Cartesian message passing neural networks for directional properties: Fast and transferable atomic multipoles.

The message passing neural network (MPNN) framework is a promising tool for modeling atomic properties but is, until recently, incompatible with directional properties, such as Cartesian tensors. We propose a modified Cartesian MPNN (CMPNN) suitable for predicting atom-centered multipoles, an essential component of ab initio force fields. The efficacy of this model is demonstrated on a newly developed dataset consisting of 46 623 chemical structures and corresponding high-quality atomic multipoles, which was deposited into the publicly available Molecular Sciences Software Institute QCArchive server. We show that the CMPNN accurately predicts atom-centered charges, dipoles, and quadrupoles and that errors in the predicted atomic multipoles have a negligible effect on multipole-multipole electrostatic energies. The CMPNN is accurate enough to model conformational dependencies of a molecule's electronic structure. This opens up the possibility of recomputing atomic multipoles on the fly throughout a simulation in which they might exhibit strong conformational dependence.

CLIFF: A component-based, machine-learned, intermolecular force field.

Computation of intermolecular interactions is a challenge in drug discovery because accurate ab initio techniques are too computationally expensive to be routinely applied to drug-protein models. Classical force fields are more computationally feasible, and force fields designed to match symmetry adapted perturbation theory (SAPT) interaction energies can remain accurate in this context. Unfortunately, the application of such force fields is complicated by the laborious parameterization required for computations on new molecules. Here, we introduce the component-based machine-learned intermolecular force field (CLIFF), which combines accurate, physics-based equations for intermolecular interaction energies with machine-learning models to enable automatic parameterization. The CLIFF uses functional forms corresponding to electrostatic, exchange-repulsion, induction/polarization, and London dispersion components in SAPT. Molecule-independent parameters are fit with respect to SAPT2+(3)δMP2/aug-cc-pVTZ, and molecule-dependent atomic parameters (atomic widths, atomic multipoles, and Hirshfeld ratios) are obtained from machine learning models developed for C, N, O, H, S, F, Cl, and Br. The CLIFF achieves mean absolute errors (MAEs) no worse than 0.70 kcal mol-1 in both total and component energies across a diverse dimer test set. For the side chain-side chain interaction database derived from protein fragments, the CLIFF produces total interaction energies with an MAE of 0.27 kcal mol-1 with respect to reference data, outperforming similar and even more expensive methods. In applications to a set of model drug-protein interactions, the CLIFF is able to accurately rank-order ligand binding strengths and achieves less than 10% error with respect to SAPT reference values for most complexes.

AP-Net: An atomic-pairwise neural network for smooth and transferable interaction potentials.

Intermolecular interactions are critical to many chemical phenomena, but their accurate computation using ab initio methods is often limited by computational cost. The recent emergence of machine learning (ML) potentials may be a promising alternative. Useful ML models should not only estimate accurate interaction energies but also predict smooth and asymptotically correct potential energy surfaces. However, existing ML models are not guaranteed to obey these constraints. Indeed, systemic deficiencies are apparent in the predictions of our previous hydrogen-bond model as well as the popular ANI-1X model, which we attribute to the use of an atomic energy partition. As a solution, we propose an alternative atomic-pairwise framework specifically for intermolecular ML potentials, and we introduce AP-Net-a neural network model for interaction energies. The AP-Net model is developed using this physically motivated atomic-pairwise paradigm and also exploits the interpretability of symmetry adapted perturbation theory (SAPT). We show that in contrast to other models, AP-Net produces smooth, physically meaningful intermolecular potentials exhibiting correct asymptotic behavior. Initially trained on only a limited number of mostly hydrogen-bonded dimers, AP-Net makes accurate predictions across the chemically diverse S66x8 dataset, demonstrating significant transferability. On a test set including experimental hydrogen-bonded dimers, AP-Net predicts total interaction energies with a mean absolute error of 0.37 kcal mol-1, reducing errors by a factor of 2-5 across SAPT components from previous neural network potentials. The pairwise interaction energies of the model are physically interpretable, and an investigation of predicted electrostatic energies suggests that the model "learns" the physics of hydrogen-bonded interactions.

Approaches for machine learning intermolecular interaction energies and application to energy components from symmetry adapted perturbation theory.

Accurate prediction of intermolecular interaction energies is a fundamental challenge in electronic structure theory due to their subtle character and small magnitudes relative to total molecular energies. Symmetry adapted perturbation theory (SAPT) provides rigorous quantum mechanical means for computing such quantities directly and accurately, but for a computational cost of at least O(N5), where N is the number of atoms. Here, we report machine learned models of SAPT components with a computational cost that scales asymptotically linearly, O(N). We use modified multi-target Behler-Parrinello neural networks and specialized intermolecular symmetry functions to address the idiosyncrasies of the intermolecular problem, achieving 1.2 kcal mol-1 mean absolute errors on a test set of hydrogen bound complexes including structural data extracted from the Cambridge Structural Database and Protein Data Bank, spanning an interaction energy range of 20 kcal mol-1. Additionally, we recover accurate predictions of the physically meaningful SAPT component energies, of which dispersion and induction/polarization were the easiest to predict and electrostatics and exchange-repulsion are the most difficult.

In-Silico Extraction of Design Ideas Using MMPA-by-QSAR and its Application on ADME Endpoints.

Matched molecular pair analysis (MMPA) has emerged as a powerful approach to mine and extract tacit knowledge from measured databases of small molecules. Extracted knowledge from past experimentation can assist future lead optimization as an idea generation tool and, hence, reduce the number of design-synthesis-test cycles. While attractive and intuitive, MMPA still presents several limitations. Analyses of internal absorption, distribution, metabolism, and excretion (ADME) databases of measured compounds show that chemical transformations with 10 pairs or more represent less than 1% of the total transforms identified by MMPA. A great wealth of design ideas remains effectively untapped and underutilized as the lack of measured data hinders extraction of robust trends. In this study we report the use of a quantitative structure-activity relationship (QSAR) model augmented MMPA approach (MMPA-by-QSAR) to infer the overall effect of chemical transformations on two essential ADME endpoints-lipophilicity and metabolic clearance. First, QSAR models are employed to predict compound activities, and subsequently, MMPA is used to identify and to extract virtual trends. Results obtained from retrospective analyses showed the ability to predict magnitudes of change close to experimental ones for the majority of transforms from each ADME data set. In the case of the lipophilicity endpoint (SFLogD) 73.7%, 87.85%, and 99% of transforms were predicted within 0.1, 0.15, and 0.3 units of the actual change. In the case of the clearance endpoint (HLM) 67.2%, 82.3%, and 99.5% of transforms were predicted within 0.08, 0.11, and 0.3 log units, respectively. Prospective application of MMPA-by-QSAR on untested compounds identified several novel transforms not observed in our measured data sets. When MMPs from these transforms were screened in our internal assays, it was found that the correct directionality of change was predicted for all but one of the tested transforms, and the predicted magnitudes of change have varying errors between predicted and measured mean changes ranging from 0.01 to 0.24 units for SFLogD and from 0.0 to 0.38 log units for HLM. This proposed MMPA-by-QSAR modeling approach has the potential to allow exploration of infrequent transforms or even identify completely novel transforms where no measured MMP is available.

How diverse are diversity assessment methods? A comparative analysis and benchmarking of molecular descriptor space.

Chemical diversity is a widely applied approach to select structurally diverse subsets of molecules, often with the objective of maximizing the number of hits in biological screening. While many methods exist in the area, few systematic comparisons using current descriptors in particular with the objective of assessing diversity in bioactivity space have been published, and this shortage is what the current study is aiming to address. In this work, 13 widely used molecular descriptors were compared, including fingerprint-based descriptors (ECFP4, FCFP4, MACCS keys), pharmacophore-based descriptors (TAT, TAD, TGT, TGD, GpiDAPH3), shape-based descriptors (rapid overlay of chemical structures (ROCS) and principal moments of inertia (PMI)), a connectivity-matrix-based descriptor (BCUT), physicochemical-property-based descriptors (prop2D), and a more recently introduced molecular descriptor type (namely, "Bayes Affinity Fingerprints"). We assessed both the similar behavior of the descriptors in assessing the diversity of chemical libraries, and their ability to select compounds from libraries that are diverse in bioactivity space, which is a property of much practical relevance in screening library design. This is particularly evident, given that many future targets to be screened are not known in advance, but that the library should still maximize the likelihood of containing bioactive matter also for future screening campaigns. Overall, our results showed that descriptors based on atom topology (i.e., fingerprint-based descriptors and pharmacophore-based descriptors) correlate well in rank-ordering compounds, both within and between descriptor types. On the other hand, shape-based descriptors such as ROCS and PMI showed weak correlation with the other descriptors utilized in this study, demonstrating significantly different behavior. We then applied eight of the molecular descriptors compared in this study to sample a diverse subset of sample compounds (4%) from an initial population of 2587 compounds, covering the 25 largest human activity classes from ChEMBL and measured the coverage of activity classes by the subsets. Here, it was found that "Bayes Affinity Fingerprints" achieved an average coverage of 92% of activity classes. Using the descriptors ECFP4, GpiDAPH3, TGT, and random sampling, 91%, 84%, 84%, and 84% of the activity classes were represented in the selected compounds respectively, followed by BCUT, prop2D, MACCS, and PMI (in order of decreasing performance). In addition, we were able to show that there is no visible correlation between compound diversity in PMI space and in bioactivity space, despite frequent utilization of PMI plots to this end. To summarize, in this work, we assessed which descriptors select compounds with high coverage of bioactivity space, and can hence be used for diverse compound selection for biological screening. In cases where multiple descriptors are to be used for diversity selection, this work describes which descriptors behave complementarily, and can hence be used jointly to focus on different aspects of diversity in chemical space.

Chemogenomics approaches to rationalizing the mode-of-action of traditional Chinese and Ayurvedic medicines.

Traditional Chinese medicine (TCM) and Ayurveda have been used in humans for thousands of years. While the link to a particular indication has been established in man, the mode-of-action (MOA) of the formulations often remains unknown. In this study, we aim to understand the MOA of formulations used in traditional medicine using an in silico target prediction algorithm, which aims to predict protein targets (and hence MOAs), given the chemical structure of a compound. Following this approach we were able to establish several links between suggested MOAs and experimental evidence. In particular, compounds from the 'tonifying and replenishing medicinal' class from TCM exhibit a hypoglycemic effect which can be related to activity of the ingredients against the Sodium-Glucose Transporters (SGLT) 1 and 2 as well as Protein Tyrosine Phosphatase (PTP). Similar results were obtained for Ayurvedic anticancer drugs. Here, both primary anticancer targets (those directly involved in cancer pathogenesis) such as steroid-5-alpha-reductase 1 and 2 were predicted as well as targets which act synergistically with the primary target, such as the efflux pump P-glycoprotein (P-gp). In addition, we were able to elucidate some targets which may point us to novel MOAs as well as explain side effects. Most notably, GPBAR1, which was predicted as a target for both 'tonifying and replenishing medicinal' and anticancer classes, suggests an influence of the compounds on metabolism. Understanding the MOA of these compounds is beneficial as it provides a resource for NMEs with possibly higher efficacy in the clinic than those identified by single-target biochemical assays.

In silico target predictions: defining a benchmarking data set and comparison of performance of the multiclass Naïve Bayes and Parzen-Rosenblatt window.

In this study, two probabilistic machine-learning algorithms were compared for in silico target prediction of bioactive molecules, namely the well-established Laplacian-modified Naïve Bayes classifier (NB) and the more recently introduced (to Cheminformatics) Parzen-Rosenblatt Window. Both classifiers were trained in conjunction with circular fingerprints on a large data set of bioactive compounds extracted from ChEMBL, covering 894 human protein targets with more than 155,000 ligand-protein pairs. This data set is also provided as a benchmark data set for future target prediction methods due to its size as well as the number of bioactivity classes it contains. In addition to evaluating the methods, different performance measures were explored. This is not as straightforward as in binary classification settings, due to the number of classes, the possibility of multiple class memberships, and the need to translate model scores into "yes/no" predictions for assessing model performance. Both algorithms achieved a recall of correct targets that exceeds 80% in the top 1% of predictions. Performance depends significantly on the underlying diversity and size of a given class of bioactive compounds, with small classes and low structural similarity affecting both algorithms to different degrees. When tested on an external test set extracted from WOMBAT covering more than 500 targets by excluding all compounds with Tanimoto similarity above 0.8 to compounds from the ChEMBL data set, the current methodologies achieved a recall of 63.3% and 66.6% among the top 1% for Naïve Bayes and Parzen-Rosenblatt Window, respectively. While those numbers seem to indicate lower performance, they are also more realistic for settings where protein targets need to be established for novel chemical substances.

Computer-aided (in silico) approaches in the mode-of-action analysis and safety assessment of ostarine and 4-methylamphetamine.

OBJECTIVE: This study exemplifies computer-aided (in silico) approaches in assessing the risks of new psychoactive substances emerging in the European Union. In this work, we (i) consider the potential of Ostarine exhibiting psychoactivity and (ii) anticipate potential activities and toxicities of 4-methylamphetamine. METHOD: The approach, termed in silico target prediction, suggests potential protein targets modulated by compounds given their chemical structure. This is achieved by first establishing the associations between chemical structure and protein targets using data from the bioactivity database, ChEMBL, via the use of two different computational algorithms. On the basis of the associations, protein targets and consequently the mode of action of novel compounds were predicted. RESULTS: For Ostarine, none of the targets anticipated are currently known to elicit psychoactivity. Furthermore, Ostarine is unlikely to cross the blood-brain barrier to reach relevant target sites on the basis of its physicochemical properties. For 4-methylamphetamine, toxicities were anticipated, that is, serotonin syndrome (based on the prediction of SERT) and other effects similar to related substances, that is, methamphetamine. CONCLUSION: From the two case studies, we showed that in silico target prediction appears to have potential in assessing new psychoactive compounds where experimental data are scarce. The applicability domain of target predictions when applied to psychoactive compounds needs to be established in future work.

Deep-learning: investigating deep neural networks hyper-parameters and comparison of performance to shallow methods for modeling bioactivity data.

BACKGROUND: In recent years, research in artificial neural networks has resurged, now under the deep-learning umbrella, and grown extremely popular. Recently reported success of DL techniques in crowd-sourced QSAR and predictive toxicology competitions has showcased these methods as powerful tools in drug-discovery and toxicology research. The aim of this work was dual, first large number of hyper-parameter configurations were explored to investigate how they affect the performance of DNNs and could act as starting points when tuning DNNs and second their performance was compared to popular methods widely employed in the field of cheminformatics namely Naïve Bayes, k-nearest neighbor, random forest and support vector machines. Moreover, robustness of machine learning methods to different levels of artificially introduced noise was assessed. The open-source Caffe deep-learning framework and modern NVidia GPU units were utilized to carry out this study, allowing large number of DNN configurations to be explored. RESULTS: We show that feed-forward deep neural networks are capable of achieving strong classification performance and outperform shallow methods across diverse activity classes when optimized. Hyper-parameters that were found to play critical role are the activation function, dropout regularization, number hidden layers and number of neurons. When compared to the rest methods, tuned DNNs were found to statistically outperform, with p value <0.01 based on Wilcoxon statistical test. DNN achieved on average MCC units of 0.149 higher than NB, 0.092 than kNN, 0.052 than SVM with linear kernel, 0.021 than RF and finally 0.009 higher than SVM with radial basis function kernel. When exploring robustness to noise, non-linear methods were found to perform well when dealing with low levels of noise, lower than or equal to 20%, however when dealing with higher levels of noise, higher than 30%, the Naïve Bayes method was found to perform well and even outperform at the highest level of noise 50% more sophisticated methods across several datasets.

Improving the prediction of organism-level toxicity through integration of chemical, protein target and cytotoxicity qHTS data.

Prediction of compound toxicity is essential because covering the vast chemical space requiring safety assessment using traditional experimentally-based, resource-intensive techniques is impossible. However, such prediction is nontrivial due to the complex causal relationship between compound structure and in vivo harm. Protein target annotations and in vitro experimental outcomes encode relevant bioactivity information complementary to chemicals' structures. This work tests the hypothesis that utilizing three complementary types of data will afford predictive models that outperform traditional models built using fewer data types. A tripartite, heterogeneous descriptor set for 367 compounds was comprised of (a) chemical descriptors, (b) protein target descriptors generated using an algorithm trained on 190 000 ligand-protein interactions from ChEMBL, and (c) descriptors derived from in vitro cell cytotoxicity dose-response data from a panel of human cell lines. 100 random forests classification models for predicting rat LD50 were built using every combination of descriptors. Successive integration of data types improved predictive performance; models built using the full dataset had an average external correct classification rate of 0.82, compared to 0.73-0.80 for models built using two data types and 0.67-0.78 for models built using one. Pairwise comparisons of models trained on the same data showed that including a third data domain on top of chemistry improved average correct classification rate by 1.4-2.4 points, with p-values <0.01. Additionally, the approach enhanced the models' applicability domains and proved useful for generating novel mechanism hypotheses. The use of tripartite heterogeneous bioactivity datasets is a useful technique for improving toxicity prediction. Both protein target descriptors - which have the practical value of being derived in silico - and cytotoxicity descriptors derived from experiment are suitable contributors to such datasets.

Comparing global and local likelihood score thresholds in multiclass laplacian-modified Naive Bayes protein target prediction.

The increase of publicly available bioactivity data has led to the extensive development and usage of in silico bioactivity prediction algorithms. A particularly popular approach for such analyses is the multiclass Naïve Bayes, whose output is commonly processed by applying empirically-derived likelihood score thresholds. In this work, we describe a systematic way for deriving score cut-offs on a per-protein target basis and compare their performance with global thresholds on a large scale using both 5-fold cross-validation (ChEMBL 14, 189k ligand-protein pairs over 477 protein targets) and external validation (WOMBAT, 63k pairs, 421 targets). The individual protein target cut-offs derived were compared to global cut-offs ranging from -10 to 40 in score bouts of 2.5. The results indicate that individual thresholds had equal or better performance in all comparisons with global thresholds, ranging from 95% of protein targets to 57.96%. It is shown that local thresholds behave differently for particular families of targets (CYPs, GPCRs, Kinases and TFs). Furthermore, we demonstrate the discrepancy in performance when we move away from the training dataset chemical space, using Tanimoto similarity as a metric (from 0 to 1 in steps of 0.2). Finally, the individual protein score cut-offs derived for the in silico bioactivity application used in this work are released, as well as the reproducible and transferable KNIME workflows used to carry out the analysis.

Extending in silico mechanism-of-action analysis by annotating targets with pathways: application to cellular cytotoxicity readouts.

BACKGROUND: An in silico mechanism-of-action analysis protocol was developed, comprising molecule bioactivity profiling, annotation of predicted targets with pathways and calculation of enrichment factors to highlight targets and pathways more likely to be implicated in the studied phenotype. RESULTS: The method was applied to a cytotoxicity phenotypic endpoint, with enriched targets/pathways found to be statistically significant when compared with 100 random datasets. Application on a smaller apoptotic set (10 molecules) did not allowed to obtain statistically relevant results, suggesting that the protocol requires modification such as analysis of the most frequently predicted targets/annotated pathways. CONCLUSION: Pathway annotations improved the mechanism-of-action information gained by target prediction alone, allowing a better interpretation of the predictions and providing better mapping of targets onto pathways.

Anti-cancer drug development: computational strategies to identify and target proteins involved in cancer metabolism.

Cancer remains a fundamental burden to public health despite substantial efforts aimed at developing effective chemotherapeutics and significant advances in chemotherapeutic regimens. The major challenge in anti-cancer drug design is to selectively target cancer cells with high specificity. Research into treating malignancies by targeting altered metabolism in cancer cells is supported by computational approaches, which can take a leading role in identifying candidate targets for anti-cancer therapy as well as assist in the discovery and optimisation of anti-cancer agents. Natural products appear to have privileged structures for anti-cancer drug development and the bulk of this particularly valuable chemical space still remains to be explored. In this review we aim to provide a comprehensive overview of current strategies for computer-guided anti-cancer drug development. We start with a discussion of state-of-the art bioinformatics methods applied to the identification of novel anti-cancer targets, including machine learning techniques, the Connectivity Map and biological network analysis. This is followed by an extensive survey of molecular modelling and cheminformatics techniques employed to develop agents targeting proteins involved in the glycolytic, lipid, NAD+, mitochondrial (TCA cycle), amino acid and nucleic acid metabolism of cancer cells. A dedicated section highlights the most promising strategies to develop anti-cancer therapeutics from natural products and the role of metabolism and some of the many targets which are under investigation are reviewed. Recent success stories are reported for all the areas covered in this review. We conclude with a brief summary of the most interesting strategies identified and with an outlook on future directions in anti-cancer drug development.

Linking Ayurveda and Western medicine by integrative analysis.

In this article, we discuss our recent work in elucidating the mode-of-action of compounds used in traditional medicine including Ayurvedic medicine. Using computational ('in silico') approach, we predict potential targets for Ayurvedic anti-cancer compounds, obtained from the Indian Plant Anticancer Database given its chemical structure. In our analysis, we observed that: (i) the targets predicted can be connected to cancer pathogenesis i.e. steroid-5-alpha reductase 1 and 2 and estrogen receptor-ß, and (ii) predominantly hormone-dependent cancer targets were predicted for the anti-cancer compounds. Through the use of our in silico target prediction, we conclude that understanding how traditional medicine such as Ayurveda work through linking with the 'western' understanding of chemistry and protein targets can be a fruitful avenue in addition to bridging the gap between the two different schools of thinking. Given that compounds used in Ayurveda have been tested and used for thousands of years (although not in the same approach as Western medicine), they can potentially be developed into potential new drugs. Hence, to further advance the case of Ayurvedic medicine, we put forward some suggestions namely: (a) employing and integrating novel analytical methods given the advancements of 'omics' and (b) sharing experimental data and clinical results on studies done on Ayurvedic compounds in an easy and accessible way.

Diversity selection of compounds based on 'protein affinity fingerprints' improves sampling of bioactive chemical space.

Diversity selection is a frequently applied strategy for assembling high-throughput screening libraries, making the assumption that a diverse compound set increases chances of finding bioactive molecules. Based on previous work on experimental 'affinity fingerprints', in this study, a novel diversity selection method is benchmarked that utilizes predicted bioactivity profiles as descriptors. Compounds were selected based on their predicted activity against half of the targets (training set), and diversity was assessed based on coverage of the remaining (test set) targets. Simultaneously, fingerprint-based diversity selection was performed. An original version of the method exhibited on average 5% and an improved version on average 10% increase in target space coverage compared with the fingerprint-based methods. As a typical case, bioactivity-based selection of 231 compounds (2%) from a particular data set ('Cutoff-40') resulted in 47.0% and 50.1% coverage, while fingerprint-based selection only achieved 38.4% target coverage for the same subset size. In conclusion, the novel bioactivity-based selection method outperformed the fingerprint-based method in sampling bioactive chemical space on the data sets considered. The structures retrieved were structurally more acceptable to medicinal chemists while at the same time being more lipophilic, hence bioactivity-based diversity selection of compounds would best be combined with physicochemical property filters in practice.

From in silico target prediction to multi-target drug design: current databases, methods and applications.

Given the tremendous growth of bioactivity databases, the use of computational tools to predict protein targets of small molecules has been gaining importance in recent years. Applications span a wide range, from the 'designed polypharmacology' of compounds to mode-of-action analysis. In this review, we firstly survey databases that can be used for ligand-based target prediction and which have grown tremendously in size in the past. We furthermore outline methods for target prediction that exist, both based on the knowledge of bioactivities from the ligand side and methods that can be applied in situations when a protein structure is known. Applications of successful in silico target identification attempts are discussed in detail, which were based partly or in whole on computational target predictions in the first instance. This includes the authors' own experience using target prediction tools, in this case considering phenotypic antibacterial screens and the analysis of high-throughput screening data. Finally, we will conclude with the prospective application of databases to not only predict, retrospectively, the protein targets of a small molecule, but also how to design ligands with desired polypharmacology in a prospective manner.