Predictive Biomarkers

Transcription

1 Uğur Sezerman Evolutionary Selection of Near Optimal Number of Features for Classification of Gene Expression Data Using Genetic Algorithms Predictive Biomarkers Biomarker: A gene, protein, or other change that heralds a biomedical phenotype before that phenotype is clinically apparent. Biomarkers used for: Diagnosing Disease Establishing Disease Risk Establishing Drug Response Establishing Drug Toxicity Predicting Drug Response Predicting Drug Toxicity 1

2 Introduction Selecting minimum number of genes among thousands of genes have crucial importance in identification of a disease since they can be used as biomarkers in diagnostics. If the features do not capture the information that is necessary for classification, accuracy of the classification method will be limited by the lack of this information regardless of the methodology used. Abundance of irrelevant attributes would unnecessarily increase the search space while decreasing the accuracy of the classification algorithm Microarray Technology I: 2 channels Hybridization: A T C G T A G T A G C A T C 2

3 Microarray Microarray Technology II 3

4 AIM Determine a set of informative tests (features) for a specific disease from microarray data. Lower the number of features while keeping the accuracy so that the number of diagnostic tests (cost) are decreased. Biological Applications of Feature Selection Problem Punch et. al. applied genetic algorithm and K-nearest neighbor algorithm to large biological data sets. (31 Feature; Accuracy=83%) Guyon et. al. points out another approach that uses Support Vector Machine methods based on recursive feature elimination to classify cancer patients and normal patients. (16 Feature; Accuracy = 90%) Jirapech-Umpai and Aitken used evolutionary algorithm, a stochastic search and optimization technique, to find the near optimal set of predictive genes in order to classify a leukemia dataset. (30 Feature; Accuracy = 72%) Furey et. al. They used SVM classification and verification of microarray expression data and They found miss-labeled tissue samples. Alon et. al. used clustering algorithms to determine highly correlated expression levels of genes that can be used for diagnosis.(90% Accuracy) Fröhlich et. al worked on the same data set and used Genetic algorithms in combination with Support Vector Classifiers. They found a set of 30 genes that can distinguish the cancer data from the control set with 85 % accuracy 4

5 Methodology Our algorithm involves a basic genetic algorithm with roulette wheel based selection strategy. Following parameter settings are used at each run: P1: Binary Encoding Population Size: 20 *Our fitness function is The # Generations: 160 the classification Crossover Rate: 0.9 accuracy of a classifier such as SVM. Mutation Rate: 0.05 Basic GA Each parent P i is randomly generated and represents selected m features. (m = 30) P1: 14, 11, 9, 5, 16, 1, P1: P2: 2, 8, 6, 4, 12, 10, P2: Fitness function of each P i, F(P i ) is calculated using the Support Vector Classification. Crossover Operator Same features are eliminated from the parents. A random number is generated for the remaining features and that number of features are switched between the mating parents. P1: 14, 11, 9, 5, 16, 1, P2: 2, 8, 6, 4, 12, 10, P1: P2: Mutation Operator Two random numbers are generated, one for 1s one for 0s. And 1 -> 0, 0->1. 5

6 Basics of Our Algorithm Feature # Total Score Parents are randomly generated Basic GA is runing for N(Initial=30) steps Each feature (gene) is assigned the fitness value of its parent. Total fitness score of each feature is calculated at the end of all the steps. (Population table) Generate new parents using roulette wheel according to population table. (Since we are using roullete wheel same gene can be selected several times depending on fitness score thus dynamically reducing number of features) Run basic GA for M(Repeat=10) steps more. Basic GA Dynamic Parent Generation Procedure (Population GA) Generate N Parents Randomly Find fitness score of each parent using Support Vector Machines (SVM) which give the prediction error rate as the output that we try to minimize Calculate the fitness score of new generation If Random number less then Crossover rate No Randomly pair all the parents with each other Allow a feature to be selected several times in an individual if it has high score Generate N parents for next generation by generating m random numbers between and then choosing the corresponding features to be included in individual Yes Apply crossover operation by changing 1 s between two parents while keeping the number of features in an individual constant If Random number less then Mutation rate Assign a numerical region to each feature depending on their relative fitness score amongst all features Assign a probability to be selected in generation of new individuals to each feature depending on its fitness score from population table Apply mutation by eliminating one feature while adding another to keep the number of features constant Calculate the fitness score of new generation Yes No Find the average fitness score of each feature by dividing the total fitness score of the feature to the number of times that feature was chosen in an individual Rank the fitness scores of the parents and the offsprings Yes Replace the worst scoring n offsprings with the best scoring n parents if the parents fitness scores are better than the scores of children s No If (Generation number = First-point) or (Generation number > First-point and mode Generation number, Repeat-num) = 0) Offsprings become our new parents in this generation For each individual, assign the fitness value of that individual to the features present in the individual Update the population table by adding these fitness values for each feature 6

7 Dataset In the first data set we used a colon cancer data obtained by Alon et. al. 1 using Affymetrics oligonucleotide arrays. Data showed gene expression levels of 2000 genes for 40 tumor and 22 normal colon tissue samples. Second data set consisted of gene expression levels of genes in 162 ovarian cancer and 91 control patients are obtained from Petricoin et. al. 2 The third data set consisted of gene expression levels of 12,600 genes taken from 52 prostate cancer and 50 control samples are obtained from Singh et. al. 3 1.U. Alon, N. Barkai, D. Notterman, K. Gish, S. Ybarra, D. Mack, A. Levine, Broad patterns of gene expression revealed by clustering analysis of tumor and normal colon cancer tissues probed by oligonucleotide arrays, Cell Biology, 96: , Petricoin EF, Ardekani AM, Hitt BA, Levine PJ, Fusaro VA, Steinberg SM, Mills GB, Simone C, Fishman DA, Kohn EC, Liotta LA: Use of proteomic patterns in serum to identify ovarian cancer. Lancet 2002, 359: Singh D, Febbo PG, Ross K, Jackson DG, Manola J, Ladd C, Tamayo P, Renshaw AA, D'Amico AV, Richie JP, Lander ES, Loda M, Kantoff PW, Golub TR, Sellers WR: Gene expression correlates of clinical prostate cancer behavior. Cancer Cell 2002, 1: Results Table 1. Classification Accuracy using ten fold cross validation and BSVM tool Colon Ovarian Prostate Acc. #Feat. Acc. #Feat. Acc. #Feat. All Features Our Features Others Results 91.94* 30* 100** 17** 97.06* * Results from Bing et. al 1, ** Results from Liu et.al. 2 30* 1. Liu, Bing; Cui, Qinghua; Jiang, Tianzi; Ma, Songde (2004) " A combinational feature selection and ensemble neural network method for classification of gene expression data" BMC Bioinformatics Liu H, Li J, Wong L: A comparative study on feature selection and classification methods using gene expression profiles and proteomic Patterns. Genome Inform Ser Workshop Geonome Inform 2002, 13:51-6 7

8 Colon Cancer 1 Figure 1. Average error rates of the population and the best individual scores of 10 Basic GA runs for Colon Data Figure 4. Average error rates of 10 runs of our algorithm for Colon Data. Green lines indicate the average number of minimum features in each iteration.. 1.U. Alon, N. Barkai, D. Notterman, K. Gish, S. Ybarra, D. Mack, A. Levine, Broad patterns of gene expression revealed by clustering analysis of tumor and normal colon cancer tissues probed by oligonucleotide arrays, Cell Biology, 96: , 1999 Ovarian Cancer 1 Figure 3. Average error rates of the population and the best individual scores of 10 Basic GA runs for Ovarian Data. Figure 4. Average error rates of 10 runs of our algorithm for Ovarian Data. Green lines indicate the average number of minimum features in each iteration. 1.Petricoin EF, Ardekani AM, Hitt BA, Levine PJ, Fusaro VA, Steinberg SM, Mills GB, Simone C, Fishman DA, Kohn EC, Liotta LA: Use of proteomic patterns in serum to identify ovarian cancer. Lancet 2002, 359:

9 Prostate Cancer 1 Figure 5. Average error rates of the population and the best individual scores of 10 Basic GA runs for Prostate Cancer Data Figure 6. Average error rates of 10 runs of our algorithm for Prostate Cancer Data. Green lines indicate the average number of minimum features in each iteration. 1.Singh D, Febbo PG, Ross K, Jackson DG, Manola J, Ladd C, Tamayo P, Renshaw AA, D'Amico AV, Richie JP, Lander ES, Loda M, Kantoff PW, Golub TR, Sellers WR: Gene expression correlates of clinical prostate cancer behavior. Cancer Cell 2002, 1: Randomly Chosen Features Test Figure 7. Classification accuracies of 40 experiments of randomly chosen features from each dataset 9

10 Conclusion Dynamic parent generation step based on fitness score of individual genes is inspired by Richard Dawkins s, selfish gene idea. Parents are mere transporters of fit genes. The idea of a fitter and fewer genes (features) make-up for fitter and more evolved efficient parents enabled us to dynamically reduce number of genes. In this way, The computational time is decreased. The number of selected features is decreased. The classification accuracy is increased up to 100% for different dataset. Each run ended up different subset of features however the correlation between the features were higher than 76%. The selected set of genes for the colon cancer data includes oncogenes, cell adhesion molecules and collagens which were shown to be involved in colon cancer by experimental studies. Similar set of features were also found by Guyon et. al. Acknowledgment Thanks to Alper Kucukural Suveyda Yeniterzi Reyyan yeniterzi 10