PEffect - prediction of bacterial type III effector proteins
The type III secretion system mediates a wide range of bacterial infections in human, animals and plants. This system comprises a hollow needle-like structure localized on the surface of bacterial cells that injects specific bacterial proteins, effectors, directly into the cytoplasm of a host cell. During infection, effectors act in concert to convert host resources to their advantage and promote pathogenicity.
We -Tatyana Goldberg, Burkhard Rost and Yana Bromberg- at BrombergLab and RostLab developed a novel method, pEffect that predicts bacterial type III effector proteins. In our method, we combine sequence-based homology searches and advanced machine learning to accurately predict effector proteins. We use information encoded in the entire protein sequence for a prediction.
pEffect requires a .fasta file and a .profile file for each sequence to be predicted as input. A profile file can be obtained by e.g. using PSI-BLAST (Position-Specific Iterated BLAST). We built our profiles using an 80% non-redundant database combining SWISS-PROT, TrEMBL and PDB.
Example input files can be viewed here.
In order to assign a localization class to a query protein Support Vector Machines (SVMs; implemented as internal nodes in LocTree2 decision trees) take a protein in its input space (as a sequence-profile tuple) and map it to a higher dimensional feature space where it is represented as a feature vector. The vector of a query protein is then 'compared' to the vectors of proteins used for the training. The predicted localization class of a query protein is then the class of the most 'similar' vector of a protein used for training.
In our case, the mapping and the comparison are carried out by a string kernel function, called the Profile Kernel. A feature vector built by the The Profile Kernel is indexed by all possible subsequences of length k from the alphabet of 20 amino acids. Each element in the vector represents one particular k-mer with a score below a user-defined threshold sigma. This score is calculated as the ungapped cumulative substitution score in the corresponding sequence profile. The similarity between training and a test protein is calculated as the dot product between their k-mer vector representations and is given as a single positive integer number.
In short, the Profile Kernel identifies sets of k-mers (stretches of k adjacent residues) that are most informative for the prediction of localization and then matches these in a query protein.
LocTree2 combines three different systems of decision trees, one for each domain of life (Figure 1). The trees were built by incorporating a hierarchical ontology of localization classes modeled onto the biological sorting mechanism in that domain. In eukaryotes pathways for membrane and non-membrane proteins are treated separately. The branches represent paths of the protein sorting, the leaves (rectangles) the final prediction of one localization class, and the internal nodes (circles) are the decision points along the path.
As in LocTree, biological similarities were incorporated from the description of cellular components provided by the Gene Ontology Consortium (GO). In cases of ambiguous relations (e.g. PER, MIT, CHL) we explored different trees in which these classes were placed at different levels in the hierarchy and selected the hierarchy with the highest prediction performance. LocTree2 was extremely successful at learning evolutionary similarities among subcellular localization classes and was significantly more accurate than other traditional networks at predicting localization.
Hierarchy of SVMs
The decision points along the path in the hierarchical trees were implemented as binary Support Vector Machines (SVMs). As the SVM model, we chose the WEKA version of Sequential Minimal Optimization. Each SVM was trained on a different set of proteins. For example, the SVM at the root node in the archaeal tree (Fig. 1a) was trained on the full set of proteins (comprising cytoplasmic and non-cytoplasmic classes), while the SVM at a lower level in the tree was trained on plasma-membrane and extra-cellular proteins only.
In addition to the predicted localization class LocTree2 provides a Reliability Index (RI) measuring the strength of a prediction. For a predicted class (leaf node) the RI is compiled as the product over the reliabilities of all parental nodes. The RI is a value between 0 and 100, with 100 denoting the most confident predictions.
We rigorously evaluated the reliability of LocTree2 predictions on a non-redundant test set of proteins (Fig. 2). We observed that 50% of proteins with the highest reliability reached levels of overall accuracy Q6=98% for Bacteria (gray arrow) and Q18=92% for Eukaryota (black arrow). To pick another point, almost 40% of all eukaryotic proteins were predicted at RI greater than 85; for these, Q18 was above 95%. Thus, two in the top 40 predictions in 100 were wrong in one of 18 states (e.g. nuclear instead of nuclear membrane).
- Q6 is six-state accuracy for predicting localization to six classes
- Q18 is eighteen-state accuracy
Accuracy of localization prediction
We evaluated the performance of LocTree2 in a stratified five-fold cross-validation using non-redundant test sets of archaeal, bacterial and eukaryotic proteins. In each fold we learned a new classification model with four training splits and tested on the fifth. In doing so, we never used any information of the test split during a training phase.
LocTree2 achieved a sustained level of 65% accuracy for predicting eighteen localization classes for eukaryotes. The first decision that is made for eukaryotic proteins is whether it is a membrane-spanning protein or not. This decision was correct for 94% of all proteins. This performance is comparable to what best methods, designed explicitly for the prediction of integral membrane helices, achieve. The most accurately predicted class for eukaryotes was extra-cellular, followed by nucleus. Overall, LocTree2 tended to predict membrane-bound classes better than the corresponding non-membrane bound classes (e.g. MITM vs. MIT).
LocTree2 performed extremely well also for prokaryotes, predicting six classes at 84% accuracy for Bacteria and three classes at 100% accuracy for Archaea. We assume, that 100% is an over-estimate of the performance for Archaea due to the limited data we had. For Bacteria, the most accurate predictions were made for plasma membrane followed by cytosolic proteins.
We rigorously benchmarked LocTree2 in comparison to the best alternative methods for localization prediction. LocTree2 outperformed all other methods in nearly all benchmarks. We could show on a few examples that LocTree2 may discover annotation mistakes of high-throughput experiments. Finally, we suggest using our tool for large-scale genome projects as it proved to sustain high levels of performance and to surpass its competitors even for protein fragments.
Additional information can be found in the LocTree2 manuscript.
The runtime was measured on a Dell M605 machine with a Six-Core AMD Opteron processor (2.4 GHz, 6MB and 75W ACP) running on Linux.
|1 Sequence||100 Sequences||500 Sequences||1000 Sequences||3000 Sequences||5000 Sequences||10000 Sequences|
- The program can be accessed online via the PredictProtein service or LocTree2 server
- Standalone version can be downloaded as a Debian package here
Data sets used for development and evaluation of LocTree2 can be accessed here.
LocTree2 predicts localization for all domains of life
Tatyana Goldberg; Tobias Hamp; Burkhard Rost
For questions, please contact firstname.lastname@example.org