MolScore: a scoring, evaluation and benchmarking framework for generative models in de novo drug design.

Generative models are undergoing rapid research and application to de novo drug design. To facilitate their application and evaluation, we present MolScore. MolScore already contains many drug-design-relevant scoring functions commonly used in benchmarks such as, molecular similarity, molecular docking, predictive models, synthesizability, and more. In addition, providing performance metrics to evaluate generative model performance based on the chemistry generated. With this unification of functionality, MolScore re-implements commonly used benchmarks in the field (such as GuacaMol, MOSES, and MolOpt). Moreover, new benchmarks can be created trivially. We demonstrate this by testing a chemical language model with reinforcement learning on three new tasks of increasing complexity related to the design of 5-HT2a ligands that utilise either molecular descriptors, 266 pre-trained QSAR models, or dual molecular docking. Lastly, MolScore can be integrated into an existing Python script with just three lines of code. This framework is a step towards unifying generative model application and evaluation as applied to drug design for both practitioners and researchers. The framework can be found on GitHub and downloaded directly from the Python Package Index.Scientific ContributionMolScore is an open-source platform to facilitate generative molecular design and evaluation thereof for application in drug design. This platform takes important steps towards unifying existing benchmarks, providing a platform to share new benchmarks, and improves customisation, flexibility and usability for practitioners over existing solutions.

Augmented Hill-Climb increases reinforcement learning efficiency for language-based de novo molecule generation.

A plethora of AI-based techniques now exists to conduct de novo molecule generation that can devise molecules conditioned towards a particular endpoint in the context of drug design. One popular approach is using reinforcement learning to update a recurrent neural network or language-based de novo molecule generator. However, reinforcement learning can be inefficient, sometimes requiring up to 105 molecules to be sampled to optimize more complex objectives, which poses a limitation when using computationally expensive scoring functions like docking or computer-aided synthesis planning models. In this work, we propose a reinforcement learning strategy called Augmented Hill-Climb based on a simple, hypothesis-driven hybrid between REINVENT and Hill-Climb that improves sample-efficiency by addressing the limitations of both currently used strategies. We compare its ability to optimize several docking tasks with REINVENT and benchmark this strategy against other commonly used reinforcement learning strategies including REINFORCE, REINVENT (version 1 and 2), Hill-Climb and best agent reminder. We find that optimization ability is improved ~ 1.5-fold and sample-efficiency is improved ~ 45-fold compared to REINVENT while still delivering appealing chemistry as output. Diversity filters were used, and their parameters were tuned to overcome observed failure modes that take advantage of certain diversity filter configurations. We find that Augmented Hill-Climb outperforms the other reinforcement learning strategies used on six tasks, especially in the early stages of training or for more difficult objectives. Lastly, we show improved performance not only on recurrent neural networks but also on a reinforcement learning stabilized transformer architecture. Overall, we show that Augmented Hill-Climb improves sample-efficiency for language-based de novo molecule generation conditioning via reinforcement learning, compared to the current state-of-the-art. This makes more computationally expensive scoring functions, such as docking, more accessible on a relevant timescale.

Identification of anti-schistosomal, anthelmintic and anti-parasitic compounds curated and text-mined from the scientific literature.

More than a billion people are infected with parasitic worms, including nematodes, such as hookworms, and flatworms, such as blood flukes. Few drugs are available to treat worm infections, but high-throughput screening approaches hold promise to identify novel drug candidates. One problem for researchers who find an interesting 'hit' from a high-throughput screen is to identify whether that compound, or a similar compound has previously been published as having anthelmintic or anti-parasitic activity. Here, we present (i) data sets of 2,828 anthelmintic compounds, and 1,269 specific anti-schistosomal compounds, manually curated from scientific papers and books, and (ii) a data set of 24,335 potential anthelmintic and anti-parasitic compounds identified by text-mining PubMed abstracts. We provide their structures in simplified molecular-input line-entry system (SMILES) format so that researchers can easily compare 'hits' from their screens to these anthelmintic compounds and anti-parasitic compounds and find previous literature on them to support/halt their progression in drug discovery pipelines.

Comparison of structure- and ligand-based scoring functions for deep generative models: a GPCR case study.

Deep generative models have shown the ability to devise both valid and novel chemistry, which could significantly accelerate the identification of bioactive compounds. Many current models, however, use molecular descriptors or ligand-based predictive methods to guide molecule generation towards a desirable property space. This restricts their application to relatively data-rich targets, neglecting those where little data is available to sufficiently train a predictor. Moreover, ligand-based approaches often bias molecule generation towards previously established chemical space, thereby limiting their ability to identify truly novel chemotypes. In this work, we assess the ability of using molecular docking via Glide-a structure-based approach-as a scoring function to guide the deep generative model REINVENT and compare model performance and behaviour to a ligand-based scoring function. Additionally, we modify the previously published MOSES benchmarking dataset to remove any induced bias towards non-protonatable groups. We also propose a new metric to measure dataset diversity, which is less confounded by the distribution of heavy atom count than the commonly used internal diversity metric. With respect to the main findings, we found that when optimizing the docking score against DRD2, the model improves predicted ligand affinity beyond that of known DRD2 active molecules. In addition, generated molecules occupy complementary chemical and physicochemical space compared to the ligand-based approach, and novel physicochemical space compared to known DRD2 active molecules. Furthermore, the structure-based approach learns to generate molecules that satisfy crucial residue interactions, which is information only available when taking protein structure into account. Overall, this work demonstrates the advantage of using molecular docking to guide de novo molecule generation over ligand-based predictors with respect to predicted affinity, novelty, and the ability to identify key interactions between ligand and protein target. Practically, this approach has applications in early hit generation campaigns to enrich a virtual library towards a particular target, and also in novelty-focused projects, where de novo molecule generation either has no prior ligand knowledge available or should not be biased by it.

Updates to the Symbol Nomenclature for Glycans guidelines.

The Symbol Nomenclature for Glycans (SNFG) is a community-curated standard for the depiction of monosaccharides and complex glycans using various colored-coded, geometric shapes, along with defined text additions. It is hosted by the National Center for Biotechnology Information (NCBI) at the NCBI-Glycans Page (www.ncbi.nlm.nih.gov/glycans/snfg.html). Several changes have been made to the SNFG page in the past year to update the rules for depicting glycans using the SNFG, to include more examples of use, particularly for non-mammalian organisms, and to provide guidelines for the depiction of ambiguous glycan structures. This Glycoforum article summarizes these recent changes.

Practical applications of matched series analysis: SAR transfer, binding mode suggestion and data point validation.

AIM: The assumption in scaffold hopping is that changing the scaffold does not change the binding mode and the same structure-activity relationships (SARs) are seen for substituents decorating each scaffold. Results/methodology: We present the use of matched series analysis, an extension of matched molecular pair analysis, to automate the analysis of a project's data and detect the presence or absence of comparable SAR between chemical series. CONCLUSION: The presence of SAR transfer can confirm the perceived binding mode overlay of different chemotypes or suggest new arrangements between scaffolds that may have gone unnoticed. The absence of series correlation can highlight the presence of inconsistent data points where assay values should be reconfirmed, or provide challenge to any project dogma.

Comparing structural fingerprints using a literature-based similarity benchmark.

BACKGROUND: The concept of molecular similarity is one of the central ideas in cheminformatics, despite the fact that it is ill-defined and rather difficult to assess objectively. Here we propose a practical definition of molecular similarity in the context of drug discovery: molecules A and B are similar if a medicinal chemist would be likely to synthesise and test them around the same time as part of the same medicinal chemistry program. The attraction of such a definition is that it matches one of the key uses of similarity measures in early-stage drug discovery. If we make the assumption that molecules in the same compound activity table in a medicinal chemistry paper were considered similar by the authors of the paper, we can create a dataset of similar molecules from the medicinal chemistry literature. Furthermore, molecules with decreasing levels of similarity to a reference can be found by either ordering molecules in an activity table by their activity, or by considering activity tables in different papers which have at least one molecule in common. RESULTS: Using this procedure with activity data from ChEMBL, we have created two benchmark datasets for structural similarity that can be used to guide the development of improved measures. Compared to similar results from a virtual screen, these benchmarks are an order of magnitude more sensitive to differences between fingerprints both because of their size and because they avoid loss of statistical power due to the use of mean scores or ranks. We measure the performance of 28 different fingerprints on the benchmark sets and compare the results to those from the Riniker and Landrum (J Cheminf 5:26, 2013. doi:10.1186/1758-2946-5-26) ligand-based virtual screening benchmark. CONCLUSIONS: Extended-connectivity fingerprints of diameter 4 and 6 are among the best performing fingerprints when ranking diverse structures by similarity, as is the topological torsion fingerprint. However, when ranking very close analogues, the atom pair fingerprint outperforms the others tested. When ranking diverse structures or carrying out a virtual screen, we find that the performance of the ECFP fingerprints significantly improves if the bit-vector length is increased from 1024 to 16,384.Graphical abstractAn example series from one of the benchmark datasets. Each fingerprint is assessed on its ability to reproduce a specific series order.

Efficient chemical-disease identification and relationship extraction using Wikipedia to improve recall.

Awareness of the adverse effects of chemicals is important in biomedical research and healthcare. Text mining can allow timely and low-cost extraction of this knowledge from the biomedical literature. We extended our text mining solution, LeadMine, to identify diseases and chemical-induced disease relationships (CIDs). LeadMine is a dictionary/grammar-based entity recognizer and was used to recognize and normalize both chemicals and diseases to Medical Subject Headings (MeSH) IDs. The disease lexicon was obtained from three sources: MeSH, the Disease Ontology and Wikipedia. The Wikipedia dictionary was derived from pages with a disease/symptom box, or those where the page title appeared in the lexicon. Composite entities (e.g. heart and lung disease) were detected and mapped to their composite MeSH IDs. For CIDs, we developed a simple pattern-based system to find relationships within the same sentence. Our system was evaluated in the BioCreative V Chemical-Disease Relation task and achieved very good results for both disease concept ID recognition (F1-score: 86.12%) and CIDs (F1-score: 52.20%) on the test set. As our system was over an order of magnitude faster than other solutions evaluated on the task, we were able to apply the same system to the entirety of MEDLINE allowing us to extract a collection of over 250 000 distinct CIDs.

Using matched molecular series as a predictive tool to optimize biological activity.

A matched molecular series is the general form of a matched molecular pair and refers to a set of two or more molecules with the same scaffold but different R groups at the same position. We describe Matsy, a knowledge-based method that uses matched series to predict R groups likely to improve activity given an observed activity order for some R groups. We compare the Matsy predictions based on activity data from ChEMBLdb to the recommendations of the Topliss tree and carry out a large scale retrospective test to measure performance. We show that the basis for predictive success is preferred orders in matched series and that this preference is stronger for longer series. The Matsy algorithm allows medicinal chemists to integrate activity trends from diverse medicinal chemistry programs and apply them to problems of interest as a Topliss-like recommendation or as a hypothesis generator to aid compound design.

Towards a Universal SMILES representation - A standard method to generate canonical SMILES based on the InChI.

BACKGROUND: There are two line notations of chemical structures that have established themselves in the field: the SMILES string and the InChI string. The InChI aims to provide a unique, or canonical, identifier for chemical structures, while SMILES strings are widely used for storage and interchange of chemical structures, but no standard exists to generate a canonical SMILES string. RESULTS: I describe how to use the InChI canonicalisation to derive a canonical SMILES string in a straightforward way, either incorporating the InChI normalisations (Inchified SMILES) or not (Universal SMILES). This is the first description of a method to generate canonical SMILES that takes stereochemistry into account. When tested on the 1.1 m compounds in the ChEMBL database, and a 1 m compound subset of the PubChem Substance database, no canonicalisation failures were found with Inchified SMILES. Using Universal SMILES, 99.79% of the ChEMBL database was canonicalised successfully and 99.77% of the PubChem subset. CONCLUSIONS: The InChI canonicalisation algorithm can successfully be used as the basis for a common standard for canonical SMILES. While challenges remain - such as the development of a standard aromatic model for SMILES - the ability to create the same SMILES using different toolkits will mean that for the first time it will be possible to easily compare the chemical models used by different toolkits.

Open Babel: An open chemical toolbox.

BACKGROUND: A frequent problem in computational modeling is the interconversion of chemical structures between different formats. While standard interchange formats exist (for example, Chemical Markup Language) and de facto standards have arisen (for example, SMILES format), the need to interconvert formats is a continuing problem due to the multitude of different application areas for chemistry data, differences in the data stored by different formats (0D versus 3D, for example), and competition between software along with a lack of vendor-neutral formats. RESULTS: We discuss, for the first time, Open Babel, an open-source chemical toolbox that speaks the many languages of chemical data. Open Babel version 2.3 interconverts over 110 formats. The need to represent such a wide variety of chemical and molecular data requires a library that implements a wide range of cheminformatics algorithms, from partial charge assignment and aromaticity detection, to bond order perception and canonicalization. We detail the implementation of Open Babel, describe key advances in the 2.3 release, and outline a variety of uses both in terms of software products and scientific research, including applications far beyond simple format interconversion. CONCLUSIONS: Open Babel presents a solution to the proliferation of multiple chemical file formats. In addition, it provides a variety of useful utilities from conformer searching and 2D depiction, to filtering, batch conversion, and substructure and similarity searching. For developers, it can be used as a programming library to handle chemical data in areas such as organic chemistry, drug design, materials science, and computational chemistry. It is freely available under an open-source license from http://openbabel.org.

Open Data, Open Source and Open Standards in chemistry: The Blue Obelisk five years on.

BACKGROUND: The Blue Obelisk movement was established in 2005 as a response to the lack of Open Data, Open Standards and Open Source (ODOSOS) in chemistry. It aims to make it easier to carry out chemistry research by promoting interoperability between chemistry software, encouraging cooperation between Open Source developers, and developing community resources and Open Standards. RESULTS: This contribution looks back on the work carried out by the Blue Obelisk in the past 5 years and surveys progress and remaining challenges in the areas of Open Data, Open Standards, and Open Source in chemistry. CONCLUSIONS: We show that the Blue Obelisk has been very successful in bringing together researchers and developers with common interests in ODOSOS, leading to development of many useful resources freely available to the chemistry community.

Confab - Systematic generation of diverse low-energy conformers.

BACKGROUND: Many computational chemistry analyses require the generation of conformers, either on-the-fly, or in advance. We present Confab, an open source command-line application for the systematic generation of low-energy conformers according to a diversity criterion. RESULTS: Confab generates conformations using the 'torsion driving approach' which involves iterating systematically through a set of allowed torsion angles for each rotatable bond. Energy is assessed using the MMFF94 forcefield. Diversity is measured using the heavy-atom root-mean-square deviation (RMSD) relative to conformers already stored. We investigated the recovery of crystal structures for a dataset of 1000 ligands from the Protein Data Bank with fewer than 1 million conformations. Confab can recover 97% of the molecules to within 1.5 Å at a diversity level of 1.5 Å and an energy cutoff of 50 kcal/mol. CONCLUSIONS: Confab is available from http://confab.googlecode.com.

Testing assumptions and hypotheses for rescoring success in protein-ligand docking.

In protein-ligand docking, the scoring function is responsible for identifying the correct pose of a particular ligand as well as separating ligands from nonligands. Recently there has been considerable interest in schemes that combine results from several scoring functions in an effort to achieve improved performance in virtual screens. One such scheme is consensus scoring, which involves combining the results from several rescoring experiments. Although there have been a number of studies that have investigated factors affecting success in consensus scoring, these studies have not addressed the question of why a rescoring strategy works in the first place. Here we propose and test two alternative hypotheses for why rescoring has the potential to improve results, using GOLD 4.0. The "consensus" hypothesis is that rescoring is a way of combining results from two scoring functions such that only true positives are likely to score highly. The "complementary" hypothesis is that the two scoring functions used in rescoring have complementary strengths; one is better at ranking actives with respect to inactives while the other is better at ranking poses of actives. We find that in general it is this hypothesis that explains success in a rescoring experiment. We also test an assumption of any rescoring method, which is that the scores obtained are representative of the fitness of the docked pose. We find that although rescored poses tended to have slightly higher clash values than their docked equivalents, in general the scores were representative.

Curcumin-induced degradation of PKC delta is associated with enhanced dentate NCAM PSA expression and spatial learning in adult and aged Wistar rats.

Polysialylation of the neural cell adhesion molecule (NCAM PSA) is necessary for the consolidation processes of hippocampus-based learning. Previously, we have found inhibition of protein kinase C delta (PKCdelta) to be associated with increased polysialyltransferase (PST) activity, suggesting inhibitors of this kinase might ameliorate cognitive deficits. Using a rottlerin template, a drug previously considered an inhibitor of PKCdelta, we searched the Compounds Available for Purchase (CAP) database with the Accelrys((R)) Catalyst programme for structurally similar molecules and, using the available crystal structure of the phorbol-binding domain of PKCdelta, found that diferuloylmethane (curcumin) docked effectively into the phorbol site. Curcumin increased NCAM PSA expression in cultured neuro-2A neuroblastoma cells and this was inversely related to PKCdelta protein expression. Curcumin did not directly inhibit PKCdelta activity but formed a tight complex with the enzyme. With increasing doses of curcumin, the Tyr(131) residue of PKCdelta, which is known to direct its degradation, became progressively phosphorylated and this was associated with numerous Tyr(131)-phospho-PKCdelta fragments. Chronic administration of curcumin in vivo also increased the frequency of polysialylated cells in the dentate infragranular zone and significantly improved the acquisition and consolidation of a water maze spatial learning paradigm in both adult and aged cohorts of Wistar rats. These results further confirm the role of PKCdelta in regulating PST and NCAM PSA expression and provide evidence that drug modulation of this system enhances the process of memory consolidation.

Cinfony--combining Open Source cheminformatics toolkits behind a common interface.

BACKGROUND: Open Source cheminformatics toolkits such as OpenBabel, the CDK and the RDKit share the same core functionality but support different sets of file formats and forcefields, and calculate different fingerprints and descriptors. Despite their complementary features, using these toolkits in the same program is difficult as they are implemented in different languages (C++ versus Java), have different underlying chemical models and have different application programming interfaces (APIs). RESULTS: We describe Cinfony, a Python module that presents a common interface to all three of these toolkits, allowing the user to easily combine methods and results from any of the toolkits. In general, the run time of the Cinfony modules is almost as fast as accessing the underlying toolkits directly from C++ or Java, but Cinfony makes it much easier to carry out common tasks in cheminformatics such as reading file formats and calculating descriptors. CONCLUSION: By providing a simplified interface and improving interoperability, Cinfony makes it easy to combine complementary features of OpenBabel, the CDK and the RDKit.

Simultaneous feature selection and parameter optimisation using an artificial ant colony: case study of melting point prediction.

BACKGROUND: We present a novel feature selection algorithm, Winnowing Artificial Ant Colony (WAAC), that performs simultaneous feature selection and model parameter optimisation for the development of predictive quantitative structure-property relationship (QSPR) models. The WAAC algorithm is an extension of the modified ant colony algorithm of Shen et al. (J Chem Inf Model 2005, 45: 1024-1029). We test the ability of the algorithm to develop a predictive partial least squares model for the Karthikeyan dataset (J Chem Inf Model 2005, 45: 581-590) of melting point values. We also test its ability to perform feature selection on a support vector machine model for the same dataset. RESULTS: Starting from an initial set of 203 descriptors, the WAAC algorithm selected a PLS model with 68 descriptors which has an RMSE on an external test set of 46.6 degrees C and R2 of 0.51. The number of components chosen for the model was 49, which was close to optimal for this feature selection. The selected SVM model has 28 descriptors (cost of 5, epsilon of 0.21) and an RMSE of 45.1 degrees C and R2 of 0.54. This model outperforms a kNN model (RMSE of 48.3 degrees C, R2 of 0.47) for the same data and has similar performance to a Random Forest model (RMSE of 44.5 degrees C, R2 of 0.55). However it is much less prone to bias at the extremes of the range of melting points as shown by the slope of the line through the residuals: -0.43 for WAAC/SVM, -0.53 for Random Forest. CONCLUSION: With a careful choice of objective function, the WAAC algorithm can be used to optimise machine learning and regression models that suffer from overfitting. Where model parameters also need to be tuned, as is the case with support vector machine and partial least squares models, it can optimise these simultaneously. The moving probabilities used by the algorithm are easily interpreted in terms of the best and current models of the ants, and the winnowing procedure promotes the removal of irrelevant descriptors.

Using buriedness to improve discrimination between actives and inactives in docking.

A continuing problem in protein-ligand docking is the correct relative ranking of active molecules versus inactives. Using the ChemScore scoring function as implemented in the GOLD docking software, we have investigated the effect of scaling hydrogen bond, metal-ligand, and lipophilic interactions based on the buriedness of the interaction. Buriedness was measured using the receptor density, the number of protein heavy atoms within 8.0 A. Terms in the scaling functions were optimized using negative data, represented by docked poses of inactive molecules. The objective function was the mean rank of the scores of the active poses in the Astex Diverse Set (Hartshorn et al. J. Med. Chem., 2007, 50, 726) with respect to the docked poses of 99 inactives. The final four-parameter model gave a substantial improvement in the average rank from 18.6 to 12.5. Similar results were obtained for an independent test set. Receptor density scaling is available as an option in the recent GOLD release.