ADS Capstone Chronicles Revised
6
(KEGG) Drug (Kanehisa & Goto, 2000). Benchmark datasets were Pauwels et al.’s (2011), Mizutanietal.’s(2012),andLiu’et al.’s (2012). Drugs and side effects were representedasan n x m binarycodedmatrix. The recommendation systems were trained with 20 times 5-fold cross validation and average performance on the test folds was evaluated with area under the precision-recall curve (AUPR), area under the receiver operating characteristic (ROC) curve (AUC), sensitivity, specificity, precision, accuracy, and F scores. Overall, the best performing model was the ensemble,butthesmallgaininperformance over using the neighborhood-based or machine-based methodalonewasnotworth the added complexity. Even though the ensemble was the best, the model had low performanceonallthreebenchmarkdatasets for AUPR (.662 average), sensitivity (.623 average), precision (.614 average), and Fs scores (.618 average). AUC, accuracy, and specificity were all greater than 0.90. The results show that the training could benefit from fixing class imbalance. The models need morefinetuningforthetargetclassto beuseful.Theintendedimplicationsanduse of the model would require performance above 0.600 for all metrics. The authors offer no insight into the interpretability of the model beyond the benchmark datasets. 3.4 Deep Neural Network Wang et al. (2019) created a deep neural network with “chemical, biological, and biomedical information of drugs” from biomedicalliteraturetopredictadversedrug reactions based on Word2Vec word-embeddings (p. 1). Data about 746 drugswassourcedfromSIDER(Kuhnetal., 2016), PubChem (NCBI, n.d.), DrugBank (Wishartetal.,2018),and2.3millionpapers from MEDLINE (National Library of Medicine, n.d.) about each drug in the dataset - case studies, clinical trials, and
observational studies -toassessprogressof surveillance in the literature from 2009 to 2012. The multilayer perceptron had 1325 hidden nodes in the last layer which correspondedtothenumberofknownADR side effects in thedataset.Theperformance oftheneuralnetworkwascomparedtoother models - probability matrix factorization, linearsupportvectorclassifier,andGaussian Naive Bayes. All models were trained with five-fold cross validation andwithdifferent sets of input features (chemical properties, biological properties, word-embeddings). The models were trained as identifiers and classifiers of adverse drugreactions.Model performancewasassessedusingROCcurve and mean average precision (MAP) on the testfolds.Thedeepneuralnetworkwithtwo hidden layers performed the best as classification (AUC 0.844, MAP 0.721). The biological feature set from DrugBank carried most of the variance, with the literature word-embeddings adding slight performance improvement. The chemical features from PubChem were found to be noninformative for classifying outcomes. 3.5 Matrix Decomposition Galeano et al. (2020) generated a matrix decomposition algorithm to predict the frequencies of drug side effects. They obtained data from SIDER (Kuhn et al., 2018) to obtain side effect frequencies,and generated 5 frequency classes: very rare (=1), rare (=2) , infrequent (=3), frequent (=4), and very frequent (=5). A matrix, R , composed of 994 various side effects and 759 diverse drugs, with 37,441 known associationswasgenerated.Itisimportantto note that about 95% of associations were unobserved, and therefore coded as zero. The algorithm decomposed the matrix into twomatricesW(numberofdrugsxnumber of latent features) and H (number of latent features x number of side effects), which weremultipliedtoobtain ,themodelof R .
156
Made with FlippingBook - Online Brochure Maker