Task-similarity is a crucial factor for few-shot meta-learning of structure-activity relationships.

Machine learning models support computer-aided molecular design and compound optimization. However, the initial phases of drug discovery often face a scarcity of training data for these models. Meta-learning has emerged as a potentially promising strategy, harnessing the wealth of structure-activity data available for known targets to facilitate efficient few-shot model training for the specific target of interest. In this study, we assessed the effectiveness of two different meta-learning methods, namely model-agnostic meta-learning (MAML) and adaptive deep kernel fitting (ADKF), specifically in the regression setting. We investigated how factors such as dataset size and the similarity of training tasks impact predictability. The results indicate that ADKF significantly outperformed both MAML and a single-task baseline model on the inhibition data. However, the performance of ADKF varied across different test tasks. Our findings suggest that considerable enhancements in performance can be anticipated primarily when the task of interest is similar to the tasks incorporated in the meta-learning process.

Integrating Reaction Schemes, Reagent Databases, and Virtual Libraries into Fragment-Based Design by Reinforcement Learning.

Lead optimization supported by artificial intelligence (AI)-based generative models has become increasingly important in drug design. Success factors are reagent availability, novelty, and the optimization of multiple properties. Directed fragment-replacement is particularly attractive, as it mimics medicinal chemistry tactics. Here, we present variations of fragment-based reinforcement learning using an actor-critic model. Novel features include freezing fragments and using reagents as the fragment source. Splitting molecules according to reaction schemes improves synthesizability, while tuning network output probabilities allows us to balance novelty versus diversity. Combining fragment-based optimization with virtual library encodings allows the exploration of large chemical spaces with synthesizable ideas. Collectively, these enhancements influence design toward high-quality molecules with favorable profiles. A validation study using 15 pharmaceutically relevant targets reveals that novel structures are obtained for most cases, which are identical or related to independent validation sets for each target. Hence, these modifications significantly increase the value of fragment-based reinforcement learning for drug design. The code is available on GitHub: https://github.com/Sanofi-Public/IDD-papers-fragrl.

Impact of Applicability Domains to Generative Artificial Intelligence.

Molecular generative artificial intelligence is drawing significant attention in the drug design community, with several experimentally validated proof of concepts already published. Nevertheless, generative models are known for sometimes generating unrealistic, unstable, unsynthesizable, or uninteresting structures. This calls for methods to constrain those algorithms to generate structures in drug-like portions of the chemical space. While the concept of applicability domains for predictive models is well studied, its counterpart for generative models is not yet well-defined. In this work, we empirically examine various possibilities and propose applicability domains suited for generative models. Using both public and internal data sets, we use generative methods to generate novel structures that are predicted to be actives by a corresponding quantitative structure-activity relationships model while constraining the generative model to stay within a given applicability domain. Our work looks at several applicability domain definitions, combining various criteria, such as structural similarity to the training set, similarity of physicochemical properties, unwanted substructures, and quantitative estimate of drug-likeness. We assess the structures generated from both qualitative and quantitative points of view and find that the applicability domain definitions have a strong influence on the drug-likeness of generated molecules. An extensive analysis of our results allows us to identify applicability domain definitions that are best suited for generating drug-like molecules with generative models. We anticipate that this work will help foster the adoption of generative models in an industrial context.

Optimizing interactions to protein binding sites by integrating docking-scoring strategies into generative AI methods.

The identification and optimization of promising lead molecules is essential for drug discovery. Recently, artificial intelligence (AI) based generative methods provided complementary approaches for generating molecules under specific design constraints of relevance in drug design. The goal of our study is to incorporate protein 3D information directly into generative design by flexible docking plus an adapted protein-ligand scoring function, thereby moving towards automated structure-based design. First, the protein-ligand scoring function RFXscore integrating individual scoring terms, ligand descriptors, and combined terms was derived using the PDBbind database and internal data. Next, design results for different workflows are compared to solely ligand-based reward schemes. Our newly proposed, optimal workflow for structure-based generative design is shown to produce promising results, especially for those exploration scenarios, where diverse structures fitting to a protein binding site are requested. Best results are obtained using docking followed by RFXscore, while, depending on the exact application scenario, it was also found useful to combine this approach with other metrics that bias structure generation into "drug-like" chemical space, such as target-activity machine learning models, respectively.

Interpretation of Structure-Activity Relationships in Real-World Drug Design Data Sets Using Explainable Artificial Intelligence.

In silico models based on Deep Neural Networks (DNNs) are promising for predicting activities and properties of new molecules. Unfortunately, their inherent black-box character hinders our understanding, as to which structural features are important for activity. However, this information is crucial for capturing the underlying structure-activity relationships (SARs) to guide further optimization. To address this interpretation gap, "Explainable Artificial Intelligence" (XAI) methods recently became popular. Herein, we apply and compare multiple XAI methods to projects of lead optimization data sets with well-established SARs and available X-ray crystal structures. As we can show, easily understandable and comprehensive interpretations are obtained by combining DNN models with some powerful interpretation methods. In particular, SHAP-based methods are promising for this task. A novel visualization scheme using atom-based heatmaps provides useful insights into the underlying SAR. It is important to note that all interpretations are only meaningful in the context of the underlying models and associated data.

Artificial Intelligence in Compound Design.

Artificial intelligence has seen an incredibly fast development in recent years. Many novel technologies for property prediction of drug molecules as well as for the design of novel molecules were introduced by different research groups. These artificial intelligence-based design methods can be applied for suggesting novel chemical motifs in lead generation or scaffold hopping as well as for optimization of desired property profiles during lead optimization. In lead generation, broad sampling of the chemical space for identification of novel motifs is required, while in the lead optimization phase, a detailed exploration of the chemical neighborhood of a current lead series is advantageous. These different requirements for successful design outcomes render different combinations of artificial intelligence technologies useful. Overall, we observe that a combination of different approaches with tailored scoring and evaluation schemes appears beneficial for efficient artificial intelligence-based compound design.

Application of Deep Neural Network Models in Drug Discovery Programs.

In silico driven optimization of compound properties related to pharmacokinetics, pharmacodynamics, and safety is a key requirement in modern drug discovery. Nowadays, large and harmonized datasets allow to implement deep neural networks (DNNs) as a framework for leveraging predictive models. Nevertheless, various available model architectures differ in their global applicability and performance in lead optimization projects, such as stability over time and interpretability of the results. Here, we describe and compare the value of established DNN-based methods for the prediction of key ADME property trends and biological activity in an industrial drug discovery environment, represented by microsomal lability, CYP3A4 inhibition and factor Xa inhibition. Three architectures are exemplified, our earlier described multilayer perceptron approach (MLP), graph convolutional network-based models (GCN) and a vector representation approach, Mol2Vec. From a statistical perspective, MLP and GCN were found to perform superior over Mol2Vec, when applied to external validation sets. Interestingly, GCN-based predictions are most stable over a longer period in a time series validation study. Apart from those statistical observations, DNN prove of value to guide local SAR. To illustrate this important aspect in pharmaceutical research projects, we discuss challenging applications in medicinal chemistry towards a more realistic picture of artificial intelligence in drug discovery.

Automated De Novo Design in Medicinal Chemistry: Which Types of Chemistry Does a Generative Neural Network Learn?

Artificial intelligence offers promising solutions for property prediction, compound design, and retrosynthetic planning, which are expected to significantly accelerate the search for pharmacologically relevant molecules. Here, we investigate aspects of artificial intelligence based de novo design pertaining to its integration into real-life workflows. First, different chemical spaces were used as training sets for reinforcement learning (RL) in combination with different reward functions. With the trained neuronal networks different biologically active molecules could be regenerated. Excluding molecules with substructures such as five-membered rings from training spaces nevertheless produced results containing these moieties. Furthermore, different scoring functions in RL were investigated and produced different design ensembles. In summary, some of these design proposals are close in chemical space to the query, thus supporting lead optimization, while 3D-shape or QSAR (quantitative structure-activity relationship) models produced significantly different proposals by sampling a broader region of the chemical space, thus supporting lead generation. Therefore, RL provides a good framework to tailored design approaches for different discovery phases.

Virtual Screening in the Cloud: How Big Is Big Enough?

Virtual screening is a standard tool in Computer-Assisted Drug Design (CADD). Early in a project, it is typical to use ligand-based similarity search methods to find suitable hit molecules. However, the number of compounds which can be screened and the time required are usually limited by computational resources. We describe here a high-throughput virtual screening project using 3D similarity (FastROCS) and automated evaluation workflows on Orion, a cloud computing platform. Cloud resources make this approach fully scalable and flexible, allowing the generation and search of billions of virtual molecules, and give access to an explicit 3D virtual chemistry space not available before. We discuss the impact of the size of the search space with respect to finding novel chemical hits and the size of the required hit list, as well as computational and economical aspects of resource scaling.

PELE-MSM: A Monte Carlo Based Protocol for the Estimation of Absolute Binding Free Energies.

In this study, we present a fully automatic platform based on our Monte Carlo algorithm, the Protein Energy Landscape Exploration method (PELE), for the estimation of absolute protein-ligand binding free energies, one of the most significant challenges in computer aided drug design. Based on a ligand pathway approach, an initial short enhanced sampling simulation is performed to identify reasonable starting positions for more extended sampling. This stepwise approach allows for a significant faster convergence of the free energy estimation using the Markov State Model (MSM) technique. PELE-MSM was applied on four diverse protein and ligand systems, successfully ranking compounds for two systems. Based on the results, current limitations and challenges with physics-based methods in computational structural biology are discussed. Overall, PELE-MSM constitutes a promising step toward computing absolute binding free energies and in their application into drug discovery projects.

Structural insight into allosteric modulation of protease-activated receptor 2.

Protease-activated receptors (PARs) are a family of G-protein-coupled receptors (GPCRs) that are irreversibly activated by proteolytic cleavage of the N terminus, which unmasks a tethered peptide ligand that binds and activates the transmembrane receptor domain, eliciting a cellular cascade in response to inflammatory signals and other stimuli. PARs are implicated in a wide range of diseases, such as cancer and inflammation. PARs have been the subject of major pharmaceutical research efforts but the discovery of small-molecule antagonists that effectively bind them has proved challenging. The only marketed drug targeting a PAR is vorapaxar, a selective antagonist of PAR1 used to prevent thrombosis. The structure of PAR1 in complex with vorapaxar has been reported previously. Despite sequence homology across the PAR isoforms, discovery of PAR2 antagonists has been less successful, although GB88 has been described as a weak antagonist. Here we report crystal structures of PAR2 in complex with two distinct antagonists and a blocking antibody. The antagonist AZ8838 binds in a fully occluded pocket near the extracellular surface. Functional and binding studies reveal that AZ8838 exhibits slow binding kinetics, which is an attractive feature for a PAR2 antagonist competing against a tethered ligand. Antagonist AZ3451 binds to a remote allosteric site outside the helical bundle. We propose that antagonist binding prevents structural rearrangements required for receptor activation and signalling. We also show that a blocking antibody antigen-binding fragment binds to the extracellular surface of PAR2, preventing access of the tethered ligand to the peptide-binding site. These structures provide a basis for the development of selective PAR2 antagonists for a range of therapeutic uses.

Exploring Binding Mechanisms in Nuclear Hormone Receptors by Monte Carlo and X-ray-derived Motions.

In this study, we performed an extensive exploration of the ligand entry mechanism for members of the steroid nuclear hormone receptor family (androgen receptor, estrogen receptor α, glucocorticoid receptor, mineralocorticoid receptor, and progesterone receptor) and their endogenous ligands. The exploration revealed a shared entry path through the helix 3, 7, and 11 regions. Examination of the x-ray structures of the receptor-ligand complexes further showed two distinct folds of the helix 6-7 region, classified as "open" and "closed", which could potentially affect ligand binding. To improve sampling of the helix 6-7 loop, we incorporated motion modes based on principal component analysis of existing crystal structures of the receptors and applied them to the protein-ligand sampling. A detailed comparison with the anisotropic network model (an elastic network model) highlights the importance of flexibility in the entrance region. While the binding (interaction) energy of individual simulations can be used to score different ligands, extensive sampling further allows us to predict absolute binding free energies and analyze reaction kinetics using Markov state models and Perron-cluster cluster analysis, respectively. The predicted relative binding free energies for three ligands binding to the progesterone receptor are in very good agreement with experimental results and the Perron-cluster cluster analysis highlighted the importance of a peripheral binding site. Our analysis revealed that the flexibility of the helix 3, 7, and 11 regions represents the most important factor for ligand binding. Furthermore, the hydrophobicity of the ligand influences the transition between the peripheral and the active binding site.

3D-Lab: a collaborative web-based platform for molecular modeling.

AIM: The use of 3D information has shown impact in numerous applications in drug design. However, it is often under-utilized and traditionally limited to specialists. We want to change that, and present an approach making 3D information and molecular modeling accessible and easy-to-use 'for the people'. METHODOLOGY/RESULTS: A user-friendly and collaborative web-based platform (3D-Lab) for 3D modeling, including a blazingly fast virtual screening capability, was developed. 3D-Lab provides an interface to automatic molecular modeling, like conformer generation, ligand alignments, molecular dockings and simple quantum chemistry protocols. 3D-Lab is designed to be modular, and to facilitate sharing of 3D-information to promote interactions between drug designers. Recent enhancements to our open-source virtual reality tool Molecular Rift are described. CONCLUSION: The integrated drug-design platform allows drug designers to instantaneously access 3D information and readily apply advanced and automated 3D molecular modeling tasks, with the aim to improve decision-making in drug design projects.

Enhancing backbone sampling in Monte Carlo simulations using internal coordinates normal mode analysis.

Normal mode methods are becoming a popular alternative to sample the conformational landscape of proteins. In this study, we describe the implementation of an internal coordinate normal mode analysis method and its application in exploring protein flexibility by using the Monte Carlo method PELE. This new method alternates two different stages, a perturbation of the backbone through the application of torsional normal modes, and a resampling of the side chains. We have evaluated the new approach using two test systems, ubiquitin and c-Src kinase, and the differences to the original ANM method are assessed by comparing both results to reference molecular dynamics simulations. The results suggest that the sampled phase space in the internal coordinate approach is closer to the molecular dynamics phase space than the one coming from a Cartesian coordinate anisotropic network model. In addition, the new method shows a great speedup (∼5-7×), making it a good candidate for future normal mode implementations in Monte Carlo methods.

Binding Mode and Induced Fit Predictions for Prospective Computational Drug Design.

Computer-aided drug design plays an important role in medicinal chemistry to obtain insights into molecular mechanisms and to prioritize design strategies. Although significant improvement has been made in structure based design, it still remains a key challenge to accurately model and predict induced fit mechanisms. Most of the current available techniques either do not provide sufficient protein conformational sampling or are too computationally demanding to fit an industrial setting. The current study presents a systematic and exhaustive investigation of predicting binding modes for a range of systems using PELE (Protein Energy Landscape Exploration), an efficient and fast protein-ligand sampling algorithm. The systems analyzed (cytochrome P, kinase, protease, and nuclear hormone receptor) exhibit different complexities of ligand induced fit mechanisms and protein dynamics. The results are compared with results from classical molecular dynamics simulations and (induced fit) docking. This study shows that ligand induced side chain rearrangements and smaller to medium backbone movements are captured well in PELE. Large secondary structure rearrangements, however, remain challenging for all employed techniques. Relevant binding modes (ligand heavy atom RMSD < 1.0 Å) can be obtained by the PELE method within a few hours of simulation, positioning PELE as a tool applicable for rapid drug design cycles.

Molecular Rift: Virtual Reality for Drug Designers.

Recent advances in interaction design have created new ways to use computers. One example is the ability to create enhanced 3D environments that simulate physical presence in the real world--a virtual reality. This is relevant to drug discovery since molecular models are frequently used to obtain deeper understandings of, say, ligand-protein complexes. We have developed a tool (Molecular Rift), which creates a virtual reality environment steered with hand movements. Oculus Rift, a head-mounted display, is used to create the virtual settings. The program is controlled by gesture-recognition, using the gaming sensor MS Kinect v2, eliminating the need for standard input devices. The Open Babel toolkit was integrated to provide access to powerful cheminformatics functions. Molecular Rift was developed with a focus on usability, including iterative test-group evaluations. We conclude with reflections on virtual reality's future capabilities in chemistry and education. Molecular Rift is open source and can be downloaded from GitHub.

Ligand Binding Mechanism in Steroid Receptors: From Conserved Plasticity to Differential Evolutionary Constraints.

Steroid receptor drugs have been available for more than half a century, but details of the ligand binding mechanism have remained elusive. We solved X-ray structures of the glucocorticoid and mineralocorticoid receptors to identify a conserved plasticity at the helix 6-7 region that extends the ligand binding pocket toward the receptor surface. Since none of the endogenous ligands exploit this region, we hypothesized that it constitutes an integral part of the binding event. Extensive all-atom unbiased ligand exit and entrance simulations corroborate a ligand binding pathway that gives the observed structural plasticity a key functional role. Kinetic measurements reveal that the receptor residence time correlates with structural rearrangements observed in both structures and simulations. Ultimately, our findings reveal why nature has conserved the capacity to open up this region, and highlight how differences in the details of the ligand entry process result in differential evolutionary constraints across the steroid receptors.

CAST: a new program package for the accurate characterization of large and flexible molecular systems.

The presented program package, Conformational Analysis and Search Tool (CAST) allows the accurate treatment of large and flexible (macro) molecular systems. For the determination of thermally accessible minima CAST offers the newly developed TabuSearch algorithm, but algorithms such as Monte Carlo (MC), MC with minimization, and molecular dynamics are implemented as well. For the determination of reaction paths, CAST provides the PathOpt, the Nudge Elastic band, and the umbrella sampling approach. Access to free energies is possible through the free energy perturbation approach. Along with a number of standard force fields, a newly developed symmetry-adapted perturbation theory-based force field is included. Semiempirical computations are possible through DFTB+ and MOPAC interfaces. For calculations based on density functional theory, a Message Passing Interface (MPI) interface to the Graphics Processing Unit (GPU)-accelerated TeraChem program is available. The program is available on request.

QM/MM investigations of organic chemistry oriented questions.

About 35 years after its first suggestion, QM/MM became the standard theoretical approach to investigate enzymatic structures and processes. The success is due to the ability of QM/MM to provide an accurate atomistic picture of enzymes and related processes. This picture can even be turned into a movie if nuclei-dynamics is taken into account to describe enzymatic processes. In the field of organic chemistry, QM/MM methods are used to a much lesser extent although almost all relevant processes happen in condensed matter or are influenced by complicated interactions between substrate and catalyst. There is less importance for theoretical organic chemistry since the influence of nonpolar solvents is rather weak and the effect of polar solvents can often be accurately described by continuum approaches. Catalytic processes (homogeneous and heterogeneous) can often be reduced to truncated model systems, which are so small that pure quantum-mechanical approaches can be employed. However, since QM/MM becomes more and more efficient due to the success in software and hardware developments, it is more and more used in theoretical organic chemistry to study effects which result from the molecular nature of the environment. It is shown by many examples discussed in this review that the influence can be tremendous, even for nonpolar reactions. The importance of environmental effects in theoretical spectroscopy was already known. Due to its benefits, QM/MM can be expected to experience ongoing growth for the next decade.In the present chapter we give an overview of QM/MM developments and their importance in theoretical organic chemistry, and review applications which give impressions of the possibilities and the importance of the relevant effects. Since there is already a bunch of excellent reviews dealing with QM/MM, we will discuss fundamental ingredients and developments of QM/MM very briefly with a focus on very recent progress. For the applications we follow a similar strategy.

PathOpt--a global transition state search approach: outline of algorithm.

We propose a new algorithm to determine reaction paths and test its capability for Ar12 and Ar13 clusters. Its main ingredient is a search for the local minima on a (n-1) dimensional hyperplane (n = dimension of the complete system in Cartesian coordinates) lying perpendicular to the straight line connection between initial and final states. These minima are part of possible reaction paths and are, hence, used as starting points for an uphill search to the next transition state. First, path fragments are obtained from subsequent relaxations starting from these transition states. They can be combined with information from the straight line connection procedure to obtain complete paths. Our test computations for Ar12 and Ar13 clusters prove that PathOpt delivers several reaction paths in one round.