Dr. George Dombi of Wayne State University has developed neural networks which generalize common themes found in peptides of 25 amino acid length. The sequences were sorted into two groups: transmembrane or nontransmembrance type. 1751 training examples were used. As a result of training, bacteriohodopsin was examined to determine the position of it 7 transmembrane helices.
Using several training and testing experimental procedures, test results were obtained with up to 98% accuracy. A symbolic rather than numeric representation was used. For example, the alanine position 22 was represented by the input A22 being either on or off.
INPUTS
Ala_1
Ala_2
(etc...)
Ala_25
Cys_1
Cys_2
(etc...)
Cys_25
Asp_1
Asp_2
(etc...)
Asp_25
Glu_1
Glu_2
(etc...)
Glu_25
Phe_1
Phe_2
etc...)
Phe_25
(repeated for the other 16 amino acids)
OUTPUTS
transmembrane
nontransmembrane
Reference:
Dombi, G.W. and Lawrence, J, Analysis of protein transmembrane helical regions by a
neural network, Protein Science (1994) 3:557-566.