Table I.
Classification and short description of CASP11 QA methods. (For a more detailed description please consult CASP11 Methods Abstracts - http://predictioncenter.org/casp11/doc/CASP11_Abstracts.pdf).
| Method | C/S | L/G | Short description |
|---|---|---|---|
| BITS | S* | G | Structural quality of predicted binding sites; if no binding site is identified - global structural comparison with a model predicted using multiple templates; if no templates - scoring with a knowledge-based potential. |
| ConsMQAPsingle | S*M | G | 0.8*MQAPsingleA + 0.1*MQAPSingleC + 0.1*MetaMQAP. |
| DAVIS-QAconsensus | C | LG | Average GDT-TS score to all other models in a decoy set with subsequent target length normalization. |
| FUSION | S | LG | Probability of the torsional angles against an Input-Output Hidden Markov Model conditioned on protein sequence and predicted secondary structures of the model. |
| Keasar | S* | G | Standard energy terms, solvation, and metaterms that compare the distributions of per-atom energy values to the ones observed in native structures (using the in-house MESHI package). |
| LNCCUnB | S* | LG | Overall atomic burial similarity (mutual information) between the submitted structures and the in-house predictions. |
| ModFOLD5 | C | LG | Clustering of models in CASP datasets together with IntFOLD3 models using ModFOLDclust2 approach. |
| ModFOLD5-single | S* | LG | Comparing every model in the CASP dataset against the pool of IntFOLD3 models using a global and local scoring approach similar to that used by ModFOLDclust2. |
| ModFOLDclust2 | C | LG | Global: mean of the QA scores obtained from the ModFOLDclustQ method and the original ModFOLDclust method; local: the per-residue score taken from ModFOLDclust. |
| MQAPmulti (MQAPsingleC) | S | G | Linear regression of all the afore-mentioned single-model QA-metrics. |
| MQAPsingle | S*M | G | Compares models in the test set against models generated by GeneSilico metaserver using the MQAPmulti algorithm. |
| MQAPsingleA | S*M | G | Average GDT_TS distance of the model to GeneSilico models. |
| MQAPsingleB | S*M | G | 0.8*MQAPsingleA + 0.2*MQAPsingleC. |
| MUFOLD-QA | C | G | Average pair-wise similarity of the model to all models in a test set with complement of models generated by MUFOLD server. |
| MUFOLD-server | C | G | Combination of single scoring functions (secondary structure, solvent accessibility, torsion angles) with consensus GDT. |
| MULTICOM-clust | S | LG | An SVM-based score combining predicted secondary structure, solvent accessibility and PSICOV and DNcon contact scores. |
| MULTICOM-constr | C | LG | The features used by MULTICOM-cluster (above) + a normalized pairwise score. |
| MULTICOM-novel | S | LG | Combination of the density maps of physical-chemical features, and four single-model energy scores. |
| MULTICOM-refine | C | LG | Global: average GDT-TS score from the pairwise comparison for easy targets and the Model Evaluator score for harder ones; local: random forest based on physical-chemical features of each residue. |
| myprotein-me | S* | LG | Random forest combining agreement of top-ranked predicted contacts, secondary structure prediction by PSIPRED, and four statistical potentials: dDFIRE, RW/RW+ and ORDER_AVE. |
| NNS | S* | G | Similarity to in-house structure models in combination with a single-model score based on random forest approach. |
| PconsD | C | LG | Fast, superposition-free method based on consensus of inter-residue distance matrices. |
| Pcons-net | CM | LG | Structural consensus of models. |
| ProQ2 | S | LG | Combination of evolutionary information, multiple sequence alignment and structural features of a model using SVM |
| ProQ2-refine | S | LG | Sidechain repacking to find the optimal ProQ2 score given the current backbone. |
| Raghavagps-qaspro | SM | G | Regression model combining secondary structure and evolutionary features. |
| RFMQA | S* | G | Random forest machine learning using secondary structure, solvent accessibility and potential energy terms. |
| VoroMQA | S | LG | Comparing inter-atomic contact areas and solvent contact areas using a knowledge-based potential. |
| Wallner | C | LG | Pcomb=0.2*ProQ2+0.8*Pcons. |
| Wang_deep_1,2,3 | S | LG | Deep learning algorithm (stacked denoising autoencoders) based on PSI-BLAST profile, SS and residue-residue contact comparisons. |
Legend:
G – a global quality estimator (one score per model).
L – a local quality estimator (per-residue reliability scores).
S – a single model method capable of generating the quality estimate for a single model without relying on consensus between models or templates.
C – a clustering (consensus) method that utilizes information from a provided set of models.
S* - a quasi-single model method capable of generating the quality estimate for a single model but only by means of preliminary generation of auxiliary ensembles of models or finding evolutionary related proteins and then measuring similarity of the sought model to the structures in the ensemble.
M – a meta-method combining scores from different quality assessment methods.