1. Predicting the Cellular Localization Sites of Proteins Using Decision Tree and Neural Networks Yetian Chen 2008-12-12

2. 2008 Nobel Prize in Chemistry Roger TsienOsamu ShimomuraMartin Chalfie Green Fluorescent Protein (GFP) Use GFP to track a protein in living cells

3. The cellular Localization information of a protein is embedded in protein sequence PKKKRKV: Nuclear Localization Signal VALLAL: transmembrane segment Cellular Localization SitesAmino Acid sequence of a protein Challenge: predict

4. Extracting cellular localization information from protein sequence mcg: McGeoch's method for signal sequence recognition. gvh: von Heijne's method for signal sequence recognition. alm: Score of the ALOM membrane spanning region prediction program. mit: Score of discriminant analysis of the amino acid content of the N- terminal region (20 residues long) of mitochondrial and non-mitochondrial proteins. erl: Presence of "HDEL" substring (thought to act as a signal for retention in the endoplasmic reticulum lumen). Binary attribute. pox: Peroxisomal targeting signal in the C-terminus. vac: Score of discriminant analysis of the amino acid content of vacuolar and extracellular proteins. nuc: Score of discriminant analysis of nuclear localization signals of nuclear and non- nuclear proteins.

5. Problem Statement & Datasets Protein Name mcg gvh lip chg aac alm1 alm2 Location EMRB_ECOLI 0.71 0.52 0.48 0.50 0.64 1.00 0.99 cp ATKC_ECOLI 0.85 0.53 0.48 0.50 0.53 0.52 0.35 imS NFRB_ECOLI 0.63 0.49 0.48 0.50 0.54 0.76 0.79 im Dataset 1: 336 proteins from E.coli (Prokaryote Kingdom) http://archive.ics.uci.edu/ml/datasets/Ecoli Dataset 2: 1484 proteins from yeast (Eukaryote Kingdom) http://archive.ics.uci.edu/ml/datasets/Yeast

6. Implementation of AI algorithms Decision Tree > C5 Neural Network > Single layer feed-forwad NN: Perceptrons > Multilayer feed-forward NN: one hidden layer

7. Implementation of Decision Tree: C5 Preprocessing of Dataset > If the data point is linear and continous, divide the data range to 5 equal- width bins: tiny, small, medium, large, huge. Then discretize the data points to these bins. > if the feature value is missing (?), replace ? with tiny. Generating training set and test set > Randomly split the data set to training set and test set such that 70% will be in the training set and 30% for test set. Learning the Decision Tree > using the decision tree learning algorithm in chapter 18.3 of text book Testing

8. Implementation of Neural Networks Structure of Perceptrons and two-layer NN Protein Name mcg gvh lip chg aac alm1 alm2 Location EMRB_ECOLI 0.71 0.52 0.48 0.50 0.64 1.00 0.99 cp ATKC_ECOLI 0.85 0.53 0.48 0.50 0.53 0.52 0.35 imS NFRB_ECOLI 0.63 0.49 0.48 0.50 0.54 0.76 0.79 im input Att 1 Att 2 Att 3 Att 4 output cp imS im 1 0 0 Desired output input Att 1 Att 2 Att 3 Att 4 output cp imS im 1 0 0 Desired output PerceptronsTwo-layer NN

9. Implementation of Perceptrons & Two-layer NN: Algorithms Function PERCEPTRONS-LEARNING (examples, network) initially set correct=0 initialize the weight matrix w[i][j] with randomized number within[-0.5,0.5] While(correct < threshold) //threshold =0.0, 0.1, 0.2…, 1.0 for each e in the example do calculate output for each output node //g() is sigmoid function prediction = r such that if r != y(e) for each output node j for i=1,…,m endfor endif endfor endwhile Return w[i][j] 2-layer NN(example,network) Using the Back-Prop-Learning in Chap 20.5 of textbook

10. Results Accuracy comparison DatasetDecision TreePerceptronsTwo-layer NN (hidden nodes:5) Majority E.coli68.04±5.03%66.76±6.34% (Threshold=0.7) 65.68±6.09% (Threshold=0.7) 45.05% Yeast46.63±2.55%50.41±2.74% (Threshold=0.5) 50.28±2.23% (Threshold=0.55) 28.82% The statistics for Decision Tree are average over 100 runs The statistics for Perceptrons and Two-layer NN are average over 50 runs Threshold is the termination condition for training the neural networks

11. Conclusions The two datasets are linearly inseparable. For the E.coli dataset, DT, Perceptrons, Two-layer NN achieve similar accuracy For the yeast dataset, Perceptrons, Two-layer NN achieve slightly better accuracy than DT All the three AI algorithms have much better accuracy than the simple majority algorithm

12. Future work Probabilistic model Bayesian network K-Nearest Neighbor ……

13. A protein localization sites prediction scheme mcggvhalmmiterlpoxvacnuc Classifiers prediction Guide the experimental design and biological research, save much labor and time!