Table 1.
Publicly available fold recognition servers
| Code | |
|---|---|
| Sequence-only methods (no structural information required) | |
| PDBb | PDB-BLAST is based on the PSI-BLAST (21) program. PSI-BLAST is iterated five times on the non-redundant protein sequence database clustered at 70% identity and masked with low complexity filters. Before the fifth iteration, the sequence profile is saved and used as query against sequences of proteins with known structures (from PDB). This server is a default reference most fold recognition servers are compared with. |
| FFAS FFA3 | FFAS (45) is a profile–profile comparison method. Profiles are generated for protein families in a different way than in PSI-BLAST, but PSI-BLAST is used to collect the sequences of the families. The old version FFAS is now obsolete and replaced with the new version FFAS-03, which uses vector times matrix times vector multiplication when aligning two positions and improved transformation of raw alignment scores into Z-scores. FFAS is one of the first profile–profile comparison servers. |
| ORFs ORF2 | ORFeus (51) is a meta-profile with meta-profile comparison method (meta profiles include sequence profiles and predicted secondary structure). It uses vector times vector multiplication. The old version returns the raw alignment score, while the new version ORFeus-2 translates the score into a Z-score. |
| mBAS BasD BasP | Meta-BASIC (mBAS) (69) is a local meta predictor, which uses six different versions of meta-profile alignment methods, including two versions of ORFeus. Distal-BASIC (BasD) uses two versions of low stringency meta profiles (five PSI-BLAST iterations) aligned with vector times vector and vector times matrix times vector multiplication. Proximal-BASIC (BasP) uses high stringency meta profiles (only three PSI-BLAST iterations). The strongest asset of these algorithms is their high specificity. |
| ST99 | Sam-T99 (44) builds a multiple alignment (the SAM-T99 alignment) by iterated search using HMMs. It uses the alignment to predict secondary structure (with various methods) and to build an HMM for searching PDB for similar proteins. Also, a library of HMMs built by similar methods from PDB sequences is used to score the target sequence. This server has a long tradition and was one of the best servers in CAFASP-1. |
| SFAM SFPP | SUPERFAMILY (102) is a library of HMMs based on SCOP. The server uses HMMs and the SAM methodology as does Sam-T99. SUPFAM_PP is the next generation of SUPERFAMILY. Both servers are capable of generating hybrid models using partial alignments to various templates. The top 10 generated models are sometimes quite redundant. |
| FRT1 | FORTE-1 (103) is a profile–profile comparison method. The correlation coefficient is used as similarity measure of two aligned profile positions. The profiles are generated using PSI-BLAST. |
| Hybrid methods (use structural information of the template) | |
| ST02 | SAM-T2K (104) iterated search procedure is used to create a multiple alignment of homologs. Templates are aligned with three different target HMMs (using different secondary structure predictions and also no secondary structure prediction at all) and the target is aligned with template HMMs. Many alignments are made and the top five distinctly different ones are reported. This server has a higher accuracy than Sam-T99. |
| 3DPS | 3D-PSSM (105) is based on a hybrid threading approach using 1D and 3D sequence profiles coupled with secondary structure prediction and solvation potential. 3D-PSSM is one of the first fold recognition servers. It was rated as very sensitive in LiveBench-2. |
| GETH MGTH | GenTHREADER (GETH) (106) uses a combination of various methods, including sequence alignment with structure-based scoring functions as well as a neural network-based jury system to calculate the final score for the alignment. mGenTHREADER (MGTH) is an enhanced version of GenTHREADER. It takes as input a PSI-BLAST profile calculated for the target sequence. Both versions took part in the first CAFASP evaluation and have a long history. mGenTHREADER was rated as very specific in LiveBench-2. |
| FUG2 FUG3 | In FUGUE (107), environment-specific substitution tables were derived from the structure-based alignments in the HOMSTRAD database. Each alignment in HOMSTRAD was converted into a scoring template (profile) using the environment-specific substitution tables with environment-dependent gap penalties and enhanced by homologous sequences. FUGUE takes a sequence or sequence alignment and searches against the library of profiles. FUGUE is a relatively new server. |
| RAPT | RAPTOR (108) uses a threading technique for fold recognition. It minimizes an energy function consisting of mutation, singleton, pair-wise and secondary structure terms. The method is formulated as a large-scale integer programming problem. Support Vector Machine technique is used to assess the alignment reliability. RAPTOR is quite new and was very successful in CASP-5. |
| SPKS | SPARKS (Sequence, secondary structure Profiles And Residue-level Knowledge-based Score for fold recognition) (109) uses single-body residue-level knowledge-based energy score combined with sequence profile and secondary structure information for fold recognition. |
| PRO2 | PROSPECT (PROtein Structure Prediction and Evaluation Computer Toolkit) (110) is a threading-based protein structure prediction system. The system uses the following terms: mutation energy (including position-specific score matrix derived from multiple sequence alignments), singleton energy (including matching scores to the predicted secondary structures), pairwise contact potential (distance dependent or independent) and alignment gap penalties. |
| INBG | INBGU (111) is a combination of five methods, which exploit sequence and structure information in different ways and produces one consensus prediction of the five. It uses predicted versus observed secondary structure and sequence profiles for both the target and for the folds in the library. The precursor of INBGU (frsvr) was one of the best performing servers in CAFASP-1. |
| SHGU | ShotGun-INBGU (68) uses the ShotGun consensus layer to create alternative consensus models from the INBGU components. The server is much more accurate than INBGU. It uses only in-house components, but it is almost as accurate as structure meta predictors. |
| Structure meta predictors (build consensus form other servers) | |
| PCO2 PCO3 PCO4 PCO5 PMOD PMO3 PMO4 | Pcons (37) comes in various versions and uses various sets of component servers to generate consensus predictions. The largest set includes SUPFAM_PP, FFAS-03, FFAS, SAM-T2K, FUGUE-3, PROSPECT, mGenTHREADER, INBGU, 3D-PSSM, ORFeus, FORTE-1 and PDB-BLAST. Pcons (PCO…) returns one of the models obtained from component servers while Pmodeller (PMO…) runs Modeller (65) using the alignments collected from the servers. Pcons is the first automated structure meta predictor and receives very good scores since LiveBench-2. |
| 3DS3 3DS5 | ShotGun (68) is a consensus predictor which utilizes the results of FFAS-03, 3D-PSSM and INBGU (3DS3) and in the larger version also FUGUE and mGenTHREADER (3DS5). It compiles a hybrid model from the models produced by the component servers by combining partial structures. The generated structures are sometimes unphysical but the server has very high sensitivity and specificity (reliability estimation). |
| 3JAa 3JBa 3JCa 3JA1 3JB1 3JC1 | 3D-Jury (64) is an interactive meta predictor. The user can select the set of servers used for consensus building. 3D-Jury can also include other meta predictors making this server a ‘meta–meta predictor’ (the 3JB and 3JC version). It can operate in single model (one model per server, suffix ‘1’) or multiple model (suffix ‘a’) modes. The default 3JA1 version uses 8 component servers and the single model mode. |
| Ab initio meta predictors (use meta predictors and ab initio modules) | |
| RBTA | Robetta (66) produces full chain models with the Rosetta de novo and comparative modeling methods. De novo models are built by fragment insertion simulated annealing. Comparative models are built by detecting a parental PDB structure with PSI-BLAST or Pcons2, aligning the query to the template with the K*SYNC alignment method, and modeling variable regions with a modified version of the de novo protocol. Robetta is one of the best performing servers as evaluated in CASP-5. |
| PRCM | PROTINFO-CM (67) uses 3D-Jury as initial alignment provider. Initial models are then built for each alignment and scored. Loops and side-chains are built on the best scoring models using a frozen approximation. The relatively slow method (not available publicly at present) does a sophisticated graph–theory search to mix and match between various main-chain and side-chain conformations. Good results have been obtained in LiveBench-7 but due to limited computational resources it was withdrawn from LiveBench-8. |
The table provides a short description of selected, publicly available servers that took part in LiveBench-7 or LiveBench-8 and gives a short description of the underlying algorithm. The new ab initio meta predictors are quite slow and because of this not yet available for public use. Their additional weakness is the lack of a confidence score. Sequence-only methods neglect by definition any information about the structure of the template and in contrast to hybrid methods can be used as general homology inference methods between any protein families. Structure meta predictors offer currently the highest utility by producing accurate models with reliable confidence assessment.