Faster and more diverse de novo molecular optimization with double-loop reinforcement learning using augmented SMILES.

Using generative deep learning models and reinforcement learning together can effectively generate new molecules with desired properties. By employing a multi-objective scoring function, thousands of high-scoring molecules can be generated, making this approach useful for drug discovery and material science. However, the application of these methods can be hindered by computationally expensive or time-consuming scoring procedures, particularly when a large number of function calls are required as feedback in the reinforcement learning optimization. Here, we propose the use of double-loop reinforcement learning with simplified molecular line entry system (SMILES) augmentation to improve the efficiency and speed of the optimization. By adding an inner loop that augments the generated SMILES strings to non-canonical SMILES for use in additional reinforcement learning rounds, we can both reuse the scoring calculations on the molecular level, thereby speeding up the learning process, as well as offer additional protection against mode collapse. We find that employing between 5 and 10 augmentation repetitions is optimal for the scoring functions tested and is further associated with an increased diversity in the generated compounds, improved reproducibility of the sampling runs and the generation of molecules of higher similarity to known ligands.

Design and Evaluation of a Low Hydrogen Bond Donor Count Fragment Screening Set to Aid Hit Generation of PROTACs Intended for Oral Delivery.

The development of orally bioavailable PROTACs presents a significant challenge due to the inflated physicochemical properties of such heterobifunctional molecules. Molecules occupying this "beyond rule of five" space often demonstrate limited oral bioavailability due to the compounding effects of elevated molecular weight and hydrogen bond donor count (among other properties), but it is possible to achieve sufficient oral bioavailability through physicochemical optimization. Herein, we disclose the design and evaluation of a low hydrogen bond donor count (≤1 HBD) fragment screening set to aid hit generation of PROTACs intended for an oral route of delivery. We demonstrate that application of this library can enhance fragment screens against PROTAC proteins of interest and ubiquitin ligases, yielding fragment hits containing ≤1 HBD suitable for optimizing toward orally bioavailable PROTACs.

Automated relative binding free energy calculations from SMILES to ΔΔG.

In drug discovery, computational methods are a key part of making informed design decisions and prioritising experiments. In particular, optimizing compound affinity is a central concern during the early stages of development. In the last 10 years, alchemical free energy (FE) calculations have transformed our ability to incorporate accurate in silico potency predictions in design decisions, and represent the 'gold standard' for augmenting experiment-driven drug discovery. However, relative FE calculations are complex to set up, require significant expert intervention to prepare the calculation and analyse the results or are provided only as closed-source software, not allowing for fine-grained control over the underlying settings. In this work, we introduce an end-to-end relative FE workflow based on the non-equilibrium switching approach that facilitates calculation of binding free energies starting from SMILES strings. The workflow is implemented using fully modular steps, allowing various components to be exchanged depending on licence availability. We further investigate the dependence of the calculated free energy accuracy on the initial ligand pose generated by various docking algorithms. We show that both commercial and open-source docking engines can be used to generate poses that lead to good correlation of free energies with experimental reference data.

Icolos: a workflow manager for structure-based post-processing of de novo generated small molecules.

SUMMARY: We present Icolos, a workflow manager written in Python as a tool for automating complex structure-based workflows for drug design. Icolos can be used as a standalone tool, for example in virtual screening campaigns, or can be used in conjunction with deep learning-based molecular generation facilitated for example by REINVENT, a previously published molecular de novo design package. In this publication, we focus on the internal structure and general capabilities of Icolos, using molecular docking experiments as an illustrative example. AVAILABILITY AND IMPLEMENTATION: The source code is freely available at https://github.com/MolecularAI/Icolos under the Apache 2.0 license. Tutorial notebooks containing minimal working examples can be found at https://github.com/MolecularAI/IcolosCommunity. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Implications of Additivity and Nonadditivity for Machine Learning and Deep Learning Models in Drug Design.

Matched molecular pairs (MMPs) are nowadays a commonly applied concept in drug design. They are used in many computational tools for structure-activity relationship analysis, biological activity prediction, or optimization of physicochemical properties. However, until now it has not been shown in a rigorous way that MMPs, that is, changing only one substituent between two molecules, can be predicted with higher accuracy and precision in contrast to any other chemical compound pair. It is expected that any model should be able to predict such a defined change with high accuracy and reasonable precision. In this study, we examine the predictability of four classical properties relevant for drug design ranging from simple physicochemical parameters (log D and solubility) to more complex cell-based ones (permeability and clearance), using different data sets and machine learning algorithms. Our study confirms that additive data are the easiest to predict, which highlights the importance of recognition of nonadditivity events and the challenging complexity of predicting properties in case of scaffold hopping. Despite deep learning being well suited to model nonlinear events, these methods do not seem to be an exception of this observation. Though they are in general performing better than classical machine learning methods, this leaves the field with a still standing challenge.

Pandemic, Epidemic, Endemic: B Cell Repertoire Analysis Reveals Unique Anti-Viral Responses to SARS-CoV-2, Ebola and Respiratory Syncytial Virus.

Immunoglobulin gene heterogeneity reflects the diversity and focus of the humoral immune response towards different infections, enabling inference of B cell development processes. Detailed compositional and lineage analysis of long read IGH repertoire sequencing, combining examples of pandemic, epidemic and endemic viral infections with control and vaccination samples, demonstrates general responses including increased use of IGHV4-39 in both Zaire Ebolavirus (EBOV) and COVID-19 patient cohorts. We also show unique characteristics absent in Respiratory Syncytial Virus or yellow fever vaccine samples: EBOV survivors show unprecedented high levels of class switching events while COVID-19 repertoires from acute disease appear underdeveloped. Despite the high levels of clonal expansion in COVID-19 IgG1 repertoires there is a striking lack of evidence of germinal centre mutation and selection. Given the differences in COVID-19 morbidity and mortality with age, it is also pertinent that we find significant differences in repertoire characteristics between young and old patients. Our data supports the hypothesis that a primary viral challenge can result in a strong but immature humoral response where failures in selection of the repertoire risk off-target effects.

DockStream: a docking wrapper to enhance de novo molecular design.

Recently, we have released the de novo design platform REINVENT in version 2.0. This improved and extended iteration supports far more features and scoring function components, which allows bespoke and tailor-made protocols to maximize impact in small molecule drug discovery projects. A major obstacle of generative models is producing active compounds, in which predictive (QSAR) models have been applied to enrich target activity. However, QSAR models are inherently limited by their applicability domains. To overcome these limitations, we introduce a structure-based scoring component for REINVENT. DockStream is a flexible, stand-alone molecular docking wrapper that provides access to a collection of ligand embedders and docking backends. Using the benchmarking and analysis workflow provided in DockStream, execution and subsequent analysis of a variety of docking configurations can be automated. Docking algorithms vary greatly in performance depending on the target and the benchmarking and analysis workflow provides a streamlined solution to identifying productive docking configurations. We show that an informative docking configuration can inform the REINVENT agent to optimize towards improving docking scores using public data. With docking activated, REINVENT is able to retain key interactions in the binding site, discard molecules which do not fit the binding cavity, harness unused (sub-)pockets, and improve overall performance in the scaffold-hopping scenario. The code is freely available at https://github.com/MolecularAI/DockStream .

REINVENT 2.0: An AI Tool for De Novo Drug Design.

In the past few years, we have witnessed a renaissance of the field of molecular de novo drug design. The advancements in deep learning and artificial intelligence (AI) have triggered an avalanche of ideas on how to translate such techniques to a variety of domains including the field of drug design. A range of architectures have been devised to find the optimal way of generating chemical compounds by using either graph- or string (SMILES)-based representations. With this application note, we aim to offer the community a production-ready tool for de novo design, called REINVENT. It can be effectively applied on drug discovery projects that are striving to resolve either exploration or exploitation problems while navigating the chemical space. It can facilitate the idea generation process by bringing to the researcher's attention the most promising compounds. REINVENT's code is publicly available at https://github.com/MolecularAI/Reinvent.

UDP-N-acetyl-α-D-galactosamine:polypeptide N-acetylgalactosaminyltransferase from the snail Biomphalaria glabrata - structural reflections.

UDP-GalNAc:polypeptide GalNAc transferase (ppGalNAcT; EC 2.4.1.41) is the initiating enzyme for mucin-type O-glycosylation in animals. Members of this highly conserved glycosyltransferase family catalyse a single glycosidic linkage. They transfer an N-acetylgalactosamine (GalNAc) residue from an activated donor (UDP-GalNAc) to a serine or threonine of an acceptor polypeptide chain. A ppGalNAcT from the freshwater snail Biomphalaria glabrata is the only characterised member of this enzyme family from mollusc origin. In this work, we interpret previously published experimental characterization of this enzyme in the context of in silico models of the enzyme and its acceptor substrates. A homology model of the mollusc ppGalNAcT is created and various substrate peptides are modelled into the active site. We hypothesize about possible molecular interpretations of the available experimental data and offer potential explanations for observed substrate and cofactor specificity. Here, we review and synthesise the current knowledge of Bge-ppGalNAcT, supported by a molecular interpretation of the available data.

Tissue-specific shaping of the TCR repertoire and antigen specificity of iNKT cells.

Tissue homeostasis is critically dependent on the function of tissue-resident lymphocytes, including lipid-reactive invariant natural killer T (iNKT) cells. Yet, if and how the tissue environment shapes the antigen specificity of iNKT cells remains unknown. By analysing iNKT cells from lymphoid tissues of mice and humans we demonstrate that their T cell receptor (TCR) repertoire is highly diverse and is distinct for cells from various tissues resulting in differential lipid-antigen recognition. Within peripheral tissues iNKT cell recent thymic emigrants exhibit a different TCR repertoire than mature cells, suggesting that the iNKT population is shaped after arrival to the periphery. Consistent with this, iNKT cells from different organs show distinct basal activation, proliferation and clonal expansion. Moreover, the iNKT cell TCR repertoire changes following immunisation and is shaped by age and environmental changes. Thus, post-thymic modification of the TCR-repertoire underpins the distinct antigen specificity for iNKT cells in peripheral tissues.

Lipid Head Group Parameterization for GROMOS 54A8: A Consistent Approach with Protein Force Field Description.

Membranes are a crucial component of both bacterial and mammalian cells, being involved in signaling, transport, and compartmentalization. This versatility requires a variety of lipid species to tailor the membrane's behavior as needed, increasing the complexity of the system. Molecular dynamics simulations have been successfully applied to study model membranes and their interactions with proteins, elucidating some crucial mechanisms at the atomistic detail and thus complementing experimental techniques. An accurate description of the functional interplay of the diverse membrane components crucially depends on the selected parameters that define the adopted force field. A coherent parameterization for lipids and proteins is therefore needed. In this work, we propose and validate new lipid head group parameters for the GROMOS 54A8 force field, making use of recently published parametrizations for key chemical moieties present in lipids. We make use additionally of a new canonical set of partial charges for lipids, chosen to be consistent with the parameterization of soluble molecules such as proteins. We test the derived parameters on five phosphocholine model bilayers, composed of lipid patches four times larger than the ones used in previous studies, and run 500 ns long simulations of each system. Reproduction of experimental data like area per lipid and deuterium order parameters is good and comparable with previous parameterizations, as well as the description of liquid crystal to gel-phase transition. On the other hand, the orientational behavior of the head groups is more realistic for this new parameter set, and this can be crucial in the description of interactions with other polar molecules. For that reason, we tested the interaction of the antimicrobial peptide lactoferricin with two model membranes showing that the new parameters lead to a weaker peptide-membrane binding and give a more realistic outcome in comparing binding to antimicrobial versus mammal membranes.

BRepertoire: a user-friendly web server for analysing antibody repertoire data.

Antibody repertoire analysis by high throughput sequencing is now widely used, but a persisting challenge is enabling immunologists to explore their data to discover discriminating repertoire features for their own particular investigations. Computational methods are necessary for large-scale evaluation of antibody properties. We have developed BRepertoire, a suite of user-friendly web-based software tools for large-scale statistical analyses of repertoire data. The software is able to use data preprocessed by IMGT, and performs statistical and comparative analyses with versatile plotting options. BRepertoire has been designed to operate in various modes, for example analysing sequence-specific V(D)J gene usage, discerning physico-chemical properties of the CDR regions and clustering of clonotypes. Those analyses are performed on the fly by a number of R packages and are deployed by a shiny web platform. The user can download the analysed data in different table formats and save the generated plots as image files ready for publication. We believe BRepertoire to be a versatile analytical tool that complements experimental studies of immune repertoires. To illustrate the server's functionality, we show use cases including differential gene usage in a vaccination dataset and analysis of CDR3H properties in old and young individuals. The server is accessible under http://mabra.biomed.kcl.ac.uk/BRepertoire.

MDplot: Visualise Molecular Dynamics.

The MDplot package provides plotting functions to allow for automated visualisation of molecular dynamics simulation output. It is especially useful in cases where the plot generation is rather tedious due to complex file formats or when a large number of plots are generated. The graphs that are supported range from those which are standard, such as RMsD/RMsF (root-mean-square deviation and root-mean-square fluctuation, respectively) to less standard, such as thermodynamic integration analysis and hydrogen bond monitoring over time. All told, they address many commonly used analyses. In this article, we set out the MDplot package's functions, give examples of the function calls, and show the associated plots. Plotting and data parsing is separated in all cases, i.e. the respective functions can be used independently. Thus, data manipulation and the integration of additional file formats is fairly easy. Currently, the loading functions support GROMOS, GROMACS, and AMBER file formats. Moreover, we also provide a Bash interface that allows simple embedding of MDplot into Bash scripts as the final analysis step. AVAILABILITY: The package can be obtained in the latest major version from CRAN (https://cran.r-project.org/package=MDplot) or in the most recent version from the project's GitHub page at https://github.com/MDplot/MDplot, where feedback is also most welcome. MDplot is published under the GPL-3 license.

Update on phosphate and charged post-translationally modified amino acid parameters in the GROMOS force field.

In this study, we propose newly derived parameters for phosphate ions in the context of the GROMOS force field parameter sets. The non-bonded parameters used up to now lead to a hydration free energy, which renders the dihydrogen phosphate ion too hydrophobic when compared to experimentally derived values, making a reparametrization of the phosphate moiety necessary. Phosphate species are of great importance in biomolecular simulations not only because of their crucial role in the backbone of nucleic acids but also as they represent one of the most important types of post-translational modifications to protein side-chains and are an integral part in many lipids. Our re-parametrization of the free dihydrogen phosphate (H 2PO 4-) and three derivatives (methyl phosphate, dimethyl phosphate, and phenyl phosphate) leads, in conjunction with the previously updated charged side-chains in the GROMOS parameter set 54A8, to new nucleic acid backbone parameters and a 54A8 version of the widely used GROMOS protein post-translational modification parameter set. © 2017 Wiley Periodicals, Inc.

Optimization of Protein Backbone Dihedral Angles by Means of Hamiltonian Reweighting.

Molecular dynamics simulations depend critically on the accuracy of the underlying force fields in properly representing biomolecules. Hence, it is crucial to validate the force-field parameter sets in this respect. In the context of the GROMOS force field, this is usually achieved by comparing simulation data to experimental observables for small molecules. In this study, we develop new amino acid backbone dihedral angle potential energy parameters based on the widely used 54A7 parameter set by matching to experimental J values and secondary structure propensity scales. In order to find the most appropriate backbone parameters, close to 100 000 different combinations of parameters have been screened. However, since the sheer number of combinations considered prohibits actual molecular dynamics simulations for each of them, we instead predicted the values for every combination using Hamiltonian reweighting. While the original 54A7 parameter set fails to reproduce the experimental data, we are able to provide parameters that match significantly better. However, to ensure applicability in the context of larger peptides and full proteins, further studies have to be undertaken.

Antibody humanization by molecular dynamics simulations-in-silico guided selection of critical backmutations.

Monoclonal antibodies represent the fastest growing class of biotherapeutic proteins. However, as they are often initially derived from rodent organisms, there is a severe risk of immunogenic reactions, hampering their applicability. The humanization of these antibodies remains a challenging task in the context of rational drug design. "Superhumanization" describes the direct transfer of the complementarity determining regions to a human germline framework, but this humanization approach often results in loss of binding affinity. In this study, we present a new approach for predicting promising backmutation sites using molecular dynamics simulations of the model antibody Ab2/3H6. The simulation method was developed in close conjunction with novel specificity experiments. Binding properties of mAb variants were evaluated directly from crude supernatants and confirmed using established binding affinity assays for purified antibodies. Our approach provides access to the dynamical features of the actual binding sites of an antibody, based solely on the antibody sequence. Thus we do not need structural data on the antibody-antigen complex and circumvent cumbersome methods to assess binding affinities. © 2016 The Authors Journal of Molecular Recognition Published by John Wiley & Sons Ltd.

A systematic framework for molecular dynamics simulations of protein post-translational modifications.

By directly affecting structure, dynamics and interaction networks of their targets, post-translational modifications (PTMs) of proteins play a key role in different cellular processes ranging from enzymatic activation to regulation of signal transduction to cell-cycle control. Despite the great importance of understanding how PTMs affect proteins at the atomistic level, a systematic framework for treating post-translationally modified amino acids by molecular dynamics (MD) simulations, a premier high-resolution computational biology tool, has never been developed. Here, we report and validate force field parameters (GROMOS 45a3 and 54a7) required to run and analyze MD simulations of more than 250 different types of enzymatic and non-enzymatic PTMs. The newly developed GROMOS 54a7 parameters in particular exhibit near chemical accuracy in matching experimentally measured hydration free energies (RMSE=4.2 kJ/mol over the validation set). Using this tool, we quantitatively show that the majority of PTMs greatly alter the hydrophobicity and other physico-chemical properties of target amino acids, with the extent of change in many cases being comparable to the complete range spanned by native amino acids.

Vienna-PTM web server: a toolkit for MD simulations of protein post-translational modifications.

Post-translational modifications (PTMs) play a key role in numerous cellular processes by directly affecting structure, dynamics and interaction networks of target proteins. Despite their importance, our understanding of protein PTMs at the atomistic level is still largely incomplete. Molecular dynamics (MD) simulations, which provide high-resolution insight into biomolecular function and underlying mechanisms, are in principle ideally suited to tackle this problem. However, because of the challenges associated with the development of novel MD parameters and a general lack of suitable computational tools for incorporating PTMs in target protein structures, MD simulations of post-translationally modified proteins have historically lagged significantly behind the studies of unmodified proteins. Here, we present Vienna-PTM web server (http://vienna-ptm.univie.ac.at), a platform for automated introduction of PTMs of choice to protein 3D structures (PDB files) in a user-friendly visual environment. With 256 different enzymatic and non-enzymatic PTMs available, the server performs geometrically realistic introduction of modifications at sites of interests, as well as subsequent energy minimization. Finally, the server makes available force field parameters and input files needed to run MD simulations of modified proteins within the framework of the widely used GROMOS 54A7 and 45A3 force fields and GROMACS simulation package.