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.

Autoencoders for dimensionality reduction in molecular dynamics: Collective variable dimension, biasing, and transition states.

The heat shock protein 90 (Hsp90) is a molecular chaperone that controls the folding and activation of client proteins using the free energy of ATP hydrolysis. The Hsp90 active site is in its N-terminal domain (NTD). Our goal is to characterize the dynamics of NTD using an autoencoder-learned collective variable (CV) in conjunction with adaptive biasing force Langevin dynamics. Using dihedral analysis, we cluster all available experimental Hsp90 NTD structures into distinct native states. We then perform unbiased molecular dynamics (MD) simulations to construct a dataset that represents each state and use this dataset to train an autoencoder. Two autoencoder architectures are considered, with one and two hidden layers, respectively, and bottlenecks of dimension k ranging from 1 to 10. We demonstrate that the addition of an extra hidden layer does not significantly improve the performance, while it leads to complicated CVs that increase the computational cost of biased MD calculations. In addition, a two-dimensional (2D) bottleneck can provide enough information of the different states, while the optimal bottleneck dimension is five. For the 2D bottleneck, the 2D CV is directly used in biased MD simulations. For the five-dimensional (5D) bottleneck, we perform an analysis of the latent CV space and identify the pair of CV coordinates that best separates the states of Hsp90. Interestingly, selecting a 2D CV out of the 5D CV space leads to better results than directly learning a 2D CV and allows observation of transitions between native states when running free energy biased dynamics.

Explaining and avoiding failure modes in goal-directed generation of small molecules.

Despite growing interest and success in automated in-silico molecular design, questions remain regarding the ability of goal-directed generation algorithms to perform unbiased exploration of novel chemical spaces. A specific phenomenon has recently been highlighted: goal-directed generation guided with machine learning models produce molecules with high scores according to the optimization model, but low scores according to control models, even when trained on the same data distribution and the same target. In this work, we show that this worrisome behavior is actually due to issues with the predictive models and not the goal-directed generation algorithms. We show that with appropriate predictive models, this issue can be resolved, and molecules generated have high scores according to both the optimization and the control models.

Contact Map Fingerprints of Protein-Ligand Unbinding Trajectories Reveal Mechanisms Determining Residence Times Computed from Scaled Molecular Dynamics.

The binding kinetic properties of potential drugs may significantly influence their subsequent clinical efficacy. Predictions of these properties based on computer simulations provide a useful alternative to their expensive and time-consuming experimental counterparts, even at an early drug discovery stage. Herein, we perform scaled molecular dynamics (ScaledMD) simulations on a set of 27 ligands of HSP90 belonging to more than seven chemical series to estimate their relative residence times. We introduce two new techniques for the analysis and the classification of the simulated unbinding trajectories. The first technique, which helps in estimating the limits of the free energy well around the bound state, and the second one, based on a new contact map fingerprint, allow the description and the comparison of the paths that lead to unbinding. Using these analyses, we find that ScaledMD's relative residence time generally enables the identification of the slowest unbinders. We propose an explanation for the underestimation of the residence times of a subset of compounds, and we investigate how the biasing in ScaledMD can affect the mechanistic insights that can be gained from the simulations.

Protein loops with multiple meta-stable conformations: A challenge for sampling and scoring methods.

Flexible regions in proteins, such as loops, cannot be represented by a single conformation. Instead, conformational ensembles are needed to provide a more global picture. In this context, identifying statistically meaningful conformations within an ensemble generated by loop sampling techniques remains an open problem. The difficulty is primarily related to the lack of structural data about these flexible regions. With the majority of structural data coming from x-ray crystallography and ignoring plasticity, the conception and evaluation of loop scoring methods is challenging. In this work, we compare the performance of various scoring methods on a set of eight protein loops that are known to be flexible. The ability of each method to identify and select all of the known conformations is assessed, and the underlying energy landscapes are produced and projected to visualize the qualitative differences obtained when using the methods. Statistical potentials are found to provide considerable reliability despite their being designed to tradeoff accuracy for lower computational cost. On a large pool of loop models, they are capable of filtering out statistically improbable states while retaining those that resemble known (and thus likely) conformations. However, computationally expensive methods are still required for more precise assessment and structural refinement. The results also highlight the importance of employing several scaffolds for the protein, due to the high influence of small structural rearrangements in the rest of the protein over the modeled energy landscape for the loop.

Scaffold-Constrained Molecular Generation.

One of the major applications of generative models for drug discovery targets the lead-optimization phase. During the optimization of a lead series, it is common to have scaffold constraints imposed on the structure of the molecules designed. Without enforcing such constraints, the probability of generating molecules with the required scaffold is extremely low and hinders the practicality of generative models for de novo drug design. To tackle this issue, we introduce a new algorithm, named SAMOA (Scaffold Constrained Molecular Generation), to perform scaffold-constrained in silico molecular design. We build on the well-known SMILES-based Recurrent Neural Network (RNN) generative model, with a modified sampling procedure to achieve scaffold-constrained generation. We directly benefit from the associated reinforcement learning methods, allowing to design molecules optimized for different properties while exploring only the relevant chemical space. We showcase the method's ability to perform scaffold-constrained generation on various tasks: designing novel molecules around scaffolds extracted from SureChEMBL chemical series, generating novel active molecules on the Dopamine Receptor D2 (DRD2) target, and finally, designing predicted actives on the MMP-12 series, an industrial lead-optimization project.

Conformational changes in antibody Fab fragments upon binding and their consequences on the performance of docking algorithms.

BACKGROUND: The existence of conformational changes in antibodies upon binding has been previously established. However, existing analyses focus on individual cases and no quantitative study provides a more global view of potential moves and repacking, especially on recent data. The present study focuses on analyzing the conformational changes in various antibodies upon binding, providing quantitative observations to be exploited for antibody-related modeling. METHODS: Cartesian and dihedral Root-Mean-Squared Deviations were calculated for different subparts of 27 different antibodies, for which X-ray structures in the bound and unbound states are available. Elbow angle variations were also calculated. Previously reported results of four docking algorithms were condensed into one score giving overall docking success for each of 16 antibody-antigen cases. RESULTS: Very diverse movements are observed upon binding. While many loops stay very rigid, several others display side-chain repacking or backbone rearrangements, or both, at many different levels. Large conformational changes restricted to one or more antibody hypervariable loops were found to be a better indicator of docking difficulty than overall conformational variation at the antibody-antigen interface. However, the failure of docking algorithms on some almost-rigid cases shows that scoring is still a major bottleneck in docking pose prediction. CONCLUSIONS: This study is aimed to help scientists working on antibody analysis and design by giving insights into the nature and the extent of conformational changes at different levels upon antigen binding.

Estimation of Drug-Target Residence Times by τ-Random Acceleration Molecular Dynamics Simulations.

Drug-target residence time (τ), one of the main determinants of drug efficacy, remains highly challenging to predict computationally and, therefore, is usually not considered in the early stages of drug design. Here, we present an efficient computational method, τ-random acceleration molecular dynamics (τRAMD), for the ranking of drug candidates by their residence time and obtaining insights into ligand-target dissociation mechanisms. We assessed τRAMD on a data set of 70 diverse drug-like ligands of the N-terminal domain of HSP90α, a pharmaceutically important target with a highly flexible binding site, obtaining computed relative residence times with an accuracy of about 2.3τ for 78% of the compounds and less than 2.0τ within congeneric series. Analysis of dissociation trajectories reveals features that affect ligand unbinding rates, including transient polar interactions and steric hindrance. These results suggest that τRAMD will be widely applicable as a computationally efficient aid to improving drug residence times during lead optimization.

Inhibition of tumor angiogenesis and growth by a small-molecule multi-FGF receptor blocker with allosteric properties.

Receptor tyrosine kinases (RTK) are targets for anticancer drug development. To date, only RTK inhibitors that block orthosteric binding of ligands and substrates have been developed. Here, we report the pharmacologic characterization of the chemical SSR128129E (SSR), which inhibits fibroblast growth factor receptor (FGFR) signaling by binding to the extracellular FGFR domain without affecting orthosteric FGF binding. SSR exhibits allosteric properties, including probe dependence, signaling bias, and ceiling effects. Inhibition by SSR is highly conserved throughout the animal kingdom. Oral delivery of SSR inhibits arthritis and tumors that are relatively refractory to anti-vascular endothelial growth factor receptor-2 antibodies. Thus, orally-active extracellularly acting small-molecule modulators of RTKs with allosteric properties can be developed and may offer opportunities to improve anticancer treatment.

Molecular mechanism of SSR128129E, an extracellularly acting, small-molecule, allosteric inhibitor of FGF receptor signaling.

The fibroblast growth factor (FGF)/fibroblast growth factor receptor (FGFR) signaling network plays an important role in cell growth, survival, differentiation, and angiogenesis. Deregulation of FGFR signaling can lead to cancer development. Here, we report an FGFR inhibitor, SSR128129E (SSR), that binds to the extracellular part of the receptor. SSR does not compete with FGF for binding to FGFR but inhibits FGF-induced signaling linked to FGFR internalization in an allosteric manner, as shown by crystallography studies, nuclear magnetic resonance, Fourier transform infrared spectroscopy, molecular dynamics simulations, free energy calculations, structure-activity relationship analysis, and FGFR mutagenesis. Overall, SSR is a small molecule allosteric inhibitor of FGF/FGFR signaling, acting via binding to the extracellular part of the FGFR.

Free-energy-based methods for binding profile determination in a congeneric series of CDK2 inhibitors.

Free-energy pathway methods show great promise in computing the mode of action and the free energy profile associated with the binding of small molecules with proteins, but are generally very computationally demanding. Here we apply a novel approach based on metadynamics and path collective variables. We show that this combination is able to find an optimal reaction coordinate and the free energy profile of binding with explicit solvent and full flexibility, while minimizing human intervention and computational costs. We apply it to predict the binding affinity of a congeneric series of 5 CDK2 inhibitors. The predicted binding free energy profiles are in accordance with experiment.

Influence of heparin mimetics on assembly of the FGF.FGFR4 signaling complex.

Fibroblast growth factor (FGF) signaling regulates mammalian development and metabolism, and its dysregulation is implicated in many inherited and acquired diseases, including cancer. Heparan sulfate glycosaminoglycans (HSGAGs) are essential for FGF signaling as they promote FGF.FGF receptor (FGFR) binding and dimerization. Using novel organic synthesis protocols to prepare homogeneously sulfated heparin mimetics (HM), including hexasaccharide (HM(6)), octasaccharide (HM(8)), and decasaccharide (HM(10)), we tested the ability of these HM to support FGF1 and FGF2 signaling through FGFR4. Biological assays show that both HM(8) and HM(10) are significantly more potent than HM(6) in promoting FGF2-mediated FGFR4 signaling. In contrast, all three HM have comparable activity in promoting FGF1.FGFR4 signaling. To understand the molecular basis for these differential activities in FGF1/2.FGFR4 signaling, we used NMR spectroscopy, isothermal titration calorimetry, and size-exclusion chromatography to characterize binding interactions of FGF1/2 with the isolated Ig-domain 2 (D2) of FGFR4 in the presence of HM, and binary interactions of FGFs and D2 with HM. Our data confirm the existence of both a secondary FGF1.FGFR4 interaction site and a direct FGFR4.FGFR4 interaction site thus supporting the formation of the symmetric mode of FGF.FGFR dimerization in solution. Moreover, our results show that the observed higher activity of HM(8) relative to HM(6) in stimulating FGF2.FGFR4 signaling correlates with the higher affinity of HM(8) to bind and dimerize FGF2. Notably FGF2.HM(8) exhibits pronounced positive binding cooperativity. Based on our findings we propose a refined symmetric FGF.FGFR dimerization model, which incorporates the differential ability of HM to dimerize FGFs.

Reactivity of phosphate monoester monoanions in aqueous solution. 1. Quantum mechanical calculations support the existence of "anionic zwitterion" MeO(+)(H)PO(3)(2-) as a key intermediate in the dissociative hydrolysis of the methyl phosphate anion.

The dissociative hydrolysis reaction of the methyl phosphate monoanion has been studied for the reactant species CH(3)OPO(3)H(-) (1) and CH(3)OPO(3)H(-) x H(2)O (1a) in the gas and aqueous phases by density functional theory (B3LYP) calculations. Nonspecific solvation effects were taken into account with the polarizable continuum model PCM either by solvating the gas-phase reaction paths or by performing geometry searches directly in the presence of the solvation correction. In agreement with previous theoretical studies, our gas-phase calculations indicate that proton transfer to the methoxy group of 1 is concerted with P-O bond cleavage. In contrast, optimizations performed with the PCM solvation model establish the existence of the tautomeric form CH(3)O(+)(H)PO(3)(2-) (2) as an intermediate, indicating that proton transfer and P-O bond cleavage become uncoupled in aqueous solution. The dissociative pathway of 1a is energetically favored over the dissociative pathway of 1 only when the added water molecule plays an active catalytic role in the prototropic rearrangement 1 <--> 2. In that case, it is found that the collapse (via P-O bond cleavage) of the hydrated zwitterionic form CH(3)O(+)(H)PO(3)(2-) x H(2)O (2a) is rate-determining. This collapse may occur by a stepwise mechanism through a very short-lived metaphosphate intermediate (PO(3)(-)), or by a concerted S(N)2-like displacement through a loose metaphosphate-like transition state. The present calculations do not allow a distinction to be made between these two alternatives, which are both in excellent agreement with experiment. The present study also reveals that PO(3)(-) reacts selectively with CH(3)OH and H(2)O nucleophiles in aqueous solution. However, the observed selectivity of metaphosphate is governed by solvation effects, not nucleophilicity (water being much more effective than methanol in capturing PO(3)(-)). This arises from a better solvation of the addition product H(2)O(+)PO(3)(2-) as compared to CH(3)O(+)(H)PO(3)(2-).