NoD: a Nucleolar localization sequence detector for eukaryotic and viral proteins

Implementation

The NoD web server provides an easy way to predict NoLSs within a protein sequence. NoD predictions are obtained by entering a protein sequence in fasta format on the NoD webserver http://www.compbio.dundee.ac.uk/nod. Protein sequences are encoded as previously described [10]. Briefly, sliding windows of size 13 are sparsely encoded in a binary format using a reduced alphabet of size 12 for submission to an artificial neural network (ANN). The current implementation of NoD uses a local version of Batchman from the Stuttgart Neural Network Simulator [12] and the human-trained NoLS prediction model developed previously [10] to provide the prediction for each encoded subsequence. The Batchman output is then processed and NoLSs are predicted if the average score output by the ANN of 8 consecutive windows is at least 0.8 [10]. Finally, the prediction is displayed as shown in Figure 1 if at least one NoLS is identified. Otherwise, the user is informed that no NoLS is predicted in the input protein. As shown in Figure 1, for proteins predicted to contain NoLS(s), the output consists of 3 sections:

The NoLS window-based score graph can be useful to guide experimental design of nucleolar targeting. The graph gives an overview of the entire protein and shows the proportion of the protein with putative nucleolar targeting capabilities as well as regions of the protein that are near the cut-off threshold and therefore almost predicted as NoLSs.

When entering a protein sequence, the user is provided with the option of also running JPred secondary structure predictions [11] to include as input to the NoLS neural network. If JPred is selected, the accuracy of prediction is slightly higher [10] but the computation time is increased.

For users who want predictions for whole proteomes there is a command line version of NoD called clinod. Clinod produces the same results as a web server but it is more suitable for processing of multiple sequences and is convenient to use within software pipelines.

Finally, for users preferring to run and visualize their predictions locally, there is a virtual appliance version of NoD, which can easily be deployed on a variety of operating systems by a non-specialist user. The virtual appliance version of NoD offers the same functionality as our public server, with the exception of JPred predictions. However, in the near future we intend to release a version which will support JPred.

Prediction of NoLSs in non-human eukaryotes

Because more NoLSs have been reported in human than in all other organisms combined, the NoLS predictor was originally trained and tested only on human sequences [10]. More precisely, as described previously [10], the predictor was trained on a manually curated positive set of 46 human experimentally validated NoLSs and a negative set consisting of several hundred human proteins chosen because they are believed not to localize to the nucleolus. After training, ten of the NoLS predictions were chosen for experimental validation and all were confirmed as positives [10].

As shown in Table 3, mammalian NoLSs and their viral counterparts are well predicted, with sensitivity and positive predictive values ranging from 0.75 to 1.0. This is not surprising because of the close evolutionary distance between humans and other mammals and the constant adaptation of their viruses. Amongst the non-mammalian proteins considered, the Dictyostelium discoideum protein investigated has two NoLSs, one of which is well-predicted. The NoLS that was not correctly identified consists of only 7 amino acids and is likely too short for the predictor to find. The two mollusc NoLSs are entirely well-predicted but low numbers of examples in this group of organisms prevents robust conclusions. Similarly, plant and plant-infecting virus NoLSs are generally well-predicted but also suffer from small numbers of examples. However, the human-trained predictor is entirely incapable of identifying the NoLSs experimentally detected in trypanosomes. This agrees well with experiments in which the NoLS of a Trypanosome brucei protein, ESAG8, was fused to a fluorescent reporter protein and tested for nucleolar localization in human cells. With an intact trypanosome NoLS, the fusion protein was found to be nuclear but not nucleolar in human cells [14]. This observation and our predictions suggest that nucleolar targeting mechanisms differ significantly between humans and trypanosomes and are not interchangeable. Although a larger sample would be needed to confirm this difference, trypanosomal nucleolar targeting mechanisms might represent good drug targets.

While no experimentally generated negative dataset is available for testing the predictor in non-human organisms, we note that the non-NoLS regions of NoLS-containing proteins provide such a set. For example, the human adenovirus 2 protein VII encodes three basic regions at positions 47-54, 93-112 and 127-141 which represent possible nuclear/nucleolar localization sequences [15]. Deletion constructs demonstrate that only the 93-112 segment targets a reporter protein to the nucleolus [15]. This segment matches very closely the NoD NoLS predictions (see Table 2), providing not only an accurate test example but also perfect negative controls (the two other basic regions are not predicted as NoLSs). Thus, the positive predictive values shown in Table 3 give an indication of the false positive rate of prediction in different organisms. However, while some false positives probably represent prediction errors, others might have occurred because NoLSs were not experimentally mapped with enough precision or more than one NoLS exists in the protein. Larger experimental datasets will undoubtedly help to clarify this problem.

Of the 31 eukaryotic and viral NoLSs considered for independent testing, 22 were correctly identified, resulting in an overall sensitivity of 71%. In addition, 6 non-NoLS regions were also identified as positives (and thus are considered here as false positives) yielding an overall positive predictive value of 79%. Finally, of the 27 proteins considered, 6 were predicted to encode NoLSs in regions not experimentally shown to harbour a NoLS resulting in a specificity of 78% (although we note that some of these false positives might represent NoLSs that have yet to be experimentally identified).

Availability and requirements