On the Frustration to Predict Binding Affinities from Protein-Ligand Structures with Deep Neural Networks

Geometric deep learning-guided Suzuki reaction conditions assessment for applications in medicinal chemistry

Suzuki cross-coupling reactions are considered a valuable tool for constructing carbon-carbon bonds in small molecule drug discovery. However, the synthesis of chemical matter often represents a time-consuming and labour-intensive bottleneck. We demonstrate how machine learning methods trained on high-throughput experimentation (HTE) data can be leveraged to enable fast reaction condition selection for novel coupling partners. We show that the trained models support chemists in determining suitable catalyst-solvent-base combinations for individual transformations including an evaluation of the need for HTE screening. We introduce an algorithm for designing 96-well plates optimized towards reaction yields and discuss the model performance of zero- and few-shot machine learning. The best-performing machine learning model achieved a three-category classification accuracy of 76.3% (±0.2%) and an F 1-score for a binary classification of 79.1% (±0.9%). Validation on eight reactions revealed a receiver operating characteristic (ROC) curve (AUC) value of 0.82 (±0.07) for few-shot machine learning. On the other hand, zero-shot machine learning models achieved a mean ROC-AUC value of 0.63 (±0.16). This study positively advocates the application of few-shot machine learning-guided reaction condition selection for HTE campaigns in medicinal chemistry and highlights practical applications as well as challenges associated with zero-shot machine learning.

Proximity Graph Networks: Predicting Ligand Affinity with Message Passing Neural Networks

Message passing neural networks (MPNNs) on molecular graphs generate continuous and differentiable encodings of small molecules with state-of-the-art performance on protein-ligand complex scoring tasks. Here, we describe the proximity graph network (PGN) package, an open-source toolkit that constructs ligand-receptor graphs based on atom proximity and allows users to rapidly apply and evaluate MPNN architectures for a broad range of tasks. We demonstrate the utility of PGN by introducing benchmarks for affinity and docking score prediction tasks. Graph networks generalize better than fingerprint-based models and perform strongly for the docking score prediction task. Overall, MPNNs with proximity graph data structures augment the prediction of ligand-receptor complex properties when ligand-receptor data are available.

The application of artificial intelligence to accelerate G protein-coupled receptor drug discovery

The application of artificial intelligence (AI) approaches to drug discovery for G protein-coupled receptors (GPCRs) is a rapidly expanding area. Artificial intelligence can be used at multiple stages during the drug discovery process, from aiding our understanding of the fundamental actions of GPCRs to the discovery of new ligand-GPCR interactions or the prediction of clinical responses. Here, we provide an overview of the concepts behind artificial intelligence, including the subfields of machine learning and deep learning. We summarise the published applications of artificial intelligence to different stages of the GPCR drug discovery process. Finally, we reflect on the benefits and limitations of artificial intelligence and share our vision for the exciting potential for further development of applications to aid GPCR drug discovery. In addition to making the drug discovery process "faster, smarter and cheaper," we anticipate that the application of artificial intelligence will create exciting new opportunities for GPCR drug discovery. LINKED ARTICLES: This article is part of a themed issue Therapeutic Targeting of G Protein-Coupled Receptors: hot topics from the Australasian Society of Clinical and Experimental Pharmacologists and Toxicologists 2021 Virtual Annual Scientific Meeting. To view the other articles in this section visit http://onlinelibrary.wiley.com/doi/10.1111/bph.v181.14/issuetoc.

Deep learning for low-data drug discovery: Hurdles and opportunities

Deep learning is becoming increasingly relevant in drug discovery, from de novo design to protein structure prediction and synthesis planning. However, it is often challenged by the small data regimes typical of certain drug discovery tasks. In such scenarios, deep learning approaches-which are notoriously 'data-hungry'-might fail to live up to their promise. Developing novel approaches to leverage the power of deep learning in low-data scenarios is sparking great attention, and future developments are expected to propel the field further. This mini-review provides an overview of recent low-data-learning approaches in drug discovery, analyzing their hurdles and advantages. Finally, we venture to provide a forecast of future research directions in low-data learning for drug discovery.

DiffBindFR: an SE(3) equivariant network for flexible protein-ligand docking

Molecular docking, a key technique in structure-based drug design, plays pivotal roles in protein-ligand interaction modeling, hit identification and optimization, in which accurate prediction of protein-ligand binding mode is essential. Conventional docking approaches perform well in redocking tasks with known protein binding pocket conformation in the complex state. However, in real-world docking scenario without knowing the protein binding conformation for a new ligand, accurately modeling the binding complex structure remains challenging as flexible docking is computationally expensive and inaccurate. Typical deep learning-based docking methods do not explicitly consider protein side chain conformations and fail to ensure the physical plausibility and detailed atomic interactions. In this study, we present DiffBindFR, a full-atom diffusion-based flexible docking model that operates over the product space of ligand overall movements and flexibility and pocket side chain torsion changes. We show that DiffBindFR has higher accuracy in producing native-like binding structures with physically plausible and detailed interactions than available docking methods. Furthermore, in the Apo and AlphaFold2 modeled structures, DiffBindFR demonstrates superior advantages in accurate ligand binding pose and protein binding conformation prediction, making it suitable for Apo and AlphaFold2 structure-based drug design. DiffBindFR provides a powerful flexible docking tool for modeling accurate protein-ligand binding structures.

MISATO: machine learning dataset of protein-ligand complexes for structure-based drug discovery

Large language models have greatly enhanced our ability to understand biology and chemistry, yet robust methods for structure-based drug discovery, quantum chemistry and structural biology are still sparse. Precise biomolecule-ligand interaction datasets are urgently needed for large language models. To address this, we present MISATO, a dataset that combines quantum mechanical properties of small molecules and associated molecular dynamics simulations of ~20,000 experimental protein-ligand complexes with extensive validation of experimental data. Starting from the existing experimental structures, semi-empirical quantum mechanics was used to systematically refine these structures. A large collection of molecular dynamics traces of protein-ligand complexes in explicit water is included, accumulating over 170 μs. We give examples of machine learning (ML) baseline models proving an improvement of accuracy by employing our data. An easy entry point for ML experts is provided to enable the next generation of drug discovery artificial intelligence models.

SadNet: a novel multimodal fusion network for protein-ligand binding affinity prediction

Protein-ligand binding affinity prediction plays an important role in the field of drug discovery. Existing deep learning-based approaches have significantly improved the efficiency of protein-ligand binding affinity prediction through their excellent inductive bias capability. However, these methods only focus on fragmented three-dimensional data, which truncates the integrity of pocket data, leading to the neglect of potential long-range interactions. In this paper, we propose a dual-stream framework, with amino acid sequence assisting the atomic data fusion for graph neural network (termed SadNet), to fuse both 3D atomic data and sequence data for more accurate prediction results. In detail, SadNet consists of a pocket module and a sequence module. The sequence module expands the "receptive field" of the pocket module through a mid-term virtual node fusion. To better integrate sequence-level information from the sequence module and 3D structural information from the pocket module, we incorporate structural information for each amino acid within the sequence module. Besides, to better understand the intrinsic relationship between sequences and 3D atomic information, our SadNet utilizes information stacking from both the early stage and later stage. Experimental results on publicly available benchmark datasets demonstrate the superiority of the proposed dual-stream approach over the state-of-the-art alternatives. The code of this work is available online at https://github.com/wardhong/SadNet.

Prospective de novo drug design with deep interactome learning

De novo drug design aims to generate molecules from scratch that possess specific chemical and pharmacological properties. We present a computational approach utilizing interactome-based deep learning for ligand- and structure-based generation of drug-like molecules. This method capitalizes on the unique strengths of both graph neural networks and chemical language models, offering an alternative to the need for application-specific reinforcement, transfer, or few-shot learning. It enables the "zero-shot" construction of compound libraries tailored to possess specific bioactivity, synthesizability, and structural novelty. In order to proactively evaluate the deep interactome learning framework for protein structure-based drug design, potential new ligands targeting the binding site of the human peroxisome proliferator-activated receptor (PPAR) subtype gamma are generated. The top-ranking designs are chemically synthesized and computationally, biophysically, and biochemically characterized. Potent PPAR partial agonists are identified, demonstrating favorable activity and the desired selectivity profiles for both nuclear receptors and off-target interactions. Crystal structure determination of the ligand-receptor complex confirms the anticipated binding mode. This successful outcome positively advocates interactome-based de novo design for application in bioorganic and medicinal chemistry, enabling the creation of innovative bioactive molecules.

A High-Quality Data Set of Protein-Ligand Binding Interactions Via Comparative Complex Structure Modeling

High-quality protein-ligand complex structures provide the basis for understanding the nature of noncovalent binding interactions at the atomic level and enable structure-based drug design. However, experimentally determined complex structures are scarce compared with the vast chemical space. In this study, we addressed this issue by constructing the BindingNet data set via comparative complex structure modeling, which contains 69,816 modeled high-quality protein-ligand complex structures with experimental binding affinity data. BindingNet provides valuable insights into investigating protein-ligand interactions, allowing visual inspection and interpretation of structural analogues' structure-activity relationships. It can also be used for evaluating machine-learning-based scoring functions. Our results indicate that machine learning models trained on BindingNet could reduce the bias caused by buried solvent-accessible surface area, as we previously found for models trained on the PDBbind data set. We also discussed strategies to improve BindingNet and its potential utilization for benchmarking the molecular docking methods and ligand binding free energy calculation approaches. The BindingNet complements PDBbind in constructing a sufficient and unbiased protein-ligand binding data set and is freely available at http://bindingnet.huanglab.org.cn.

From Proteins to Ligands: Decoding Deep Learning Methods for Binding Affinity Prediction

Accurate in silico prediction of protein-ligand binding affinity is important in the early stages of drug discovery. Deep learning-based methods exist but have yet to overtake more conventional methods such as giga-docking largely due to their lack of generalizability. To improve generalizability, we need to understand what these models learn from input protein and ligand data. We systematically investigated a sequence-based deep learning framework to assess the impact of protein and ligand encodings on predicting binding affinities for commonly used kinase data sets. The role of proteins is studied using convolutional neural network-based encodings obtained from sequences and graph neural network-based encodings enriched with structural information from contact maps. Ligand-based encodings are generated from graph-neural networks. We test different ligand perturbations by randomizing node and edge properties. For proteins, we make use of 3 different protein contact generation methods (AlphaFold2, Pconsc4, and ESM-1b) and compare these with a random control. Our investigation shows that protein encodings do not substantially impact the binding predictions, with no statistically significant difference in binding affinity for KIBA in the investigated metrics (concordance index, Pearson's R Spearman's Rank, and RMSE). Significant differences are seen for ligand encodings with random ligands and random ligand node properties, suggesting a much bigger reliance on ligand data for the learning tasks. Using different ways to combine protein and ligand encodings did not show a significant change in performance.

BigBind: Learning from Nonstructural Data for Structure-Based Virtual Screening

Deep learning methods that predict protein-ligand binding have recently been used for structure-based virtual screening. Many such models have been trained using protein-ligand complexes with known crystal structures and activities from the PDBBind data set. However, because PDBbind only includes 20K complexes, models typically fail to generalize to new targets, and model performance is on par with models trained with only ligand information. Conversely, the ChEMBL database contains a wealth of chemical activity information but includes no information about binding poses. We introduce BigBind, a data set that maps ChEMBL activity data to proteins from the CrossDocked data set. BigBind comprises 583 K ligand activities and includes 3D structures of the protein binding pockets. Additionally, we augmented the data by adding an equal number of putative inactives for each target. Using this data, we developed Banana (basic neural network for binding affinity), a neural network-based model to classify active from inactive compounds, defined by a 10 μM cutoff. Our model achieved an AUC of 0.72 on BigBind's test set, while a ligand-only model achieved an AUC of 0.59. Furthermore, Banana achieved competitive performance on the LIT-PCBA benchmark (median EF1% 1.81) while running 16,000 times faster than molecular docking with Gnina. We suggest that Banana, as well as other models trained on this data set, will significantly improve the outcomes of prospective virtual screening tasks.

PLANET: A Multi-objective Graph Neural Network Model for Protein-Ligand Binding Affinity Prediction

Predicting protein-ligand binding affinity is a central issue in drug design. Various deep learning models have been published in recent years, where many of them rely on 3D protein-ligand complex structures as input and tend to focus on the single task of reproducing binding affinity. In this study, we have developed a graph neural network model called PLANET (Protein-Ligand Affinity prediction NETwork). This model takes the graph-represented 3D structure of the binding pocket on the target protein and the 2D chemical structure of the ligand molecule as input. It was trained through a multi-objective process with three related tasks, including deriving the protein-ligand binding affinity, protein-ligand contact map, and ligand distance matrix. Besides the protein-ligand complexes with known binding affinity data retrieved from the PDBbind database, a large number of non-binder decoys were also added to the training data for deriving the final model of PLANET. When tested on the CASF-2016 benchmark, PLANET exhibited a scoring power comparable to the best result yielded by other deep learning models as well as a reasonable ranking power and docking power. In virtual screening trials conducted on the DUD-E benchmark, PLANET's performance was notably better than several deep learning and machine learning models. As on the LIT-PCBA benchmark, PLANET achieved comparable accuracy as the conventional docking program Glide, but it only spent less than 1% of Glide's computation time to finish the same job because PLANET did not need exhaustive conformational sampling. Considering the decent accuracy and efficiency of PLANET in binding affinity prediction, it may become a useful tool for conducting large-scale virtual screening.

Enhancing Generalizability in Protein-Ligand Binding Affinity Prediction with Multimodal Contrastive Learning

Improving the generalization ability of scoring functions remains a major challenge in protein-ligand binding affinity prediction. Many machine learning methods are limited by their reliance on single-modal representations, hindering a comprehensive understanding of protein-ligand interactions. We introduce a graph-neural-network-based scoring function that utilizes a triplet contrastive learning loss to improve protein-ligand representations. In this model, three-dimensional complex representations and the fusion of two-dimensional ligand and coarse-grained pocket representations converge while distancing from decoy representations in latent space. After rigorous validation on multiple external data sets, our model exhibits commendable generalization capabilities compared to those of other deep learning-based scoring functions, marking it as a promising tool in the realm of drug discovery. In the future, our training framework can be extended to other biophysical- and biochemical-related problems such as protein-protein interaction and protein mutation prediction.

Directional ΔG Neural Network (DrΔG-Net): A Modular Neural Network Approach to Binding Free Energy Prediction

The protein-ligand binding free energy is a central quantity in structure-based computational drug discovery efforts. Although popular alchemical methods provide sound statistical means of computing the binding free energy of a large breadth of systems, they are generally too costly to be applied at the same frequency as end point or ligand-based methods. By contrast, these data-driven approaches are typically fast enough to address thousands of systems but with reduced transferability to unseen systems. We introduce DrΔG-Net (or simply Dragnet), an equivariant graph neural network that can blend ligand-based and protein-ligand data-driven approaches. It is based on a 3D fingerprint representation of the ligand alone and in complex with the protein target. Dragnet is a global scoring function to predict the binding affinity of arbitrary protein-ligand complexes, but can be easily tuned via transfer learning to specific systems or end points, performing similarly to common 2D ligand-based approaches in these tasks. Dragnet is evaluated on a total of 28 validation proteins with a set of congeneric ligands derived from the Binding DB and one custom set extracted from the ChEMBL Database. In general, a handful of experimental binding affinities are sufficient to optimize the scoring function for a particular protein and ligand scaffold. When not available, predictions from physics-based methods such as absolute free energy perturbation can be used for the transfer learning tuning of Dragnet. Furthermore, we use our data to illustrate the present limitations of data-driven modeling of binding free energy predictions.

Generative artificial intelligence in drug discovery: basic framework, recent advances, challenges, and opportunities

There are two main ways to discover or design small drug molecules. The first involves fine-tuning existing molecules or commercially successful drugs through quantitative structure-activity relationships and virtual screening. The second approach involves generating new molecules through de novo drug design or inverse quantitative structure-activity relationship. Both methods aim to get a drug molecule with the best pharmacokinetic and pharmacodynamic profiles. However, bringing a new drug to market is an expensive and time-consuming endeavor, with the average cost being estimated at around $2.5 billion. One of the biggest challenges is screening the vast number of potential drug candidates to find one that is both safe and effective. The development of artificial intelligence in recent years has been phenomenal, ushering in a revolution in many fields. The field of pharmaceutical sciences has also significantly benefited from multiple applications of artificial intelligence, especially drug discovery projects. Artificial intelligence models are finding use in molecular property prediction, molecule generation, virtual screening, synthesis planning, repurposing, among others. Lately, generative artificial intelligence has gained popularity across domains for its ability to generate entirely new data, such as images, sentences, audios, videos, novel chemical molecules, etc. Generative artificial intelligence has also delivered promising results in drug discovery and development. This review article delves into the fundamentals and framework of various generative artificial intelligence models in the context of drug discovery via de novo drug design approach. Various basic and advanced models have been discussed, along with their recent applications. The review also explores recent examples and advances in the generative artificial intelligence approach, as well as the challenges and ongoing efforts to fully harness the potential of generative artificial intelligence in generating novel drug molecules in a faster and more affordable manner. Some clinical-level assets generated form generative artificial intelligence have also been discussed in this review to show the ever-increasing application of artificial intelligence in drug discovery through commercial partnerships.

Exploring protein-ligand binding affinity prediction with electron density-based geometric deep learning

Rational structure-based drug design relies on accurate predictions of protein-ligand binding affinity from structural molecular information. Although deep learning-based methods for predicting binding affinity have shown promise in computational drug design, certain approaches have faced criticism for their potential to inadequately capture the fundamental physical interactions between ligands and their macromolecular targets or for being susceptible to dataset biases. Herein, we propose to include bond-critical points based on the electron density of a protein-ligand complex as a fundamental physical representation of protein-ligand interactions. Employing a geometric deep learning model, we explore the usefulness of these bond-critical points to predict absolute binding affinities of protein-ligand complexes, benchmark model performance against existing methods, and provide a critical analysis of this new approach. The models achieved root-mean-squared errors of 1.4-1.8 log units on the PDBbind dataset, and 1.0-1.7 log units on the PDE10A dataset, not indicating significant advantages over benchmark methods, and thus rendering the utility of electron density for deep learning models context-dependent. The relationship between intermolecular electron density and corresponding binding affinity was analyzed, and Pearson correlation coefficients r > 0.7 were obtained for several macromolecular targets.

Augmenting a training dataset of the generative diffusion model for molecular docking with artificial binding pockets

This study introduces the PocketCFDM generative diffusion model, aimed at improving the prediction of small molecule poses in the protein binding pockets. The model utilizes a novel data augmentation technique, involving the creation of numerous artificial binding pockets that mimic the statistical patterns of non-bond interactions found in actual protein-ligand complexes. An algorithmic method was developed to assess and replicate these interaction patterns in the artificial binding pockets built around small molecule conformers. It is shown that the integration of artificial binding pockets into the training process significantly enhanced the model's performance. Notably, PocketCFDM surpassed DiffDock in terms of non-bond interaction and steric clash numbers, and the inference speed. Future developments and optimizations of the model are discussed. The inference code and final model weights of PocketCFDM are accessible publicly via the GitHub repository: https://github.com/vtarasv/pocket-cfdm.git.

Binding Thermodynamics of Fourth-Generation EGFR Inhibitors Revealed by Absolute Binding Free Energy Calculations

The overexpression or mutation of the kinase domain of the epidermal growth factor receptor (EGFR) is strongly associated with non-small-cell lung cancer (NSCLC). EGFR tyrosine kinase inhibitors (TKIs) have proven to be effective in treating NSCLC patients. However, EGFR mutations can result in drug resistance. To elucidate the mechanisms underlying this resistance and inform future drug development, we examined the binding affinities of BLU-945, a recently reported fourth-generation TKI, to wild-type EGFR (EGFRWT) and its double-mutant (L858R/T790M; EGFRDM) and triple-mutant (L858R/T790M/C797S; EGFRTM) forms. We compared the binding affinities of BLU-945, BLU-945 analogues, CH7233163 (another fourth-generation TKI), and erlotinib (a first-generation TKI) using absolute binding free energy calculations. Our findings reveal that BLU-945 and CH7233163 exhibit binding affinities to both EGFRDM and EGFRTM stronger than those of erlotinib, corroborating experimental data. We identified K745 and T854 as the key residues in the binding of fourth-generation EGFR TKIs. Electrostatic forces were the predominant driving force for the binding of fourth-generation TKIs to EGFR mutants. Furthermore, we discovered that the incorporation of piperidinol and sulfone groups in BLU-945 substantially enhanced its binding capacity to EGFR mutants. Our study offers valuable theoretical insights for optimizing fourth-generation EGFR TKIs.

Advances in ligand-specific biosensing for structurally similar molecules

The specificity of biological systems makes it possible to develop biosensors targeting specific metabolites, toxins, and pollutants in complex medical or environmental samples without interference from structurally similar compounds. For the last two decades, great efforts have been devoted to creating proteins or nucleic acids with novel properties through synthetic biology strategies. Beyond augmenting biocatalytic activity, expanding target substrate scopes, and enhancing enzymes' enantioselectivity and stability, an increasing research area is the enhancement of molecular specificity for genetically encoded biosensors. Here, we summarize recent advances in the development of highly specific biosensor systems and their essential applications. First, we describe the rational design principles required to create libraries containing potential mutants with less promiscuity or better specificity. Next, we review the emerging high-throughput screening techniques to engineer biosensing specificity for the desired target. Finally, we examine the computer-aided evaluation and prediction methods to facilitate the construction of ligand-specific biosensors.

ResBiGAAT: Residual Bi-GRU with attention for protein-ligand binding affinity prediction

Protein-ligand interaction plays a crucial role in drug discovery, facilitating efficient drug development and enabling drug repurposing. Several computational algorithms, such as Graph Neural Networks and Convolutional Neural Networks, have been proposed to predict the binding affinity using the three-dimensional structure of ligands and proteins. However, there are limitations due to the need for experimental characterization of the three-dimensional structure of protein sequences, which is still lacking for some proteins. Moreover, these models often suffer from unnecessary complexity, resulting in extraneous computations. This study presents ResBiGAAT, a novel deep learning model that combines a deep Residual Bidirectional Gated Recurrent Unit with two-sided self-attention mechanisms. ResBiGAAT leverages protein and ligand sequence-level features and their physicochemical properties to efficiently predict protein-ligand binding affinity. Through rigorous evaluation using 5-fold cross-validation, we demonstrate the performance of our proposed approach. The model exhibits competitive performance on an external dataset, highlighting its generalizability. Our publicly available web interface, located at resbigaat.streamlit.app, allows users to conveniently input protein and ligand sequences to estimate binding affinity.

From intuition to AI: evolution of small molecule representations in drug discovery

Within drug discovery, the goal of AI scientists and cheminformaticians is to help identify molecular starting points that will develop into safe and efficacious drugs while reducing costs, time and failure rates. To achieve this goal, it is crucial to represent molecules in a digital format that makes them machine-readable and facilitates the accurate prediction of properties that drive decision-making. Over the years, molecular representations have evolved from intuitive and human-readable formats to bespoke numerical descriptors and fingerprints, and now to learned representations that capture patterns and salient features across vast chemical spaces. Among these, sequence-based and graph-based representations of small molecules have become highly popular. However, each approach has strengths and weaknesses across dimensions such as generality, computational cost, inversibility for generative applications and interpretability, which can be critical in informing practitioners' decisions. As the drug discovery landscape evolves, opportunities for innovation continue to emerge. These include the creation of molecular representations for high-value, low-data regimes, the distillation of broader biological and chemical knowledge into novel learned representations and the modeling of up-and-coming therapeutic modalities.

The Impact of Data on Structure-Based Binding Affinity Predictions Using Deep Neural Networks

Artificial intelligence (AI) has gained significant traction in the field of drug discovery, with deep learning (DL) algorithms playing a crucial role in predicting protein-ligand binding affinities. Despite advancements in neural network architectures, system representation, and training techniques, the performance of DL affinity prediction has reached a plateau, prompting the question of whether it is truly solved or if the current performance is overly optimistic and reliant on biased, easily predictable data. Like other DL-related problems, this issue seems to stem from the training and test sets used when building the models. In this work, we investigate the impact of several parameters related to the input data on the performance of neural network affinity prediction models. Notably, we identify the size of the binding pocket as a critical factor influencing the performance of our statistical models; furthermore, it is more important to train a model with as much data as possible than to restrict the training to only high-quality datasets. Finally, we also confirm the bias in the typically used current test sets. Therefore, several types of evaluation and benchmarking are required to understand models' decision-making processes and accurately compare the performance of models.

CurvAGN: Curvature-based Adaptive Graph Neural Networks for Predicting Protein-Ligand Binding Affinity

Accurately predicting the binding affinity between proteins and ligands is crucial for drug discovery. Recent advances in graph neural networks (GNNs) have made significant progress in learning representations of protein-ligand complexes to estimate binding affinities. To improve the performance of GNNs, there frequently needs to look into protein-ligand complexes from geometric perspectives. While the "off-the-shelf" GNNs could incorporate some basic geometric structures of molecules, such as distances and angles, through modeling the complexes as homophilic graphs, these solutions seldom take into account the higher-level geometric attributes like curvatures and homology, and also heterophilic interactions.To address these limitations, we introduce the Curvature-based Adaptive Graph Neural Network (CurvAGN). This GNN comprises two components: a curvature block and an adaptive attention guided neural block (AGN). The curvature block encodes multiscale curvature informaton, then the AGN, based on an adaptive graph attention mechanism, incorporates geometry structure including angle, distance, and multiscale curvature, long-range molecular interactions, and heterophily of the graph into the protein-ligand complex representation. We demonstrate the superiority of our proposed model through experiments conducted on the PDBbind-V2016 core dataset.

Rational design of small-molecule responsive protein switches

Small-molecule responsive protein switches are powerful tools for controlling cellular processes. These switches are designed to respond rapidly and specifically to their inducer. They have been used in numerous applications, including the regulation of gene expression, post-translational protein modification, and signal transduction. Typically, small-molecule responsive protein switches consist of two proteins that interact with each other in the presence or absence of a small molecule. Recent advances in computational protein design already contributed to the development of protein switches with an expanded range of small-molecule inducers and increasingly sophisticated switch mechanisms. Further progress in the engineering of small-molecule responsive switches is fueled by cutting-edge computational design approaches, which will enable more complex and precise control over cellular processes and advance synthetic biology applications in biotechnology and medicine. Here, we discuss recent milestones and how technological advances are impacting the development of chemical switches.

Exploring the ability of machine learning-based virtual screening models to identify the functional groups responsible for binding

Many recently proposed structure-based virtual screening models appear to be able to accurately distinguish high affinity binders from non-binders. However, several recent studies have shown that they often do so by exploiting ligand-specific biases in the dataset, rather than identifying favourable intermolecular interactions in the input protein-ligand complex. In this work we propose a novel approach for assessing the extent to which machine learning-based virtual screening models are able to identify the functional groups responsible for binding. To sidestep the difficulty in establishing the ground truth importance of each atom of a large scale set of protein-ligand complexes, we propose a protocol for generating synthetic data. Each ligand in the dataset is surrounded by a randomly sampled point cloud of pharmacophores, and the label assigned to the synthetic protein-ligand complex is determined by a 3-dimensional deterministic binding rule. This allows us to precisely quantify the ground truth importance of each atom and compare it to the model generated attributions. Using our generated datasets, we demonstrate that a recently proposed deep learning-based virtual screening model, PointVS, identified the most important functional groups with 39% more efficiency than a fingerprint-based random forest, suggesting that it would generalise more effectively to new examples. In addition, we found that ligand-specific biases, such as those present in widely used virtual screening datasets, substantially impaired the ability of all ML models to identify the most important functional groups. We have made our synthetic data generation framework available to facilitate the benchmarking of new virtual screening models. Code is available at https://github.com/tomhadfield95/synthVS .

Benchmarking Cross-Docking Strategies for Structure-Informed Machine Learning in Kinase Drug Discovery

In recent years machine learning has transformed many aspects of the drug discovery process including small molecule design for which the prediction of the bioactivity is an integral part. Leveraging structural information about the interactions between a small molecule and its protein target has great potential for downstream machine learning scoring approaches, but is fundamentally limited by the accuracy with which protein:ligand complex structures can be predicted in a reliable and automated fashion. With the goal of finding practical approaches to generating useful kinase:inhibitor complex geometries for downstream machine learning scoring approaches, we present a kinase-centric docking benchmark assessing the performance of different classes of docking and pose selection strategies to assess how well experimentally observed binding modes are recapitulated in a realistic cross-docking scenario. The assembled benchmark data set focuses on the well-studied protein kinase family and comprises a subset of 589 protein structures co-crystallized with 423 ATP-competitive ligands. We find that the docking methods biased by the co-crystallized ligand-utilizing shape overlap with or without maximum common substructure matching-are more successful in recovering binding poses than standard physics-based docking alone. Also, docking into multiple structures significantly increases the chance to generate a low RMSD docking pose. Docking utilizing an approach that combines all three methods (Posit) into structures with the most similar co-crystallized ligands according to shape and electrostatics proofed to be the most efficient way to reproduce binding poses achieving a success rate of 66.9 % across all included systems. The studied docking and pose selection strategies-which utilize the OpenEye Toolkit-were implemented into pipelines of the KinoML framework allowing automated and reliable protein:ligand complex generation for future downstream machine learning tasks. Although focused on protein kinases, we believe the general findings can also be transferred to other protein families.

Ligand binding affinity prediction with fusion of graph neural networks and 3D structure-based complex graph

Accurate prediction of protein-ligand binding affinity is pivotal for drug design and discovery. Here, we proposed a novel deep fusion graph neural networks framework named FGNN to learn the protein-ligand interactions from the 3D structures of protein-ligand complexes. Unlike 1D sequences for proteins or 2D graphs for ligands, the 3D graph of protein-ligand complex enables the more accurate representations of the protein-ligand interactions. Benchmark studies have shown that our fusion models FGNN can achieve more accurate prediction of binding affinity than any individual algorithm. The advantages of fusion strategies have been demonstrated in terms of expressive power of data, learning efficiency and model interpretability. Our fusion models show satisfactory performances on diverse data sets, demonstrating their generalization ability. Given the good performances in both binding affinity prediction and virtual screening, our fusion models are expected to be practically applied for drug screening and design. Our work highlights the potential of the fusion graph neural network algorithm in solving complex prediction problems in computational biology and chemistry. The fusion graph neural networks (FGNN) model is freely available in https://github.com/LinaDongXMU/FGNN.

Similarity-based pairing improves efficiency of siamese neural networks for regression tasks and uncertainty quantification

Siamese networks, representing a novel class of neural networks, consist of two identical subnetworks sharing weights but receiving different inputs. Here we present a similarity-based pairing method for generating compound pairs to train Siamese neural networks for regression tasks. In comparison with the conventional exhaustive pairing, it reduces the algorithm complexity from O(n2) to O(n). It also results in a better prediction performance consistently on the three physicochemical datasets, using a multilayer perceptron with the circular fingerprint as a proof of concept. We further include into a Siamese neural network the transformer-based Chemformer, which extracts task-specific features from the simplified molecular-input line-entry system representation of compounds. Additionally, we propose a means to measure the prediction uncertainty by utilizing the variance in predictions from a set of reference compounds. Our results demonstrate that the high prediction accuracy correlates with the high confidence. Finally, we investigate implications of the similarity property principle in machine learning.

3DDPDs: describing protein dynamics for proteochemometric bioactivity prediction. A case for (mutant) G protein-coupled receptors

Proteochemometric (PCM) modelling is a powerful computational drug discovery tool used in bioactivity prediction of potential drug candidates relying on both chemical and protein information. In PCM features are computed to describe small molecules and proteins, which directly impact the quality of the predictive models. State-of-the-art protein descriptors, however, are calculated from the protein sequence and neglect the dynamic nature of proteins. This dynamic nature can be computationally simulated with molecular dynamics (MD). Here, novel 3D dynamic protein descriptors (3DDPDs) were designed to be applied in bioactivity prediction tasks with PCM models. As a test case, publicly available G protein-coupled receptor (GPCR) MD data from GPCRmd was used. GPCRs are membrane-bound proteins, which are activated by hormones and neurotransmitters, and constitute an important target family for drug discovery. GPCRs exist in different conformational states that allow the transmission of diverse signals and that can be modified by ligand interactions, among other factors. To translate the MD-encoded protein dynamics two types of 3DDPDs were considered: one-hot encoded residue-specific (rs) and embedding-like protein-specific (ps) 3DDPDs. The descriptors were developed by calculating distributions of trajectory coordinates and partial charges, applying dimensionality reduction, and subsequently condensing them into vectors per residue or protein, respectively. 3DDPDs were benchmarked on several PCM tasks against state-of-the-art non-dynamic protein descriptors. Our rs- and ps3DDPDs outperformed non-dynamic descriptors in regression tasks using a temporal split and showed comparable performance with a random split and in all classification tasks. Combinations of non-dynamic descriptors with 3DDPDs did not result in increased performance. Finally, the power of 3DDPDs to capture dynamic fluctuations in mutant GPCRs was explored. The results presented here show the potential of including protein dynamic information on machine learning tasks, specifically bioactivity prediction, and open opportunities for applications in drug discovery, including oncology.

Protein-ligand binding affinity prediction exploiting sequence constituent homology

MOTIVATION

Molecular docking is a commonly used approach for estimating binding conformations and their resultant binding affinities. Machine learning has been successfully deployed to enhance such affinity estimations. Many methods of varying complexity have been developed making use of some or all the spatial and categorical information available in these structures. The evaluation of such methods has mainly been carried out using datasets from PDBbind. Particularly the Comparative Assessment of Scoring Functions (CASF) 2007, 2013, and 2016 datasets with dedicated test sets. This work demonstrates that only a small number of simple descriptors is necessary to efficiently estimate binding affinity for these complexes without the need to know the exact binding conformation of a ligand.

RESULTS

The developed approach of using a small number of ligand and protein descriptors in conjunction with gradient boosting trees demonstrates high performance on the CASF datasets. This includes the commonly used benchmark CASF2016 where it appears to perform better than any other approach. This methodology is also useful for datasets where the spatial relationship between the ligand and protein is unknown as demonstrated using a large ChEMBL-derived dataset.

AVAILABILITY AND IMPLEMENTATION

Code and data uploaded to https://github.com/abbiAR/PLBAffinity.

Improving drug-target affinity prediction via feature fusion and knowledge distillation

Rapid and accurate prediction of drug-target affinity can accelerate and improve the drug discovery process. Recent studies show that deep learning models may have the potential to provide fast and accurate drug-target affinity prediction. However, the existing deep learning models still have their own disadvantages that make it difficult to complete the task satisfactorily. Complex-based models rely heavily on the time-consuming docking process, and complex-free models lacks interpretability. In this study, we introduced a novel knowledge-distillation insights drug-target affinity prediction model with feature fusion inputs to make fast, accurate and explainable predictions. We benchmarked the model on public affinity prediction and virtual screening dataset. The results show that it outperformed previous state-of-the-art models and achieved comparable performance to previous complex-based models. Finally, we study the interpretability of this model through visualization and find it can provide meaningful explanations for pairwise interaction. We believe this model can further improve the drug-target affinity prediction for its higher accuracy and reliable interpretability.

The Impact of Supervised Learning Methods in Ultralarge High-Throughput Docking

Structure-based virtual screening methods are, nowadays, one of the key pillars of computational drug discovery. In recent years, a series of studies have reported docking-based virtual screening campaigns of large databases ranging from hundreds to thousands of millions compounds, further identifying novel hits after experimental validation. As these larg-scale efforts are not generally accessible, machine learning-based protocols have emerged to accelerate the identification of virtual hits within an ultralarge chemical space, reaching impressive reductions in computational time. Herein, we illustrate the motivation and the problem behind the screening of large databases, providing an overview of key concepts and essential applications of machine learning-accelerated protocols, specifically concerning supervised learning methods. We also discuss where the field stands with these novel developments, highlighting possible insights for future studies.

Chemical language models for de novo drug design: Challenges and opportunities

Generative deep learning is accelerating de novo drug design, by allowing the generation of molecules with desired properties on demand. Chemical language models - which generate new molecules in the form of strings using deep learning - have been particularly successful in this endeavour. Thanks to advances in natural language processing methods and interdisciplinary collaborations, chemical language models are expected to become increasingly relevant in drug discovery. This minireview provides an overview of the current state-of-the-art of chemical language models for de novo design, and analyses current limitations, challenges, and advantages. Finally, a perspective on future opportunities is provided.

Computational approaches streamlining drug discovery

Computer-aided drug discovery has been around for decades, although the past few years have seen a tectonic shift towards embracing computational technologies in both academia and pharma. This shift is largely defined by the flood of data on ligand properties and binding to therapeutic targets and their 3D structures, abundant computing capacities and the advent of on-demand virtual libraries of drug-like small molecules in their billions. Taking full advantage of these resources requires fast computational methods for effective ligand screening. This includes structure-based virtual screening of gigascale chemical spaces, further facilitated by fast iterative screening approaches. Highly synergistic are developments in deep learning predictions of ligand properties and target activities in lieu of receptor structure. Here we review recent advances in ligand discovery technologies, their potential for reshaping the whole process of drug discovery and development, as well as the challenges they encounter. We also discuss how the rapid identification of highly diverse, potent, target-selective and drug-like ligands to protein targets can democratize the drug discovery process, presenting new opportunities for the cost-effective development of safer and more effective small-molecule treatments.

Structure-based drug design with geometric deep learning

Structure-based drug design uses three-dimensional geometric information of macromolecules, such as proteins or nucleic acids, to identify suitable ligands. Geometric deep learning, an emerging concept of neural-network-based machine learning, has been applied to macromolecular structures. This review provides an overview of the recent applications of geometric deep learning in bioorganic and medicinal chemistry, highlighting its potential for structure-based drug discovery and design. Emphasis is placed on molecular property prediction, ligand binding site and pose prediction, and structure-based de novo molecular design. The current challenges and opportunities are highlighted, and a forecast of the future of geometric deep learning for drug discovery is presented.

Geometric Interaction Graph Neural Network for Predicting Protein-Ligand Binding Affinities from 3D Structures (GIGN)

Predicting protein-ligand binding affinities (PLAs) is a core problem in drug discovery. Recent advances have shown great potential in applying machine learning (ML) for PLA prediction. However, most of them omit the 3D structures of complexes and physical interactions between proteins and ligands, which are considered essential to understanding the binding mechanism. This paper proposes a geometric interaction graph neural network (GIGN) that incorporates 3D structures and physical interactions for predicting protein-ligand binding affinities. Specifically, we design a heterogeneous interaction layer that unifies covalent and noncovalent interactions into the message passing phase to learn node representations more effectively. The heterogeneous interaction layer also follows fundamental biological laws, including invariance to translations and rotations of the complexes, thus avoiding expensive data augmentation strategies. GIGN achieves state-of-the-art performance on three external test sets. Moreover, by visualizing learned representations of protein-ligand complexes, we show that the predictions of GIGN are biologically meaningful.

Predicting Potent Compounds Using a Conditional Variational Autoencoder Based upon a New Structure-Potency Fingerprint

Prediction of the potency of bioactive compounds generally relies on linear or nonlinear quantitative structure-activity relationship (QSAR) models. Nonlinear models are generated using machine learning methods. We introduce a novel approach for potency prediction that depends on a newly designed molecular fingerprint (FP) representation. This structure-potency fingerprint (SPFP) combines different modules accounting for the structural features of active compounds and their potency values in a single bit string, hence unifying structure and potency representation. This encoding enables the derivation of a conditional variational autoencoder (CVAE) using SPFPs of training compounds and apply the model to predict the SPFP potency module of test compounds using only their structure module as input. The SPFP-CVAE approach correctly predicts the potency values of compounds belonging to different activity classes with an accuracy comparable to support vector regression (SVR), representing the state-of-the-art in the field. In addition, highly potent compounds are predicted with very similar accuracy as SVR and deep neural networks.

Identifying Protein-Ligand Interactions via a Novel Distance Self-Feedback Biomolecular Interaction Network

Efficient and accurate characterizations of protein-ligand interactions are key to understanding biology at the molecular level. They are particularly useful in pharmaceutical industry applications. They are usually computationally demanding for those widely applied dynamics-based methods in identifying important residues or calculating ligand binding free energy. In this work, we proposed a graph deep learning (DL) framework, namely, the distance self-feedback biomolecular interaction network (DSBIN), in which the relationship between the complex structure and binding affinity can be established by means of a carefully designed distance self-feedback module and interaction layer. Our model can directly provide a quantitative evaluation of inhibitor binding affinities (pK_d). More importantly, the DSBIN model efficiently identifies key interactions for inhibitor binding and thus intrinsically bears the interpretability. Its generalization performance was further verified using 1405 unseen structures. The predicted binding free energies' deviations were calculated to be less than 1.37 kcal/mol for more than 55% structures. Moreover, we also compared the DSBIN model with a commonly used theoretical method in calculating the substrate binding free energy, MM/GBSA. Our results show that the current DL model has generally better performance in predicting the binding free energy. For a specific complex system, mannopentaose/TmCBM27, the DSBIN predicted binding free energy is -8.21 kcal/mol, which is very close to experimentally measured -7.76 kcal/mol and MM/GBSA calculated -7.16 kcal/mol. Meanwhile, all important aromatic residues around the binding pocket can be identified by our DL model. Considering the accuracy and efficiency of the newly developed DL model, it may be very helpful in the field of drug design and molecular recognition.

New avenues in artificial-intelligence-assisted drug discovery

Over the past decade, the amount of biomedical data available has grown at unprecedented rates. Increased automation technology and larger data volumes have encouraged the use of machine learning (ML) or artificial intelligence (AI) techniques for mining such data and extracting useful patterns. Because the identification of chemical entities with desired biological activity is a crucial task in drug discovery, AI technologies have the potential to accelerate this process and support decision making. In addition, the advent of deep learning (DL) has shown great promise in addressing diverse problems in drug discovery, such as de novo molecular design. Herein, we will appraise the current state-of-the-art in AI-assisted drug discovery, discussing the recent applications covering generative models for chemical structure generation, scoring functions to improve binding affinity and pose prediction, and molecular dynamics to assist in the parametrization, featurization and generalization tasks. Finally, we will discuss current hurdles and the strategies to overcome them, as well as potential future directions.

Beyond sequence: Structure-based machine learning

Recent breakthroughs in protein structure prediction demarcate the start of a new era in structural bioinformatics. Combined with various advances in experimental structure determination and the uninterrupted pace at which new structures are published, this promises an age in which protein structure information is as prevalent and ubiquitous as sequence. Machine learning in protein bioinformatics has been dominated by sequence-based methods, but this is now changing to make use of the deluge of rich structural information as input. Machine learning methods making use of structures are scattered across literature and cover a number of different applications and scopes; while some try to address questions and tasks within a single protein family, others aim to capture characteristics across all available proteins. In this review, we look at the variety of structure-based machine learning approaches, how structures can be used as input, and typical applications of these approaches in protein biology. We also discuss current challenges and opportunities in this all-important and increasingly popular field.

Exposing the Limitations of Molecular Machine Learning with Activity Cliffs

Machine learning has become a crucial tool in drug discovery and chemistry at large, e.g., to predict molecular properties, such as bioactivity, with high accuracy. However, activity cliffs─pairs of molecules that are highly similar in their structure but exhibit large differences in potency─have received limited attention for their effect on model performance. Not only are these edge cases informative for molecule discovery and optimization but also models that are well equipped to accurately predict the potency of activity cliffs have increased potential for prospective applications. Our work aims to fill the current knowledge gap on best-practice machine learning methods in the presence of activity cliffs. We benchmarked a total of 24 machine and deep learning approaches on curated bioactivity data from 30 macromolecular targets for their performance on activity cliff compounds. While all methods struggled in the presence of activity cliffs, machine learning approaches based on molecular descriptors outperformed more complex deep learning methods. Our findings highlight large case-by-case differences in performance, advocating for (a) the inclusion of dedicated "activity-cliff-centered" metrics during model development and evaluation and (b) the development of novel algorithms to better predict the properties of activity cliffs. To this end, the methods, metrics, and results of this study have been encapsulated into an open-access benchmarking platform named MoleculeACE (Activity Cliff Estimation, available on GitHub at: https://github.com/molML/MoleculeACE). MoleculeACE is designed to steer the community toward addressing the pressing but overlooked limitation of molecular machine learning models posed by activity cliffs.

Graph-Convolutional Neural Net Model of the Statistical Torsion Profiles for Small Organic Molecules

We present a graph-convolutional neural network (GCNN)-based method for learning and prediction of statistical torsional profiles (STP) in small organic molecules based on the experimental X-ray structure data. A specialized GCNN torsion profile model is trained using the structures in the Crystallography Open Database (COD). The GCNN-STP model captures torsional preferences over a wide range of torsion rotor chemotypes and correctly predicts a variety of effects from the vicinal atoms and moieties. GCNN-STP statistical profiles also show good agreement with quantum chemically (DFT) calculated torsion energy profiles. Furthermore, we demonstrate the application of the GCNN-STP statistical profiles for conformer generation. A web server that allows interactive profile prediction and viewing is made freely available at https://www.molsoft.com/tortool.html.

Chemical Space Exploration with Active Learning and Alchemical Free Energies

Drug discovery can be thought of as a search for a needle in a haystack: searching through a large chemical space for the most active compounds. Computational techniques can narrow the search space for experimental follow up, but even they become unaffordable when evaluating large numbers of molecules. Therefore, machine learning (ML) strategies are being developed as computationally cheaper complementary techniques for navigating and triaging large chemical libraries. Here, we explore how an active learning protocol can be combined with first-principles based alchemical free energy calculations to identify high affinity phosphodiesterase 2 (PDE2) inhibitors. We first calibrate the procedure using a set of experimentally characterized PDE2 binders. The optimized protocol is then used prospectively on a large chemical library to navigate toward potent inhibitors. In the active learning cycle, at every iteration a small fraction of compounds is probed by alchemical calculations and the obtained affinities are used to train ML models. With successive rounds, high affinity binders are identified by explicitly evaluating only a small subset of compounds in a large chemical library, thus providing an efficient protocol that robustly identifies a large fraction of true positives.

On the Choice of Active Site Sequences for Kinase-Ligand Affinity Prediction

Recent work showed that active site rather than full-protein-sequence information improves predictive performance in kinase-ligand binding affinity prediction. To refine the notion of an “active site”, we here propose and compare multiple definitions. We report significant evidence that our novel definition is superior to previous definitions and better models of ATP-noncompetitive inhibitors. Moreover, we leverage the discontiguity of the active site sequence to motivate novel protein-sequence augmentation strategies and find that combining them further improves performance.

Proteome-Wide Profiling of the Covalent-Druggable Cysteines with a Structure-Based Deep Graph Learning Network

Covalent ligands have attracted increasing attention due to their unique advantages, such as long residence time, high selectivity, and strong binding affinity. They also show promise for targets where previous efforts to identify noncovalent small molecule inhibitors have failed. However, our limited knowledge of covalent binding sites has hindered the discovery of novel ligands. Therefore, developing in silico methods to identify covalent binding sites is highly desirable. Here, we propose DeepCoSI, the first structure-based deep graph learning model to identify ligandable covalent sites in the protein. By integrating the characterization of the binding pocket and the interactions between each cysteine and the surrounding environment, DeepCoSI achieves state-of-the-art predictive performances. The validation on two external test sets which mimic the real application scenarios shows that DeepCoSI has strong ability to distinguish ligandable sites from the others. Finally, we profiled the entire set of protein structures in the RCSB Protein Data Bank (PDB) with DeepCoSI to evaluate the ligandability of each cysteine for covalent ligand design, and made the predicted data publicly available on website.