Diverse Training Datasets Research Articles

Similarity-based machine learning support vector machine predictor of drug-drug interactions with improved accuracies.

Drug-drug interactions (DDI) are frequent causes of adverse clinical drug reactions. Efforts have been directed at the early stage to achieve accurate identification of DDI for drug safety assessments, including the development of in silico predictive methods. In particular, similarity-based in silico methods have been developed to assess DDI with good accuracies, and machine learning methods have been employed to further extend the predictive range of similarity-based approaches. However, the performance of a developed machine learning method is lower than expectations partly because of the use of less diverse DDI training data sets and a less optimal set of similarity measures. In this work, we developed a machine learning model using support vector machines (SVMs) based on the literature-reported established set of similarity measures and comprehensive training data sets. The established similarity measures include the 2D molecular structure similarity, 3D pharmacophoric similarity, interaction profile fingerprint (IPF) similarity, target similarity and adverse drug effect (ADE) similarity, which were extracted from well-known databases, such as DrugBank and Side Effect Resource (SIDER). A pairwise kernel was constructed for the known and possible drug pairs based on the five established similarity measures and then used as the input vector of the SVM. The 10-fold cross-validation studies showed a predictive performance of AUROC >0.97, which is significantly improved compared with the AUROC of 0.67 of an analogously developed machine learning model. Our study suggested that a similarity-based SVM prediction is highly useful for identifying DDI. in silico methods based on multifarious drug similarities have been suggested to be feasible for DDI prediction in various studies. In this way, our pairwise kernel SVM model had better accuracies than some previous works, which can be used as a pharmacovigilance tool to detect potential DDI.

ISEE: Interface structure, evolution, and energy-based machine learning predictor of binding affinity changes upon mutations.

Quantitative evaluation of binding affinity changes upon mutations is crucial for protein engineering and drug design. Machine learning‐based methods are gaining increasing momentum in this field. Due to the limited number of experimental data, using a small number of sensitive predictive features is vital to the generalization and robustness of such machine learning methods. Here we introduce a fast and reliable predictor of binding affinity changes upon single point mutation, based on a random forest approach. Our method, iSEE, uses a limited number of interface Structure, Evolution, and Energy‐based features for the prediction. iSEE achieves, using only 31 features, a high prediction performance with a Pearson correlation coefficient (PCC) of 0.80 and a root mean square error of 1.41 kcal/mol on a diverse training dataset consisting of 1102 mutations in 57 protein‐protein complexes. It competes with existing state‐of‐the‐art methods on two blind test datasets. Predictions for a new dataset of 487 mutations in 56 protein complexes from the recently published SKEMPI 2.0 database reveals that none of the current methods perform well (PCC < 0.42), although their combination does improve the predictions. Feature analysis for iSEE underlines the significance of evolutionary conservations for quantitative prediction of mutation effects. As an application example, we perform a full mutation scanning of the interface residues in the MDM2–p53 complex.

A universal strategy for the creation of machine learning-based atomistic force fields

Emerging machine learning (ML)-based approaches provide powerful and novel tools to study a variety of physical and chemical problems. In this contribution, we outline a universal strategy to create ML-based atomistic force fields, which can be used to perform high-fidelity molecular dynamics simulations. This scheme involves (1) preparing a big reference dataset of atomic environments and forces with sufficiently low noise, e.g., using density functional theory or higher-level methods, (2) utilizing a generalizable class of structural fingerprints for representing atomic environments, (3) optimally selecting diverse and non-redundant training datasets from the reference data, and (4) proposing various learning approaches to predict atomic forces directly (and rapidly) from atomic configurations. From the atomistic forces, accurate potential energies can then be obtained by appropriate integration along a reaction coordinate or along a molecular dynamics trajectory. Based on this strategy, we have created model ML force fields for six elemental bulk solids, including Al, Cu, Ti, W, Si, and C, and show that all of them can reach chemical accuracy. The proposed procedure is general and universal, in that it can potentially be used to generate ML force fields for any material using the same unified workflow with little human intervention. Moreover, the force fields can be systematically improved by adding new training data progressively to represent atomic environments not encountered previously.

Random Subclasses Ensembles by Using 1-Nearest Neighbor Framework

A classifier ensemble is a combination of diverse and accurate classifiers. Generally, a classifier ensemble performs better than any single classifier in the ensemble. Naive Bayes classifiers are simple but popular classifiers for many applications. As it is difficult to create diverse naive Bayes classifiers, naive Bayes ensembles are not very successful. In this paper, we propose Random Subclasses (RS) ensembles for Naive Bayes classifiers. In the proposed method, new subclasses for each class are created by using 1-Nearest Neighbor (1-NN) framework that uses randomly selected points from the training data. A classifier considers each subclass as a class of its own. As the method to create subclasses is random, diverse datasets are generated. Each classifier in an ensemble learns on one dataset from the pool of diverse datasets. Diverse training datasets ensure diverse classifiers in the ensemble. New subclasses create easy to learn decision boundaries that in turn create accurate naive Bayes classifiers. We developed two variants of RS, in the first variant RS(2), two subclasses per class were created whereas in the second variant RS(4), four subclasses per class were created. We studied the performance of these methods against other popular ensemble methods by using naive Bayes as the base classifier. RS(4) outperformed other popular ensemble methods. A detailed study was carried out to understand the behavior of RS ensembles.

Cross-platform normalization of microarray and RNA-seq data for machine learning applications.

Large, publicly available gene expression datasets are often analyzed with the aid of machine learning algorithms. Although RNA-seq is increasingly the technology of choice, a wealth of expression data already exist in the form of microarray data. If machine learning models built from legacy data can be applied to RNA-seq data, larger, more diverse training datasets can be created and validation can be performed on newly generated data. We developed Training Distribution Matching (TDM), which transforms RNA-seq data for use with models constructed from legacy platforms. We evaluated TDM, as well as quantile normalization, nonparanormal transformation, and a simple log2 transformation, on both simulated and biological datasets of gene expression. Our evaluation included both supervised and unsupervised machine learning approaches. We found that TDM exhibited consistently strong performance across settings and that quantile normalization also performed well in many circumstances. We also provide a TDM package for the R programming language.

Estimation of the chemical-induced eye injury using a weight-of-evidence (WoE) battery of 21 artificial neural network (ANN) c-QSAR models (QSAR-21): Part I: Irritation potential

Evaluation of potential chemical-induced eye injury through irritation and corrosion is required to ensure occupational and consumer safety for industrial, household and cosmetic ingredient chemicals. The historical method for evaluating eye irritant and corrosion potential of chemicals is the rabbit Draize test. However, the Draize test is controversial and its use is diminishing - the EU 7th Amendment to the Cosmetic Directive (76/768/EEC) and recast Regulation now bans marketing of new cosmetics having animal testing of their ingredients and requires non-animal alternative tests for safety assessments. Thus, in silico and/or in vitro tests are advocated. QSAR models for eye irritation have been reported for several small (congeneric) data sets; however, large global models have not been described. This report describes FDA/CFSAN’s development of 21 ANN c-QSAR models (QSAR-21) to predict eye irritation using the ADMET Predictor™ program and a diverse training data set of 2928 chemicals. The 21 models had external (20% test set) and internal validation and average training/verification/test set statistics were: 88/88/85(%) sensitivity and 82/82/82(%) specificity, respectively. The new method utilized multiple artificial neural network (ANN) molecular descriptor selection functionalities to maximize the applicability domain of the battery. The eye irritation models will be used to provide information to fill the critical data gaps for the safety assessment of cosmetic ingredient chemicals.

Random Ordinality Ensembles: Ensemble methods for multi-valued categorical data

Data with multi-valued categorical attributes can cause major problems for decision trees. The high branching factor can lead to data fragmentation, where decisions have little or no statistical support. In this paper, we propose a new ensemble method, Random Ordinality Ensembles (ROE), that reduces this problem, and provides significantly improved accuracies over current ensemble methods. We perform a random projection of the categorical data into a continuous space. As the transformation to continuous data is a random process, each dataset has a different imposed ordinality. A decision tree that learns on this new continuous space is able to use binary splits, hence reduces the data fragmentation problem. Generally, these binary trees are accurate. Diverse training datasets ensure diverse decision trees in the ensemble. We created two variants of the technique, ROE. In the first variant, we used decision trees as the base models for ensembles. In the second variant, we combined the attribute randomisation of Random Subspaces with Random Ordinality. These methods match or outperform other popular ensemble methods. Different properties of these ensembles were studied. The study suggests that random ordinality trees are generally more accurate and smaller than multi-way split decision trees. It is also shown that random ordinality attributes can be used to improve Bagging and AdaBoost.M1 ensemble methods.

Random Projection Random Discretization Ensembles—Ensembles of Linear Multivariate Decision Trees

In this paper, we present a novel ensemble method random projection random discretization ensembles (RPRDE) to create ensembles of linear multivariate decision trees by using a univariate decision tree algorithm. The present method combines the better computational complexity of a univariate decision tree algorithm with the better representational power of linear multivariate decision trees. We develop random discretization (RD) method that creates random discretized features from continuous features. Random projection (RP) is used to create new features that are linear combinations of original features. A new dataset is created by augmenting discretized features (created by using RD) with features created by using RP. Each decision tree of a RPRD ensemble is trained on one dataset from the pool of these datasets by using a univariate decision tree algorithm. As these multivariate decision trees (because of features created by RP) have more representational power than univariate decision trees, we expect accurate decision trees in the ensemble. Diverse training datasets ensure diverse decision trees in the ensemble. We study the performance of RPRDE against other popular ensemble techniques using C4.5 tree as the base classifier. RPRDE matches or outperforms other popular ensemble methods. Experiments results also suggest that the proposed method is quite robust to the class noise.

INTEGRATION OF RANDOM SAMPLE SELECTION, SUPPORT VECTOR MACHINES AND ENSEMBLES FOR FINANCIAL RISK FORECASTING WITH AN EMPIRICAL ANALYSIS ON THE NECESSITY OF FEATURE SELECTION

SUMMARYFinancial risk forecasting (FRF) is an effective tool to help people forecast whether or not a company will fail in future. Among all techniques of FRF, the support vector machine (SVM) is the most newly developed, and one of the most accurate and effective techniques. This study is devoted to investigating an ensemble model of FRF by integrating bagging with an SVM to generate a data‐driven SVM ensemble. Bagging is used to produce diverse training datasets on which multiple SVM classifiers are trained to make FRF for a target company. Simple voting is employed to produce a final decision from the SVM model committee. The empirical study has two objectives. One is to verify whether the data‐driven SVM ensemble can produce a more dominating performance than the most frequently used techniques in the area of FRF, i.e. multivariate discriminant analysis, logistics regression and a single SVM. The other is to verify whether feature selection is necessary to help the SVM make more precise FRF, although the SVM can handle high‐dimensional data. The results indicate that the data‐driven SVM ensemble significantly improves the predictive ability of SVM‐based FRF. Meanwhile, feature selection can effectively help the SVM achieve better predictive performance, which means that use of feature selection is necessary in SVM‐based FRF. Copyright © 2012 John Wiley &amp; Sons, Ltd.

Identification of DNA adduct formation of small molecules by molecular descriptors and machine learning methods

In this study, we developed new computational DNA adduct prediction models by using significantly more diverse training data-set of 217 DNA adducts and 1024 non-DNA adducts, and applying five machine learning methods which include support vector machine (SVM), k-nearest neighbour, artificial neural networks, logistic regression and continuous kernel discrimination. The molecular descriptors used for DNA adduct prediction were selected from a pool of 548 descriptors by using a multi-step hybrid feature selection method combining Fischer-score and Monte Carlo simulated annealing method. Some of the selected descriptors are consistent with the structural and physicochemical properties reported to be important for DNA adduct formation. The y-scrambling method was used to test whether there is a chance correlation in the developed SVM model. In the meantime, fivefold cross-validation of these machine learning methods results in the prediction accuracies of 64.1–82.5% for DNA adducts and 95.1–97.6% for non-DNA adducts, and the prediction accuracies for external test set are 78.2–100% for DNA adducts and 92.6–98.4% for non-DNA adducts. Our study suggested that the tested machine learning methods are potentially useful for DNA adducts identification.

Machine Vision Tools for Quantitatively Measuring Animal Behavior in Large Scale Experiments

Machine Vision Tools for Quantitatively Measuring Animal Behavior in Large Scale Experiments

An in silico ensemble method for lead discovery: decision forest

Recent progress in combinatorial chemistry and parallel synthesis has radically changed the approach to drug discovery in the pharmaceutical industry. At present, thousands of compounds can be made in a short period, creating a need for fast and effective in silico methods to select the most promising lead candidates. Decision forest is a novel pattern recognition method, which combines the results of multiple distinct but comparable decision tree models to reach a consensus prediction. In this article, a decision forest model was developed using a structurally diverse training data set containing 232 compounds whose estrogen receptor binding activity was tested at the U.S. Food and Drug Administration (FDA)'s National Center for Toxicological Research (NCTR). The model was subsequently validated using a test data set of 463 compounds selected from the literature, and then applied to a large data set with 57,145 compounds as a screening example. The results show that the decision forest method is a fast, reliable and effective in silico approach, which could be useful in drug discovery.