1. Artificial Neural Networks Thomas Nordahl Petersen & Morten Nielsen

2. Use of artificial neural networks
A data-driven method to predict a feature, given a set of training data
In biology input features could be amino acid sequence or nucleotides
Secondary structure prediction
Signal peptide prediction
Surface accessibility
Propeptide prediction
O- and N-glycosylation C N
Signal
peptide
Propeptide
Mature/active protein

3. Neural network prediction methods http://www.cbs.dtu.dk/services/

4. Pattern recognition

5. Biological Neural network

6. Biological neuron structure

7. Diversity of interactions in a network enables complex calculations
Similar in biological and artificial systems
Excitatory (+) and inhibitory (-) relations
between compute units 1
fire

8. Transfer of biological principles to artificial neural network algorithms
Non-linear relation between input and output
Massively parallel information processing
Data-driven construction of algorithms
Ability to generalize to new data items
google translate (

9. Sparse encoding of amino acid sequence windows

10. Sparse encoding
Inp Neuron
AAcid A R N D C Q E

11. BLOSUM encoding (Blosum50 matrix)
A R N D C Q E G H I L K M F P S T W Y V A R N D C Q E G H I L K M F P S T W Y V

12. Sequence encoding (continued)
Sparse encoding V: L:
V.L=0 (unrelated)
Blosum encoding V: L:
V.L = (highly related)
V.R = (close to unrelated)

13. Simplified feed-forward neural network
no bias neurons are shown I1 I2 I3
Input
w1,2
w3,1
w2,2
w2,1
w1,1
w3,2 h1 h2
hidden
v1,1
v2,1 O1
output
h1 
h1
 = 1/ (1+e-x)
o=H1*v1,1 + H2*v2,1
O1 = (o)
Error = O - True

14. Sigmodial or logistic function

16. Training and error reduction

17. Training and error reduction

18. Training and error reduction
Size matters 

19. Secondary Structure Elements
ß-strand
Helix
Bend
Turn

20. Neural Network Architecture
Weights
Input Layer I K H E
Output Layer E E C H V I I Q A E
Hidden Layer
Window
IKEEHVIIQAEFYLNPDQSGEF…..

21. Predictions and reliability of a prediction
Normally the best prediction is obtained by averaging
results from several predictions - “wisdom of the crowd
Two types of neural networks
Prediction of features in classes/bins e.g. H, E or C (1,0,0)
Values close to 1 or 0 are more accurate than values close to 1/2
Prediction of real values e.g. Surface accessibility (0.43)
Reliability of a prediction is more difficult to estimate

22. Signal peptide

23. Eukaryotic SP & TM Signal peptide cleavage 1523 seq C-terminal end of
TM-regions
669 seq 23

24. Signal peptide is present
Signal peptide prediction
Signal pepdide likeness
Cleavage site
Combined information
Signal peptide is present
If D-score is above
threshold

25. Protein Engineering, Design and Selection: 17: 107-112, 2004.
Propeptide prediction
Many secretory proteins and peptides are synthesized as inactive precursors that in
addition to signal peptide cleavage undergo post-translational processing to become
biologically active polypeptides. Precursors are usually cleaved at sites composed of
single or paired basic amino acid residues by members of the subtilisin/kexin-like
proprotein convertase (PC) family. In mammals, seven members have been identified,
with furin being the one first discovered and best characterized. Recently, the
involvement of furin in diseases ranging from Alzheimer's disease and cancer to
anthrax and Ebola fever has created additional focus on proprotein processing.
We have developed a method for prediction of cleavage sites for PCs based on
artificial neural networks. Two different types of neural networks have been
constructed: a furin-specific network based on experimental results derived from
the literature, and a general PC-specific network trained on data from the Swiss-Prot
protein database. The method predicts cleavage sites in independent sequences with
a sensitivity of 95% for the furin neural network and 62% for the general PC network.
Protein Engineering, Design and Selection: 17: , 2004.
General cleavage: R/K-Xn-R/K , n=0, 2, 4, 6
Furin cleavage: R-X-R/K-R

26. Propeptide prediction
Furin cleavage

28. NetSurfP

29. a b c
a = a-helix
b=b-strand
c=coil

30. O- and N-glycosylation
O-glycosylation of Ser and Thr (mucin type glycolylation)
N-acetylgalactosamine (GalNAc)
N-glycosylation of Asn-X-Ser/Thr, where X differ from Pro
N-acetylglycosamine (GlcNAc)