Modeling of Biomolecular Systems Interactions, Dynamics, and Allostery: Bridging Experiments and Computations - September 10-14, 2014, Istanbul, Turkey - page 100

94
Modeling of Biomolecular Systems Interactions, Dynamics, and Allostery Poster Session I
41-POS Board 41
The ProteinModelPortal - How Good is My Prediction? –First Results from The
Continuous Automated Model EvaluatiOn (CAMEO)
Juergen Haas1,2, Alessandro Barbato1,2, Steven Roth1,2, Tobias Schmidt1,2, Khaled Mostaguir1,2,
Stefan Bienert1,2, Andrew M. Waterhouse1,2, Tiziano Gallo-Cassarino1,2, Valentina Romano 1,2,
Lorenza Bordoli1,2, Torsten Schwede1,2.
1Swiss Institute of Bioinformatics, Basel, Basel, Switzerland, 2Biozentrum, Universitaet Basel,
Basel, Switzerland.
Protein structure modeling is widely used in the life science community to build models for
proteins, where no experimental structures are available. The ProteinModelPortal (PMP,
http://www.proteinmodelportal.org)contains more than 21 million models from various
modeling servers for more than 5.1 million distinct UniProt sequences and is regularly updated.
Depending on the difficulty of the target protein the various modeling approaches differ in
performance and we established the Continuous Automated Model EvaluatiOn (CAMEO,
http://www.cameo3d.org)platform assessing the performance of servers predicting protein
structures in its 3D category. Based on the weekly PDB pre-release, protein sequences (targets)
are submitted to participating servers. After the release of the 3D coordinates four days later, the
returned predictions are then compared to the corresponding PDB structures (references). Within
121 weeks 49837 protein structure predictions by 34 servers for 2129 targets were assessed in
the 3D category. Besides 3D predictions, CAMEO assesses ligand binding residues predictions
in proteins (16 servers registered). Since the quality of models ultimately determine their utility,
a new category "Quality Estimation of protein structures models" is being introduced to
CAMEO. So far two servers and 4 stand-alone versions of the most popular quality estimation
methods along with a custom naïve predictor based on CAMEO 3D data are being evaluated.
Continuous assessment of prediction servers allows to retrospectively analyze the performance
of a given server - with implications for PMP, as the quality of the predictions may vary
significantly among different servers depending on the specific target protein and the chosen
approach. The blind predictions obtained by CAMEO are directly suited for publication and the
straightforward modular extension of CAMEO allows new categories and scores to be added on
demand of the respective communities.