Integrative relational machine-learning for understanding drug side-effect profiles

Chemical space: drugs and their properties

The SIDER database contains drug side-effect relationships [3]. DrugBank is used to collect data such as categories and targets [11]. The join between SIDER and DrugBank is based on the PubChem Compound identifier given by SIDER and DrugBank. A total of 554 drugs from SIDER are referenced in DrugBank v3.0.

Each drug is described by its category and a set of clusters it belongs to. In fact, various structural representations and associated similarity measures were used to cluster drugs. The first similarity measure is based on SMILES representation. The SMILES codes are converted thanks to Open Babel program into fingerprints which allows linear and ring substructures to be identified [12]. Then, the structural similarity between two molecular fingerprints is calculated using the Tanimoto measure. In addition, we calculated three other similarity scores using spherical harmonics representation of molecules. This parametric representation of macromolecular surface was originally proposed and applied by Ritchie and Kemp [13] and Cai et al. [14]. The proprietary program HPCC (Harmonic Pharma) supports three variants of the spherical harmonic representation. HPCCgeo uses spherical harmonic coefficients (shape information) to calculate similarity between drugs, HPCCchem is based on chemical properties mapped on the spherical harmonic representation, and HPCCcombo combines shape and chemical information. Ward’s method is used to perform four hierarchical clusterings of drugs [15]. The optimal numbers of clusters is determined by the method of Kelley and al. [16]. Thus, 60 clusters are obtained with Tanimoto, 53 with HPCCgeo, 21 with HPCCchem and 34 with HPCCcombo measures.

Drug categories are retrieved from DrugBank. These categories are mapped on the descendants of three MeSH concepts, namely “Molecular Mechanisms of Pharmacological Action” (D27.505.519), “Physiological Effects of Drugs” (D27.505.696) and “Therapeutic Uses” (D27.505.954).

Biological space: proteins and their properties

Drug targets are extracted from both DrugBank and PDB [17]. The outer join between PDB and DrugBank (retaining all DrugBank targets) is based on SMILES code identity. Drug targets are associated with their UniProt accession numbers. Thus, 768 targets are collected, representing an average of four targets per drug. Then, target annotations are retrieved from different databases. Protein-protein interactions are retrieved from the IntAct database [18] and 5959 interactions were collected which correspond to 2827 new proteins. For all the proteins (drug targets and their interactants), 1403 pathway names are extracted from the KEGG database and the Pathway Interaction Database which integrates data from NCI-Nature, BioCarta and Reactome [19, 20]. For the same proteins, GO terms are also collected from QuickGO database [21]. Thus, 6494 GO terms annotating the 3595 proteins are stored in NetworkDB. Moreover, the “is_a” and “part_of” relationships between GO terms are stored in NetworkDB. Finally, 4650 protein domains associated with the targets and their interactants are retrieved from InterPro [22].

Association of drugs with side effects

The association between drugs and TCs is an important step for the characterization of drugs sharing side effects. As the TC size varies from 2 to 59 terms, it seems consistent to use a heuristic procedure depending on the TC size. Let k _i be the number of terms in T C _i and n _i be the minimal number of side effects required for assigning T C _i to a drug. Considering n _i  = 1 for any T C _i results in a very loose association yielding a very dense binary table hampering further computation, whereas considering n _i  = k _i for any T C _i results in a very stringent association which might skip over important drug side effects. In fact a trade-off between these two extreme solutions is required. Grouping the k _i values into 5-range intervals with the last interval from 21 to 59 allows to set up a simple association procedure ranging n _i from 1 to 5. The resulting association between drugs and TCs is shown in Figure 3 where each row represents the side-effect binary fingerprint associated with a drug. This binary table (drug ×TC) is then used to discover interesting side-effect profiles defined here as the longest combinations of TCs shared by significant sets of drugs.

Single-table datasets

Single table datasets designed for DT learning represent each drug by an attribute-value vector. Four types of descriptors retrieved from NetworkDB are used to generate these attributes: the first is the class information, i.e. the studied SEP, the second one includes drug categories, the third one lists all drug targets with for each target, three attributes referring to the type of action of the drug (activation, inhibition and other) and the fourth concerns clusters of similar drugs according to the four similarity measures described above. Because of target and category multiplicity, the total dimension of this dataset varies between 741 and 924 depending on the SEP.

Relational datasets

Relational datasets designed for Inductive Logic Programming (ILP) consist in a set of tables extracted from NetworkDB describing drugs properties and background knowledge. Drugs properties are the same as in the single-table dataset, i.e. categories, targets and clusters. Background knowledge includes GO annotations, domain composition, interactants and pathways of each drug target. Relationships between GO terms constitute an additional table.

Maximal frequent itemsets

In a binary table (object ×attribute), a frequent itemset is a group of attributes shared by a number of objects greater than a threshold support. A frequent itemset is considered as a maximal frequent itemset (MFI) if all its proper supersets are not frequent [25]. It follows that two maximal frequent itemsets (MFIs) cannot be shared by a number of objects greater than the threshold support. In our case, MFIs are the largest combinations of TCs shared by a number of drugs greater than 100. This threshold was chosen as a trade-off between high values yielding short MFIs limited to one or two TCs and low values yielding numerous MFIs covering only a few molecules. MFIs are extracted from the binary table (Figure 2) using the Coron program [26] after excluding TCs which cover more than 50% of the molecules.

Decision trees

Decision tree (DT) construction is a machine-learning method which uses (object ×attribute) table to classify objects. Results given by this method are easily readable. Decision trees are built here with the J48 implementation of C4.5 tree learner in the Weka toolbox using single table datasets converted into the ARFF format [27]. We use the default parameters except for two of them: we use minNumObj = 5 and binarySplits = true.

Inductive Logic Programming (ILP)

ILP is a machine-learning method which uses relational data as input and has been successfully applied to various areas including bioinformatics [28-30]. It allows us to learn a concept definition from observations, i.e, a set of positive examples (E+) and a set of negative examples (E-), and background knowledge (B) [31]. The ILP experiments produce theories as sets of first-order logic rules. They where conducted here with the Aleph Program [32]. Many parameters can be tuned for theory construction. The three main parameters are the min-pos, the noise and the induce-type. The min-pos parameter is the minimal number of positive examples that a rule must cover. The noise corresponds to the maximal number of negative examples that an acceptable rule may cover (in our case, one is never sure that a drug does not have a given side effect). The third parameter is induce-type which directs theory construction. When this parameter is set to induce-cover, overlapping rules are produced (i.e., a drug can be covered by several rules). Based on previous experience [33], we used the following settings: min-pos = 5, noise = 1 and induce-type = induce-cover.

Cross-validation

Both ILP theories and decision trees are evaluated with 10 runs of a 10-fold stratified cross-validation. DT cross-validation is performed with the Weka experimenter interface. For ILP, we took advantage of our recent integration of Aleph into the KNIME platform [34]. KNIME cross-validation meta-node is adapted for theory evaluation. An example is predicted as positive if it is covered by at least one rule. Each cross-validation assay yields a confusion matrix counting true and false positives, as well as true and false negatives. Each assay is then evaluated by the calculation of accuracy (ratio of correctly classified instances), specificity (true negative rate) and sensitivity (true positive rate).

Direct test

Theories and decision trees are also evaluated by direct test. Drugs used for testing are those present in SIDER 2 and DrugBank (v3.0) but not present in SIDER. For these drugs all descriptors are retrieved and stored in the NetworkDB. Furthermore, the reports of FAERS from 2004 to 2011 were imported as a database and used as an external information source for checking the false positives predicted by our models. We consider that a molecule is associated with a SEP in FAERS if for each TC of the SEP there is at least one report that states that the molecule is the primary suspect of an observed side effect belonging to the TC. Our checking procedure is just an anticipation as it relies on the fact that updating the package insert of a drug (stored in SIDER) requires that sufficient amount of adverse effect incidents occur (especially for new drugs).