QA-RecobineIt

A web server for Quality Assessment and Recombination of protein models.

The purpose of using QA-RecombineIt

Many computational methods have been developed to predict protein structure. However, to generate models of the target protein, the typical users of structure modeling methods (e.g., biologists willing to predict a structure of a protein) use only a few fully automated servers, usually the ones that have well performed in the most recent CASP experiment. Then, they choose the model that is highest scored by a model quality assessment program (MQAP). Even though, the model of the best overall score is selected, it does not always mean that its local conformation over the entire length of the target sequence is always closest to the native structure compared to the conformations of the remaining models. Thus, even if there are regions more accurately modeled in the remaining models, such information will not be used by non-expert users. To address this problem, we developed the QA-RecombineIt web server.

When I can use QA-RecombineIt?

In general, our method can be successfully applied to any case of homology modeling (see official CASP9 benchmarkfor TMB and TMB/FM modelling, out method - recombineIt was ranked the 4th).

We would strongly recommend it for using to those who are working on difficult modeling problems, and have e.g. a set of different homology models plus additional de novo models (e.g. from I-TASSER or ROSETTA or REFINER or CABS or whatever else), and would like to get one model that incorporates the best parts into one structure that is internally coherent (see offical CASP9 benchmark for TBM/FM modelling, out method - recombineIt was ranked the 1st).

What is QA-RecombineIt?

Starting from protein sequence and computational models of the protein,

at the first stage -QA- our server predicts both global and local accuracy of these models. To do so, the server run three model quality programs: MQAPsingle, MetaMQAP and ProQ2 (true MQAPs), and MQAPmulti (a clustering MQAP).

Then, at the second stage - RecombineIt- the server generates hybrid models. Some of these new models are likely to be more accurate compared to the initial models. The merging of finest (according to the predicted model accuracy) fragments of input models is done by our recombination algorithm (RecombineIt). Next, our server executes the MODELLER program to build models of the 100 most promising combinations of fragments. Finally, the 100 models are scored and ranked by abovementioned MQAPs.

The novelty of QA-RecombineIt is that it is composed from two independent, but fully integrated components: model quality assessment (QA) and fragment recombination algorithm (RecombineIt). Each component can be executed independently, however the naïve user having a set of alterative models can automatically evaluate and recombine these models to produce models closer to the native structure

How does the QA-RecombieneIt works?

 

 

 

Benchmark

QA-RecombineIt was tested in CASP9 experiment, which was a community-wide blind assessment of the computational methods for protein structure prediction. During CASP9 experiment out method operates in fully automatic mode, using only the server models as an input, as presented:

RESULTS

TEMPLATE BASE MODELING

RecombineIt was ranked 4th (174 groups participated) for cases where there was at least one template to predict protein structure (55 targets, TBM or TMB/FM). Noteworthy, the best performing server-RAPTORX-was ranked 16th.

For cases where there was at least one template to predict protein structure, QA-RecombineIt - our combined approach for assessment and recombination of protein models - can significantly improve the quality of models yielded by an user.

 


This picture shows the overall performance of the predictors in the CASP9 experiment. Only data for TBM and TBM/FM (hard template based modelling) modelling are shown. Red bars correspond to server (automatic) methods, blue to human predictors. Although our method was a "human" predictor, it automatically generated models by using only the models that had been submitted by the servers (red bars). On average, our automatic procedure performed much better (4th position) than any server method, including the best server - RaptorX ranked 16th by the CASP assessors.
source: http://predictioncenter.org/casp9/groups_analysis.cgi?type=all&tbm=on&tbmfm=on&submit=Filter

In case of hard template base modelling (TMB/FM) our method was the best performing method, nevertheless this benchmark included only 3 protein targets.
source: http://predictioncenter.org/casp9/groups_analysis.cgi?type=all&tbmfm=on&submit=Filter

Valid HTML 4.01!