Download
1 / 21
210 likes | 228 Views
This study aims to predict optimal drug therapies for cancer patients by analyzing gene expression profiles using a novel support vector machine (SVM)-based algorithm. The research shows promising results in correlating drug sensitivities in ovarian cancer patients. The methodology involves identifying informative features and building models to enhance prediction accuracy. By incorporating machine learning, this approach offers a potential alternative to traditional trial-and-error strategies in cancer treatment. Results indicate high positive predictive values, highlighting the clinical benefits of this computational method in personalized cancer therapy selection.
E N D
Published November 6 , 2018 I.F. 4.122 PhD John F. McDonald 6 Publications Georgia Institute of Technology 16 Citations PhD John F. McDonald 174 Publications 47 h-index Georgia Institute of Technology 4211 Citations
Introduction • A primary goal of precision cancer medicine is the accurate prediction of optimal drug therapies based upon the personalized molecular profiles of patient tumors. • Such predictions are based upon well-established molecular cause-and-effect relationships that are disrupted in cancer cells. • Unfortunately, the molecular processes underlying most cancers, and especially solid tumors, are currently not as well understood. • An alternative path to accurate predictions is based simply on observed, significant correlations, even when the underlying causal connections are unknown or incompletely understood.
Introduction They recently introduced an “open source” support vector machine (SVM)-based algorithm that inputs gene expression profiles of cancer cells to predict the response of individual cancers to chemotherapeutic drugs. They employed the algorithm to predict the sensitivities of 273 ovarian cancer patients to 7 commonly prescribed chemotherapeutic drugs. These predictions were shown to correlate significantly (Linear regression p value = 0.0031, R2 = 0.8201) While the importance of the initial testing of drug prediction algorithms in well-characterized cancer cell lines cannot be overstated, eventual adoption of this computational approach into clinical practice will require extensive testing in human cancer patients.
Methodology • To determine viable candidate drugs for the analysis, they required a sample size of at least 30 patients for the drug of interest with at least 15 for each type of response. and only patients on a single candidate drug at a time were retained for evaluation obtaining gemcitabine and fluorouracil. • GDC provides UQ-FPKMs to facilitate cross-sample comparison and differential expression analysis. • Genes were removed from the dataset if >25% of the samples displayed a zero expression value. We applied SVM on training data to get weights for each feature, and sorted the features based on the weights.
Methodology • Recursive feature elimination (RFE) is employed to determine the minimum set of features that maximize accuracy on the test dataset. The approach starts by removing the 100 features with the lowest ranked weights in the sorted feature list. • An SVM model is subsequently built using the remaining features and this process proceeds recursively until the number of remaining features reaches 100. • Thereafter, features are removed one at a time until the most informative set of features is obtained. • Leave one out cross-validation (LOOCV) is subsequently used to evaluate the performance of each of the models as previously described.
Result • The minimum number of informative features associated with optimally predicted responsiveness to GEM was 81 and for 5-FU was 31. • Although the majority of these genes remain functionally unannotated, a number have either directly or indirectly been previously associated with apoptosis, which is consistent with the DNA damaging action of both of these drugs.
Results • Patients observed to respond positively to the drug therapy are represented in the figures by blue dots and those observed not to respond to the therapy by red dots. • The overall accuracies (GEM 81.5%; 5-FU 81.7%; PPV GEM 77.8%; 5-FU 83.3%; NPV GEM 83.9%; 5-FU 79.2%), sensitivities (GEM 75.7%; 5-FU 85.7%), and specificities (GEM 85.5; 5-FU 76.0%)
Results • The computational predictions resulted in 17 TP, 2 TN, 3 FP and 1 FN. • This equates to a positive predictive value (PPV) of 85% (sensitivity 94.4%), the negative predictive value (NPV) was 66.7% (specificity of 40%) equating to an overall accuracy of 82.6%.
Results • Algorithms with high positive predictive value (PPV) may be of particular clinical benefit in the selection of alternative second-line chemotherapies. • An estimated 20–30% of all ovarian cancer patients treated with this standard-of-care combination therapy similarly fail to respond to treatmentleaving physicians with the decision as to what to try next. • ML-based models with validated high positive predictive values, such as reported here, may provide physicians with a useful alternative to the traditional trial-and-error strategies. • Patient 545 was predicted and observed not to respond to standard-of-care carboplatin/paclitaxel therapy. Of possible second-line therapies, gemcitabine is predicted to be the preferred choice.
Thesis Advances Cody Eduardo Evans Trejo MCC – 2do Semester Thesis: Predicting Drug Responses in Cancer Cells using Genomic Features and machine learning Advisor: Victor Trevino, PhD Coadvisor: Emmanuel Martinez Ledesma, PhD
Problem definition • The molecular processes underlying most cancers are currently not as well understood. • There are a multitude of possible molecular paths to developing even the same type of cancer explaining why the response to any given chemotherapeutic drug can be highly variable across patients. • Clinical data sets in cancer research are high-dimensional data that reduce predictive power of a classifier or regressor • Classical regression methods will fail in high-dimensional data (least squares, logistic, and Cox-PH regressions). • The bottleneck of data dimensionality in cancer research lies in the unique difficulties in sample collection and annotation.
Objectives • Verify the performance of different machine learning algorithms predicting responses in different features of cancer therapies. • Design and implement feature selection methods capable of processing and simplifying various types of variables (continuous and discrete). • Design and implement strategies of feature engineering able to find new types of biomarkers to determine the responses of different types of drugs. • Verify the performance of the prediction system against those presented in the literature.
Overview Modeling Data • Testing models • Design Ensemble model • Training model • Performance Measure Dataset creation • Extracting data • Clean up data • Filter data • Normalization Feature Engineering • Feature selection • Feature construction • Processed Data
CCLE Drug response Mutations (Binary) K-fold cross-validation Copy-Number (Binary) Genomic Data Training the model Feature Engineering Ensemble model Expression (Continuous) Elastic-Net, SVM, Random forest Prediction model
Data • Cell Description • miRNA • RNAseq • Methylation • Mutation • Copy Number • Drug response • Protein Expression • Fusion Genes
Data Creation • Elimination of unnecessary information • Reclassification of columns • Uniformity in cell lines (1321-N-1, 131321N1_CENTRAL_NERVOUS_SYSTEM, 131321N1) • Transform FPKM (miRNA and RNAseq) • Binarization of Mutation and Fusion Gene Dataset • GDSC dataset transformation • Filtration data that does not respond to drugs
Radom forest in mirna • Training=80% • Testing=20% Root Mean Square Error • Usando líneas celulares de cáncer de pulmón RMSE=1.5917 • Usando todas líneas celulares RMSE=2.58
