Various overlapping symmetrical sequence patterns of residues length 21 that included central glycosylated residues were constructed according to previous studies [14], [15]. A sequence pattern was considered positive if central residue was glycosylated otherwise the same was assigned as a negative pattern. To generate a pattern corresponding to the terminal residues in a protein sequence of length L, dummy residues “X” in number (L-1)/2 were added at both the termini of the protein [29], [30].

The SVM is a supervised machine-learning technique based on the structural risk minimization principle [34]. In this study, we have used freely available SVM^light classifier v 6.01 (http://svmlight.joachims.org/) where we could adjust the parameters and kernel (linear, polynomial, radial basis function, sigmoid) functions. The advantage of SVM over other machine learning techniques is that it can be trained on small dataset (as in this study) with minimum over-optimization. SVM based approach has been successfully employed in developing both N- and O-glycosylation prediction tools for mammalian glycoproteins in past [13], [14]. Using different sequence properties like identity and position of residues (BPP), percentage composition of residues (CPP), residue conservation information (PSSM) along with structural features like secondary structure and surface accessibility several SVM classifiers have been trained and optimized for this study. Our group has successfully used one or more of these features in predicting GTP interacting residues, Mannose interacting residues, in predicting Cyclin protein sequences and in identification of conformational B-cell Epitopes from primary sequences of proteins, previously [29], [30], [35], [36]. In this study, a 5-fold cross-validation procedure has been used to develop the prediction model, where five subsets were constructed randomly from the main datasets. At a given point of time, the models were trained on four sets of the training dataset and the performance was measured on the remaining fifth set. This process is repeated five times in such a way that each set was used once for testing. The final performance was obtained by averaging the performances of all five sets. The models thus obtained were evaluated for performance using threshold dependent parameters namely, sensitivity (Sn), Specificity (Sp), Accuracy (Acc), Matthews correlation coefficient (MCC) as well as using threshold dependent parameters Area Under Curve (AUC) values.

Evaluation parameters employed in this study are described briefly as below:

Sensitivity is the percentage of glycosites that are correctly predicted as glycosylated:

Specificity is the percentage of non-glycosylated sites that are correctly predicted as non-glycosylated:

Accuracy is the percentage of correct prediction out of total number of predictions:

Matthews correlation coefficient (MCC) is a measure of both sensitivity and specificity. MCC value would range from 0 (indicating completely random prediction) to 1 (indicating perfect prediction):[Where TP- true positive; FN- false negative; TN- true negative; FP- false positive]

Threshold selection is important criteria for checking the consistency of prediction results. In our study, we have varied threshold in the range of –1 to +1, normally we selected “0” as default threshold to achieve balance between sensitivity and specificity.

Area Under Curve (AUC) a threshold independent parameter describes inherent trade-off between sensitivity and specificity. Receiver Operating Characteristic (ROC) plots were drawn between TP rate (sensitivity) and FP rate (1-specificity) using R-package v 2.14.1 (http://www.r-project.org/) to calculate AUC values. Finally, the best performing models in terms of accuracy & MCC values were validated using an independent dataset of prokaryotic glycoproteins for final implementation at GlycoPP webserver (Figure 1).

In order to evaluate the suitability of models trained on eukaryotic glycoproteins for predicting glycosites in prokaryotic proteins, the proteins of main datasets were run on three of the well-known prediction tools for prediction of N- and O-glycosites. Against the experimentally validated glycoproteins of prokaryotes, the performances of these tools were found very poor and are detailed in Table 1. From this, we conclude that the methods that are trained using eukaryotic glycoprotein are not optimum for prediction of potential glycosites in bacterial and archaeal proteins. This also suggests that the sequence or structural contexts around prokaryotic glycosites could be different from what is known in eukaryotic glycosites. This is logical as several OSTs with novel mechanisms of sugar transfer on to the acceptor proteins are now known in bacteria as well as archaea. This prompted us to develop a number of new algorithms to recognize and differentiate glycosylated and unglycosylated sequence contexts of known glycosites of archaeal and bacterial proteins representing aforementioned six different phyla. These algorithms are trained using different input features and described in this study.

In an attempt to understand the general preferences for different amino acids around prokaryotic glycosites as well as the differences from the corresponding sequences in eukaryotes, we have generated a number of one sample and two sample weblogos (http://weblogo.berkeley.edu/& http://www.twosamplelogo.org/) for N- and O-glycosites of archaeal and bacterial glycoproteins in an organism specific, phylum specifc as well as domain specific manner, respectively. The interesting existing knowledge as well as our statistically significant observations for the purposes of a prediction model are discussed here, briefly. Similar to eukaryotic glycoproteins, the minimal sequon NX(S/T)(where X≠P) is essential for N-glycosylation in prokaryotic glycoproteins. For example in all archaeal glycoproteins (Figure S1), [25], in HmcA protein of Desulfovibrio [37], adhesin protein HMW1 of Haemophilus influenzae and Actinobacillus pleuropneumoniae (where glycosylation is sequential and mediated by a novel cytoplasmic glycosyltransferase, HMW1C of family GT41, Figure S2), [38]. However, as known already, the sequon is extended as (D/E)X₁NX(S/T)(where X₁ & X≠P) but not stringent in case of PglB (OST of Campylobacter) mediated en bloc N-glycosylation in Campylobacter and Helicobacter (Figure S2), [39]. As discussed before, the first-ever defined sequon D(S/T)(A/I/L/V/M/T) for O-glycosites is indeed conserved across available glycoproteins from three representative classes including Bacteroidia, Flavobacteria and Sphingobacteria of phylum Bacteroidetes (Figure S3). Further, the two-sample logos comparing prokaryotic and eukaryotic N-and O- glycosites clearly illustrate the differences in the amino acid preferences around these glycosites (Figure 2), indicating a necessity for independent prediction tool for prokaryotes. With respect to glycosylated Asn (if at position 0) the positions at -1 and -2 have previously been stated to be enriched in aromatic amino acids in eukaryotic N-glycosites [25]–[27]. However, in prokaryotic N-glycosites instead we observe a marked preference for polar residues like Asp/Glu/Thr/Asn and lysine at different positions preceding glycosylated Asn (Logo C, Figure 2). Similarly, at positions -2 and -6 occurrence of polar residues is higher around NX(S/T) motif in validated N-glycosites of prokaryotes in contrast to randomly selected equal number of NX(S/T) motifs with unglycosylated Asn from prokaryotic glycoproteins (Logo A, Figure 2). An analysis of eukaryotic N-glycosites by Pertescu and co-workers had suggested a preference for small hydrophobic residue at positions +1 and large hydrophobic residue at +3 in eukaryotes, previously [27]. Similarly, in case of prokaryotic glycoproteins hydrophobic residues are though present at +1 position yet preference for large or small residues are not very clear (Figure 2, Figure S1), [25]. Furthermore, increased instances of Pro near the glycosylated residues are not observed in bacterial and archaeal glycoproteins as found in eukaryotic glycoproteins. Instead Pro is one of the significantly depleted amino acids at +4 and +5 positions here [27]. Likewise, sequence surrounding all prokaryotic O-glycosites (Figure 2) is different in having higher instances of Gly, Ala, Val and a significant depletion of Pro at almost all positions (except in mannosylated glycoproteins of Mycobacterium spp, Figure S2), [8] in comparison to the eukaryotic mucin type O-glycosites that are rich in Ser, Ala and Pro (Logo D, Figure 2), [12]. In prokaryotic O-glycosites, apart from this general presence of small hydrophobic amino acids around 10 residues on either sides of glycosylated Ser/Thr residues (at position 0), a marked preference for negatively charged Asp at -1 that in fact is a part of potential sequon for O-glycosites in Bacteroidetes is observed (Logo B, Figure 2).

Previous statistical analysis of all available crystal structures of eukaryotic glycoproteins by Petrescu et al. had suggested that the probability of finding N-glycosites was higher at positions where there was a secondary structure change [27]. Upon analysis of 12 eukaryotic glycoproteins, Julenius et al had also concluded that O-glycosites mainly occurred in coil region of mucin type of O-glycosylated proteins [12]. The secondary structure and surface accessibility of a residue therefore are considered important criteria in prediction of glycosites in eukaryotes. Some of the existing eukaryotic glycosite prediction models have employed these features successfully [11], [12]. Unlike eukaryotic N-glycosylation that is a co-translational event, the glycosylation is considered a true post-translational modification in bacteria where the folding state of a polypeptide/protein could dictate availability of a sequon/site for glycan attachment on to a protein [21]. Although limited, yet most of the X-ray crystal structures and NMR structures of bacterial glycoproteins (as listed at http://www.proglycprot.org/CrystalStructure.aspx) show that the glycosylated residues are indeed primarily located in surface-exposed flexible loops/turns/bends that then should be accessible to bacterial OSTs/GTs. The structural contexts for 20 glycosites (13 N- & 7 O-glycosites) extracted from available structures of N- and O-glycoproteins of Archaea and Bacteria, reveal that at least 65% (13 out of 20) of these glycosites are located in aforementioned flexible regions and primarily in intra-domain region (Table 3). Incidentally, at least three of the N-glycosylated proteins namely, PotD of Escherichia coli, AcrA and PEB3 of Campylobacter jejuni are glycosylated (in vitro/in vivo) by OST of Campylobacter (PglB) that has previously been shown to transfer sugars post-translationally to locally flexible structures in folded proteins [21], [28], [40]. Similarly, glycosylated Ser/Thr residues in Endo-β-N-acetylglucosaminidase F3 (Flavobacterium meningosepticum), Chondroitinase-AC and Chondroitinase-B (Pedobacter heparinus) lie in the similar loops/bends in their respective crystal structures (Table 3).

Our analysis of the predicted secondary structure indicates that 55.92% of the validated glycosites are found in coils, 15.51% in helix and 28.57% in sheets whereas non-validated glycosites or their sequence contexts are found correspondingly less in coil (47.23%), more in helix (24.42%) and almost equally in sheets (28.35%). Similarly, 17.36% of validated O-glycosites are situated in helix, 62.63% in coils and and 20.4% in sheets in contrast to non-validated O-glycosites that are found more often in helix (22.99%) and less in coils (53.98) and almost equally in sheets (23.03), respectively (Figure 3). Similarly, predicted surface accessibility profile of glycosylated Asn residues suggest them to be much more surface accessible than the corresponding non-glycosylated sequence contexts as shown in Figure 4. The glycosylated Ser/Thr are again, more accessible (80%) compared to the non-glycosylated residues (60%). To summarize, most of the prokaryotic glycosites (both N as well as O) indeed seems to be present in flexible and exposed regions. Further, not only the central glycosylated-residues but also their surrounding residues are highly accessible and surface exposed.

SVM models based on BPP, CPP and PSSM profiles are well recognized for their notable performances in predicting a variety of motifs and interactions in biomolecules and have been used effectively in the past for glycosites predictions as well [29]–[31], [35], [36]. Accordingly, we have generated several SVM models using BPP, CPP and PPP profiles as input features. The performance measures were calculated at different thresholds of SVM scores ranging from −1.0 to 1.0 and the best performing thresholds were selected for further optimization. The prediction of the N- glycosites were best achieved by the SVM models developed using BPP profile achieving 79.91% accuracy and 0.60 MCC (Table 4) whereas O-glycosites were best predicted by SVM model developed using PPP with 74.57% accuracy and 0.49 MCC (Table 5). As discussed before, sequence features namely secondary structure (SS) and accessible surface area (ASA) could play an important role in correct predictions of sites of glycoslation in a protein. Therefore, we have developed prediction models with these features in following three combinations: (i) composition profile of patterns with either secondary structure or surface accessibility or both (ii) Binary profile of patterns with either secondary structure or surface accessibility or both (iii) PPP with either secondary structure or surface accessibility or both. In general, inclusion of SS and SAS profiles in prediction models helped improvise predictions (Table 4, Table 5, Figure S4). The hybrid model of BPP+ASA proved as good as BPP+SS+ASA improving the maximum MCC of prediction from 0.60 to 0.65 and accuracy of prediction from 79.91% to 82.24% for N-glycosites (Table 4). Similarly, predictions of O-glycosites were improvised slightly using the hybrid model (based on combination of PPP+SS+ASA profiles) giving MCC of 0.48 and accuracy value of 71.73% in comparison with PPP alone derived 73.28% accuracy and 0.47 MCC (Table 5).

For any machine learning technique, learning of datasets is very easy when both positive and negative instances are equal in number. Nevertheless, in case of glycoproteins, the negative instances could be much more than the positive instances in a protein sequence. Therefore, in order to judge accuracy and applicability of our SVM prediction schemes on realistic datasets of users, in parallel we have calculated the performances of aforementioned SVM models using realistic datasets. As was seen in case of models optimized with balanced datasets, BPP based SVM models performed better in case of realistic datasets and could achieve a maximum MCC 0.48 and 0.51 with accuracy value of 82.03% and 86.39% for prediction of N-glycosites using solo feature based and hybrid models (BPP+ASA), respectively. Similarly, O-glycosites could also be predicted with reasonably high accuracy of 70.24% and 89.69% (with corresponding maximum MCC values of 0.19 and 0.50) using CPP and CPP+ASA based models, respectively. Surprisingly, while using realistic datasets, predictions for O-glycosites were better with CPP based models in contrast to PPP based models that fared well in case of balanced datasets (Table 5, Table S1, Figure S4). Infact inclusion of surface accessibility features in combination with CPP in O-glycosites prediction scheme could enhance maximum MCC value of prediction by 2.5 fold (Table S1, Figure S4) indicating that surface accessibility alone could indeed be a useful criterion in glycosites prediction models discussed here. Further, the observed poorer performance of SVM with realistic datasets than with the balanced datasets of course is due to the inherent learning biases of realistic datasets. However, overall our SVM models optimized with realistic datasets fared reasonably well in predicting both N- and O- glycosites from realistic datasets (Table S1).

Finally, the performance of best-optimized models as discussed before (Table 4, Table 5) were evaluated and compared with performances of NetOGlyc v 3.0, NetNGlyc, EnsembleGly against an independent set of experimentally verified prokaryotic glycoproteins (Table 6). Models developed and discussed in this study not only provided reasonably high accuracy (86.84% for N-glycosites with maximum MCC of 0.74 and 76.23% for O-glycosites with maximum MCC value of 0.53, respectively) but have convincingly outperformed performances of at least three of the well-known existing glycosites prediction tools as detailed in Table 6 in the context of prokaryotic glycosites prediction.

The overall best performing models described in Table 6 are implemented in the form of a web-server GlycoPP available freely at http://www.imtech.res.in/raghava/glycopp/. The common gateway interface of GlycoPP is written using CGI/PERL script. This server allows for prediction of N- and O-glycosites in prokaryotic protein sequences. Predictions can be performed by the users at any of the user-defined thresholds ranging from −1.0 to 1.0 for optimizing SVM scores. Input is acceptable as single or multiple sequences in standard FASTA format.

The various datasets used in this study are available in downloadable format at http://www.imtech.res.in/raghava/glycopp/suppli.html.