Genbank were combined to constitute a master dataset of 373 sequences. After removing the redundant sequences at >95% sequence identity, 325 representative sequences were selected (see Dataset S1). It is the largest sequence dataset ever used to analyse V_HH domains. Our study focuses on V_HH domains, since others have compared a more limited number of sequences (90) and always compare with V_H as a priority (Mitchell & Colwell, 2018b).

Multiple sequence alignment

A first step in the analysis of a specific protein family is to look at amino acid sequence characteristics using a multiple sequence alignment (MSA). Such an alignment was generated with sequence dataset using Clustal Omega (see Fig. S1). Figure 2 shows the analysis of this MSA represented as a sequence logo, where residue conservation at each position was calculated as information content (bits). As expected, FR positions strikingly appear as conserved sequence blocks evidenced by high bit scores. The interspersed CDRs, on the other hand, have less information content in terms of bits. This is mainly due to inherent sequence variability. Results are in good correspondence with the recent analyses of Mitchell & Colwell (2018a), but show more information in terms of bits for the CDRs than this study. This can be due to the use of an alignment tool that can provide less efficient alignment in variable regions, or simply to a larger number of analysed V_HH domains, since both free and complexed forms are studied. For the first time, this analysis was also performed for each genus (see Fig. S2 for a summary and Figs. S4–S5 for complete sequence profiles). Interestingly, even with the divergence of species, sequence conservation also shows similar conservation for all the different regions with no significant differences.

V_HH sequences have a median length value of 123 amino acids (aa) with minimal and maximal length of 109 aa and 137 aa, respectively (see Fig. S6). The amino acid length distribution in different regions of V_HH (see Fig. S7) shows diversity in CDR lengths, especially in CDR3. The median values for CDR lengths are 8 aa, 8 aa and 16 aa for CDR1, CDR2 and CDR3, respectively. The average length of CDR3 in V_HH is higher than conventional human or mouse V_H sequences (Muyldermans et al., 2009). However, an important point is that FRs are not of absolute invariant length: FR1 is 25 aa in length, FR2 is 18 aa, and FR3 is 37 aa, but with some differences of 2–3 residues. These differences are often forgotten, i.e., in the recent (Mitchell & Colwell, 2018a; Mitchell & Colwell, 2018b) studies, but must be taken into account, or else the analyses could be biased.

The pairwise sequence identity between sequences in the dataset has a median value of 62% and is always above 35% (see Fig. S8A). Thus, it can be considered a relevant case for homology modelling. The variability of amino acids is not constant in these domains. FRs are more conserved, with sequence identity 84, 72, 81 and 90% (median values) for FRs 1, 2, 3 and 4, respectively (see Figs. S8E–S8H). In the case of CDRs, low sequence identities are observed (below 30%); CDR3 is the lowest with 18%, followed by CDR2 with 25% and CDR1 with 28% (see Figs. S8B–S8D). These results also underline why it is so important to keep a high sequence threshold for building of the dataset, since FRs are highly redundant. Hence, a threshold of 30% would have selected only one V_HH.

Sequence characteristics

At the time of their discovery, V_HH sequences were known to have unique amino acid substitutions in FR2 (Vu et al., 1997) compared to V_H found in camelids at positions V42F/Y, G49E/K, L50R/C, W52G/L (IMGT^® numbering scheme, see Fig. S1). It was thought that there is a single type of V_HH germline sequence, which has sequence similarity to Clan III of V_Hgermline sequences from humans. However, the discovery of V_HH without this tetrad added exceptions to the former theory. These V_HH domains are assumed to be derived from a different clan similar to Clan II of V_H of human germline sequences (Deschacht et al., 2010), and represent 23 sequences (8 from PDB and 15 from NCBI Genbank). Hence this new analysis underlines that the amino acid substitutions of V_HH of FR1 and FR2 are entirely non-specific: (a) V(F/Y) also has V at the alignment position 51 in Fig 2.3, in 20% of cases, (b) in G(E/K), E is majority, but K is not found in our dataset and D can be also found (no G at all) at position 58, (c) in L(R), R is majority at position 59, but in a few cases L is found, like for classical V_H, and finally (d) W(G/L) alignment position 61 is explicit, since G and L represent 50% of the residues while F represents the rest, with V_H typically another aromatic W. It is the same thing with L11S, where S is found in 1/3 of the cases and the rest is L –supposedly not typical.

A second kind of unique characteristics are those of the Cystine residues. The conserved Cys23 (FR1)/Cys104 (FR3) disulphide bond is observed in all the variable regions irrespective of the type of chain or species (highlighted in green in Figs. S1). Some of the camelid V_HH sequences are known to possess an additional disulphide bond between CDR3 or CDR1, CDR3 or CDR2/FR2. In sequences from llamas and camels (see Figs. S3, S4) the presence of an extra cysteine in CDR3 is compensated by another cysteine in CDR1 or CDR2 or CDR2/FR2 boundary. In case of V_HH from llama, 8 out of 68 structures from PDB show this signature, whereas in camels this number is slightly higher, with 15 out of 34 structures from PDB. In the case of V_HH sequences from alpacas, this extra cysteine bond is formed between cysteines of CDR3 and FR2/CDR2 boundary with 3 out of 11 structures from PDB (see Figs. S5).

Structural analysis of V_HH domain

As mentioned previously, V_HH adopts the immunoglobulin fold formed by the anti-parallel arrangement of 9 β-stands connected by loops. Each V_HH domain has three CDR loops. The fold is held together by the inter strand hydrogen bonds and disulphide bond(s). The conserved disulphide bond is between Cys23 of FR1 and Cys104 of FR3 (IMGT numbering). In few cases the presence of a second disulphide bond is also observed, which is known to increase the overall stability in the respective V_HH structures (Zabetakis et al., 2014). A non-redundant set of 105 V_HH domains was assessed for structural domain similarity, using RMSD. The median RMSD value for the whole domain was at 2.63 Å (Figs. S9A–S9H), all the FRs showed a median value <0.8 Å (Figs. S9E to S9H), while they rose to 2.01, 1.56 and 3.51 Å for CDR1, CDR2 and CDR3 respectively (see Figs. S9B to S9D).

Secondary structure analysis of V_HH domain

At a more local conformational level, Fig. 3 is a representation of the secondary structures when aligned as per sequence-aligned positions. The nine conserved β-strands (in alignment positions 3–7, 10–13, 18–26, 44–50, 58–62, 85–90, 95–99, 109–119, 144–148) can easily be observed. The CDR regions represented in the alignment at positions 29-41, 64-75 and 116-139, are mainly associated with turns, bends and coil secondary structure elements (SSEs). Interestingly, the connecting loops between some β-strands, encompassed in FRs, are composed of turns, bends and helical SSEs, i.e., positions 13–17 in FR1, 51–59 in FR2, 78–84, 91–94 and 100–108 in the FR3 region. This suggests some of them have different backbone constraints compared to the rest. The middle region of FR4 is an interesting region with almost all coils, which are the irregular SSEs for alignment positions 140–143.

The noteworthy point in the analysis of the different canonical β-strands that are in FRs is that some are close to pure β-strands such as β-strand F (for which it is 99%), but for most of them some positions supposed to be only β-strands are not. They are mainly N-terminus first residue(s). For instance, the first position of β-strand A1 is only 70% for β-strand; it is 80% for β-strand C or A2. The most striking case is the C-terminus of β-strand F for which many residues are either β-strand or β-turns, i.e., clearly different types of local conformations. These results (i) are in concordance with previous analyses using PBs (Noël, Malpertuy & de Brevern, 2016), and (ii) show that FRs are not as conserved in terms of secondary structure, i.e., this finding can so have a direct impact on the proposition of structural models.

Analysis of disulphide bond geometry in V_HH domains

An additional disulphide bridge is present in some V_HH domains along with the conserved-canonical disulphide bond between Cys 23 and Cys 104 (IMGT numbering system). These additional disulphide bridges were observed between Cysteine residues found (i) CDR1 and CDR3, namely type 1, (ii) FR2 and CDR3, namely type 2 and (iii) both in CDR3 (see Figs. S10 to S10C for schematic representation). In our initial dataset, 25 V_HH domains were observed to have the type 1 (12 V_HH domains) and type 2 (13 V_HH domains) additional disulphide bridges. Figure 4 illustrates their position in V_HH structure, with both types involved in bending CDR3 onto the β-sheet surface.

Using a recent classification of disulphide bridges (Schmidt, Ho & Hogg, 2006) (see Material and Methods section and Figs. S11 for schematic representation) we have analysed the geometry of disulphide bonds in subset of V_HH structures which contain two disulphide bonds. The sign pattern of the canonical and the additional disulphide bonds of both types have been characterized (see Table S3). The greater the number of negative values a dihedral angle pattern has, the lower is the strain energy, suggestive of a stable disulphide bond. Although, the canonical disulphide bond is conserved in terms of position of cysteines, the sign pattern of the dihedral angles varies much more compared to the additional disulphide bond which is surprising. Interestingly, most of the additional disulphide bonds all have negative sign patterns, suggesting a lower energy bond which might increase stability. Two V_HH domains (PDB id 1YC8 from type 1 and 4Y7M from type 2 class in Table S3) have the dihedral angle pattern characteristic of allosteric disulphide bonds according to the classification. This analysis shows that both canonical and the additional disulphide bridges do not play a simple structural role, as many are not favourable, and can be or are associated to functional roles of V_HH domains.

Analysis of local conformations of CDR1 and CDR2

The CDR1 region is defined between the first conserved cysteine (Cys 22) and the conserved tryptophan (Trp 41). The recent PyIgClassify database has 30 clusters of CDRH1 and 19 of these clusters have V_HH structures. Out of the 19 clusters, 13 are very sparsely populated in V_HH structures (<10 structures). Clusters H1-13-1, H1-13-3, and H1-13-5 were analysed, as they were associated to a correct number of occurrences (>20 structures). PB analyses of these clusters show variations in PB assignment in the CDR1 region (see PB frequency maps in Figs. 5A to 4C). The residue region from 21 to 33 is the CDR1 according to PyIgClassify. The PB motif includes 3 residues flanking the CDR1, representing the transition from the β-strand region represented by PB d to loop and back to β-strand. The first cluster (H1-13-1, see Fig. 5A) has 21 out of 37 (57%) structures sharing a strict common PB signature ddddehiafklpccdddddd; the remaining structures of this cluster are close to this PB series as shown with low N_eq (see Fig. 5G). The highest N_eq value for H1-13-1 cluster is not more than 2 for residue position 32, with PB p either changing to c or g. In the other two clusters no strict common PB signature was observed. The third cluster (H1-13-5, see Fig. 5C) has the highest N_eq value. The PBs series of the first cluster dehiafkl is replaced by dehjfklp — a striking difference. The observed structural words in the first cluster PBs dehia, fklpc and cdddd are highly recurrent and reported to have an RMSD below 1 Å, while in the third cluster, the transitions are from the most to the less frequent ones (see the most frequent PB series in de Brevern et al., 2002). As seen in Figs. 5B and 5G, the cluster H1-13-3 has the highest N_eq values and does not show specific PB series like the two other clusters. We can notice small tendencies towards α helical PBs like PB m. A similar in-depth analysis was performed for three CDR2 clusters that also show heterogeneity in the three clusters (see Figs. 5D to 5E and Data S1). In summary, the analyses of CDR1 and CDR2 V_HH in light of PyIgClassify clusters using PBs show a large diversity.

Inter-cluster comparison

The previous analyses showed the variations within a cluster. Using Δ PB and ΔN_eq, it is possible to compare the clusters directly (see ‘Materials and Methods’ section and Fig. S12 to S12D). ΔPB profile allows a detailed description of the conformational diversity between any two sets of protein structural regions. For H1-13-1 and H1-13-5, Δ PB is always higher than 1.5, showing a completely different PB series occurrence. Hence, these two clusters sample two different conformational spaces.

For H2-9-1, H2-10-1 and H2-10-2, ΔPB shows that PB series are closely related in the N-terminal regions between H2-9-1 and H2-10-2, and in C-ter for H2-10-1 and H2-10-2, leading to the idea of composite series with a common PB at position 26 (all ΔPB values of 0.1), information that cannot be provided using RMSD quantification. This analysis shows the relevancy of PBs to compare V_HH structures and V_HH structural models in the following sections. It allows a precise comparison of V_HH structures, which can be considered as highly similar but are not identical, and which quantify this local distance.

V_HH structure prediction survey

V_HH structures are essential to understand binding with partners. Molecular modelling is a simple idea to propose structural models. We performed a complete survey of the V_HH structure prediction studies published so far (see Table 1). Two main methodologies have been used. (i) generic homology modelling like Modeller and threading approaches like i-TASSER, and (ii) hybrid modelling where FRs and CDRs are modelled separately and then assembled together, like RosettaAntibody (Sircar, Kim & Gray, 2009), ABodyBuilder (Dunbar et al., 2014; Leem et al., 2016), or BioLuminate from Schrödinger^® (Beard et al., 2013; Zhu et al., 2014; Salam et al., 2014). Of the above-mentioned approaches, Modeller was used in 50% of the cases in the literature.

All the different structure prediction studies are presented in brief in Table 1 and in detail in Supplemental Information 2). We have presented and classified all the 22 studies into three groups based on the method followed; the reader can, if he wishes, go directly to the following section where we test critically an example of V_HH structure prediction.

Case study

In case of antibodies, the acceptable range of RMSD is different for CDRs and FRs, often leading to global RMSD <3 Å, but it may not be the only near-native structure/conformation possible (Kufareva & Abagyan, 2012). Improvement of model quality was detailed by using multiple templates (Chakravarty et al., 2004; Larsson et al., 2008). Here, we decided to reproduce the study of Steeland and co-workers (Steeland et al., 2015) to assess the interest and impact of such an approach. They predicted structural models of V_HH domains (named Nb70 and Nb96) against Tumour Necrosis Factor receptor 1α using multiple templates: (i) Single chain variable fragment (ScFv) from mouse against Interleukin 1 β (PDBID 2KH2 (Wilkinson et al., 2009)), named temp-m), (ii) a llama V_HH used to stabilize β2 adrenoceptor (PDBID: 3P0G (Rasmussen et al., 2011), referred to as temp-l), (iii) an alpaca V_HH against Hepatitis C virus (HCV) glycoprotein E2 (PDBID: 4JVP (Tarr et al., 2013), temp-a), and (iv) a V_H from human F_ab which is Antibody-dependent cell-mediated cytotoxicity anti-HIV 1 antibody (PDBID: 4FZE (Tolbert, Wu & Pazgier, 0000), henceforth referred to as temp-h). In the following four sections we will present (a) the analysis of sequence relationship, (b) analysis of the structures, (c) the modelling results and finally (d) how the different structural templates have different impacts on the proposition of structural models.

Structure analysis of templates

Sequence similarity from the above analysis suggests good conservation, but structural similarity between the templates provides a different view (see Table S1C). Temp-m is the closest to all three templates, with RMSD values mainly from 1.9 to 2.4 Å, while temp-a is further away with a RMSD value of 3.7 Å with temp-l and of 4.5 Å with temp-h. A large difference between structural templates is observed and does not correspond directly to the information seen in the previous section (see Figs. 6A & 6B). The most conspicuous region in the superposed structures is CDR3 for temp-a and temp-h; in the structure of temp-a, the torso of the CDR3 is bent towards the FR2 region due to the presence of additional disulphide bond. Whereas in the case of temp-h, the CDR3 loop, though long, has a protruding tip (upwards). The RMSD values range from 0.7 to 1.5 Å for FR1, 1.7 to 3.1 Å  for CDR1, 0.4 to 0.9 Å for FR2, 0.3 to 2.2 Å for CDR2, 0.8 to 1.3 Å for FR3, and 0.4 to 1.0 Å for FR4, while CDR3 had segmented values due to difficulty in superimposition. As expected, the general trend of high RMSD in CDRs compared to FRs is observed in all the pairwise comparisons (see Table S2). It is interesting to note the deviations in FR regions between temp-l, temp-a and temp-h, that might be unexpected.

(d) Analysis of local conformational sampling by Modeller

Analyses of best models only provided information for best DOPE score selected models; it does not provide information about potential sampling proposed by the comparative modelling approach. Using PBxplore, Protein Blocks were assigned to each of the models and N_eq entropy computed at each position. A summary of N_eq values >1 is provided in Table 2; in all the cases, modelling with multiple templates is the one with the highest number of residue positions with N_eq>1 (49 residue positions), suggesting the largest conformational diversity. The least number of residue positions with N_eq>1 is for models generated using temp-m (23 residue positions). Notably, models generated from temp-h had two residue positions with high N_eq values in the CDR3 region. Figure 7A is a position-wise distribution of N_eq values in each case. The interesting results are that (i) the N_eq of template temp-m is most often low compared to the multi-template case, and, (ii) the models from multiple templates case show maximum N_eq in the CDR2 region (see also Supplemental Information 2).

Table 2:

Distribution of N_eq in each modelling scenario.

Neq	Temp-m	Temp-l	Temp-a	Temp-h	All templates
1–2	19	34	35	23	39
2–3	4	1	3	4	6
3–4			3	2	3
4–5			1
5–6				1	1
6–7				1

DOI: 10.7717/peerj.8408/table-2

Analysis of PB distribution (see Fig. S15) shows preferred PB ‘d ‘for most residues of FRs in β-sheets in all the cases. The three preceding and succeeding residues of CDRs were considered anchor residues in each loop. Surprisingly, no PB variability was observed for CDR1 anchors. In the case of CDR2, only the position 51 from models generated from temp-l showed a slight variation of N_eq to 1.22. For CDR3, the preceding anchors showed no PB variation; however, the succeeding anchor positions in the modelling scenario of temp-h and temp-a showed variations.

CDRs (aa regions 26–34, 52–58, 97–108), as expected, have more diverse conformations than FRs. Amongst the five scenarios, the most PB diversity in CDR3 is seen in 11 out of 12 residue positions from models generated from temp-a, followed by models generated from temp-l (9/12 residues positions), multiple templates (8/12 residue positions), temp-h (7/12 residue positions) and temp-m (1/12 residue position). This observation suggests that adding more information in the case of multiple templates, and poor template target alignments can cause unexpected conformational diversity.

The local conformational diversity in terms of PBs between mono template(s) and multi-template structural models can be understood by analysing the differences in PBs, quantified by ΔPB. Figure 7B shows ΔPB calculated between multi-template models and each individual template model. Among all four cases of comparison, the case of modelling with temp-m and multi-template shows the least change in PBs at each residue. The differences in local conformations are expected in the CDRs, while the changes in FRs are the ones least expected. These comparisons may also suggest that in case of multi-template modelling, temp-m mostly influenced model conformations due to better alignment quality in CDR regions, and a shared second-best sequence identity with the query. Please note that Nb96 produced roughly the same results, while CDR3 is longer and so more complex to analyse.