Introduction

     FOLDpro is a web server to predict protein 3D structure using a machine learning fold recognition approach. It makes predictions in three steps.

• Step 1: Use a machine learning information retrieval approach to rank template proteins for the query protein, integrating a variety of similarity features including sequence/family information, sequence-sequence alignment, sequence-profile (profile-sequence) alignment, profile-profile alignment, and structural features (see Table 1 for all the features).



• Step 2: Generate profile-profile alignments between the query protein and the top ranked template proteins. Multiple templates are used to improve both the alignment and the structure modeling if necessary.



• Step 3: Based on the query-template alignments and 3D structures of the templates, Modeller (Sali and Blundell, 1993) is used to generate structure models for the query protein. Five models are generated for the query. The models ranked higher are generated based on the templates ranked higher. So they are presumably, but not always, better than the models ranked lower (e.g., fold1.pdb is likely better than fold2.pdb. fold2.pdb is likely better than fold3.pdb).

Input and Output

• Inputs to the web server include target name, sequence, and email address. It usually takes several hours to process one query, depending on the server load and sequence length. Sequence must be entered as a plain sequence of amino acids. Maximum sequence length allowed is 500.



• Outputs include the top 20 ranked templates with their ranking scores and the predicted 3D structures. The ranking scores are generated by Support Vector Machine Algorithm. A positive score usually indicates the query and template are significantly related. The higher the score, the more significant the match is. The PDB files and their corresponding alignment files used to generate the structures are attached with the returned email. The alignment files are in the PIR format. In addition, a ranking list of all templates for the query is also returned as a file: query.rank.

Download the feature dataset used in FOLDpro

FOLDpro dataset (zipped file is about 500M) and the description of the dataset.

Reference

• J. Cheng and P. Baldi. A Machine Learning Information Retrieval Approach to Protein Fold Recognition. Bioinformatics, vol. 22, no. 12, pp. 1456-1463, 2006. [PDF] [PDF at Bioinformatics web site]