Introduction

Since 1958, when Kendrew et al reported the first atomic-level resolution of a protein structure (myoglobin), the structural biology field dramatically expanded with the development of new tools and methods to gain access into atomic details of a protein or a nucleic acid [1]. This opened a completely new world of knowledge and understanding to the scientific community, as the analysis of protein structures provides fundamental insight into most biochemical functions and consequently into the cause and treatment of diseases. Structural biology is now recognized as a fundamental step in our quest to understanding life at the molecular level.

Finding the structures of all proteins is currently a bottleneck for genomics studies. In this matter, the Protein Structure Initiative (PSI) aims at the determination of the three-dimensional (3D) structure of approximately 100,000 structures in 10 years. However, the protein sequence databank (UniProt/TrEMBL) is growing at a much faster rate, with more than 10 millions sequences available to date (March 2010). At the same time point, the Protein Data Bank (PDB) includes 64,100 structures, out of which only approximately 4300 are "unique" at chain level (i.e. once we remove "redundant" proteins whose sequences have more than 95% sequence identity with another protein in the PDB). It should be noted that these structures only represent a biased sample of the protein universe. For example, the PDB includes only 220 unique membrane proteins which is very little since membrane proteins constitute around 20-30% of most proteomes [2]. Noteworthy, the human genome has ~21,000 protein-encoding genes for a proteome of ~1,000,000 proteins when combining the complexity induced by alternative slicing events [3]. In addition, due to experimental limitations, the vast majority of the solved structures are below the 50 KDa threshold excluding numerous larger proteins. Large proteins however represent a significant fraction of the proteins present in an organism; for instance, proteins found in yeast Saccharomyces cerevisiae have five hundred amino acid residues on average and their lengths can reach two thousand eight hundred residues [4]. The structure of these large proteins, as well as of even larger assembly can be solved by electron microscopy at a somewhat low-resolution. While this field is expanding very fast and a growing number of structures solved at atomic-level resolution have been reported [5, 6], its impact with respect to the size of the protein sequence databank remains limited. Many more protein structures have been solved by either X-ray crystallography or nuclear magnetic resonance (NMR). It remains however that most proteins are out-of-reach because of technical difficulties. There is clearly a huge gap between the world of known structures and the universe of known protein sequences. Structural genomic projects are unable to keep up with newly discovered genes.

One way to work around this problem is to use computational methods to predict proteins structures. In silico protein structure prediction techniques can be divided into two categories: the ab-initio folding methods and homology modeling. In this paper we focus on the latter. We note that both approaches have been shown to yield astounding results, as shown in the successive CASP contests [7]. However, they do require caution: while predicting the structure of a protein is an intellectual challenge that requires solving many practical issues, it is often considered as an art in essence.

The growth rate of structures deposited in the PDB is slowing down since 2004, along with the number of new superfamilies or folds discovered [8, 9]. One possible explanation is that many proteins still evade the structural biology pipelines at this time because of the technical difficulties described above. In 1992 Chothia hypothesized that the number of protein folds in nature is probably finite and around 1,000 [9, 10]. The latest analysis of the PDB and of the structural classification of proteins (SCOP) showed that we have not yet reached a plateau (currently estimated to be around 1,500) [8]. The current rate at which proteins are added in the PDB is far too slow to match with the number of new protein sequences discovered every year. The situation is however not so negative. There is a definite hope that the current content of the PDB will allow us to predict reliably the correct scaffold of more than 70% of the whole proteome using in silico methods [11]. This is the rationale for using homology modeling to complement experimental techniques.

Homology modeling predicts the structure of a protein by inference from a homologous protein whose structure is known (see Figure 1 for a schematic illustration of the technique). Its success rests on (i) the existence of a homologue with known structure, (ii) our ability to detect this homologue, and (iii) the quality of the model building process once the homologue is detected. Steps (i) and (ii) have greatly benefited from the different genomics projects: with the additional structures generated by the structural genomics projects, sampling of the protein structure space is becoming finer, improving the chance that a structural homologue exists for any given protein sequence. With the parallel increase in the number of sequences available in genomic databases and the development of meta-servers to analyze and query sequence databases, there has been significant improvement in the detection of homology [12]. In the recent CASP experiments, targets with sequence identity of 6% to their templates were included in the Homology Modeling category [13–15]. There is hope that in the near future homology modeling will reach its ultimate goal: the generation of model protein structures as accurate as those determined by high-resolution experimental studies.

Biologists unanimously consider X-ray crystallography as the prime source of structural information on proteins and the "gold standard" in term of accuracy: they base their confidence on its long list of published successes. The vast majority of structures deposited in the PDB were determined by X-ray crystallography and 14 Nobel prizes in Chemistry and Medicine have been awarded to crystallographers [16] (For recent reviews, see Kleywegt and Jones, 2002 [17]; Wlodawer et al, 2007 [16]; Brown and Ramaswamy, 2007 [18]; Ilari and Savino, 2008 [19]). Homology modeling on its own however is not devoid of successes. The very first published homology model in 1969 was the small protein α-lactalbumin, which was modeled on the basis of the structure of hen egg white lysozyme as a template [20] with the two proteins sharing 39% sequence identity. When the structure of α-lactalbumin was later solved by X-ray crystallography [21], the model turned out to be essentially correct [22]. Since then, homology modeling has continuously extended its field of applications, including designing mutants to test hypotheses about protein functions, identification of active sites, drug design, protein-protein docking, facilitating molecular replacement in X-ray structure determination, refining models based on NMR constraints (for recent reviews of homology modeling applications, see [23–25]). Despite these successes, homology modeling is not yet a well-established alternative or complement to experimental structural biology. It remains the focus of many criticisms often coming from the structural biologists themselves as they often consider a protein model to be unreliable, not being based on experimental data. The question arises as to what needs to be done to give homology modeling its credentials.

In this paper, we focus on homology modeling and more specifically, we review how the structural biology community perceives it and what can be done to impress on the experimentalists that it can be a valuable resource to them. It is organized as follows. The next section reviews the differences and similarities between homology modeling and high-resolution experimental structural biology. In particular, we illustrate steps in the homology modeling procedure that are putative source of errors. Our goal is to provide a useful step-by step handbook for non-specialists in order to help building better model.

The following section introduces the H-factor, a new indicator for assessing the quality of homology models. The H-factor is designed to check how well a family of homology models reflects the data that were used to generate those models, in the spirit of the R-factors in X-ray crystallography. The results section that follows validates the H-factor on a series of test cases. We conclude the paper with a discussion on what remains to be done to make homology modeling a prime technique for the biologists.

Homology Modeling versus Experimental Structural Biology

This section reviews briefly the quality of protein structure models obtained either using high-resolution experimental techniques or homology modeling. Our hope is to identify common good practices as well as safeguards from which we can derive a validation tool for the latter. We start with the concept of a structural model and its meaning in the two communities of experimental and computational structural biologists. We then highlight the pros and cons of X-ray crystallography and NMR spectroscopy. An overview of the different steps involved in homology modeling follows, with emphasis on sources of errors and how they can be checked. Ultimately our goal is not to rank these methods but rather we hope to show that they all provide valuable and often complementary information, as long as the proper safeguards are applied.

*Steps 1 and 2: template selection and sequence alignment

The selection of template(s) is undoubtedly a critical step in modeling. It was long assumed that two proteins whose sequences share at least 40% identity have similar structures. If such a template exists, it is easily detected by any sequence alignment techniques. Homology modeling under such conditions is then expected to generate models whose accuracy is close to that of an experimental structure. We know however that this is not always true. Roessler et al. recently reported the discovery of two native Cro proteins sharing 40% sequence identity but with different folds [43]. Moreover, Alexander et al. were able to design two proteins with 88% sequence identity but having different structure and function [44]. Reversely, it is not uncommon for proteins, especially enzymes carrying the same function across the tree of life to share a somewhat low sequence identity and at the same time being structurally similar, with a Cα r.ms.d. ~1.5Å [45]. All these observations indicate that the selection of template is far from being a trivial task and extreme caution should be applied.

The situation is even more difficult if there is no significant sequence similarity between the sequence of the target protein and any of the known protein structures. This is one of the current challenges of the post-genomic era that is tackled by fold-recognition methods, namely to identify a suitable template for homology modeling [46–48]. Unlike sequence-only comparison, fold-recognition techniques take advantage of the information made available by 3 D structures. Despite a steady development over the years, as illustrated through the successive CASP experiments, fold recognition techniques still have a number of limitations. They are however the key to extend the domain of application of homology modeling methods.

The accuracy of the sequence alignment is another critical step in the homology modeling process [49]. It is important to keep in mind that a shift of one residue in the alignment introduces a distortion of 3.8 Å in the model for the backbone of the target protein (Figure 2A &2B). Therefore, it is of paramount importance to precisely verify and locally refine the sequence alignment used to build a model. For that matter, the areas of sequence comparison and alignment optimization have grown in parallel to the comparative modeling field [50]. There are many ways to align protein sequences and numerous methods to score the accuracy of the resulting alignments. Homology modeling usually relies on one of following three techniques: (a) standard pairwise sequence alignment using dynamic programming, (b) multiple sequence alignment when the target and template sequences belong to a large family for which many sequences and structures are known, and (c) direct alignment of the sequence on the structure of the template using a threading technique. None of these techniques is the panacea, especially when the sequence identity between the target and template sequences is low. Errors come for example from the fact that the optimal alignment between the sequences of two proteins may not match the alignment between their structures: this is caused by the sequence alignment technique that optimizes a score based on a substitution matrix and gap penalties while structural alignment techniques optimize geometric matching. There is no reason that these two different metrics are equivalent. Errors are often frequent in the loop regions between secondary structures, as well as in regions where the sequence similarity is low. Recent techniques that attempt to reduce the error rate in the sequence alignment rely on the inclusion of as many sources of information as possible, such as amino-acid variation profiles, secondary structure knowledge, structural alignment data of known remote homologues, knowledge of "anchor regions" (active-site catalytic residues or conserved motifs based on biological data). Many programs also attempt to optimize the raw sequence alignment derived from one of the techniques described above. Most of these programs have been embedded into web-servers for ease of uses, such as MUSCLE [51] and Dialign-TX [52].

* Step 5: The side-chains positioning problem

The prediction of side-chain conformations for a given backbone architecture remains a challenge. It is however the key to generating models at atomic resolution. For instance, critical side-chains forming an active site must be accurately positioned in order to support any putative catalytic mechanism. Nearly all the side-chain positioning methods are based on rotamer libraries with discrete side-chains conformations. Rotamer libraries contain a list of all the preferred conformations of the side-chains of all twenty amino acids, along with their corresponding dihedral angles [56, 57]. Some of these rotamer libraries are further refined to account for the local geometry of the backbone [58, 59]. Sidechain prediction techniques select the best rotamer for each residue of the protein under study from one of these libraries, based on a score that includes both geometric and energetic constraints. This leads to a large combinatorial problem: see Figure 3 for an example of the size of the conformational space accessible to a sidechain in a protein, even when we discretize this space by using rotamers. The combinatorial problem is solved by heuristic techniques such as mean field theory, derivatives of the dead-end elimination theorem or Monte Carlo techniques [60]. The success rates of the most successful techniques range from 78% to 89% for the χ¹ and χ² angles of residues in the core of the protein (i.e. residues whose solvent accessibility is less than 30%). It is important to note that these results usually relate to mock experiments in which the exact conformation of the backbone is known. It has been shown that the quality of sidechain prediction decreases as the deviation between the backbone used for modeling and the actual conformation of the backbone increases [58, 61]. These results emphasize the need of a good framework (and implicitly of a good sequence alignment between the target and template protein) for homology modeling. It is worth mentioning however that a 1Å accuracy over the main-chain atom position can be related to X-ray structures with 2.5 Å resolution and with a R-factor around 25%, which is fairly common [62]. In addition, multiple side chain orientations are routinely observed in crystal structures and crystal-packing forces can alter their positions as well. The side chain prediction methods are improving steadily, as reported by the CASP experiments [63].

Computing the H-factor

In this study, we introduce a novel indicator as an attempt to quantify how well a homology model correlates with the input data and its biological relevance. This indicator, the H-factor is designed to mimic the R-Factor in X-ray crystallography. It is rooted in the basics of homology modeling and it uses all data that were included in the model building process to assess its correctness, in addition to checking for good stereochemistry. More specifically, the H-factor combines information on (1) the template structure(s) (based on the corresponding PDB files); (2) the sequence alignment between the template(s) and the target sequences; (3) the structural heterogeneity of the models built; and (4) the structural neighborhood within protein families. Note that at this stage, the computation of the H-factor is based on a single framework. We plan to derive an extended H-factor that will account for multiple templates. Each of these four categories is assigned a scoring function returning a value between 0 (best) and 10 (worst). The H-factor is simply the sum of these scoring functions divided by 40. For ease of use, the H-factor is converted into a percentage with a value between 0% (ideally the perfect score) and 100% (the worst score). In par with the R-factor in X-ray crystallography, a low H-factor value indicates a trustworthy set of models, while a large H-factor would raise a flag on the correctness of the models. The H-factor focuses on the Cα-backbone as a correct tracing is a prerequisite for a valid model. The workflow of the H-factor calculation is detailed on Figure 4A and 4B.

where the sum is computed over all positions in the sequence alignment between the target and template, N is the length of the sequence alignment, p is the secondary structure prediction of the target at position i (values for p are 'H' for helix, 'S' for strands, and 'C' otherwise), c(i) is the confidence factor reported by psipred (integer value, from 1 to 10) for the secondary structure prediction at position i and s is the secondary structure type observed at position i in the template structure reported by stride. The offset coefficients a and b are set to 1.3 and 0.9, respectively, to ensure that score (1) has values between 0 and 10.

where N is the length of the sequence alignment.

n is the number of models. The offset coefficients a and b are chosen such that average RMS values of 0.1 and 7 Å correspond to scores of 1 and 10, respectively; the corresponding values are a = 1.3 and b = 0.87.

m is the number of functional domains identified in the target sequence, MA _d is the structural fragment extracted from the average structure MA corresponding to the domain d, n is the number of domains homologous to domain d found in PDB structures, and D_d,i is the i-th possible structure of the domain homologous to d. The offset coefficients a and b have been chosen such that average RMS values of 0.1 and 7 Å correspond to scores of 1 and 10, respectively; the corresponding values are a = 1.3 and b = 0.87. This usually enforces that score (4) is between 0 and 10. Note that if this procedure does not find an equivalent domain for a fragment, the fragment is ignored; if no domains are found for all fragments, score ( 4 ) is ignored.

The H-factor computation is accessible online at http://koehllab.genomecenter.ucdavis.edu/toolkit/h-factor with a simplified operating manual (Cf. additional files). The source code is available upon request.

Testing of the H-factor on CASP targets

The H-factor: detailed analysis on three CASP targets

When we deliberately introduce a shift at position 31 in the alignment between the sequence of T0295 and the sequence of its template 1ZQ9 (see Figure 2), the corresponding models generated by MODELLER show structural diversity in the loop region near the shift (i.e. near Thr31). Score (3) captures this structural diversity within a set of models. It leads to the H-factor being raised from 19% to 21% (see table 2). However, score (2) could not detect a single position shift in the alignment. The H-factor is therefore capable of detecting backbone deviation due to modeling errors, the same way the R-factor does.

The CASP7 target T0375 is a more difficult modeling case. It is a human ketohexokinase and the rigid-body domain closure of sugar kinases is known to be large, adding complexity into the modeling process [89]. Although several sugar kinase structures have been solved, the search for templates for T0375 identified only six distinct remote templates. Moreover, all six templates are needed to obtain complete sequence coverage of T0375 within one single framework with MODELLER. In addition, the template sequences have low similarity with the target sequence. This is detected by the scoring function (2), which returns a value of 8.6 (out of 10) (table 2). Note that the score (2) is not a direct measurement of the quality of the sequence alignment. It is designed to quantify the difference between the two sets of sequences: if this difference is small, the model is expected to be good, while if the difference is large, the sequence alignment most probably belongs to the twilight zone and the models should then be considered with caution. The overall H-factor for the models generated for T0375 is 41%. This mid-range value indicates that caution should be used when interpreting or using these models. Indeed, the average cRMS between these models and the actual structure of T0387 (available in the PDB in the file 2HLZ) is 3 Å, i.e. reflecting a medium-resolution agreement.

The CASP7 target T0287 is the most difficult test case we have considered. In fact, it would not be considered a homology-modeling target by many, despite the fact that a (remote) structural homologue is available in the PDB. We did decide to include it in our study to test whether the H-factor was still providing useful information when applied on a difficult test case. T0287 corresponds to CaGS, a protein from Helicobacter pilori whose function is unknown. A database search over all sequences of proteins whose structure is known identifies a unique template, 1V55, with a low sequence identity (16%, see table 1). 1V55 is a Cytochrome C Oxidase and it is not clear that T0287 and 1V55 are homologues. We did build 20 models for T0287, using 1V55 as a template, and the out-of-the-box alignment between the sequences of T0287 and 1V55 generated by ClustalW. As mentioned above, we did not try to optimize the alignment or the modeling itself as our interest is to see if the H-factor is able to assess the quality of the models we generated. In this specific case, all four scores reported high values (7.0, 8.1, 7.5 and 7.5 for scores (1), (2), (3) and (4), respectively). The overall H-factor is 75%, a valued that should raise concerns about the quality of these models. As for target T0375, this is confirmed by the experiment: the average cRMS between these models and the actual structure for T0287 (PDB code 2G3V) is 5.8 Å, indicating that the models are poor approximations of the native structure.

H-factor: Detecting bad models

H-factor: characterizing good models

Although the H-factor and the R-factor are mathematically unrelated, they have the same purpose: to assess the quality of structures, either experimental or computed from modeling experiments, where quality refers to reflecting correctly the input data used to generate these structures. The H-factor mimics the R-factor as it provides a quality-index to follow in the process of building a model, the same way crystallographers monitor the R-factor/R-free indexes during structures refinements. We compared the H-factor results with the experimental R-factor and R-free on a randomly chosen subset of the PDB containing 445 structures with 6 or more identical chains solved by X-ray crystallography. Results are shown in Figure 5. The H-factor and R-factor are not linear correlated but it remains that "good" R-factors (below 30%) correspond to "good" H-factor values (below 45%). The H-factor checks the diversity of the set of models generated for a structure, as well as their similarities with the structures of domains that share the same function, as defined by pfam. High H-factor values may be caused by structures with disordered loops or remote structural neighbors in the PDB. It remains that, in agreement with what we observed for the NMR structures, the H-factor recognizes experimentally determined structures as being valid.

Relationship between the H-factor and cRMS

The cRMS by itself is a reasonable quality indicator as long as its value remains low (say below 2 Å). It should be noted that it is an average value computed over the whole structure. As such, it is very sensitive to large structural fluctuations in disordered loops for example that can lead to large cRMS values even if the conserved domains are structurally very similar. It is well known for example that caution should be applied when using cRMS to assess the quality of a structural alignment. cRMS is implemented in score (3) of the H-factor to evaluate the heterogeneity amongst a set of models as it is directly related to both the choice of the template and the quality of the sequence alignment. However, the score (3) loses accuracy for cRMS values larger than 2 Å, which is not uncommon when a remote template is used. cRMS is also implemented in score (4) to quantity the modelling quality of specific individual domains by comparing them with corresponding domains in the PDB. This is a domain-based cRMS that does not take into account potential long loops between domains, making it more reliable. Taken together, the scores (3) and (4) alleviate most of the limitations of cRMS while retaining its major properties. The H-factor is therefore expected to be more reliable than a sole cRMS to judge the accuracy of a wide range of models, as seen in Table 3.

Comparing the H-factor with ProSA, DOPE and QMEAN

The structural biology community as well as the protein structure modeling community have always been reasonably good at setting safeguards to estimate the validity of both experimental and computational models for protein structures. The first validation tests imposed on experimental structures focused on the stereochemistry of the molecules. It was found, however that this was by far not sufficient, as "good stereochemistry" can be misleading. A "clean" Ramachandran plot for example does not necessary prove that a given model is valid. Also, we observed that all the models listed in Table 1 have good stereochemistry, including the T0287 model that has a Cα backbone deviation of almost 6 Å from the actual experimental structure. Experimentalists have access to another essential set of validation tools that check the consistency of the structures with the experimental data that were used to derive them. Such tools include the R factors in X-ray crystallography and NMR. Models derived from modeling experiments are more difficult to verify, unless they include some external information. For example, the SWISS-model server provides a confidence score along with the models it built and this confidence score is based on the amount of structural information that supports each part of the model [38, 90]. In general, however, validation of computer-generated models is based on comparison with known protein structures. The idea is that basic properties of protein structures can be inferred from the PDB, and translated into database-derived scoring functions. This is the main rational for the statistical potentials, also referred to as mean-field potentials that have become very popular validation tools [67, 91, 92]. For example, ProSA-web implement two statistical potentials: a pairwise potential based on Cß atoms and a surface term that models the protein-solvent interactions [93]. Here we compare the output of ProSA and H-factor analyses on the three CASP7 test cases given in table 1. The ProSA results are given if Figure 6. Both the H-factor and ProSA estimate that the models generated for T0295 are valid. When the modeling process is deliberately altered (test case T0295*, see above for details), again both ProSA and the H-factor analyses reveal a loss of quality. ProSA however considers the models generated for the targets T0287 and T0375 as statistically correct (i.e. within range of what is observed in native proteins), while we know that these models are not correct (see table 2). In addition, it is striking to see that the PDB structure 1US0, an ultra high-resolution crystal structure at 0.66Å resolution, has a ProSA Z-score (-12.27) statistically as correct as T0287 (data not shown).

The statistical potential Discrete Optimized Protein Energy (DOPE) is another measure of model quality that has been introduced in MODELLER-8 [94]. DOPE is a statistical potential with an improved reference state that accounts for the compact shape of native protein structures. The DOPE score is designed such that large, negative scores are usually indicators of good models. In their original study, Shi and Sali [94] found that the accuracy of DOPE to asses a homology model improves as the accuracy of the models improve. We observe a similar behaviour for targets T0295 and T0295* (Table 2). These two targets correspond to the same protein and it is therefore possible to compare the DOPE scores of their models. The model generates for T0295*, based on an incorrect alignment, has a much lower DOPE score (-24317) that the model generated with the correct alignment (T0295;-33940). Note that we cannot compare DOPE scores for proteins of different size, as these scores are not normalized. DOPE scores are therefore relative, and designed to pick a "good" model among poorer model. DOPE scores do not assess directly the quality of the model that is picked, i.e. if it is likely to be similar to the actual structure. The H-factor is a better indicator in that respect.

QMEAN, which stands for Qualitative Model Energy ANalysis, is a composite scoring function for homology models that describes the major geometrical aspects of protein structures (including a torsion angle potential over three consecutive amino acids, a secondary structure-specific residue-based statistical potential, a solvation potential for the burial of residues) as well as the agreement between the predicted and calculated secondary structure and solvent accessibility, respectively [95]. As such, it includes a term similar to the score (1) of the H-factor, as well as terms that assess different properties such as residue accessibility. The score QMEANnorm is a normalised version of the QMEAN score in which all terms are divided by the number of interactions/residue in order to avoid a size-bias of the score [95]. QMEANnorm scores vary between 0 and 1, with larger scores expected to correspond to better models. Unlike the DOPE score, both the H-factor and QMEANnorm scores allow for the comparison of proteins of different sizes. The QMEANscore is as effective as PROSA or DOPE for detecting errors in a model that result from errors in the sequence alignment between the template and target protein: T0295* has a QMEANnorm score of 0.196 while the score forT0295 is 0.735 (Table 2). Interestingly, T0295* (0.196) has a less favorable QMEANnorm score than the erroneous model generated for the CASP target T0287 (0.285) (see table 3). We have observed however that the QMEANnorm score is prone to fail: some of the erroneous models generated for the CASP target T0307 have QMEANscores of 0.4 to 0.6, i.e. they are evaluated to be almost as correct as the positive control T0295 (0.735). Unlike ProSA and QMEAN, the H-factor did detect that these models were to be considered with caution. Because it analyzes a set of models, we believe that the H-factor score is more robust as an absolute measure of the quality of a model. It lacks however the ability to discriminate among a set of models generated for the same target; PROSA and DOPE are better potentials for this specific task.

These results emphasize the essential differences in the nature of the ProSA, DOPE, QMEANnorm and H-factor scores. ProSA, DOPE and QMEAN check the quality of a model, independently of the context in which it was generated. The H-factor on the other hand checks the quality of a set of models with respect to a context that includes for example the sequence alignment assessed by the score (2) . The modeler however should use these differences to extend his/her assessment of the model his/she generates. We believe that ProSA, DOPE, QMEAN and H-Factor analyses are needed to provide a better overview of the quality of models derived by homology modeling.

Current limitations and originalities of the H-factor

The H-factor is not the panacea, and does not provide a universal solution to the problem of asserting the quality of a model generated by homology modeling. Firstly, the H-factor has some technical limitations. Our current implementation does not take into account multiple templates, but rather only one single framework. The structural components included in the H-factor (i.e. scores (3) and (4)) are based on the backbone of the models, and do not take into account sidechains and possible errors in their modeling. Second, the scoring function (3) of the H-factor measures the heterogeneity of a set of models generated with the same input. It means, that the H-factor cannot be computed on a singular model. In homology modeling the heterogeneity of models can be seen as a quality indicator and building only one single model is not recommended. Similarly to NMR structures where only one of the models can be chosen for analysis, the best model in homology modeling regime is chosen based on the MODELLER energy function for instance. Third, the H-factor does not include any external information. For example, if some biological data are available, such as the knowledge of the residues involved in the active site, or standard biophysical data such as melting temperature, or secondary structure content derived for circular dichroism, these data are currently not included in the H-factor analysis.

The R-factor is a measure of the agreement between the crystallographic model and the experimental X-ray diffraction data. Despite the lack of 'experimental' data to compare with, the modelling community has been searching for a similar indicator for homology modeling. Both QMEAN and the H-factor are designed to be 'absolute' indicators that assess the quality of homology models in a way that mimics the R-factor in X-ray crystallography. Both QMEAN and H-factor provide an easy-to-use estimate of the quality of models based on scoring functions assessing various aspects of the modelling process as well as the model itself [95].

In vivo macromolecular structures oscillate between numerous conformers, some more than others. While X-ray structures correspond to snapshots of a limited numbers of conformers, NMR structures tend to describe more accurately flexibility. Indeed, NMR "structures" are usually provided as a family of conformers that are meant to sample the conformational space accessible to the molecule of interest. In homology modeling on the other hand, the heterogeneity of models is a quality indicator. A good set of models will have a cRMS very close to their framework. Moreover, if errors are being made in the template choice or in the sequence alignment, then the models will be heterogeneous. The scoring function (3) is designed to quantify this assertion. It also means that the H-factor cannot be computed on a single model.

One of the originality of the H-factor is the scoring function (4). It has been designed to evaluate the biological relevance of the models by comparing the model conformations of all the functional domains in the protein considered with the existing sibling deposited in the Protein Data Bank.

We acknowledge that there is room for improvement. However, It remains that the H-factor we have introduced here is a first step in the direction of validating homology models for the biologists in addition to existing methods, as proved in the examples shown above.