Osteoporosis is a degenerative bone condition characterized by decreased bone density which currently affects 40 million Americans. The only current FDA-approved anabolic treatment for low bone density, that is capable of forming new healthy bone, is teriparatide. However, this drug carries risks for development of osteosarcomas, and as such is recommended for no more than 2 years of treatment. Characterizing the proteins and pathways responsible for osteoblast development and subsequent formation of new bone would be valuable to discover new anabolic osteoporosis treatments. Machine learning and classification methods hold great promise to determine protein function given the volume of genome-scale data now available. We have collected information from 661 gene expression experiments (both microarray and RNA-seq based) in laboratory mouse and created a machine-learning model utilizing Support Vector Machines (SVMs) to predict novel genes related to osteoporosis and bone formation. Microarray data was primarily obtained from the Gene Expression Omnibus, and RNA-seq data was generated by our lab sampling a time course of cells differentiating in culture from mesenchymal stem cells to mature osteoblasts derived from 5 different mouse strains. These data span approximately 120M measurements suitable for downstream analysis. Data was converted into approximately 500B pairwise data points using multiple distance metrics in order to predict relationships between all gene pairs. We manually curated a training set of known osteoporosis and bone density gene relationships, which was used to train SVM classifiers to produce models, which were then applied to unclassified gene pairs to predict the likelihood of them being involved in bone development. One of the major challenges of machine learning in biology is the high levels of correlation between data due to reuse of biological pathways and components in different contexts. Our network construction algorithm accounts for this and finds highly connected cliques (3.1) that have some degree of predicted functional relationship to existing gene pathways known to be related to osteoblast development (2 and 3). These results demonstrate the accuracy and value of our method and suggest many novel osteoporosis genes suitable for further laboratory study pending further verification via SVMs trained only on osteoblast related experiments and learning models based on Bayesian data integration. 4 1 KEY 1: PR and ROC Curves from 4 fold cross-validated SVM model building. 2: Functional Protein Association Network Representing known gene pair relationships generated from manual curation (demarcated in red) union predictions for all possible binary combinations of genes that are in GS. 3: Network of all nodes with at least degree 10 for the 100 highest ranked pairs genome wide for each unary gene involved in a highly ranked match in (2). Known genes pairs from GS demarcated in red. 3.1: Clique of gene pairs from (3) that model finds extremely likely to be related to osteoblast development not in GS. 4: Most correlated GO term enrichment cluster for the clique represented in (3.1) generated using DAVID. 2 3 3.1 Identification of Osteoporosis-related Genes using Support Vector Machines Jacob M. Luber 1 , Catherine Sharp 2 , KB Choi 2 , Cheryl Ackert-Bicknell 2,3 , Matthew Hibbs 1,2 1 Department of Computer Science, Trinity University, San Antonio, TX 2 The Jackson Laboratory, Bar Harbor, ME 3 University of Rochester Medical Center, Rochester, NY
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Osteoporosis is a degenerative bone condition characterized by decreased bone density which currently affects 40 million Americans. The only current FDA-approved anabolic treatment for low bone density, that is capable of forming new healthy bone, is teriparatide. However, this drug carries risks for development of osteosarcomas, and as such is recommended for no more than 2 years of treatment. Characterizing the proteins and pathways responsible for osteoblast development and subsequent formation of new bone would be valuable to discover new anabolic osteoporosis treatments. Machine learning and classification methods hold great promise to determine protein function given the volume of genome-scale data now available. We have collected information from 661 gene expression experiments (both microarray and RNA-seq based) in laboratory mouse and created a machine-learning model utilizing Support Vector Machines (SVMs) to predict novel genes related to osteoporosis and bone formation.
Microarray data was primarily obtained from the Gene Expression Omnibus, and RNA-seq data was generated by our lab sampling a time course of cells differentiating in culture from mesenchymal stem cells to mature osteoblasts derived from 5 different mouse strains. These data span approximately 120M measurements suitable for downstream analysis. Data was converted into approximately 500B pairwise data points using multiple distance metrics in order to predict relationships between all gene pairs. We manually curated a training set of known osteoporosis and bone density gene relationships, which was used to train SVM classifiers to produce models, which were then applied to unclassified gene pairs to predict the likelihood of them being involved in bone development. One of the major challenges of machine learning in biology is the high levels of correlation between data due to reuse of biological pathways and components in different contexts. Our network construction algorithm accounts for this and finds highly connected cliques (3.1) that have some degree of predicted functional relationship to existing gene pathways known to be related to osteoblast development (2 and 3). These results demonstrate the accuracy and value of our method and suggest many novel osteoporosis genes suitable for further laboratory study pending further verification via SVMs trained only on osteoblast related experiments and learning models based on Bayesian data integration.
4
1KEY
1: PR and ROC Curves from 4 fold cross-validated SVM model building. 2: Functional Protein Association Network Representing known gene pair relationships generated from manual curation (demarcated in red) union predictions for all possible binary combinations of genes that are in GS.3: Network of all nodes with at least degree 10 for the 100 highest ranked pairs genome wide for each unary gene involved in a highly ranked match in (2). Known genes pairs from GS demarcated in red.3.1: Clique of gene pairs from (3) that model finds extremely likely to be related to osteoblast development not in GS.4: Most correlated GO term enrichment cluster for the clique represented in (3.1) generated using DAVID.
2
3
3.1
Identification of Osteoporosis-related Genes using Support Vector MachinesJacob M. Luber1, Catherine Sharp2, KB Choi2, Cheryl Ackert-Bicknell2,3, Matthew Hibbs1,2 1Department of Computer Science, Trinity University, San Antonio, TX 2The Jackson Laboratory, Bar Harbor, ME 3University of Rochester Medical Center, Rochester, NY
Related Documents
