همایش ملی بیوانفورماتیک ایران

صفحه اصلی / 4th international edition and 13th Iranian Conference on Bioinformatics

Machine-learning based biomarker discovery for Striga resistance in sorghum

نویسندگان :

Leyla Nazari¹ Afshar Estakhr²

1- مرکز تحقیقات و آموزش کشاورزی و منابع طبیعی فارس، سازمان تحقیقات، آموزش و ترویج کشاورزی، شیراز، ایران 2- مرکز تحقیقات و آموزش کشاورزی و منابع طبیعی فارس، سازمان تحقیقات، آموزش و ترویج کشاورزی، شیراز، ایران

کلمات کلیدی :

Striga hermonthica،ReliefF،Information gain،Gain Ratio

چکیده :

Sorghum bicolor is the fifth most important cereal crop in the world, cultivated across the globe in almost 110 countries, predominantly in Asia and Africa but also in Europe, America, and Oceania. Despite its outstanding resilience to abiotic stresses, approximately 20% of sorghum yield loss is annually due to infestation with the parasitic weed Striga hermonthica. Identifying sources of striga resistance gene within sorghum is imperative to developing resistant sorghum cultivars. Feature selection algorithms are frequently employed in preprocessing machine learning pipelines applied to biological data to identify relevant features. The objective of this study is about selecting the important features along with improving the prediction accuracy. Therefore, we propose to use an integrated strategy including Information gain, Gain Ratio, and ReliefF to filter important genes involved in striga resistance in sorghum. For this, were searched in the public database, National Center for Biotechnology Information Gene Expression Omnibus (NCBI GEO) and found a study with accession number GSE216351. The DESeq2 package (v 1.34.0) was employed and the genes less than 10 counts across all samples were filtered out. The resulting matrix including 25037 genes and 31 samples were submitted to the feature selection algorithms and the top 50 genes ranked by Information gain, Gain Ratio, and ReliefF were selected as input for venn diagram. Seven genes including Sobic.001G429700, Sobic.002G087200, Sobic.010G134900, Sobic.005G063800, Sobic.005G192400, Sobic.006G053500, and Sobic.002G307700 were found to be common identified by the three methods. To validate the accuracy of our feature selection methods, we tested different algorithms from classifiers bayes, functions, lazy, meta, rules, and trees and the best performance algorithm from each classifier was selected. Modeling was performed using 10-fold cross-validation. NaiveBayes, SGD, IBK, AdaBoostM1, PART (rules), and j48 prediction models were the best algorithms from each classifier to discriminate control and infected crops. The highest performance was obtained by NaiveBayes with 96.7742% accuracy. Therefore, considering the high efficiency of these seven genes to classify control and infected crops, they could be suggested as biomarkers for striga resistant in sorghum.