Use of Artificial Neural Networks and effects of amino acid encodings in the membrane protein predic

1
Use of Artificial Neural Networks and effects of
amino acid encodings in the membrane protein
prediction problem

• Subrata K Bose1, Antony Browne2, Hassan Kazemian3
  and Kenneth White4
• 1 MRC Clinical Sciences Centre, Faculty of
  Medicine, Imperial College London, UK
• 2 Department of Computing, School of Engineering
  and Physical Sciences, University of
• Surrey, Guildford
• 3 Intelligent Systems Research Centre,
  Department of Computing, Communication
• Technology and Mathematics, London Metropolitan
  University, London, UK
• 4 Institute for Health Research and Policy,
  Departments of Health and Human Sciences,
• London Metropolitan University, London, UK
• Email subrata.bose_at_ic.ac.uk

2

• Proteins
• Proteins are essential macromolecules of life
  that constitute about half of a living cells dry
  weight
• They are polymers of one or more uncleaved chains
  of amino acids ranging typically from two
  hundreds to few thousands
• Proteins play crucial roles in almost in
  all-biological processes
• Example
• Almost all known enzymes are proteins
• Hormones, are also proteins

• .

3

• Amino Acids
• There are 20 different amino acids
• Amino acids have similar structures (in Fig)
• Side chain groups (the R group) may vary in size,
  shape, charge, hydrophobicity, and reactivity
• The amino acids could be considered as the
  alphabet from which the proteins are written
• Amino acids form long chains of polymer via
  peptide bond to form peptides
• Long chains of a-amino acids make Proteins

• .

Fig Amino Acid's Structure
4

• Structure of Protein
• The structure of the protein can be classified in
  the following fashion

• .

Level one
Primary structure
Level two Secondary structure
Level three Tertiary structure
Level Four Quaternary structure
5

• Membrane Protein (MP)
• Proteins can be classified as being either
  soluble, such as the proteins found in blood or
  in the liquid compartments of cells (cytosol), or
  bound to cell membranes
• All living cells are enclosed by a cytoplasmic
  membrane to separate the cytoplasm from the
  outside environment
• A MP is attached to, or associated with the
  membrane of a cell or an organelle
• The part of the protein, which makes contact with
  the cell membrane, is called the transmembrane
  domain (TMs) or the membrane spanning region
  (MSR)

• .

6

• Pharma-economical Importance of MPs
• MPs are accountable for a number of key tasks in
  a living cell
• Comprise some extremely essential biochemical
  components for cell-cell signaling,
  transportations of nutrients, solvents and ions
  across the membrane, cell-cell adhesion and
  intercellular communication
• Many drugs or drug adjuvants may interfere with
  membrane dynamics and subcellular localisation of
  enzyme systems
• From a pharma-economical perspective, MPs
  constitute 75 of possible targets for novel
  drugs
• 3D structures of proteins derived by X-ray
  crystallography have been determined for about
  15000 proteins, only about 80 of these are MPs
• The human proteome is estimated to be comprised
  of about 30 proteins containing membrane
  spanning regions (MSRs)

• .

7

• System Method

• .

Fig Neural Networks (NNs) architecture
8

• Prediction Problem
• MPs are not soluble in aqueous buffer solutions
  and denature in organic solvent because of the
  presence of the hydrophobic and hydrophilic
  region on the surface of MPs
• Due to the hydrophobicity, and detergent micelles
  formation around the hydrophobic regions, it is
  hard to induce MPs to form well-ordered 3D
  crystals
• The precise and reliable prediction of membrane
  protein secondary structure is an important
  intermediate step towards a fuller understanding
  of protein folding
• No explicit methods for prediction of the
  structure or packing of transmembrane proteins
  from primary amino acids sequences exist today
• Re-evaluation studies show that available methods
  are far from achieving 95 reliability in
  prediction and overrated the accuracy of their
  methods by 15-50
• Mainly due to the lack of non-redundant membrane
  datasets which were used in the learning and
  tuning processes

9

• Dataset
• Two MP datasets with known biochemical
  characterisations of membrane topology
• For Nonmembrane Protein DB-NTMR dataset was
  chosen
• Amino acid Sequence Encodings
• Four binary encoding schemas are adopted
• 20 bits- an orthogonal distributed 20 binary bits
  string consisting of nineteen 0s and one 1
• 16 bits-a product of a genetic search algorithm
  and 30 shorter than the traditional orthogonal
  representations
• 9 bits -a feature-based grouping of amino acids
  where hydrophobicity, positive charge, negative
  charge, aromatic side chain, aliphatic side
  chain, small size, bulk size properties and two
  correction bits are used to discriminate similar
  amino acids
• 5 bits - 5 bits representation is designed based
  on a classification of 20 amino acids which
  unites hydropathy index, molecular weight , and
  pI values together

• .

10

• Architecture of Neural Networks
• The multi-layered feed-forward neural network
  used here consists of an input layer, a single
  hidden layer and an output layer
• Training was done by the scaled conjugate
  gradient algorithm
• This algorithm was applied to all the encodings
  separately to identify the best encodings
• 75 of the source data was randomly taken for
  training the network, and the remaining 25 of
  the data were selected as the test data
• To detect when the over-fitting starts and to
  prevent the over-fitting problem, the training
  set was further partitioned into two disjoint
  subsets training and validation datasets

• .

11

• Results

• .

12

• Results
• Increasing the ANNs binary input vectors number
  may improve the classification rate and
  robustness of the system slightly, but
  incorporation of physico-chemical information
  didnt improve the networks performance
  dramatically
• This could be due to these binary encodings dont
  initiate any artificial ordering uniform weight
  was given to each amino acid for the network
  learning purposes
• Binary encodings have the advantage of not
  importing any artificial correlations between the
  amino acids during the learning process
• Some earlier research suggests that binary input
  data may improve an ANNs learning but according
  to the best of our knowledge, there is no
  recorded evidence on the effect of binary input
  data on the robustness of the ANNs, especially in
  the area of membrane protein prediction

• .

13

• Acknowledgement
• London Metropolitan University
• This project is fully funded by the London
  Metropolitan University

• .

14

• References
• Baldi.P. and Pollastri, G. (2002). Machine
  Learning Structural and Functional Proteomics.
  IEEE Intelligent Systems (Intelligent Systems in
  Biology II).
• Bose, S., Kazemian, H., White, K. Browne., A.
  (2005). Use Of Neural Networks To Predict And
  Analyse Membrane Proteins in the Proteome. BMC
  Bioinformatics. ISSN 1471-2105 Vol. 6 (Suppl 3)
  P3.
• Bose S., Kazemian H.B., White K. Browne A.
  (2006), Presenting a Novel Neural Network
  Architecture for Membrane Protein Prediction.
  Proceedings 10th International Conference on
  Intelligent Engineering Systems, London, UK June
  26-28(ISBN of printed proceedings1-4244-9708-8
  and ISBN of CD proceedings 1-4244-9709-6) Brusic
  V., Rudy G. and Harrison L.C. (1995). Prediction
  of MHC binding peptides using arti- ficial neural
  networks. Complexity International volume 2, ISSN
  1320-0682.
• Chandonia, J M. and Karplus M (1996). The
  importance of larger data sets for protein
  secondary structure prediction with neural
  networks. Protein Science, 5, 768-774.
• de la Maza M (1994). Generate, Test and Explain
  Synthesizing Regulatory Exposing Attributes in
  Large Protein Databases. Proceedings of the
  Twenty-Seven Annual Hawaii International
  Conference on System Sciences.
• Ikeda, M., Arai, M., Okuno, T. and Shimizu,
  T.(2003). TMPDB a database of experimentallychara
  cterized transmembrane topologies Nucleic Acids
  Res., January 1, 2003 31(1) 406 - 409.
• Ito A. (2000). Mitochondrial processing
  peptidase multiple-site recognition of precursor
  proteins. TICB, 1025-31.
• Kihara, D., Shimizu, T. and Kanehisa, M. (1998)
  Prediction of membrane proteins based on
  classification of transmembrane segments. Protein
  Eng., 11, 961-970.
• Qian, N. and Sejnowski, T. J. (1988). Predicting
  the secondary structure of globular proteins
  using neural network models. Journal of Molecular
  Biology. 202, 865-884.
• Parris, N. Onwulata, C. (1995). Food Proteins
  and Interactions. In Molecular Biology and
  Biotechnology A comprehensive desk reference, (R.
  A. Meyers ed.) pp. 320-323, Cambridge UK VCH
  Publishers.
• Pasquier, C., and Hamodrakas, S. J. (1999a) An
  hierarchical artificial neural network system for
  the classification of transmembrane proteins,
  Protein Eng., 12(8), 631-4.
• Pasquier C, Promponas VJ, Palaios GA, Hamodrakas
  JS, Hamodrakas SJ(1999b).A novel method for
  predicting transmembrane segments in proteins
  based on a statistical analysis of the SwissProt
  database the PRED-TMR algorithm. Protein Eng.
  12(5)381-5.

• .

15
Use of Artificial Neural Networks and effects of
amino acid encodings in the membrane protein
prediction problem

• Subrata K Bose1, Antony Browne2, Hassan Kazemian3
  and Kenneth White4
• 1 MRC Clinical Sciences Centre, Faculty of
  Medicine, Imperial College London, UK
• 2 Department of Computing, School of Engineering
  and Physical Sciences, University of
• Surrey, Guildford
• 3 Intelligent Systems Research Centre,
  Department of Computing, Communication
• Technology and Mathematics, London Metropolitan
  University, London, UK
• 4 Institute for Health Research and Policy,
  Departments of Health and Human Sciences,
• London Metropolitan University, London, UK
• Email subrata.bose_at_ic.ac.uk