This has motivated the development of numerous computational methods for predicting kinase–substrate relations, for example, Scansite (4), NetphosK (5,6), predikin (7), PredPhospho (8) GPS (9), PPSP (10) and KinasePhos (11). These methods all rely on consensus sequence motifs recognized by the active site of the enzymes, represented by either position-specific scoring matrices (PSSMs), neural networks, support vector machines or other machine-learning representations. However, kinase specificity is known to also depend on other factors, such as auxiliary protein interactions, scaffolds, coexpression and colocalization (collectively referred to as ‘context’). We recently introduced a computational framework, NetworKIN, which uses a probabilistic protein association network [string (12)] to model the context of kinases and substrates; combined with consensus sequence motifs, this gave a 2.5-fold leap in prediction accuracy over previous methods (13).

The NetworKIN algorithm is designed to work from a set of experimentally identified in vivo phosphorylation sites (although the algorithm can also be used ab initio). The precomputed results in the database are based on the latest human phosphoproteome from the Phospho.ELM and PhosphoSite databases (2,14) (Figure 1). The release cycle of the database is approximately every 3 months due to the high throughput of mass-spectrometry-driven proteomics, and we intend to keep NetworKIN up-to-date with future releases of Phospho.ELM and PhosphoSite.

These data are processed through the NetworKIN algorithm, which is implemented in Python and C. The sites are classified by matching them to a motif collection (Figure 1) based on the position-specific scoring matrices from Scansite (4) (http://scansite.mit.edu) and the artificial neural networks from NetPhosK (5) (http://www.cbs.dtu.dk/services/NetPhosK). Each consensus sequence motif is considered to be a representative for a family of closely related kinases; for example, the NetPhosK Cdk5 predictor is used for predicting possible phosphorylation sites for all cyclin-dependent kinases. Within a proteome, kinases are identified and assigned to these families based on their best hit in a blastp (15) sequence similarity search against a set of 82 representative kinase domain sequences, which have been manually assigned to families. Only hits with an E-value better than 10^{− 40} and with at least 50% sequence identity are considered.

To capture the biological context of a substrate, we use a probabilistic network of functional associations extracted from the string database (12) (http://string.embl.de, Figure 1). This network is based on four fundamentally different types of evidence: genomic context (gene fusion, gene neighbourhood and phylogentic profiles), primary experimental evidence (physical protein interactions and gene co-expression), manually curated pathway databases, and automatic literature mining. We showed that the three latter evidence types are of comparable importance, whereas genomic context methods contribute very little towards the predictions made by NetworKIN (13). As the curated pathway databases generally contain few errors, a confidence score of 0.9 is assigned to this type of evidence. The best candidate kinases within the appropriate kinase families are identified from a protein network of functional associations [generated using the string database (12)] by calculating the proximity to the substrate for all kinases, defined as the probability of the most probable path connecting them (Floyd–Warshall algorithm). The context is thus used as a filter that eliminates many of the false-positive predictions obtained from the sequence motifs and hence improves the prediction accuracy. However, the current algorithm is unable to recover sites that are missed by the sequence motifs (i.e. false-negative predictions).

For all other users, the primary entry point to NetworKIN is its search interface shown in Figure 2A. The user can select a specific substrate and/or kinase to view the corresponding subset of predictions; in our example, we query for 53BP1 as the substrate and use the wildcard * to obtain predictions for all kinases. The web interface also offers an advanced search form, which enables the user to pose much more refined queries. In either case, the search results will be presented as a table in which each row shows a predicted relation between a kinase and a specific phosphorylation site in a substrate. In case of 53BP1, we get a list of 78 predictions for 39 sites and 12 kinases; the first 10 of these predictions are shown in Figure 2B. For each prediction we list two scores, namely the context score and the motif score, both of which should preferably be high. It should be noted that the motif scores for different kinase families are not comparable; in particular, motif scores from NetPhosK should not be compared with motif scores from Scansite. For this reason, the predictions for a given phosphorylation site are sorted by their context score. As the results of a single query may be extensive, the results can also be downloaded in the formats mentioned previously.

Furthermore, the user can investigate the predictions in greater detail via the web interface. For each substrate, we link to Phospho.ELM or PhosphoSite where the user can find manually curated information on in vivo phosphorylation sites including, when known, the kinase(s) involved (Figure 2C). To allow the user to investigate how a specific prediction was made by NetworKIN, we provide a link to the string network viewer, in which the most probable path connecting the kinase and the substrate will be highlighted (Figure 2D). Alternatively, the user can select multiple predictions and display the network context for all the proteins involved. From the network viewer, the evidence underlying each individual association can be inspected in further detail. This ability to thoroughly investigate individual predictions is particularly useful for interpreting non-obvious cases, which are often based on indirect links between the kinase and the substrate.

Although Phospho.ELM, PhosphoSite and hence NetworKIN are kept up-to-date with new published phosphorylation sites, many researchers will be interested in predictions for their own, unpublished sites. We thus allow users to submit protein sequences and a corresponding set of phosphorylation sites for analysis; although possible, we discourage submitting sequences without prior knowledge on phosphorylation. After uploading the data, the user will be presented with a confirmation page where potential data entry errors can be detected and fixed. The final predictions will be presented in a tabular format similar to the one used when querying the precomputed results in the database.