Search

Deep neural networks (DNNs) are complex computational models that have found great success in many artificial intelligence applications, such as computer vision1,2 and natural language processing.3,4 In the past four years, DNNs have also generated promising results for quantitative structure–activity relationship (QSAR) tasks.5,6 Previous work showed that DNNs can routinely make better predictions than traditional methods, such as random forests, on a diverse collection of QSAR data sets. It was also found that multitask DNN modelsthose trained on and predicting multiple QSAR properties simultaneouslyoutperform DNNs trained separately on the individual data sets in many, but not all, tasks. To date there has been no satisfactory explanation of why the QSAR of one task embedded in a multitask DNN can borrow information from other unrelated QSAR tasks. Thus, using multitask DNNs in a way that consistently provides a predictive advantage becomes a challenge. In this work, we explored why multitask DNNs make a difference in predictive performance. Our results show that during prediction a multitask DNN does borrow “signal” from molecules with similar structures in the training sets of the other tasks. However, whether this borrowing leads to better or worse predictive performance depends on whether the activities are correlated. On the basis of this, we have developed a strategy to use multitask DNNs that incorporate prior domain knowledge to select training sets with correlated activities, and we demonstrate its effectiveness on several examples.

In QSAR, a statistical model is generated from a training set of molecules (represented by chemical descriptors) and their biological activities. We will call this traditional type of QSAR model an “activity model”. The activity model can be used to predict the activities of molecules not in the training set. A relatively new subfield for QSAR is domain applicability. The aim is to estimate the reliability of prediction of a specific molecule on a specific activity model. A number of different metrics have been proposed in the literature for this purpose. It is desirable to build a quantitative model of reliability against one or more of these metrics. We can call this an “error model”. A previous publication from our laboratory (Sheridan J. Chem. Inf. Model.,2012, 52, 814–823.) suggested the simultaneous use of three metrics would be more discriminating than any one metric. An error model could be built in the form of a three-dimensional set of bins. When the number of metrics exceeds three, however, the bin paradigm is not practical. An obvious solution for constructing an error model using multiple metrics is to use a QSAR method, in our case random forest. In this paper we demonstrate the usefulness of this paradigm, specifically for determining whether a useful error model can be built and which metrics are most useful for a given problem. For the ten data sets and for the seven metrics we examine here, it appears that it is possible to construct a useful error model using only two metrics (TREE_SD and PREDICTED). These do not require calculating similarities/distances between the molecules being predicted and the molecules used to build the activity model, which can be rate-limiting.

The idea of a “transformation”, making a small change to a chemical structure, for instance removing or replacing a substituent, is familiar to chemists. We suggest two ways of representing a transformation in silico, as a substructure descriptor difference vector, and as the set of atoms remaining once a maximum common substructure is eliminated. Such transformations can be filtered sensibly, and it is easy to compare one transformation to another. These representations have two applications. First, we can use these methods to automatically organize and display sets of closely related compounds such that any consistent local QSAR in a data set can be easily seen, the T-ANALYZE application. Second, we can suggest to a chemist how to change a molecule “on hand” to a more active one based on local QSAR for that activity, the T-MORPH application.

How well can a QSAR model predict the activity of a molecule not in the training set used to create the model? A set of retrospective cross-validation experiments using 20 diverse in-house activity sets were done to find a good discriminator of prediction accuracy as measured by root-mean-square difference between observed and predicted activity. Among the measures we tested, two seem useful:  the similarity of the molecule to be predicted to the nearest molecule in the training set and/or the number of neighbors in the training set, where neighbors are those more similar than a user-chosen cutoff. The molecules with the highest similarity and/or the most neighbors are the best-predicted. This trend holds true for narrow training sets and, to a lesser degree, for many diverse training sets and does not depend on which QSAR method or descriptor is used. One may define the similarity using a different descriptor than that used for the QSAR model. The similarity dependence for diverse training sets is somewhat unexpected. It appears to be greater for those data sets where the association of similar activities vs similar structures (as encoded in the Patterson plot) is stronger. We propose a way to estimate the reliability of the prediction of an arbitrary chemical structure on a given QSAR model, given the training set from which the model was derived.

There are a number of licensed databases that assign biological activities to druglike compounds. The MDL Drug Data Report (MDDR), compiled from the patent literature, is a popular example. It contains several hundred distinct activities, some of which are therapeutic areas (e.g., Antihypertensive) and some of which are related to specific enzymes or receptors (e.g., ACE inhibitor). There are several data mining applications where it would be useful to calculate a similarity between any two activities. Two distinct activity labels can have a significant similarity for a number of reasons:  two activities can be nearly synonymous (e.g., CCK B antagonist vs Gastrin antagonist), one activity may be a subset of another (e.g., Dopamine (D2) agonist vs Dopamine agonist), or an activity can be the mechanism by which another activity works (e.g., ACE inhibitor vs Antihypertensive), etc. In an ideal world, similarities for two activities could be calculated simply by comparing the compounds they have in common, but in hand-curated databases such as the MDDR the assignment of activities to compounds are inevitably inconsistent and incomplete. We propose a number of methods of calculating activity−activity similarities that hopefully compensate for errors in hand-curation. Two of these, TIMI and trend vector, show promise. Soft clustering of the activities using a union of similarity methods shows a reasonable association of therapeutic areas with their mechanisms.

A new classification and regression tool, Random Forest, is introduced and investigated for predicting a compound's quantitative or categorical biological activity based on a quantitative description of the compound's molecular structure. Random Forest is an ensemble of unpruned classification or regression trees created by using bootstrap samples of the training data and random feature selection in tree induction. Prediction is made by aggregating (majority vote or averaging) the predictions of the ensemble. We built predictive models for six cheminformatics data sets. Our analysis demonstrates that Random Forest is a powerful tool capable of delivering performance that is among the most accurate methods to date. We also present three additional features of Random Forest:  built-in performance assessment, a measure of relative importance of descriptors, and a measure of compound similarity that is weighted by the relative importance of descriptors. It is the combination of relatively high prediction accuracy and its collection of desired features that makes Random Forest uniquely suited for modeling in cheminformatics.

We have developed a method, given a database of molecules and associated activities, to identify molecular substructures that are associated with many different biological activities. These may be therapeutic areas (e.g. antihypertensive) and/or mechanism-based activities (e.g. renin inhibitor). This information helps us avoid chemical classes that are likely to have unanticipated side effects and also can suggest combinatorial libraries that might have activity on a variety of receptor targets. The method was applied to the USPDI and MDDR databases. There are clearly substructures in each database that occur in many compounds and span a variety of therapeutic categories. Some of these are expected, but some are not.

We have written a method that extracts one-to-one replacements of chemical groups in pairs of drug-like molecules with the same biological activity and counts the frequency of the replacements in a large collection of such molecules. There are two variations on the method that differ in their treatment of replacements in rings. This method is one possible approach to systematically identify candidate bioisosteres. Here we look at the MDDR database because it has a large diversity of drug-like compounds in a large number of therapeutic areas. The most frequent replacements in MDDR seem generally consistent with medicinal chemistry intuition about what chemical groups are equivalent or with groups that are easily converted by synthetic or metabolic pathways. This method can be applied to any set of molecules wherein the molecules can be paired by similar biological activity.

Similarity searches based on chemical descriptors have proven extremely useful in aiding large-scale drug screening. Typically an investigator starts with a “probe”, a drug-like molecule with an interesting biological activity, and searches a database to find similar compounds. In some projects, however, the only known actives are peptides, and the investigator needs to identify drug-like actives. 3D similarity methods are able to help in this endeavor but suffer from the necessity of having to specify the active conformation of the probe, something that is not always possible at the beginning of a project. Also, 3D methods are slow and are complicated by the need to generate low-energy conformations. In contrast, topological methods are relatively rapid and do not depend on conformation. However, unmodified topological similarity methods, given a peptide probe, will preferentially select other peptides from a database. In this paper we show some simple protocols that, if used with a standard topological similarity search method, are sufficient to select nonpeptide actives given a peptide probe. We demonstrate these protocols by using 10 peptide-like probes to select appropriate nonpeptide actives from the MDDR database.

Compounds are often synthesized and tested as mixtures. We propose the idea that the descriptor representation of a mixture may be approximated as the descriptor average of its individual component molecules. This centroid approximation has several potential advantages:  the representation is very compact, calculating similarities and deriving structure−activity relationships (SARs) of mixtures involves very little computation, and existing software can be directly applied to mixtures as if they were single molecules. Here we use the atom pair and topological torsion descriptors. We run several types of simulations using mixtures composed of druglike molecules from the MDL Drug Data Report database. We show that similarity searches using mixtures as queries and/or database entries yield reasonable results, with the caveat that a correction is necessary for mixture−mixture comparisons where at least one of the mixtures contains very diverse molecules. We also show that predictive SARs in the form of trend vectors can be derived from mixtures.

Similarity searches using topological descriptors have proved extremely useful in aiding large-scale screening. In this paper we describe the geometric atom pair, the 3D analog of the topological atom pair descriptor (Carhart et al. J. Chem. Inf. Comput. Sci. 1985, 25, 64−73). We show the results of geometric similarity searches using the CONCORD-build structures of typical small druglike molecules as probes. The database to be searched is a 3D version of the Derwent Standard Drug File that contains an average of 10 explicit conformations per compound. Using objective criteria for determining how good a descriptor is in selecting active compounds from large databases, we compare the results using the geometric versus the topological atom pair. We find that geometric and topological atom pairs are about equally effective in selecting active compounds from large databases. How the two types of descriptors rank active compounds is generally about the same as well, but occasionally active compounds will be seen as very similar to a probe in geometric descriptors, but as fairly dissimilar in topological descriptors. These are of two types:  (1) compounds where equivalent groups are in the same spatial arrangement as in the probe but are connected by very different bond paths and (2) compounds that can superimpose onto the probe when they are in a folded conformation.

Similarity searches using topological descriptors have proved extremely useful in aiding large-scale screening. We describe alternative forms of the atom pair (Carhart et al. J. Chem. Inf. Comput. Sci. 1985, 25, 64−73.) and topological torsion (Nilakantan et al. J. Chem. Inf. Comput. Sci. 1987, 27, 82−85.) descriptors that use physiochemical atom types. These types are based on binding property class, atomic log Pcontribution, and partial atomic charges. The new descriptors are meant to be more “fuzzy” than the original descriptors. We propose objective criteria for determining how effective one descriptor is versus another in selecting active compounds from large databases. Using these criteria, we run similarity searches over the Derwent Standard Drug File with ten typical druglike probes. The new descriptors are not as good as the original descriptors in selecting actives if one considers the average over all probes, but the new descriptors do better for several individual probes. Generally we find that whether one descriptor does better than another varies from probe to probe in a way almost impossible to predict a priori. Most importantly, we find that different descriptors typically select very different sets of actives. Thus it is advantageous to run similarity probes with several types of descriptors.

We present a method for detecting meaningful common topological substructures among sets of active compounds. A clique-based subgraph detection method is used to find the highest scoring common substructure (HSCS) for each pair of molecules. Only those HSCSs are kept that have a much larger score than would be expected by chance for a randomly selected pair of molecules of the same size. Information on these significant HSCSs is visualized in two complementary ways. Individual HSCSs can be displayed as pairs of molecules with the atoms in the HSCS highlighted. The HSCSs are presented in order of decreasing statistical significance. Alternatively, individual molecules can be displayed such that each atom is labeled with the number of significant HSCSs in which it has appeared. These are presented in order of decreasing maximum number per molecule. Browsing these data gives an impression of what parts of molecules are conserved among actives. The molecules can be simplified by deleting parts that are not conserved. We show three examples taken from the MDDR database. One example highlights common substructures among diverse CCK antagonists. Two examples point out substructures common between actives on different receptors:  benzodiazepine agonists vs CCK antagonists and antidepressants vs antihistamines.