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.

Probing Factor Xa Protein-Ligand Interactions: Accurate Free Energy Calculations and Experimental Validations of Two Series of High-Affinity Ligands.

The accurate prediction of protein-ligand binding affinity belongs to one of the central goals in computer-based drug design. Molecular dynamics (MD)-based free energy calculations have become increasingly popular in this respect due to their accuracy and solid theoretical basis. Here, we present a combined study which encompasses experimental and computational studies on two series of factor Xa ligands, which enclose a broad chemical space including large modifications of the central scaffold. Using this integrated approach, we identified several new ligands with different heterocyclic scaffolds different from the previously identified indole-2-carboxamides that show superior or similar affinity. Furthermore, the so far underexplored terminal alkyne moiety proved to be a suitable non-classical bioisosteric replacement for the higher halogen-π aryl interactions. With this challenging example, we demonstrated the ability of the MD-based non-equilibrium free energy calculation approach for guiding crucial modifications in the lead optimization process, such as scaffold replacement and single-site modifications at molecular interaction hot spots.

Molecular basis for inhibiting human glucose transporters by exofacial inhibitors.

Human glucose transporters (GLUTs) are responsible for cellular uptake of hexoses. Elevated expression of GLUTs, particularly GLUT1 and GLUT3, is required to fuel the hyperproliferation of cancer cells, making GLUT inhibitors potential anticancer therapeutics. Meanwhile, GLUT inhibitor-conjugated insulin is being explored to mitigate the hypoglycemia side effect of insulin therapy in type 1 diabetes. Reasoning that exofacial inhibitors of GLUT1/3 may be favored for therapeutic applications, we report here the engineering of a GLUT3 variant, designated GLUT3exo, that can be probed for screening and validating exofacial inhibitors. We identify an exofacial GLUT3 inhibitor SA47 and elucidate its mode of action by a 2.3 Å resolution crystal structure of SA47-bound GLUT3. Our studies serve as a framework for the discovery of GLUTs exofacial inhibitors for therapeutic development.

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.

Using Graph Databases to Investigate Trends in Structure-Activity Relationship Networks.

Mining the steadily increasing amount of chemical and biological data is a key challenge in drug discovery. Graph databases offer viable alternatives for capturing interrelationships between molecules and for generating novel insights for design. In a graph database, molecules and their properties are mapped to nodes, while relationships are described by edges. Here, we introduce a graph database for navigation in chemical space, analogue searching, and structure-activity relationship (SAR) analysis. We illustrate this concept using hERG channel inhibitors from ChEMBL to extract SAR knowledge. This graph database is built using different relationships, namely 2D-fingerprint similarity, matched molecular pairs, topomer distances, and structure-activity landscape indices (SALI). Typical applications include retrieving analogues linked by single or multiple edge paths to the query compound as well as detection of nonadditive SAR features. Finally, we identify triplets of linked molecules for clustering. The speed of searching and analysis allows the user to interactively navigate the database and to address complex questions in real-time.

Resolving Binding Events on the Multifunctional Human Serum Albumin.

Physiological processes rely on initial recognition events between cellular components and other molecules or modalities. Biomolecules can have multiple sites or mode of interaction with other molecular entities, so that a resolution of the individual binding events in terms of spatial localization as well as association and dissociation kinetics is required for a meaningful description. Here we describe a trichromatic fluorescent binding- and displacement assay for simultaneous monitoring of three individual binding sites in the important transporter and binding protein human serum albumin. Independent investigations of binding events by X-ray crystallography and time-resolved dynamics measurements (switchSENSE technology) confirm the validity of the assay, the localization of binding sites and furthermore reveal conformational changes associated with ligand binding. The described assay system allows for the detailed characterization of albumin-binding drugs and is therefore well-suited for prediction of drug-drug and drug-food interactions. Moreover, conformational changes, usually associated with binding events, can also be analyzed.

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.

Impact of Different Automated Binding Pose Generation Approaches on Relative Binding Free Energy Simulations.

Relative binding free energy (RBFE) prediction methods such as free energy perturbation (FEP) are important today for estimating protein-ligand binding affinities. Significant hardware and algorithmic improvements now allow for simulating congeneric series within days. Therefore, RBFE calculations have an enormous potential for structure-based drug discovery. As typically only a few representative crystal structures for a series are available, other ligands and design proposals must be reliably superimposed for meaningful results. An observed significant effect of the alignment on FEP led us to develop an alignment approach combining docking with maximum common substructure (MCS) derived core constraints from the most similar reference pose, named MCS-docking workflow. We then studied the effect of binding pose generation on the accuracy of RBFE predictions using six ligand series from five pharmaceutically relevant protein targets. Overall, the MCS-docking workflow generated consistent poses for most of the ligands in the investigated series. While multiple alignment methods often resulted in comparable FEP predictions, for most of the cases herein, the MCS-docking workflow produced the best accuracy in predictions. Furthermore, the FEP analysis data strongly support the hypothesis that the accuracy of RBFE predictions depends on input poses to construct the perturbation map. Therefore, an automated workflow without manual intervention minimizes potential errors and obtains the most useful predictions with significant impact for structure-based design.

Design, Synthesis, and Pharmacological Evaluation of Potent Positive Allosteric Modulators of the Glucagon-like Peptide-1 Receptor (GLP-1R).

The therapeutic success of peptidic GLP-1 receptor agonists for treatment of type 2 diabetes mellitus (T2DM) motivated our search for orally bioavailable small molecules that can activate the GLP-1 receptor (GLP-1R) as a well-validated target for T2DM. Here, the discovery and characterization of a potent and selective positive allosteric modulator (PAM) for GLP-1R based on a 3,4,5,6-tetrahydro-1H-1,5-epiminoazocino[4,5-b]indole scaffold is reported. Optimization of this series from HTS was supported by a GLP-1R ligand binding model. Biological in vitro testing revealed favorable ADME and pharmacological profiles for the best compound 19. Characterization by in vivo pharmacokinetic and pharmacological studies demonstrated that 19 activates GLP-1R as positive allosteric modulator (PAM) in the presence of the much less active endogenous degradation product GLP1(9-36)NH2 of the potent endogenous ligand GLP-1(7-36)NH2. While these data suggest the potential of small molecule GLP-1R PAMs for T2DM treatment, further optimization is still required towards a clinical candidate.

Predictive Multitask Deep Neural Network Models for ADME-Tox Properties: Learning from Large Data Sets.

Successful drug discovery projects require control and optimization of compound properties related to pharmacokinetics, pharmacodynamics, and safety. While volume and chemotype coverage of public and corporate ADME-Tox (absorption, distribution, excretion, metabolism, and toxicity) databases are constantly growing, deep neural nets (DNN) emerged as transformative artificial intelligence technology to analyze those challenging data. Relevant features are automatically identified, while appropriate data can also be combined to multitask networks to evaluate hidden trends among multiple ADME-Tox parameters for implicitly correlated data sets. Here we describe a novel, fully industrialized approach to parametrize and optimize the setup, training, application, and visual interpretation of DNNs to model ADME-Tox data. Investigated properties include microsomal lability in different species, passive permeability in Caco-2/TC7 cells, and logD. Statistical models are developed using up to 50 000 compounds from public or corporate databases. Both the choice of DNN hyperparameters and the type and quantity of molecular descriptors were found to be important for successful DNN modeling. Alternate learning of multiple ADME-Tox properties, resulting in a multitask approach, performs statistically superior on most studied data sets in comparison to DNN single-task models and also provides a scalable method to predict ADME-Tox properties from heterogeneous data. For example, predictive quality using external validation sets was improved from R2 of 0.6 to 0.7 comparing single-task and multitask DNN networks from human metabolic lability data. Besides statistical evaluation, a new visualization approach is introduced to interpret DNN models termed "response map", which is useful to detect local property gradients based on structure fragmentation and derivatization. This method is successfully applied to visualize fragmental contributions to guide further design in drug discovery programs, as illustrated by CRCX3 antagonists and renin inhibitors, respectively.

The Role of Bridging Water and Hydrogen Bonding as Key Determinants of Noncovalent Protein-Carbohydrate Recognition.

Mechanisms of protein-carbohydrate recognition attract a lot of interest due to their roles in various cellular processes and metabolism disorders. We have performed a large-scale analysis of protein structures solved in complex with glucose, galactose and their substituted analogues. We found that, on average, sugar molecules establish five hydrogen bonds (HBs) in the binding site, including one to three HBs with bridging water molecules. The free energy contribution of bridging and direct HBs was estimated using the free energy perturbation (FEP+) methodology for mono- and disaccharides that bind to l-ABP, ttGBP, TrmB, hGalectin-1 and hGalectin-3. We show that removing hydroxy groups that are engaged in direct HBs with the charged groups of Asp, Arg and Glu residues, protein backbone amide or buried water dramatically decreases binding affinity. In contrast, all solvent-exposed hydroxy groups and hydroxy groups engaged in HBs with the solvent-exposed bridging water molecules contribute weakly to binding affinity and so can be replaced to optimize ligand potency. Finally, we rationalize an effect of binding site water replacement on the binding affinity to l-ABP.

Characterizing hydration sites in protein-ligand complexes towards the design of novel ligands.

Water is an essential part of protein binding sites and mediates interactions to ligands. Its displacement by ligand parts affects the free binding energy of resulting protein-ligand complexes. Therefore the characterization of solvation properties is important for design. Of particular interest is the propensity of localized water to be favorably displaced by a ligand. This review discusses two popular computational approaches addressing these questions, namely WaterMap based on statistical mechanics analysis of MD simulations and 3D RISM based on integral equation theory of liquids. The theoretical background and recent applications in structure-based design will be presented.

Identification and Characterization of a Single High-Affinity Fatty Acid Binding Site in Human Serum Albumin.

A single high-affinity fatty acid binding site in the important human transport protein serum albumin (HSA) is identified and characterized using an NBD (7-nitrobenz-2-oxa-1,3-diazol-4-yl)-C12 fatty acid. This ligand exhibits a 1:1 binding stoichiometry in its HSA complex with high site-specificity. The complex dissociation constant is determined by titration experiments as well as radioactive equilibrium dialysis. Competition experiments with the known HSA-binding drugs warfarin and ibuprofen confirm the new binding site to be different from Sudlow-sites I and II. These binding studies are extended to other albumin binders and fatty acid derivatives. Furthermore an X-ray crystal structure allows locating the binding site in HSA subdomain IIA. The knowledge about this novel HSA site will be important for drug depot development and for understanding drug-protein interaction, which are important prerequisites for modulation of drug pharmacokinetics.

Thermodynamic Characterization of Hydration Sites from Integral Equation-Derived Free Energy Densities: Application to Protein Binding Sites and Ligand Series.

Water molecules play an essential role for mediating interactions between ligands and protein binding sites. Displacement of specific water molecules can favorably modulate the free energy of binding of protein-ligand complexes. Here, the nature of water interactions in protein binding sites is investigated by 3D RISM (three-dimensional reference interaction site model) integral equation theory to understand and exploit local thermodynamic features of water molecules by ranking their possible displacement in structure-based design. Unlike molecular dynamics-based approaches, 3D RISM theory allows for fast and noise-free calculations using the same detailed level of solute-solvent interaction description. Here we correlate molecular water entities instead of mere site density maxima with local contributions to the solvation free energy using novel algorithms. Distinct water molecules and hydration sites are investigated in multiple protein-ligand X-ray structures, namely streptavidin, factor Xa, and factor VIIa, based on 3D RISM-derived free energy density fields. Our approach allows the semiquantitative assessment of whether a given structural water molecule can potentially be targeted for replacement in structure-based design. Finally, PLS-based regression models from free energy density fields used within a 3D-QSAR approach (CARMa - comparative analysis of 3D RISM Maps) are shown to be able to extract relevant information for the interpretation of structure-activity relationship (SAR) trends, as demonstrated for a series of serine protease inhibitors.

Discovery and Optimization of 1-Phenoxy-2-aminoindanes as Potent, Selective, and Orally Bioavailable Inhibitors of the Na+/H+ Exchanger Type 3 (NHE3).

The design, synthesis, and structure-activity relationship of 1-phenoxy-2-aminoindanes as inhibitors of the Na+/H+ exchanger type 3 (NHE3) are described based on a hit from high-throughput screening (HTS). The chemical optimization resulted in the discovery of potent, selective, and orally bioavailable NHE3 inhibitors with 13d as best compound, showing high in vitro permeability and lacking CYP2D6 inhibition as main optimization parameters. Aligning 1-phenoxy-2-aminoindanes onto the X-ray structure of 13d then provided 3D-QSAR models for NHE3 inhibition capturing guidelines for optimization. These models showed good correlation coefficients and allowed for activity estimation. In silico ADMET models for Caco-2 permeability and CYP2D6 inhibition were also successfully applied for this series. Moreover, docking into the CYP2D6 X-ray structure provided a reliable alignment for 3D-QSAR models. Finally 13d, renamed as SAR197, was characterized in vitro and by in vivo pharmacokinetic (PK) and pharmacological studies to unveil its potential for reduction of obstructive sleep apneas.