Histogram comparing the Q3 (black) and SOV (gray) scores of PROTEUS2 (left 3 bars) versus PSIPRED, JNET, PHD, SAM-T02 and GOR.

Bar graph showing the performance of PROTEUS2 in predicting the secondary structure (Q3 score) for 125 proteins that were in neither the training or testing sets.

In order to perform structure predictions most accurately and efficiently PROTEUS2 follows a strict hierarchy of analyses and database comparisons (see the flow chart above). When a query sequence is received PROTEUS2 initially performs a signal peptide identification step using a novel machine learning method based on hidden Markov modeling (HMM) techniques. If a signal peptide (and cleavage site) is identified, the sequence is truncated and sent to the next prediction step. If no signal peptide is identified, the server keeps the query sequence unchanged. In the next step, the query sequence is aligned against a specially constructed database of ~250 solved membrane proteins (with known secondary structure) using BLAST. This membrane protein database, which is updated weekly, was compiled from TMPDB, PDBTM and manual annotation efforts. If a significant (Expect < 1e-10) match is found, the resulting alignment is used to transfer the known secondary structure of the template membrane molecule to the query protein using a technique called 3D-to-2D mapping (Montgomerie et al., 2006). If no match can be found, the program applies TMHMM 2.0 to predict the membrane spanning helical regions. If no membrane helices can be found, the program usesTMB-HUNT to identify potential membrane beta-barrel proteins and uses a jury of secondary structure predictors to identify membrane spanning beta-strands.

After the signal peptides and transmembrane segments (helices or beta-strands) have been identified, PROTEUS2 proceeds to the third structure prediction step. In this third step, PROTEUS2 initially compares the query sequence to a local, non-redundant version of the PDB containing nearly 13,000 sequences of water-soluble (i.e. non-membrane) proteins. The secondary structures for these proteins were assigned using VADAR. If a significant (Expect < 1e-10) match is found, the same 3D-to-2D mapping procedure is used to assign or predict the secondary structure. If no match can be found, the PROTEUS2 applies a locally developed "jury of experts" prediction method to predict the secondary structures as described in our earlier publication (Montgomerie et al., 2006).

Once the signal peptide, membrane spanning regions and secondary structures have been predicted, the query sequence is then directed to a locally developed homology modeling program called HOMODELLER. If a PDB file with >35% sequence identity was previously found in either of the two 3D-to-2D mapping steps, this structure is used as a template for HOMODELLER to build a 3D structure. While PROTEUS2 always succeeds in generating linear (i.e. secondary structure) predictions for any query sequence, 3D structures will only be generated if the query protein passes the HOMODELLER thresholds.

In developing the hidden Markov models for signal peptide prediction program (called PSignal or predictSP), more than 2000 signal peptides and 1200 control peptides (covering the N-terminal sequence to the cleavage site) were obtained from the SignalP data set. These examples were further partitioned into signal peptides belonging to Gram+, Gram- and Eukaryotic organisms. The predictor was trained to recognize not only the existence of a signal peptide, but also its length and cleavage site. PredictSP was subject to 10-fold cross validation during the testing phase.

The transmembrane prediction program builds from two previously developed and freely available programs, TMHMM and TMB-Hunt. TMHMM uses a hidden Markov model to identify transmembrane helices, while TMB-HUNT uses amino acid composition statistics to identify potential membrane beta-barrel proteins. We enhanced the performance of both methods by exploiting the 3D-to-2D mapping methods that we previously developed for predicting the secondary structure of water soluble proteins. To do so we created a database of 250 membrane proteins of known secondary and tertiary structure, with the membrane helices and membrane beta-strands appropriately labeled. As before, secondary structures that were derived from homologous proteins in the membrane protein database were given precedence over the predictions derived from TMHMM or TMB-HUNT. The program's performance was assessed using the TMH-Benchmark server (for transmembrane helices) and a 5X cross-validated subset of 40 transmembrane beta-barrels derived from the PDB.

In developing the "jury of experts" approach for predicting the secondary structure of soluble proteins we used the same programs and methods described previously (Montgomerie et al., 2006). The one change to this method involved editing the PDB-derived secondary structure database so that it contained only water-soluble proteins. This involved removing some 250 membrane proteins from the 13,000 protein data set. In training and testing the program nearly 2000 sequence-unique sequences were analyzed, requiring some 100 hours of CPU time. The program was written such that secondary structure predictions from the consensus predictor could be over-ridden if a homologous protein could be found in PROTEUS2's secondary structure database.

The homology modeling program, HOMODELLER, is quite conventional and employs standard homology modeling techniques. It uses the BLAST to search through a non-redundant version of the PDB database (which is updated weekly) to find and align the closest matching sequence homologue. Mismatched residues in the template sequence are changed to match the query sequence, but with the same x1 angles of the template. Gaps in coil regions are handled using a loop library (consisting of >13,000 loops derived from high resolution structures in the PDB). The inserted regions are superimposed and then iteratively adjusted to fit the surrounding regions using a simplex minimizer working in torsion angle space. The resulting structure is energy minimized using a torsion angle minimizer that uses cyclic coordinate descent in combination with a simple genetic algorithm. The method has been extensively tested and refined using hundreds of known structures. More recently, HOMODELLER was used to model more than 100,000 protein structures for the BacMap project.

Neural Network architecture

The methods and underlying theory to PSIPRED and JNET have been published previously and the programs were used as received without further modification. The TRANSSEC program was developed in-house using a Java-based neural network package known as Joone.

TRANSSEC's underlying approach is relatively simple, consisting of a standard PSI-BLAST search integrated into a two-tiered neural network architecture. The first neural network operates only on the sequence, while the second operates on a 4 x N position-specific scoring matrix consisting of the secondary structure determined via the first network. The first neural net uses a window size of 19, and was trained on sequences from the PROTEUS2-2D database (independent from those used in training the other neural nets). This neural net had a 399-160-20-4 architecture (21 x 19 inputs, 2 hidden layers of 160 and 20, and four outputs). The second neural network used a position-specific scoring matrix, combining evolutionary information from a PSI-BLAST search, and structure information from the first neural network. It was also trained on a set of sequences from the non-redundant database mentioned above, and achieved a Q3 score of 70% and a SOV score of 72%. It used a window size of 9, and was based on a 36-44-4 architecture.

PROTEUS2's globular protein secondary structure method uses neural network mehtod that combines Psipred, JNet, and TRANSEC. The Jury-of-experts program, which combined the results of the three stand-alone secondary structure predictions was also developed using Joone. It consisted of a standard feed-forward network containing a single hidden layer. Using a window size of 15, the structure annotations and confidence scores from each of the three methods (JNET, PSIPRED, and TRANSSEC) were used as input.

Neural Network Training

The neural net was trained on 100 sequences chosen randomly from the non-redundant database mentioned above. Four output nodes were used, one for each of helix, strand or coil, as well as a fourth denoting the beginning and end of the sequence. A back-propagation training procedure was applied to optimize the network weights. A momentum term of 0.2 and a learning rate of 0.3 were used, and a second test set of 20 proteins was applied at the end of each epoch, to ensure that the network was trained for the most optimal number of iterations.

Merging the homologous sequence with a prediction

If a homologous sequence was found whose sequence did not cover the entire length of the query sequence, then it was necessary to merge the homologous structure with that of the predicted structure. Though some secondary structure is the result of global interactions (hydrogen bonding between different parts of the sequence), homologous regions likely have similar local structure. Given that a homologous structure is likely to be more correct than the predicted structure, the homologous structure is mapped directly onto the predicted structure, replacing the states of the predicted structure with the homologous structure.