- Split View
-
Views
-
Cite
Cite
Hannah F Löchel, Mona Riemenschneider, Dmitrij Frishman, Dominik Heider, SCOTCH: subtype A coreceptor tropism classification in HIV-1, Bioinformatics, Volume 34, Issue 15, August 2018, Pages 2575–2580, https://doi.org/10.1093/bioinformatics/bty170
- Share Icon Share
Abstract
The V3 loop of the gp120 glycoprotein of the Human Immunodeficiency Virus 1 (HIV-1) is considered to be responsible for viral coreceptor tropism. gp120 interacts with the CD4 receptor of the host cell and subsequently V3 binds either CCR5 or CXCR4. Due to the fact that the CCR5 coreceptor is targeted by entry inhibitors, a reliable prediction of the coreceptor usage of HIV-1 is of great interest for antiretroviral therapy. Although several methods for the prediction of coreceptor tropism are available, almost all of them have been developed based on only subtype B sequences, and it has been shown in several studies that the prediction of non-B sequences, in particular subtype A sequences, are less reliable. Thus, the aim of the current study was to develop a reliable prediction model for subtype A viruses.
Our new model SCOTCH is based on a stacking approach of classifier ensembles and shows a significantly better performance for subtype A sequences compared to other available models. In particular for low false positive rates (between 0.05 and 0.2, i.e. recommendation in the German and European Guidelines for tropism prediction), SCOTCH shows significantly better prediction performances in terms of partial area under the curves and diagnostic odds ratios compared to existing tools, and thus can be used to reliably predict coreceptor tropism for subtype A sequences.
SCOTCH can be downloaded/accessed at http://www.heiderlab.de.
1 Introduction
Infection of the host cells with the Human Immunodeficiency Virus 1 (HIV-1) proceeds in several steps that include the binding of the gp120 surface protein of HIV-1 to the CD4 receptor and a coreceptor, namely one of the chemokine receptors CCR5 or CXCR4 (Lee et al., 1999). Coreceptor tropism, i.e. the type of coreceptor that is used by an HIV-1 virus, has important clinical implications. First, patients with a CXCR4-tropic virus progress faster to AIDS compared to patients with CCR5-tropic viruses (Koot et al., 1993). Second, entry inhibitors that bind to the coreceptor and thus inhibit viral entry, such as Maraviroc (Dorr et al., 2005), are only available for the CCR5 coreceptor, and are thus ineffective against CXCR4-tropic viruses. Today, entry inhibitors are frequently used in antiretroviral treatment, thus the determination of coreceptor tropism has become crucial for patient therapy. The gold standard for determining coreceptor tropism is by cell-based assays (Whitcomb et al., 2007). The main disadvantages of cell-based assays are that they can only be carried out by specialized laboratories and that these assays are expensive and time-consuming. It has been shown in several studies, that computational approaches for coreceptor tropism prediction can be a viable alternative to cell-based assays. The main advantage of these predictive models is that the procedure is cheap and very fast, in particular when these algorithms are executed in a parallelized manner, e.g. on graphics cards (Olejnik et al., 2014). Due to the fact that the third variable loop of the gp120 protein (V3) is considered to be responsible for coreceptor usage, these models are typically trained on a set of V3 sequences with known tropism, and subsequently applied to new, unseen V3 sequences in order to predict tropism. Several models have been proposed, from simple rules, such as the 11/25 rule (Fouchier et al., 1992; Shioda et al., 1992), to sophisticated machine learning models. For instance, geno2pheno[coreceptor] is based on a support vector machine (Lengauer et al., 2007) trained on V3 sequences. T-CUP uses structural information for modeling the electrostatic potential and hydrophobicity of the V3 sequences in order to predict coreceptor tropism (Dybowski et al., 2010a,b; Heider et al., 2014). PhenoSeq makes use of sequence motifs and predicted charges of the sequences (Cashin et al., 2015), while WebPSSM (Jensen et al., 2003) uses scoring matrices. These models have been shown to give reliable predictions and can be used for clinical assessment of coreceptor tropism. However, HIV-1 can be subdivided into different subtypes that show different abundancies and different spatial distributions. HIV-1 subtype B is mainly found in North America, the Caribbean, Latin America, Western and Central Europe and Australia and makes up 11% of the infections worldwide (Hemelaar et al., 2011). Almost all available computational models have been trained on subtype B data. Subtype C makes up 48% of worldwide infections and is mainly found in Africa. It has been shown in several studies that the available models for coreceptor tropism can also be applied for subtype C sequences with comparable prediction accuracy (Gupta et al., 2015; Riemenschneider et al., 2016). The third major subtype, namely subtype A, is responsible for around 12% of worldwide infections and can be mainly found in Eastern Europe and Central Asia. Unfortunately, it has been demonstrated that the available models are not reliable for tropism prediction of subtype A viruses. The performance of the existing tools drops to less than 50% accuracy when applied to subtype A sequences (Riemenschneider et al., 2016). They proposed that there may be a slightly different underlying mode of binding, which could involve other parts of gp120. The potential involvement of the V2 loop, for instance, is also mentioned by others (Kitawi et al., 2017; Pastore et al., 2006). Moreover, there is an apparent selection for subtype A variants that are less glycosylated and with shorter V1-V2 loop sequences (Chohan et al., 2005). The aim of the current study was the development of a reliable subtype A specific coreceptor prediction model.
2 Materials and methods
2.1 Dataset
We used the dataset of V3 sequences of subtype A collected by Riemenschneider et al. (2016). The V3 loop sequences of HIV-1 with assigned subtype A or CRFs with a V3 region originating from subtype A were downloaded from the Los Alamos HIV sequence database (http://hiv-web.lanl.gov) in March 2015. Sequences that occur with contradictory tropism annotation in the database or contain non-canonical amino acid symbols were removed. Duplicated sequences were included only once. Additionally, nine subtype A sequences that were collected at the Institute of Virology at the University of Cologne were used as well. The final dataset consists of 182 V3 sequences of subtype A from 147 R5-tropic and 35 X4-tropic viruses.
2.2 Phylogenetic analysis of the samples
We performed a phylogenetic analysis of the V3 sequences in order to confirm the assigned subtypes. To this end, a multiple sequence alignment (MSA) of the V3 sequences was computed with MUSCLE (Edgar, 2004). The MSA was used to generate phylogenetic trees with SeaView 4 (Gouy et al., 2010) using Poisson distance and BioNJ (Gascuel, 1997). Gap sites were ignored and significance was estimated by bootstrapping with 100 replicates.
2.3 Feature encoding
It has been shown in several studies that the most crucial part in predictive modeling is the feature encoding, i.e. the encoding of the protein sequences. In order to improve ensemble diversity (Kuncheva and Whitaker, 2003), we made use of structural and sequence information of the V3 loop according to Dybowski et al. (2010a), which have been demonstrated to give reliable prediction on HIV tropism, when used as input in subsequent machine learning models. Introducing structural information into classification models has been demonstrated to improve overall prediction performance (Bozek et al., 2013; Dybowski et al., 2011; Sander et al., 2007). To this end, we used the same overall architecture as for subtype B sequences (Dybowski et al., 2010a), however, due to the higher flexibility of the subtype A V3 loop sequences, the distances needed to be adapted. We encoded the V3 sequences in two ways: (i) by building homology models of the V3 loop and calculating the electrostatic potential on the surface, and (ii) by using a hydrophobicity encoding of the V3 sequences (see Fig. 1). We employed Modeller (v 9.17) (Šali and Blundell, 1993) in order to generate homology models of the V3 sequences, using the X-ray structure of the gp120 protein (PDB: 2QAD) as a template. The V3 loop sequences were aligned pairwisely against the sequence of the template structure with the R package bio3d (Grant et al., 2006). Upon visual inspection we found that the alignments were of good quality and there was no need to manually adjust them. Ten models were generated, and for each V3 sequence, we selected the structure with the highest DOPE-Score (Elias et al., 1991) for subsequent analyses. Next, we calculated the electrostatic potential at the surface of the V3 structures using the AMBER force field and PDB2PQR (v 2.1.1) (Dolinsky et al., 2004). The solvent accessible surface of the structures was determined by APBS (v 1.4.2.1) (Baker et al., 2001) using a grid of 333 points with a spacing of 3 Å according to Dybowski et al. (2010a), which have been demonstrated to give reliable prediction on HIV resistance and tropism, and a radius of the solvent molecules of 1.4 Å. In order to find the best distance for our prediction model, we evaluated the electrostatic hull at distances of 0, 3, 6, 9 and 12 Å from the solvent accessible surface. The hydrophobicity encoding of the V3 sequences was generated with Interpol (v 1.3) (Heider and Hoffmann, 2011). V3 sequences were translated into their numerical hydrophobicity representation according to the Kyte-Doolittle hydropathy index (Kyte and Doolittle, 1982). Subsequently, the numerical hydrophobicity vectors were interpolated to a common length of 35, which represents the average length in our dataset. We also evaluated whether an additional feature encoded in Interpol (Heider and Hoffmann, 2011) could improve overall prediction performance. However, we found no significant improvements of the resulting models in terms of AUC, compared to the model based on hydrophobicity and electrostatics.
2.4 Machine learning
The dataset was not balanced prior to training. In imbalanced data, there are two opportunities in order the balance the dataset, namely upsampling of the minority class or downsampling of the majority class. Although balancing has been shown to improve the overall performance metrics, e.g. AUC, upsampling may lead to overestimation of the AUC and downsampling may lead to a significant loss of information. Due to the fact that the amount of available subtype A V3 sequences with known tropism is limited, we used an imbalanced dataset in the current study.
2.5 Comparison with other methods
We compared our novel subtype A prediction model SCOTCH with existing models, namely T-CUP (Heider et al., 2014), geno2pheno[coreceptor] (Lengauer et al., 2007), PhenoSeq (Cashin et al., 2015), WebPSSM (Jensen et al., 2003) using all available matrices (i.e. x4r5, sinsi and sinsi c), and the genotypic rules of Raymond et al. (2012) and Esbjörnsson et al. (2010).
3 Results
3.1 Overall approach
The aim of the study was to build a reliable coreceptor tropism prediction for HIV-1 subtype A. Our model SCOTCH is based on a stacking approach of two random forests (RFs) that were trained on different feature encoding in order to improve classifier diversity. The first RF was trained on the electrostatic potentials at the surface of the V3 structure models. The second RF was trained on the numerical hydrophobicity representations of the V3 sequences. The outputs of these RFs are combined via stacking. To this end, a third RF uses the outputs, i.e. pseudo-probabilities, and makes a final prediction whether a given V3 sequence belongs to a CCR5- or CXCR4-tropic virus. The performance of SCOTCH was compared with the existing models.
3.2 Electrostatics hull
The sequences were aligned using MUSCLE (Edgar, 2004) in a pairwise manner (all-against-all). 117 V3 sequences (64.3% of all sequences) share at least 90% identity of their sequence with at least one other sequence in the dataset. For 48 V3 sequences (i.e. 26.4%), there is at least one other V3 sequence with a similarity of ≥97.1%. The R5-tropic sequences show a higher average similarity compared to the X4 sequences (91.8 and 86.7%, respectively). We used Modeller (Šali and Blundell, 1993) in order to generate the V3 structures for the prediction of coreceptor tropism, and APBS Baker et al. (2001) in order to calculate the electrostatic hulls at different distances to the surface. For the discretization of the electrostatic hulls in order to be used in the subsequent classification models, we used a grid spacing of 3 Å, which was centered over each V3 loop structure. For the electrostatics hull at a distance of 0 Å, all 182 V3 loops are too close to the hull or even penetrate it, so they have at least one grid point that is not accessible by the solvent. For a distance of 3 and 6 Å, 177 and 65 sequences, respectively, still penetrate the hull. In contrast to our results with subtype B sequences (Dybowski et al., 2010a), only the electrostatic hulls at a distance of 9 Å completely enclose all V3 loop structures. These findings might imply that, on average, V3 sequences from subtype A are more disordered compared to subtype B, which have been used in former studies. These findings are in line with the notion that subtype A sequences show biochemical differences compared to subtype B sequences (Chohan et al., 2005).
3.3 Electrostatics-based classification
A homology model for each V3 loop was generated based on the template X-ray structure of the viral gp120 protein (PDB: 2QAD). Electrostatic potentials at discrete grid for each V3 loop structure were obtained by solving the non-linear Poisson-Boltzmann equation by APBS (Baker et al., 2001). We used the RFs to estimate feature importance. Figure 2 shows the twenty most important positions on the V3 structure according to the RF importance analysis. The most important positions cluster around residues 11–14 and residue 25, which is in partial agreement with the 11/25 rule (Fouchier et al., 1992; Shioda et al., 1992). Residue 25 might be less important than residues 11 to 14, given the fact that only two out of the twenty most important grid points can be assigned to this residue.
3.4 Hydrophobicity-based classification
In addition to the RF trained on the electrostatic hulls, we also trained a model based on the hydrophobicity encoding. The V3 loop sequences were encoded with the Kyte-Doolittle hydropathy index (Kyte and Doolittle, 1982) using Interpol (Heider and Hoffmann, 2011) and normalized to the average V3 length, i.e. 35 residues. Figure 3 shows the importance of the residues of the normalized V3 sequences according to the hydrophobicity-based RF. Two important clusters can be identified at positions 10–14 and 22–25. Again, these findings are in agreement with the 11/25 rule. Due to the fact that the dataset contains sequences shorter than 35 amino acids, residue positions in the interpolated sequences are again slightly shifted to the right and do not necessarily correspond to the actual residues.
3.5 Stacking approach
We used a stacking approach in order to combine both models. To this end, the pseudo-probabilities of both models were used as an input for a third RF. Figure 4 shows the ROC curves of the electrostatics- and hydrophobicity-based models. The AUC of the RF trained on the electrostatic hulls is 0.7704 ± 0.0043, while the RF trained on hydrophobicity reaches an AUC of 0.7004 ± 0.0023. However, the difference in AUC is not significant (P = 0.1899). The electrostatics-based model reaches higher true TPRs for almost all FPRs, thus it is not obvious why the use of stacking could improve overall performance. Nevertheless, the third combined RF outperformed the other two RFs significantly at low FPRs. The AUC of the combined RF (i.e. our final method SCOTCH) is 0.7031 ± 0.0046, which is not significantly higher compared to the model trained solely on the electrostatic hull (P = 0.2216). However for low FPRs (<0.2) the combined model outperforms the single models (see Fig. 4). For instance, the European guidelines recommend the use of a 10% FPR cutoff. The corrected partial AUCs for the combined model is 0.7498 for an FPR between 0.05 and 0.2, which is significantly higher compared to the corrected partial AUCs of 0.6993 and 0.5969 for the electrostatics- and hydrophobicity-based model, respectively. The sensitivity and accuracy at a specificity of 95, 90 and 85% for the different models are listed in Table 1. For a specificity of 95%, the electrostatics-based model achieves a sensitivity of 40.0% and an accuracy of 84.4%. The combined method has a higher sensitivity (47.7%) and a higher accuracy (85.9%). The DOR is also higher for the combined model at low FPRs (see Table 1).
Method . | Sensitivity . | Specificity . | Accuracy . | DOR . |
---|---|---|---|---|
Electrostatics | 40.00 | 95.00 | 84.42 | 15.67 |
Hydrophobicity | 14.00 | 95.00 | 79.42 | 3.85 |
Combined model (SCOTCH) | 47.71 | 95.00 | 85.91 | 17.07 |
Electrostatics | 43.43 | 90.00 | 81.04 | 7.92 |
Hydrophobicity | 24.57 | 90.00 | 77.42 | 3.36 |
Combined model (SCOTCH) | 56.57 | 90.00 | 83.57 | 20.08 |
Electrostatics | 51.71 | 85.00 | 78.60 | 6.45 |
Hydrophobicity | 31.43 | 85.00 | 74.70 | 2.76 |
Combined model (SCOTCH) | 57.14 | 85.00 | 79.64 | 13.37 |
Method . | Sensitivity . | Specificity . | Accuracy . | DOR . |
---|---|---|---|---|
Electrostatics | 40.00 | 95.00 | 84.42 | 15.67 |
Hydrophobicity | 14.00 | 95.00 | 79.42 | 3.85 |
Combined model (SCOTCH) | 47.71 | 95.00 | 85.91 | 17.07 |
Electrostatics | 43.43 | 90.00 | 81.04 | 7.92 |
Hydrophobicity | 24.57 | 90.00 | 77.42 | 3.36 |
Combined model (SCOTCH) | 56.57 | 90.00 | 83.57 | 20.08 |
Electrostatics | 51.71 | 85.00 | 78.60 | 6.45 |
Hydrophobicity | 31.43 | 85.00 | 74.70 | 2.76 |
Combined model (SCOTCH) | 57.14 | 85.00 | 79.64 | 13.37 |
Note: Performance measures were calculated at 95, 90 and 85% specificity.
Method . | Sensitivity . | Specificity . | Accuracy . | DOR . |
---|---|---|---|---|
Electrostatics | 40.00 | 95.00 | 84.42 | 15.67 |
Hydrophobicity | 14.00 | 95.00 | 79.42 | 3.85 |
Combined model (SCOTCH) | 47.71 | 95.00 | 85.91 | 17.07 |
Electrostatics | 43.43 | 90.00 | 81.04 | 7.92 |
Hydrophobicity | 24.57 | 90.00 | 77.42 | 3.36 |
Combined model (SCOTCH) | 56.57 | 90.00 | 83.57 | 20.08 |
Electrostatics | 51.71 | 85.00 | 78.60 | 6.45 |
Hydrophobicity | 31.43 | 85.00 | 74.70 | 2.76 |
Combined model (SCOTCH) | 57.14 | 85.00 | 79.64 | 13.37 |
Method . | Sensitivity . | Specificity . | Accuracy . | DOR . |
---|---|---|---|---|
Electrostatics | 40.00 | 95.00 | 84.42 | 15.67 |
Hydrophobicity | 14.00 | 95.00 | 79.42 | 3.85 |
Combined model (SCOTCH) | 47.71 | 95.00 | 85.91 | 17.07 |
Electrostatics | 43.43 | 90.00 | 81.04 | 7.92 |
Hydrophobicity | 24.57 | 90.00 | 77.42 | 3.36 |
Combined model (SCOTCH) | 56.57 | 90.00 | 83.57 | 20.08 |
Electrostatics | 51.71 | 85.00 | 78.60 | 6.45 |
Hydrophobicity | 31.43 | 85.00 | 74.70 | 2.76 |
Combined model (SCOTCH) | 57.14 | 85.00 | 79.64 | 13.37 |
Note: Performance measures were calculated at 95, 90 and 85% specificity.
3.6 Comparison with other methods
The prediction performance of SCOTCH was compared with existing methods, namely T-CUP (Heider et al., 2014), geno2pheno[coreceptor] (Lengauer et al., 2007), PhenoSeq (Cashin et al., 2015), WebPSSM (Jensen et al., 2003) using all available matrices (i.e. x4r5, sinsi and sinsi c), and the genotypic rules of Raymond et al. (2012) and Esbjörnsson et al. (2010). In Table 2, the results of the comparison are shown. For all existing methods, except WebPSSM with sinsi.c, the sensitivity is less than 20% at a specificity between 93.94 and 99.49%. WebPSSM with sinsi.c only reaches a sensitivity of 37.8% at a specificity of 58.59%. SCOTCH outperforms all existing models in terms of sensitivity and accuracy at comparable specificities. For instance, PhenoSeq reaches a specificity of 94.74%, which is only slightly lower compared to SCOTCH (95%). However, the resulting sensitivity of SCOTCH is 47.71%, which is significantly higher compared to the sensitivity of PhenoSeq, namely 17.7%. The accuracy of PhenoSeq is 54.39%, which is again significantly lower than the accuracy of our new model (85.91%).
Method . | Sensitivity . | Specificity . | Accuracy . |
---|---|---|---|
T-CUP | 18.18 | 96.32 | 55.39 |
geno2pheno | 15.79 | 97.89 | 54.89 |
Phenoseq | 17.70 | 94.74 | 54.39 |
WebPSSM-x4r5 | 15.31 | 93.94 | 53.56 |
WebPSSM-sinsi | 11.54 | 97.98 | 53.69 |
WebPSSM-sinsi.c | 37.80 | 58.59 | 47.91 |
Raymond | 11.00 | 98.48 | 53.56 |
Esbjörnsson | 13.40 | 99.49 | 55.28 |
SCOTCH | 47.71 | 95.00 | 85.91 |
Method . | Sensitivity . | Specificity . | Accuracy . |
---|---|---|---|
T-CUP | 18.18 | 96.32 | 55.39 |
geno2pheno | 15.79 | 97.89 | 54.89 |
Phenoseq | 17.70 | 94.74 | 54.39 |
WebPSSM-x4r5 | 15.31 | 93.94 | 53.56 |
WebPSSM-sinsi | 11.54 | 97.98 | 53.69 |
WebPSSM-sinsi.c | 37.80 | 58.59 | 47.91 |
Raymond | 11.00 | 98.48 | 53.56 |
Esbjörnsson | 13.40 | 99.49 | 55.28 |
SCOTCH | 47.71 | 95.00 | 85.91 |
Method . | Sensitivity . | Specificity . | Accuracy . |
---|---|---|---|
T-CUP | 18.18 | 96.32 | 55.39 |
geno2pheno | 15.79 | 97.89 | 54.89 |
Phenoseq | 17.70 | 94.74 | 54.39 |
WebPSSM-x4r5 | 15.31 | 93.94 | 53.56 |
WebPSSM-sinsi | 11.54 | 97.98 | 53.69 |
WebPSSM-sinsi.c | 37.80 | 58.59 | 47.91 |
Raymond | 11.00 | 98.48 | 53.56 |
Esbjörnsson | 13.40 | 99.49 | 55.28 |
SCOTCH | 47.71 | 95.00 | 85.91 |
Method . | Sensitivity . | Specificity . | Accuracy . |
---|---|---|---|
T-CUP | 18.18 | 96.32 | 55.39 |
geno2pheno | 15.79 | 97.89 | 54.89 |
Phenoseq | 17.70 | 94.74 | 54.39 |
WebPSSM-x4r5 | 15.31 | 93.94 | 53.56 |
WebPSSM-sinsi | 11.54 | 97.98 | 53.69 |
WebPSSM-sinsi.c | 37.80 | 58.59 | 47.91 |
Raymond | 11.00 | 98.48 | 53.56 |
Esbjörnsson | 13.40 | 99.49 | 55.28 |
SCOTCH | 47.71 | 95.00 | 85.91 |
4 Discussion
Almost all existing approaches for the prediction of HIV-1 coreceptor tropism are based on subtype B and have been shown to perform poorly on subtype A, which is responsible for approximately 12% of all HIV-1 infections worldwide. We therefore sought to develop a reliable subtype A specific model. To this end, two RF models were developed, one based on electrostatics and the other one based on a hydrophobicity index. These two models were combined by using a stacking approach. The resulting model SCOTCH shows significantly better performance for subtype A compared to all other methods that have been evaluated. Nevertheless, the sensitivity and accuracy of our new model still did not reach the same levels than those of the prediction methods for subtype B or C. In this study we developed a novel coreceptor tropism prediction algorithm which makes use of sequence and structural information of the V3 loop for subtype A. Combining structural and sequence information improves diversity in ensembles and thus leads to higher prediction performance compared to single models. We could demonstrate that SCOTCH outperforms existing approaches, but there is still room for improvement. Riemenschneider et al. (2016) already proposed that other regions beside the V3 loop, namely the V2 loop, might also be involved in coreceptor binding in subtype A. Involvement of V2 information might improve coreceptor tropism prediction of subtype A sequences in the future. However, although sequencing of the complete gp120 region might be useful in order to improve prediction accuracy for subtype A, it is not really practical due to length restrictions in the sequencing protocols at the moment. Nevertheless, there is little doubt that technological progress will soon make the sequencing of gp120, or even whole viral genomes, feasible for routine diagnostics.
5 Conclusion
Our new model SCOTCH was specifically developed for HIV-1 subtype A and it has been shown in our study to outperform the existing models, which have mostly been trained with subtype B sequences. However, lower prediction performance of SCOTCH compared to subtype B prediction tools implies the existence of a different binding mode of subtype A to the host cell, as already proposed by Riemenschneider et al. (2016). Other regions, in particular the V2 loop, might be also involved in this process. In the future we intend to improve our model by also incorporating other regions of the gp120 protein. However, incorporating longer reads may be impractical due to the distance between the V3 and V2 loops. In order to get V2 and V3 information simultaneously, a region of around 430 nucleotides within the env gene needs to be sequenced.
Acknowledgements
We thank Cedric Staniewski (TUM) for supporting the analyses.
Funding
This work was supported by a grant of the Federal Ministry of Education and Research (BMBF) and the German Academic Exchange Service (DAAD) under the ATN-DAAD Joint Research Cooperation Scheme to DH. DF acknowledges the support of the Russian Science Foundation (grant number 5-14-00026).
Conflict of Interest: none declared.
References