Are Deep Learning Structural Models Sufficiently Accurate for Free-Energy Calculations? Application of FEP+ to AlphaFold2-Predicted Structures

Alchemical Transformations and Beyond: Recent Advances and Real-World Applications of Free Energy Calculations in Drug Discovery

Computational methods constitute efficient strategies for screening and optimizing potential drug molecules. A critical factor in this process is the binding affinity between candidate molecules and targets, quantified as binding free energy. Among various estimation methods, alchemical transformation methods stand out for their theoretical rigor. Despite challenges in force field accuracy and sampling efficiency, advancements in algorithms, software, and hardware have increased the application of free energy perturbation (FEP) calculations in the pharmaceutical industry. Here, we review the practical applications of FEP in drug discovery projects since 2018, covering both ligand-centric and residue-centric transformations. We show that relative binding free energy calculations have steadily achieved chemical accuracy in real-world applications. In addition, we discuss alternative physics-based simulation methods and the incorporation of deep learning into free energy calculations.

Empowering AlphaFold2 for protein conformation selective drug discovery with AlphaFold2-RAVE

Small-molecule drug design hinges on obtaining co-crystallized ligand-protein structures. Despite AlphaFold2's strides in protein native structure prediction, its focus on apo structures overlooks ligands and associated holo structures. Moreover, designing selective drugs often benefits from the targeting of diverse metastable conformations. Therefore, direct application of AlphaFold2 models in virtual screening and drug discovery remains tentative. Here, we demonstrate an AlphaFold2-based framework combined with all-atom enhanced sampling molecular dynamics and Induced Fit docking, named AF2RAVE-Glide, to conduct computational model-based small-molecule binding of metastable protein kinase conformations, initiated from protein sequences. We demonstrate the AF2RAVE-Glide workflow on three different mammalian protein kinases and their type I and II inhibitors, with special emphasis on binding of known type II kinase inhibitors which target the metastable classical DFG-out state. These states are not easy to sample from AlphaFold2. Here, we demonstrate how with AF2RAVE these metastable conformations can be sampled for different kinases with high enough accuracy to enable subsequent docking of known type II kinase inhibitors with more than 50% success rates across docking calculations. We believe the protocol should be deployable for other kinases and more proteins generally.

Interpreting forces as deep learning gradients improves quality of predicted protein structures

Protein structure predictions from deep learning models like AlphaFold2, despite their remarkable accuracy, are likely insufficient for direct use in downstream tasks like molecular docking. The functionality of such models could be improved with a combination of increased accuracy and physical intuition. We propose a new method to train deep learning protein structure prediction models using molecular dynamics force fields to work toward these goals. Our custom PyTorch loss function, OpenMM-Loss, represents the potential energy of a predicted structure. OpenMM-Loss can be applied to any all-atom representation of a protein structure capable of mapping into our software package, SidechainNet. We demonstrate our method's efficacy by finetuning OpenFold. We show that subsequently predicted protein structures, both before and after a relaxation procedure, exhibit comparable accuracy while displaying lower potential energy and improved structural quality as assessed by MolProbity metrics.

Caught between a ROCK and a hard place: current challenges in structure-based drug design

The discipline of structure-based drug design (SBDD) is several decades old and it is tempting to think that the proliferation of experimental structures for many drug targets might make computer-aided drug design (CADD) straightforward. However, this is far from true. In this review, we illustrate some of the challenges that CADD scientists face every day in their work, even now. We use Rho-associated protein kinase (ROCK), and public domain structures and data, as an example to illustrate some of the challenges we have experienced during our project targeting this protein. We hope that this will help to prevent unrealistic expectations of what CADD can accomplish and to educate non-CADD scientists regarding the challenges still facing their CADD colleagues.

AlphaFold2 structures guide prospective ligand discovery

AlphaFold2 (AF2) models have had wide impact but mixed success in retrospective ligand recognition. We prospectively docked large libraries against unrefined AF2 models of the σ2 and serotonin 2A (5-HT2A) receptors, testing hundreds of new molecules and comparing results with those obtained from docking against the experimental structures. Hit rates were high and similar for the experimental and AF2 structures, as were affinities. Success in docking against the AF2 models was achieved despite differences between orthosteric residue conformations in the AF2 models and the experimental structures. Determination of the cryo-electron microscopy structure for one of the more potent 5-HT2A ligands from the AF2 docking revealed residue accommodations that resembled the AF2 prediction. AF2 models may sample conformations that differ from experimental structures but remain low energy and relevant for ligand discovery, extending the domain of structure-based drug design.

Evaluation of AlphaFold2 Structures for Hit Identification across Multiple Scenarios

The introduction of AlphaFold2 (AF2) has sparked significant enthusiasm and generated extensive discussion within the scientific community, particularly among drug discovery researchers. Although previous studies have addressed the performance of AF2 structures in virtual screening (VS), a more comprehensive investigation is still necessary considering the paramount importance of structural accuracy in drug design. In this study, we evaluate the performance of AF2 structures in VS across three common drug discovery scenarios: targets with holo, apo, and AF2 structures; targets with only apo and AF2 structures; and targets exclusively with AF2 structures. We utilized both the traditional physics-based Glide and the deep-learning-based scoring function RTMscore to rank the compounds in the DUD-E, DEKOIS 2.0, and DECOY data sets. The results demonstrate that, overall, the performance of VS on AF2 structures is comparable to that on apo structures but notably inferior to that on holo structures across diverse scenarios. Moreover, when a target has solely AF2 structure, selecting the holo structure of the target from different subtypes within the same protein family produces comparable results with the AF2 structure for VS on the data set of the AF2 structures, and significantly better results than the AF2 structures on its own data set. This indicates that utilizing AF2 structures for docking-based VS may not yield most satisfactory outcomes, even when solely AF2 structures are available. Moreover, we rule out the possibility that the variations in VS performance between the binding pockets of AF2 and holo structures arise from the differences in their biological assembly composition.

AlphaFold2 structures template ligand discovery

AlphaFold2 (AF2) and RosettaFold have greatly expanded the number of structures available for structure-based ligand discovery, even though retrospective studies have cast doubt on their direct usefulness for that goal. Here, we tested unrefined AF2 models prospectively, comparing experimental hit-rates and affinities from large library docking against AF2 models vs the same screens targeting experimental structures of the same receptors. In retrospective docking screens against the σ2 and the 5-HT2A receptors, the AF2 structures struggled to recapitulate ligands that we had previously found docking against the receptors' experimental structures, consistent with published results. Prospective large library docking against the AF2 models, however, yielded similar hit rates for both receptors versus docking against experimentally-derived structures; hundreds of molecules were prioritized and tested against each model and each structure of each receptor. The success of the AF2 models was achieved despite differences in orthosteric pocket residue conformations for both targets versus the experimental structures. Intriguingly, against the 5-HT2A receptor the most potent, subtype-selective agonists were discovered via docking against the AF2 model, not the experimental structure. To understand this from a molecular perspective, a cryoEM structure was determined for one of the more potent and selective ligands to emerge from docking against the AF2 model of the 5-HT2A receptor. Our findings suggest that AF2 models may sample conformations that are relevant for ligand discovery, much extending the domain of applicability of structure-based ligand discovery.

Optimizing Trilobatin Production via Screening and Modification of Glycosyltransferases

Trilobatin (TBL) is a key sweet compound from the traditional Chinese sweet tea plant (Rubus suavissimus S. Lee). Because of its intense sweetness, superior taste profile, and minimal caloric value, it serves as an exemplary natural dihydrochalcone sweetener. It also has various health benefits, including anti-inflammatory and glucose-lowering effects. It is primarily produced through botanical extraction, which impedes its scalability and cost-effectiveness. In a novel biotechnological approach, phloretin is used as a precursor that is transformed into TBL by the glycosyltransferase enzyme ph-4'-OGT. However, this enzyme's low catalytic efficiency and by-product formation limit the large-scale synthesis of TBL. In our study, the enzyme Mdph-4'-OGT was used to screen 17 sequences across species for TBL synthesis, of which seven exhibited catalytic activity. Notably, PT577 exhibited an unparalleled 97.3% conversion yield within 3 h. We then optimized the reaction conditions of PT577, attaining a peak TBL bioproduction of 163.3 mg/L. By employing virtual screening, we identified 25 mutation sites for PT577, thereby creating mutant strains that reduced by-products by up to 50%. This research enhances the enzymatic precision for TBL biosynthesis and offers a robust foundation for its industrial-scale production, with broader implications for the engineering and in silico analysis of glycosyltransferases.

Using AlphaFold and Experimental Structures for the Prediction of the Structure and Binding Affinities of GPCR Complexes via Induced Fit Docking and Free Energy Perturbation

Free energy perturbation (FEP) remains an indispensable method for computationally assaying prospective compounds in advance of synthesis. However, before FEP can be deployed prospectively, it must demonstrate retrospective recapitulation of known experimental data where the subtle details of the atomic ligand-receptor model are consequential. An open question is whether AlphaFold models can serve as useful initial models for FEP in the regime where there exists a congeneric series of known chemical matter but where no experimental structures are available either of the target or of close homologues. As AlphaFold structures are provided without a bound ligand, we employ induced fit docking to refine the AlphaFold models in the presence of one or more congeneric ligands. In this work, we first validate the performance of our latest induced fit docking technology, IFD-MD, on a retrospective set of public experimental GPCR structures with 95% of cross-docks producing a pose with a ligand RMSD ≤ 2.5 Å in the top two predictions. We then apply IFD-MD and FEP on AlphaFold models of the somatostatin receptor family of GPCRs. We use AlphaFold models produced prior to the availability of any experimental structure from this family. We arrive at FEP-validated models for SSTR2, SSTR4, and SSTR5, with RMSE around 1 kcal/mol, and explore the challenges of model validation under scenarios of limited ligand data, ample ligand data, and categorical data.

Basic Residues at Position 11 of α-Conotoxin LvIA Influence Subtype Selectivity between α3β2 and α3β4 Nicotinic Receptors via an Electrostatic Mechanism

Understanding the determinants of α-conotoxin (α-CTX) selectivity for different nicotinic acetylcholine receptor (nAChR) subtypes is a prerequisite for the design of tool compounds to study nAChRs. However, selectivity optimization of these small, disulfide-rich peptides is difficult not only because of an absence of α-CTX/nAChR co-structures but also because it is challenging to predict how a mutation to an α-CTX will alter its potency and selectivity. As a prototypical system to investigate selectivity, we employed the α-CTX LvIA that is 25-fold selective for the α3β2 nAChR over the related α3β4 nAChR subtype, which is a target for nicotine addiction. Using two-electrode voltage clamp electrophysiology, we identified LvIA[D11R] that is 2-fold selective for the α3β4 nAChR, reversing the subtype preference. This effect is specifically due to the change in charge and not shape of LvIA[D11R], as substitution of D11 with citrulline retains selectivity for the α3β2 nAChR. Furthermore, LvIA[D11K] shows a stronger reversal, with 4-fold selectivity for the α3β4 nAChR. Motivated by these findings, using site-directed mutagenesis, we found that β2[K79A] (I79 on β4), but not β2[K78A] (N78 on β4), largely restores the potency of basic mutants at position 11. Finally, to understand the structural basis of this effect, we used AlphaFold2 to generate models of LvIA in complex with both nAChR subtypes. Both models confirm the plausibility of an electrostatic mechanism to explain the data and also reproduce a broad range of potency and selectivity structure-activity relationships for LvIA mutants, as measured using free energy perturbation simulations. Our work highlights how electrostatic interactions can drive α-CTX selectivity and may serve as a strategy for optimizing the selectivity of LvIA and other α-CTXs.

Quality Assessment of Selected Protein Structures Derived from Homology Modeling and AlphaFold

With technology advancing, many prediction algorithms have been developed to facilitate the modeling of inherently dynamic and flexible macromolecules such as proteins. Improvements in the prediction of protein structures have attracted a great deal of attention due to the advantages they offer, e.g., in drug design. While trusted experimental methods, such as X-ray crystallography, NMR spectroscopy, and electron microscopy, are preferred structure analysis techniques, in silico approaches are also being widely used. Two computational methods, which are on opposite ends of the spectrum with respect to their modus operandi, i.e., homology modeling and AlphaFold, have been established to provide high-quality structures. Here, a comparative study of the quality of structures either predicted by homology modeling or by AlphaFold is presented based on the characteristics determined by experimental studies using structure validation servers to fulfill the purpose. Although AlphaFold is able to predict high-quality structures, high-confidence parts are sometimes observed to be in disagreement with experimental data. On the other hand, while the structures obtained from homology modeling are successful in incorporating all aspects of the experimental structure used as a template, this method may struggle to accurately model a structure in the absence of a suitable template. In general, although both methods produce high-quality models, the criteria by which they are superior to each other are different and thus discussed in detail.

Precision Oncology Comes of Age: Designing Best-in-Class Small Molecules by Integrating Two Decades of Advances in Chemistry, Target Biology, and Data Science

Small-molecule drugs have enabled the practice of precision oncology for genetically defined patient populations since the first approval of imatinib in 2001. Scientific and technology advances over this 20-year period have driven the evolution of cancer biology, medicinal chemistry, and data science. Collectively, these advances provide tools to more consistently design best-in-class small-molecule drugs against known, previously undruggable, and novel cancer targets. The integration of these tools and their customization in the hands of skilled drug hunters will be necessary to enable the discovery of transformational therapies for patients across a wider spectrum of cancers.

SIGNIFICANCE

Target-centric small-molecule drug discovery necessitates the consideration of multiple approaches to identify chemical matter that can be optimized into drug candidates. To do this successfully and consistently, drug hunters require a comprehensive toolbox to avoid following the "law of instrument" or Maslow's hammer concept where only one tool is applied regardless of the requirements of the task. Combining our ever-increasing understanding of cancer and cancer targets with the technological advances in drug discovery described below will accelerate the next generation of small-molecule drugs in oncology.

Effectively utilizing publicly available databases for cancer target evaluation

The majority of compounds designed against cancer drug targets do not progress to become approved drugs, mainly due to lack of efficacy and/or unmanageable toxicity. Robust target evaluation is therefore required before progressing through the drug discovery process to reduce the high attrition rate. There are a wealth of publicly available databases that can be mined to generate data as part of a target evaluation. It can, however, be challenging to learn what databases are available, how and when they should be used, and to understand the associated limitations. Here, we have compiled and present key, freely accessible and easy-to-use databases that house informative datasets from in vitro, in vivo and clinical studies. We also highlight comprehensive target review databases that aim to bring together information from multiple sources into one-stop portals. In the post-genomics era, a key objective is to exploit the extensive cell, animal and patient characterization datasets in order to deliver precision medicine on a patient-specific basis. Effective utilization of the highlighted databases will go some way towards supporting the cancer research community achieve these aims.

Advancing structural biology through breakthroughs in AI

The past century has witnessed an exponential increase in our atomic-level understanding of molecular and cellular mechanisms from a structural perspective, with multiple landmark achievements contributing to the field. This, coupled with recent and continuing breakthroughs in artificial intelligence methods such as AlphaFold2, and enhanced computational power, is enabling our understanding of protein structure and function at unprecedented levels of accuracy and predictivity. Here, we describe some of the major recent advances across these fields, and describe, as these technologies coalesce, the potential to utilise our enhanced knowledge of intricate cellular and molecular systems to discover novel therapeutics to alleviate human suffering.

Cation/proton antiporters: novel structure-driven pharmaceutical opportunities

Cation/proton antiporters (CPAs) regulate cells' salt concentration and pH. Their malfunction is associated with a range of human pathologies, yet only a handful of CPA-targeting therapeutics are presently in clinical development. Here, we discuss how recently published mammalian protein structures and emerging computational technologies may help to bridge this gap.

Are Deep Learning Structural Models Sufficiently Accurate for Virtual Screening? Application of Docking Algorithms to AlphaFold2 Predicted Structures

Machine learning-based protein structure prediction algorithms, such as RosettaFold and AlphaFold2, have greatly impacted the structural biology field, arousing a fair amount of discussion around their potential role in drug discovery. While there are few preliminary studies addressing the usage of these models in virtual screening, none of them focus on the prospect of hit-finding in a real-world virtual screen with a model based on low prior structural information. In order to address this, we have developed an AlphaFold2 version where we exclude all structural templates with more than 30% sequence identity from the model-building process. In a previous study, we used those models in conjunction with state-of-the-art free energy perturbation methods and demonstrated that it is possible to obtain quantitatively accurate results. In this work, we focus on using these structures in rigid receptor-ligand docking studies. Our results indicate that using out-of-the-box Alphafold2 models is not an ideal scenario for virtual screening campaigns; in fact, we strongly recommend to include some post-processing modeling to drive the binding site into a more realistic holo model.

Recent Advances in Alchemical Binding Free Energy Calculations for Drug Discovery

Rigorous physics-based methods to calculate binding free energies of protein-ligand complexes have become a valued component of structure-based drug design. Relative and absolute binding free energy calculations have been deployed prospectively in support of solving diverse drug discovery challenges. Here we review recent applications of binding free energy calculations to fragment growing and linking, scaffold hopping, binding pose validation, virtual screening, covalent enzyme inhibition, and positional analogue scanning. Furthermore, we discuss the merits of using protein models and highlight recent efforts to replace costly binding free energy calculations with predictions from machine learning models trained on a limited number of free energy perturbation or thermodynamic integration calculations thereby allowing for extended chemical space exploration.

Hic Sunt Dracones: Molecular Docking in Uncharted Territories with Structures from AlphaFold2 and RoseTTAfold

AlphaFold2 and RoseTTAfold impress with their high accuracy in protein structure prediction. However, for structure-based virtual screenings, not only the overall structure but especially the binding sites need to be accurately predicted. In this work, the docking performance for 66 targets with known ligands but without experimental structures available in the protein data bank was elucidated. The results suggest that using an experimental surrogate-ligand complex is often superior over homology models, and only at low sequence identity to the closest homologue AlphaFold2 structures show an equal performance. The generally high fluctuation of receiver operating characteristic area under the curve values obtained for different homology models suggests that multiple combinations of docking programs and homology models should be tested prior to prospective virtual screenings, and in some cases post-processing of crude models might be necessary.

How good are AlphaFold models for docking-based virtual screening?

A crucial component in structure-based drug discovery is the availability of high-quality three-dimensional structures of the protein target. Whenever experimental structures were not available, homology modeling has been, so far, the method of choice. Recently, AlphaFold (AF), an artificial-intelligence-based protein structure prediction method, has shown impressive results in terms of model accuracy. This outstanding success prompted us to evaluate how accurate AF models are from the perspective of docking-based drug discovery. We compared the high-throughput docking (HTD) performance of AF models with their corresponding experimental PDB structures using a benchmark set of 22 targets. The AF models showed consistently worse performance using four docking programs and two consensus techniques. Although AlphaFold shows a remarkable ability to predict protein architecture, this might not be enough to guarantee that AF models can be reliably used for HTD, and post-modeling refinement strategies might be key to increase the chances of success.