Henry van Den Bedem

Small molecule ligands exhibit a diverse range of conformations in solution. Upon binding to a target protein, this conformational diversity is reduced. However, ligands can retain some degree of conformational flexibility even when bound to a receptor. In the Protein Data Bank, a small number of ligands have been modeled with distinct alternative conformations that are supported by macromolecular X-ray crystallography density maps. However, the vast majority of structural models are fit to a single-ligand conformation, potentially ignoring the underlying conformational heterogeneity present in the sample. We previously developed qFit-ligand to sample diverse ligand conformations and to select a parsimonious ensemble consistent with the density. While this approach indicated that many ligands populate alternative conformations, limitations in our sampling procedures often resulted in non-physical conformations and could not model complex ligands like macrocycles. Here, we introduce several improvements to qFit-ligand, including integrating RDKit for stochastic conformational sampling. This new sampling method greatly enriches low-energy conformations of small molecules and macrocycles. We further extended qFit-ligand to identify alternative conformations in PanDDA-modified density maps from high-throughput X-ray fragment screening experiments, as well as single-particle cryo-electron microscopy density maps. The new version of qFit-ligand improves fit to electron density and reduces torsional strain relative to deposited single-conformer models and our prior version of qFit-ligand. These advances enhance the analysis of residual conformational heterogeneity present in ligand-bound structures, which can provide important insights for the rational design of therapeutic agents.

Small molecule ligands exhibit a diverse range of conformations in solution. Upon binding to a target protein, this conformational diversity is reduced. However, ligands can retain some degree of conformational flexibility even when bound to a receptor. In the Protein Data Bank (PDB), a small number of ligands have been modeled with distinct alternative conformations that are supported by macromolecular X-ray crystallography density maps. However, the vast majority of structural models are fit to a single ligand conformation, potentially ignoring the underlying conformational heterogeneity present in the sample. We previously developed qFit-ligand to sample diverse ligand conformations and to select a parsimonious ensemble consistent with the density. While this approach indicated that many ligands populate alternative conformations, limitations in our sampling procedures often resulted in non-physical conformations and could not model complex ligands like macrocycles. Here, we introduce several improvements to qFit-ligand, including integrating RDKit for stochastic conformational sampling. This new sampling method greatly enriches low energy conformations of small molecules and macrocycles. We further extended qFit-ligand to identify alternative conformations in PanDDA-modified density maps from high throughput X-ray fragment screening experiments, as well as single-particle cryo-electron microscopy (cryo-EM) density maps. The new version of qFit-ligand improves fit to electron density and reduces torsional strain relative to deposited single conformer models and our prior version of qFit-ligand. These advances enhance the analysis of residual conformational heterogeneity present in ligand-bound structures, which can provide important insights for the rational design of therapeutic agents.

Commercially available, synthesis-on-demand virtual libraries contain upward of trillions of readily synthesizable compounds for drug discovery campaigns. These libraries are a critical resource for rapid cycles of in silico discovery, property optimization and in vitro validation. However, as these libraries continue to grow exponentially in size, traditional search strategies encounter significant limitations. Here we present NeuralGenThesis (NGT), an efficient reinforcement learning approach to generate compounds from ultralarge libraries that satisfy user-specified constraints. Our method first trains a generative model over a virtual library and subsequently trains a normalizing flow to learn a distribution over latent space that decodes constraint-satisfying compounds. NGT allows multiple constraints simultaneously without dictating how molecular properties are calculated. Using NGT, we generated potent and selective inhibitors for the melanocortin-2 receptor (MC2R) from a three trillion compound library. NGT offers a powerful and scalable solution for navigating ultralarge virtual libraries, accelerating drug discovery efforts.

Small molecule ligands exhibit a diverse range of conformations in solution. Upon binding to a target protein, this conformational diversity is reduced. However, ligands can retain some degree of conformational flexibility even when bound to a receptor. In the Protein Data Bank (PDB), a small number of ligands have been modeled with distinct alternative conformations that are supported by macromolecular X-ray crystallography density maps. However, the vast majority of structural models are fit to a single ligand conformation, potentially ignoring the underlying conformational heterogeneity present in the sample. We previously developed qFit-ligand to sample diverse ligand conformations and to select a parsimonious ensemble consistent with the density. While this approach indicated that many ligands populate alternative conformations, limitations in our sampling procedures often resulted in non-physical conformations and could not model complex ligands like macrocycles. Here, we introduce several improvements to qFit-ligand, including integrating RDKit for stochastic conformational sampling. This new sampling method greatly enriches low energy conformations of small molecules and macrocycles. We further extended qFit-ligand to identify alternative conformations in PanDDA-modified density maps from high throughput X-ray fragment screening experiments, as well as single-particle cryo-electron microscopy (cryo-EM) density maps. The new version of qFit-ligand improves fit to electron density and reduces torsional strain relative to deposited single conformer models and our prior version of qFit-ligand. These advances enhance the analysis of residual conformational heterogeneity present in ligand-bound structures, which can provide important insights for the rational design of therapeutic agents.

Commercially available, synthesis-on-demand virtual libraries contain upwards of trillions of readily synthesizable compounds for drug discovery campaigns. These libraries are a critical resource for rapid cycles of in silico discovery, property optimization and in vitro validation. However, as these libraries continue to grow exponentially in size, traditional search strategies that scale linearly with the number of compounds encounter significant limitations. Here we present NeuralGenThesis (NGT), an efficient reinforcement learning approach to retrieving compounds from ultra-large libraries that satisfy a set of user-specified constraints. Our method first trains a generative model over a virtual library and subsequently trains a normalizing flow to learn a distribution over latent space that decodes constraint-satisfying compounds. NGT allows multiple constraints without dictating how molecular properties are calculated, enabling versatile searches within virtual libraries. When NGT learned a policy reporting on compound bioactivity for the melanocortin- 2 receptor (MC2R) it prospectively identified potent and selective inhibitors in a three trillion compound library. NGT offers a powerful and scalable solution for navigating ultra-large virtual libraries, accelerating drug discovery efforts.

In their folded state, biomolecules exchange between multiple conformational states that are crucial for their function. Traditional structural biology methods, such as X-ray crystallography and cryogenic electron microscopy (cryo-EM), produce density maps that are ensemble averages, reflecting molecules in various conformations. Yet, most models derived from these maps explicitly represent only a single conformation, overlooking the complexity of biomolecular structures. To accurately reflect the diversity of biomolecular forms, there is a pressing need to shift towards modeling structural ensembles that mirror the experimental data. However, the challenge of distinguishing signal from noise complicates manual efforts to create these models. In response, we introduce the latest enhancements to qFit, an automated computational strategy designed to incorporate protein conformational heterogeneity into models built into density maps. These algorithmic improvements in qFit are substantiated by superior Rfree and geometry metrics across a wide range of proteins. Importantly, unlike more complex multicopy ensemble models, the multiconformer models produced by qFit can be manually modified in most major model building software (e.g. Coot) and fit can be further improved by refinement using standard pipelines (e.g. Phenix, Refmac, Buster). By reducing the barrier of creating multiconformer models, qFit can foster the development of new hypotheses about the relationship between macromolecular conformational dynamics and function.

In their folded state, biomolecules exchange between multiple conformational states that are crucial for their function. Traditional structural biology methods, such as X-ray crystallography and cryogenic electron microscopy (cryo-EM), produce density maps that are ensemble averages, reflecting molecules in various conformations. Yet, most models derived from these maps explicitly represent only a single conformation, overlooking the complexity of biomolecular structures. To accurately reflect the diversity of biomolecular forms, there is a pressing need to shift towards modeling structural ensembles that mirror the experimental data. However, the challenge of distinguishing signal from noise complicates manual efforts to create these models. In response, we introduce the latest enhancements to qFit, an automated computational strategy designed to incorporate protein conformational heterogeneity into models built into density maps. These algorithmic improvements in qFit are substantiated by superior R _free and geometry metrics across a wide range of proteins. Importantly, unlike more complex multicopy ensemble models, the multiconformer models produced by qFit can be manually modified in most major model building software (e.g. Coot) and fit can be further improved by refinement using standard pipelines (e.g. Phenix, Refmac, Buster). By reducing the barrier of creating multiconformer models, qFit can foster the development of new hypotheses about the relationship between macromolecular conformational dynamics and function.

Virtual screening is a widely used tool for drug discovery, but its predictive power can vary dramatically depending on how much structural data is available. In the best case, crystal structures of a ligand-bound protein can help find more potent ligands. However, virtual screens tend to be less predictive when only ligand-free crystal structures are available, and even less predictive if a homology model or other predicted structure must be used. Here, we explore the possibility that this situation can be improved by better accounting for protein dynamics, as simulations started from a single structure have a reasonable chance of sampling nearby structures that are more compatible with ligand binding. As a specific example, we consider the cancer drug target PPM1D/Wip1 phosphatase, a protein that lacks crystal structures. High-throughput screens have led to the discovery of several allosteric inhibitors of PPM1D, but their binding mode remains unknown. To enable further drug discovery efforts, we assessed the predictive power of an AlphaFold-predicted structure of PPM1D and a Markov state model (MSM) built from molecular dynamics simulations initiated from that structure. Our simulations reveal a cryptic pocket at the interface between two important structural elements, the flap and hinge regions. Using deep learning to predict the pose quality of each docked compound for the active site and cryptic pocket suggests that the inhibitors strongly prefer binding to the cryptic pocket, consistent with their allosteric effect. The predicted affinities for the dynamically uncovered cryptic pocket also recapitulate the relative potencies of the compounds (τ_b = 0.70) better than the predicted affinities for the static AlphaFold-predicted structure (τ_b = 0.42). Taken together, these results suggest that targeting the cryptic pocket is a good strategy for drugging PPM1D and, more generally, that conformations selected from simulation can improve virtual screening when limited structural data is available.