Neural network system with N-gram term weighting method for molecular sequence classification and motif identification

• US 5,845,049 A
• Filed: 03/27/1996
• Issued: 12/01/1998
• Est. Priority Date: 03/27/1996
• Status: Expired due to Fees

• Current Assignee: Board of Regents of the University of Texas System (University of Texas System)
• Original Assignee: Board of Regents of the University of Texas System (University of Texas System)
• Inventors: Wu, Cathy H.
• Primary Examiner(s): Davis, George B.
• Application Number:
  US08/622,351
• Time in Patent Office: 979 Days
• Field of Search: 395/22, 395/23
• US Class Current: 706/20
• CPC Class Codes: G16B 30/00 ICT specially adapted for sequence analysis involving nucleotides or amino acids
• CPC Class Codes: G16B 30/10 Sequence alignment; Homology search
• CPC Class Codes: G16B 40/00 ICT specially adapted for biostatistics; ICT specially adapted for bioinformatics-related machine learning or data mining, e.g. knowledge discovery or pattern finding
• CPC Class Codes: G16B 40/20 Supervised data analysis
• CPC Class Codes: G16B 50/00 ICT programming tools or database systems specially adapted for bioinformatics

Abstract

A method for rapid and sensitive protein family identification is disclosed. The new designs include an n-gram term weighting algorithm for extracting local motif patterns, an enhanced n-gram method for extracting residues of long-range correlation, and integrated neural networks for combining global and motif sequence information.

Claims

1. A method for training a neural network to predict membership in a family of linear sequences comprising;

providing a training set of full length member sequences, a training set of full length non-member sequences and a training set of family motif sequences;

deriving term weights for n-gram terms by dividing the number of occurrences of each n-gram term in the motif set by the number of occurrences in the full length member set;

deriving a set of global vectors for the full length member set and a set of global vectors for the full length non-member set using an n-gram method;

deriving a set of motif vectors for the full length member set and a set of motif vectors for the fall length non-member set by multiplying each term in the global vector set by its term weighting factor;

providing a neural network with multiple output units to represent one family;

using the global vector of the member sequence set to train the positive full length output unit;

using the motif vector of the member set to train the positive motif output unit;

using the global vector of the non-member set to train the fall length negative output unit; and

using the motif vector of the non-member set to train the negative motif output unit of said neural network.

2. The method of claim 1, further defined as comprising the step of providing n-gram terms specific for said family as additional input vectors to said neural network.

3. A neural network trained by the method of claim 1 to predict family membership of a linear sequence.

4. A series of connected neural networks, wherein each network is trained for a particular family by the method of claim 1.

5. A method of obtaining term weighted n-grams for use in training a neural network comprising:

providing a training set of full length member sequences and a training set of family motif sequences; and

deriving term weights by applying the formula

space="preserve" listing-type="equation">W.sub.k =Σ

M.sub.ik /Σ

F.sub.ik

to said motif set and said full length set, where W_k is the weight factor for the k-th n-gram term in the input vector, and F_ik and M_ik are total counts of the k-th n-gram term in the i-th sequence of the full-length sequence set and motif set, respectively.