i am looking for a way to describe protein sequences by a Neural Network. However i am still missing some descriptors i can use. Do you know of free descriptors or better a descriptor package i could use?
edit: My problem is actually: Different parts of the sequence have a influence on the protein. I want to go over the sequence and then predict this influence. Influence can be an assay or anything rly. In the end it should have formed a model for this and be able to predict for a new sequence. As i do this only to get acquainted with weka and stuff, i dont really have an idea what to use as assay :).
•
written
by
Funuser •
2 answers
I assume you're talking about feature extraction? I've once written a python tool that turns sequence-based features (predicted sec. strct., solv. acc., evolutionary information, predicted PPI interfaces, PFam data, biochemical propensities...) into position-specific numerical normalized features. Output formats are weka arff and libsvm/liblinear-compliant datasets. This tool however heavily depends on predictprotein [1] which is a command line wrapper for all kinds of sequence-based predictors developed in our group. It's available as machine image (complete linux OS) or debian packages. Let me know if this sounds appealing to you.
Hi Chris!
The python tool you mention seems interesting and I would like to explore it for my work on disease/druggability gene predictions. However, I would appreciate if you could help me get started on how to use the tool in batch mode, because my current set consists of ~20k proteins with unique uniprot IDs.
You can use AAINDEX database to derive descriptors using protein sequence. AAINDEX provides amino acid indices, substitution matrices and pair-wise contact potentials.
Background on amino acid index from AAINDEX:
AAindex is a database of numerical indices representing various physicochemical and biochemical properties of amino acids and pairs of amino acids. AAindex consists of three sections now: AAindex1 for the amino acid index of 20 numerical values, AAindex2 for the amino acid mutation matrix and AAindex3 for the statistical protein contact potentials. All data are derived from published literature.
Manuscript describing current version of AAINDEX is available here.
Current version include (ver.9.1) 544 amino acid indices, 94 amino acid mutation matrices and 47 contact potential matrices
You can use this data as a normalized score for the whole protein chain or use them to derive hybrid features. You may please refer to following papers that used AAINDEX derived features/descriptors to develop Support Vector Machines and Random Forests based machine learning algorithms for prediction of 3D domain swapping.
funuser: IMHO, That should be a separate question.
Log in to answer this question.
-
Generating embeddings for antibody data
written by tom5 •
-
In string database how to get the protein interaction for the query(multiple protein sequences) acr…
written by suvethamagudapathi •
-
Protein structural features and PSSM profile
written by Rana •
-
Neural networks feature importance
written by druggable •
-
Protein data representation in neural network
written by nazza2008 •
-
Protein sequence representation in neural network
written by JDK_92 •
-
Best Representation of Protein Sequences
written by khaeuk
-
pdb to dssp
written by uday4vijay •
-
How to reduce protein sequence dataset size for machine learning?
written by rajeshkumar_vinod •
-
Looking For Features To Predict Mutation Impacts
written by kylejudita •
