Benchmarking Active Learning Protocols for Ligand-Binding Affinity Prediction.

Active learning (AL) has become a powerful tool in computational drug discovery, enabling the identification of top binders from vast molecular libraries. To design a robust AL protocol, it is important to understand the influence of AL parameters, as well as the features of the data sets on the outcomes. We use four affinity data sets for different targets (TYK2, USP7, D2R, Mpro) to systematically evaluate the performance of machine learning models [Gaussian process (GP) model and Chemprop model], sample selection protocols, and the batch size based on metrics describing the overall predictive power of the model (R2, Spearman rank, root-mean-square error) as well as the accurate identification of top 2%/5% binders (Recall, F1 score). Both models have a comparable Recall of top binders on large data sets, but the GP model surpasses the Chemprop model when training data are sparse. A larger initial batch size, especially on diverse data sets, increased the Recall of both models as well as overall correlation metrics. However, for subsequent cycles, smaller batch sizes of 20 or 30 compounds proved to be desirable. Furthermore, adding artificial Gaussian noise to the data up to a certain threshold still allowed the model to identify clusters with top-scoring compounds. However, excessive noise (<1σ) did impact the model's predictive and exploitative capabilities.

Deconstructing allostery by computational assessment of the binding determinants of allosteric PTP1B modulators.

Fragment-based drug discovery is an established methodology for finding hit molecules that can be elaborated into lead compounds. However it is currently challenging to predict whether fragment hits that do not bind to an orthosteric site could be elaborated into allosteric modulators, as in these cases binding does not necessarily translate into a functional effect. We propose a workflow using Markov State Models (MSMs) with steered molecular dynamics (sMD) to assess the allosteric potential of known binders. sMD simulations are employed to sample protein conformational space inaccessible to routine equilibrium MD timescales. Protein conformations sampled by sMD provide starting points for seeded MD simulations, which are combined into MSMs. The methodology is demonstrated on a dataset of protein tyrosine phosphatase 1B ligands. Experimentally confirmed allosteric inhibitors are correctly classified as inhibitors, whereas the deconstructed analogues show reduced inhibitory activity. Analysis of the MSMs provide insights into preferred protein-ligand arrangements that correlate with functional outcomes. The present methodology may find applications for progressing fragments towards lead molecules in FBDD campaigns.

Open Binding Pose Metadynamics: An Effective Approach for the Ranking of Protein-Ligand Binding Poses.

Predicting the correct pose of a ligand binding to a protein and its associated binding affinity is of great importance in computer-aided drug discovery. A number of approaches have been developed to these ends, ranging from the widely used fast molecular docking to the computationally expensive enhanced sampling molecular simulations. In this context, methods such as coarse-grained metadynamics and binding pose metadynamics (BPMD) use simulations with metadynamics biasing to probe the binding affinity without trying to fully converge the binding free energy landscape in order to decrease the computational cost. In BPMD, the metadynamics bias perturbs the ligand away from the initial pose. The resistance of the ligand to this bias is used to calculate a stability score. The method has been shown to be useful in reranking predicted binding poses from docking. Here, we present OpenBPMD, an open-source Python reimplementation and reinterpretation of BPMD. OpenBPMD is powered by the OpenMM simulation engine and uses a revised scoring function. The algorithm was validated by testing it on a wide range of targets and showing that it matches or exceeds the performance of the original BPMD. We also investigated the role of accurate water positioning on the performance of the algorithm and showed how the combination with a grand-canonical Monte Carlo algorithm improves the accuracy of the predictions.

Membrane Interactions of α-Synuclein Revealed by Multiscale Molecular Dynamics Simulations, Markov State Models, and NMR.

α-Synuclein (αS) is a presynaptic protein that binds to cell membranes and is linked to Parkinson's disease (PD). Binding of αS to membranes is a likely first step in the molecular pathophysiology of PD. The αS molecule can adopt multiple conformations, being largely disordered in water, adopting a ß-sheet conformation when present in amyloid fibrils, and forming a dynamic multiplicity of α-helical conformations when bound to lipid bilayers and related membrane-mimetic surfaces. Multiscale molecular dynamics simulations in conjunction with nuclear magnetic resonance (NMR) and cross-linking mass spectrometry (XLMS) measurements are used to explore the interactions of αS with an anionic lipid bilayer. The simulations and NMR measurements together reveal a break in the helical structure of the central non-amyloid-ß component (NAC) region of αS in the vicinity of residues 65-70, which may facilitate subsequent oligomer formation. Coarse-grained simulations of αS starting from the structure of αS when bound to a detergent micelle reveal the overall pattern of protein contacts to anionic lipid bilayers, while subsequent all-atom simulations provide details of conformational changes upon membrane binding. In particular, simulations and NMR data for liposome-bound αS indicate incipient ß-strand formation in the NAC region, which is supported by intramolecular contacts seen via XLMS and simulations. Markov state models based on the all-atom simulations suggest a mechanism of conformational change of membrane-bound αS via a dynamic helix break in the region of residue 65 in the NAC region. The emergent dynamic model of membrane-interacting αS advances our understanding of the mechanism of PD, potentially aiding the design of novel therapeutic approaches.

Combining Machine Learning and Enhanced Sampling Techniques for Efficient and Accurate Calculation of Absolute Binding Free Energies.

Calculating absolute binding free energies is challenging and important. In this paper, we test some recently developed metadynamics-based methods and develop a new combination with a Hamiltonian replica-exchange approach. The methods were tested on 18 chemically diverse ligands with a wide range of different binding affinities to a complex target; namely, human soluble epoxide hydrolase. The results suggest that metadynamics with a funnel-shaped restraint can be used to calculate, in a computationally affordable and relatively accurate way, the absolute binding free energy for small fragments. When used in combination with an optimal pathlike variable obtained using machine learning or with the Hamiltonian replica-exchange algorithm SWISH, this method can achieve reasonably accurate results for increasingly complex ligands, with a good balance of computational cost and speed. An additional benefit of using the combination of metadynamics and SWISH is that it also provides useful information about the role of water in the binding mechanism.

Small molecules that inhibit TNF signalling by stabilising an asymmetric form of the trimer.

Tumour necrosis factor (TNF) is a cytokine belonging to a family of trimeric proteins; it has been shown to be a key mediator in autoimmune diseases such as rheumatoid arthritis and Crohn's disease. While TNF is the target of several successful biologic drugs, attempts to design small molecule therapies directed to this cytokine have not led to approved products. Here we report the discovery of potent small molecule inhibitors of TNF that stabilise an asymmetrical form of the soluble TNF trimer, compromising signalling and inhibiting the functions of TNF in vitro and in vivo. This discovery paves the way for a class of small molecule drugs capable of modulating TNF function by stabilising a naturally sampled, receptor-incompetent conformation of TNF. Furthermore, this approach may prove to be a more general mechanism for inhibiting protein-protein interactions.

Computational Exploration of Conformational Transitions in Protein Drug Targets.

Protein drug targets vary from highly structured to completely disordered; either way dynamics governs function. Hence, understanding the dynamical aspects of how protein targets function can enable improved interventions with drug molecules. Computational approaches offer highly detailed structural models of protein dynamics which are becoming more predictive as model quality and sampling power improve. However, the most advanced and popular models still have errors owing to imperfect parameter sets and often cannot access longer timescales of many crucial biological processes. Experimental approaches offer more certainty but can struggle to detect and measure lightly populated conformations of target proteins and subtle allostery. An emerging solution is to integrate available experimental data into advanced molecular simulations. In the future, molecular simulation in combination with experimental data may be able to offer detailed models of important drug targets such that improved functional mechanisms or selectivity can be accessed.

Quantifying Secondary Structure Changes in Calmodulin Using 2D-IR Spectroscopy.

Revealing the details of biomolecular processes in solution needs tools that can monitor structural dynamics over a range of time and length scales. We assess the ability of 2D-IR spectroscopy in combination with multivariate data analysis to quantify changes in secondary structure of the multifunctional calcium-binding messenger protein Calmodulin (CaM) as a function of temperature and Ca2+ concentration. Our approach produced quantitative agreement with circular dichroism (CD) spectroscopy in detecting the domain melting transitions of Ca2+-free (apo) CaM (reduction in α-helix structure by 13% (CD) and 15% (2D)). 2D-IR also allows accurate differentiation between melting transitions and generic heating effects observed in the more thermally stable Ca2+-bound (holo) CaM. The functionally relevant random-coil-α-helix transition associated with Ca2+ uptake that involves just 7-8 out of a total of 148 amino acid residues was clearly detected. Temperature-dependent Molecular Dynamics (MD) simulations show that apo-CaM exists in dynamic equilibrium with holo-like conformations, while Ca2+ uptake reduces conformational flexibility. The ability to combine quantitative structural insight from 2D-IR with MD simulations thus offers a powerful approach for measuring subtle protein conformational changes in solution.

Allosteric mechanism of action of the therapeutic anti-IgE antibody omalizumab.

Immunoglobulin E and its interactions with receptors FcϵRI and CD23 play a central role in allergic disease. Omalizumab, a clinically approved therapeutic antibody, inhibits the interaction between IgE and FcϵRI, preventing mast cell and basophil activation, and blocks IgE binding to CD23 on B cells and antigen-presenting cells. We solved the crystal structure of the complex between an omalizumab-derived Fab and IgE-Fc, with one Fab bound to each Cϵ3 domain. Free IgE-Fc adopts an acutely bent structure, but in the complex it is only partially bent, with large-scale conformational changes in the Cϵ3 domains that inhibit the interaction with FcϵRI. CD23 binding is inhibited sterically due to overlapping binding sites on each Cϵ3 domain. Studies of omalizumab Fab binding in solution demonstrate the allosteric basis for FcϵRI inhibition and, together with the structure, reveal how omalizumab may accelerate dissociation of receptor-bound IgE from FcϵRI, exploiting the intrinsic flexibility and allosteric potential of IgE.

Understanding Cryptic Pocket Formation in Protein Targets by Enhanced Sampling Simulations.

Cryptic pockets, that is, sites on protein targets that only become apparent when drugs bind, provide a promising alternative to classical binding sites for drug development. Here, we investigate the nature and dynamical properties of cryptic sites in four pharmacologically relevant targets, while comparing the efficacy of various simulation-based approaches in discovering them. We find that the studied cryptic sites do not correspond to local minima in the computed conformational free energy landscape of the unliganded proteins. They thus promptly close in all of the molecular dynamics simulations performed, irrespective of the force-field used. Temperature-based enhanced sampling approaches, such as Parallel Tempering, do not improve the situation, as the entropic term does not help in the opening of the sites. The use of fragment probes helps, as in long simulations occasionally it leads to the opening and binding to the cryptic sites. Our observed mechanism of cryptic site formation is suggestive of an interplay between two classical mechanisms: induced-fit and conformational selection. Employing this insight, we developed a novel Hamiltonian Replica Exchange-based method "SWISH" (Sampling Water Interfaces through Scaled Hamiltonians), which combined with probes resulted in a promising general approach for cryptic site discovery. We also addressed the issue of "false-positives" and propose a simple approach to distinguish them from druggable cryptic pockets. Our simulations, whose cumulative sampling time was more than 200 µs, help in clarifying the molecular mechanism of pocket formation, providing a solid basis for the choice of an efficient computational method.

Thermodynamics calculation of protein-ligand interactions by QM/MM polarizable charge parameters.

The calculation of protein-ligand binding free energy (ΔG) is of great importance for virtual screening and drug design. Molecular dynamics (MD) simulation has been an attractive tool to investigate this scientific problem. However, the reliability of such approach is affected by many factors including electrostatic interaction calculation. Here, we present a practical protocol using quantum mechanics/molecular mechanics (QM/MM) calculations to generate polarizable QM protein charge (QMPC). The calculated QMPC of some atoms in binding pockets was obviously different from that calculated by AMBER ff03, which might significantly affect the calculated ΔG. To evaluate the effect, the MD simulations and MM/GBSA calculation with QMPC for 10 protein-ligand complexes, and the simulation results were then compared to those with the AMBER ff03 force field and experimental results. The correlation coefficient between the calculated ΔΔG using MM/GBSA under QMPC and the experimental data is .92, while that with AMBER ff03 force field is .47 for the complexes formed by streptavidin or its mutants and biotin. Moreover, the calculated ΔΔG with QMPC for the complexes formed by ERß and five ligands is positively related to experimental result with correlation coefficient of .61, while that with AMBER ff03 charge is negatively related to experimental data with correlation coefficient of .42. The detailed analysis shows that the electrostatic polarization introduced by QMPC affects the electrostatic contribution to the binding affinity and thus, leads to better correlation with experimental data. Therefore, this approach should be useful to virtual screening and drug design.

Small Molecule Targeting of Protein-Protein Interactions through Allosteric Modulation of Dynamics.

The protein-protein interaction (PPI) target class is particularly challenging, but offers potential for "first in class" therapies. Most known PPI small molecules are orthosteric inhibitors but many PPI sites may be fundamentally intractable to this approach. One potential alternative is to consider more attractive, remote small molecule pockets; however, on the whole, allostery is poorly understood and difficult to discover and develop. Here we review the literature in order to understand the basis for allostery, especially as it can apply to PPIs. We suggest that the upfront generation of sophisticated and experimentally validated dynamic models of target proteins can aid in target choice and strategy for allosteric intervention to produce the required functional effect.

Force fields and scoring functions for carbohydrate simulation.

Carbohydrate dynamics plays a vital role in many biological processes, but we are not currently able to probe this with experimental approaches. The highly flexible nature of carbohydrate structures differs in many aspects from other biomolecules, posing significant challenges for studies employing computational simulation. Over past decades, computational study of carbohydrates has been focused on the development of structure prediction methods, force field optimization, molecular dynamics simulation, and scoring functions for carbohydrate-protein interactions. Advances in carbohydrate force fields and scoring functions can be largely attributed to enhanced computational algorithms, application of quantum mechanics, and the increasing number of experimental structures determined by X-ray and NMR techniques. The conformational analysis of carbohydrates is challengeable and has gone into intensive study in elucidating the anomeric, the exo-anomeric, and the gauche effects. Here, we review the issues associated with carbohydrate force fields and scoring functions, which will have a broad application in the field of carbohydrate-based drug design.

Mapping central α-helix linker mediated conformational transition pathway of calmodulin via simple computational approach.

The effects of intrinsic structural flexibility of calmodulin protein on the mechanism of its allosteric conformational transition are investigated in this article. Using a novel in silico approach, the conformational transition pathways of intact calmodulin as well as the isolated N- and C- terminal domains are identified and energetically characterized. It is observed that the central α-helix linker amplifies the structural flexibility of intact Ca(2+)-free calmodulin, which might facilitate the transition of the two domains. As a result, the global conformational transition of Ca(2+)-free calmodulin is initiated by the barrierless transition of two domains and proceeds through the barrier associated unwinding and bending of the central α-helix linker. The binding of Ca(2+) cations to calmodulin further increases the structural flexibility of the C-terminal domain and results in a downhill transition pathway of which all regions transit in a concerted manner. On the other hand, the separation of the N- and C-terminal domains from calmodulin protein loses the mediating function of central α-helix linker, leading to more difficult conformational transitions of both domains. The present study provides novel insights into the correlation of the integrity of protein, the structural flexibility, and the mechanism of conformational transition of proteinlike calmodulin.

Exploring transition pathway and free-energy profile of large-scale protein conformational change by combining normal mode analysis and umbrella sampling molecular dynamics.

Large-scale conformational changes of proteins are usually associated with the binding of ligands. Because the conformational changes are often related to the biological functions of proteins, understanding the molecular mechanisms of these motions and the effects of ligand binding becomes very necessary. In the present study, we use the combination of normal-mode analysis and umbrella sampling molecular dynamics simulation to delineate the atomically detailed conformational transition pathways and the associated free-energy landscapes for three well-known protein systems, viz., adenylate kinase (AdK), calmodulin (CaM), and p38α kinase in the absence and presence of respective ligands. For each protein under study, the transient conformations along the conformational transition pathway and thermodynamic observables are in agreement with experimentally and computationally determined ones. The calculated free-energy profiles reveal that AdK and CaM are intrinsically flexible in structures without obvious energy barrier, and their ligand binding shifts the equilibrium from the ligand-free to ligand-bound conformation (population shift mechanism). In contrast, the ligand binding to p38α leads to a large change in free-energy barrier (ΔΔG ≈ 7 kcal/mol), promoting the transition from DFG-in to DFG-out conformation (induced fit mechanism). Moreover, the effect of the protonation of D168 on the conformational change of p38α is also studied, which reduces the free-energy difference between the two functional states of p38α and thus further facilitates the conformational interconversion. Therefore, the present study suggests that the detailed mechanism of ligand binding and the associated conformational transition is not uniform for all kinds of proteins but correlated to their respective biological functions.

Exploration of Protein Conformational Change with PELE and Meta-Dynamics.

Atomistic molecular simulation methods are now able to explore complex protein or protein-ligand dynamical space in a tractable way with methods such as meta-dynamics or adaptive biasing force. However, many of these methods either require a careful selection of reaction coordinates or the knowledge of an initial pathway of some kind. Thus, it is important that effective methods are developed to produce this pathway data in an efficient fashion. PELE, a proven protein-ligand sampling code, has been developed to provide rapid protein sampling in highly flexible cases, using a reduced network model eigen problem approach. The resulting method is able to rapidly sample configuration space with very general driving information. When applied to ubiquitin, PELE was able to reproduce RMSD and average force data found in molecular dynamics simulations. PELE was also applied to explore the opening/closing transition of T4 lysozyme. A meta-dynamics exploration using a low energy pathway validated that the configurations explored by PELE represent the most populated regions of phase space. PELE and meta-dynamics explorations also discovered a low free energy region where a large cross-domain helix of T4 lysozyme is broken in two. There is previous NMR evidence for the validity of this unfolded helix region.

A new view of the bacterial cytosol environment.

The cytosol is the major environment in all bacterial cells. The true physical and dynamical nature of the cytosol solution is not fully understood and here a modeling approach is applied. Using recent and detailed data on metabolite concentrations, we have created a molecular mechanical model of the prokaryotic cytosol environment of Escherichia coli, containing proteins, metabolites and monatomic ions. We use 200 ns molecular dynamics simulations to compute diffusion rates, the extent of contact between molecules and dielectric constants. Large metabolites spend ∼80% of their time in contact with other molecules while small metabolites vary with some only spending 20% of time in contact. Large non-covalently interacting metabolite structures mediated by hydrogen-bonds, ionic and π stacking interactions are common and often associate with proteins. Mg(2+) ions were prominent in NIMS and almost absent free in solution. Κ(+) is generally not involved in NIMSs and populates the solvent fairly uniformly, hence its important role as an osmolyte. In simulations containing ubiquitin, to represent a protein component, metabolite diffusion was reduced owing to long lasting protein-metabolite interactions. Hence, it is likely that with larger proteins metabolites would diffuse even more slowly. The dielectric constant of these simulations was found to differ from that of pure water only through a large contribution from ubiquitin as metabolite and monatomic ion effects cancel. These findings suggest regions of influence specific to particular proteins affecting metabolite diffusion and electrostatics. Also some proteins may have a higher propensity for associations with metabolites owing to their larger electrostatic fields. We hope that future studies may be able to accurately predict how binding interactions differ in the cytosol relative to dilute aqueous solution.

Assessment of nonequilibrium free energy methods.

One of the factors preventing the general application of free energy methods in rational drug design remains the lack of sufficient computational resources. Many nonequilibrium (NE) free energy methods, however, are easily made embarrassingly parallel in comparison to equilibrium methods and may be conveniently run on desktop computers using distributed computing software. In recent years, there has been a proliferation of NE methods, but the general applicability of these approaches has not been determined. In this study, a subset including only those NE methods which are easily parallelised were considered for examination, with a view to their application to the prediction of protein-ligand binding affinities. A number of test systems were examined, including harmonic oscillator (HO) systems and the calculation of relative free energies of hydration of water-methane. The latter system uses identical potentials to the protein ligand case and is therefore an appropriate model system on which methods may be tested. As well as investigating existing protocols, a replica exchange NE approach was developed, which was found to offer advantages over conventional methods. It was found that Rosenbluth-based approaches to optimizing the NE work values used in NE free energy estimates were not consistent in the improvements in accuracy achieved and that, given their computational cost, the simple approach of taking each work value in an unbiased way is to be preferred. Of the two free energy estimators examined, Bennett's acceptance ratio was the most consistent and is, therefore, to be preferred over the Jarzynski estimator. The recommended protocols may be run very efficiently within a distributed computing environment and are of similar accuracy and precision to equilibrium free energy methods.

Protein-ligand binding affinity by nonequilibrium free energy methods.

Nonequilibrium (NE) free energy methods are embarrassingly parallel and may be very conveniently run on desktop computers using distributed computing software. In recent years there has been a proliferation of NE methods, but these approaches have barely, if at all, been used in the context of calculating protein-ligand binding free energies. In a recent study by these authors, different combinations of NE methods with various test systems were compared and protocols identified which yielded results as accurate as replica exchange thermodynamic integration (RETI). The NE approaches, however, lend themselves to extensive parallelization through the use of distributed computing. Here the best performing of those NE protocols, a replica exchange method using Bennett's acceptance ratio as the free energy estimator (RENE), is applied to two sets of congeneric inhibitors bound to neuraminidase and cyclooxygenase-2. These protein-ligand systems were originally studied with RETI, giving results to which NE and RENE simulations are compared. These NE calculations were carried out on a large, highly distributed group of low-performance desktop computers which are part of a Condor pool. RENE was found to produce results of a predictive quality at least as good as RETI in less than half the wall clock time. However, non-RE NE results were found to be far less predictive. In addition, the RENE method successfully identified a localized region of rapidly changing free energy gradients without the need for prior investigation. These results suggest that the RENE protocol is appropriate for use in the context of predicting protein-ligand binding free energies and that it can offer advantages over conventional, equilibrium approaches.