Assessment of the Generalization Abilities of Machine-Learning Scoring Functions for Structure-Based Virtual Screening

Further exploration of the quantitative distance-energy and contact number-energy relationships for predicting the binding affinity of protein-ligand complexes

Accurate estimation of the strength of the protein-ligand interaction is important in the field of drug discovery. The binding strength can be determined by using experimental binding affinity assays which are both time and labor consuming and costly. Predicting the binding affinity/energy in silico is an alternative approach, particularly for virtual screening of large data sets. In general, the distance-based terms such as electrostatic and van der Waals interactions are among the key determinants of binding energy. In this work, the distance-binding energy relationships, i.e., E ∝ -d-k, are further explored, extended, and developed for protein-ligand binding affinity prediction. The contributions of different atom-type pairs were considered synthetically and jointly. Additionally, the contact number-energy relationships (E ∝ -nk) were also explored for protein-ligand binding affinity prediction. Significantly, the power exponents of the distances or contact numbers in the energy functions are not restricted by the existing theories concerning van der Waals and electrostatic energies (expressed as ar6-br12 and cr). The performances of the new distance-based or contact number-based models are better than the performances of those sophisticated non-machine-learning-based scoring functions developed before. The exploration and extension of the distance-energy and contact number-energy relationships may offer insights into the development of more effective methods for predicting the protein-ligand binding affinity accurately and analyzing the protein-ligand interactions rationally.

Narrowing the gap between machine learning scoring functions and free energy perturbation using augmented data

Machine learning offers great promise for fast and accurate binding affinity predictions. However, current models lack robust evaluation and fail on tasks encountered in (hit-to-) lead optimisation, such as ranking the binding affinity of a congeneric series of ligands, thereby limiting their application in drug discovery. Here, we address these issues by first introducing a novel attention-based graph neural network model called AEV-PLIG (atomic environment vector-protein ligand interaction graph). Second, we introduce a new and more realistic out-of-distribution test set called the OOD Test. We benchmark our model on this set, CASF-2016, and a test set used for free energy perturbation (FEP) calculations, that not only highlights the competitive performance of AEV-PLIG, but provides a realistic assessment of machine learning models with rigorous physics-based approaches. Moreover, we demonstrate how leveraging augmented data (generated using template-based modelling or molecular docking) can significantly improve binding affinity prediction correlation and ranking on the FEP benchmark (weighted mean PCC and Kendall's τ increases from 0.41 and 0.26 to 0.59 and 0.42). These strategies together are closing the performance gap with FEP calculations (FEP+ achieves weighted mean PCC and Kendall's τ of 0.68 and 0.49 on the FEP benchmark) while being  ~400,000 times faster.

Robustly interrogating machine learning-based scoring functions: what are they learning?

MOTIVATION

Machine learning-based scoring functions (MLBSFs) have been found to exhibit inconsistent performance on different benchmarks and be prone to learning dataset bias. For the field to develop MLBSFs that learn a generalizable understanding of physics, a more rigorous understanding of how they perform is required.

RESULTS

In this work, we compared the performance of a diverse set of popular MLBSFs (RFScore, SIGN, OnionNet-2, Pafnucy, and PointVS) to our proposed baseline models that can only learn dataset biases on a range of benchmarks. We found that these baseline models were competitive in accuracy to these MLBSFs in almost all proposed benchmarks, indicating these models only learn dataset biases. Our tests and provided platform, ToolBoxSF, will enable researchers to robustly interrogate MLBSF performance and determine the effect of dataset biases on their predictions.

AVAILABILITY AND IMPLEMENTATION

https://github.com/guydurant/toolboxsf.

CPIScore: A Deep Learning Approach for Rapid Scoring and Interpretation of Protein-Ligand Binding Interactions

Protein-ligand binding affinity prediction is a crucial and challenging task in the field of drug discovery. However, traditional simulation-based computational approaches are often prohibitively time-consuming, limiting their practical utility. In this study, we introduce a novel deep learning method, CPIScore, which leverages the capabilities of Transformer and Graph Convolutional Networks (GCN) to enhance the prediction of protein-ligand binding affinity. CPIScore utilizes the Transformer architecture to capture comprehensive global contexts of protein and ligand sequences, while the GCN component effectively extracts local features from small molecular graphs. Our results demonstrate that CPIScore surpasses both traditional machine learning and other deep learning models in accuracy, achieving a Pearson's r of 0.74 on our test set. Furthermore, CPIScore has been validated across multiple targets, proving its ability to discern inhibitors from a diverse compound library with high enrichment rates. Notably, when applied to a generated focused library of compounds, CPIScore successfully identified six potent small-molecule inhibitors of ATR, which were tested experimentally and four small molecules exhibited inhibitory activity below ten nanomoles. These results highlight CPIScore's potential to significantly streamline and enhance the efficiency of drug discovery processes.

From Static to Dynamic Structures: Improving Binding Affinity Prediction with Graph-Based Deep Learning

Accurate prediction of protein-ligand binding affinities is an essential challenge in structure-based drug design. Despite recent advances in data-driven methods for affinity prediction, their accuracy is still limited, partially because they only take advantage of static crystal structures while the actual binding affinities are generally determined by the thermodynamic ensembles between proteins and ligands. One effective way to approximate such a thermodynamic ensemble is to use molecular dynamics (MD) simulation. Here, an MD dataset containing 3,218 different protein-ligand complexes is curated, and Dynaformer, a graph-based deep learning model is further developed to predict the binding affinities by learning the geometric characteristics of the protein-ligand interactions from the MD trajectories. In silico experiments demonstrated that the model exhibits state-of-the-art scoring and ranking power on the CASF-2016 benchmark dataset, outperforming the methods hitherto reported. Moreover, in a virtual screening on heat shock protein 90 (HSP90) using Dynaformer, 20 candidates are identified and their binding affinities are further experimentally validated. Dynaformer displayed promising results in virtual drug screening, revealing 12 hit compounds (two are in the submicromolar range), including several novel scaffolds. Overall, these results demonstrated that the approach offer a promising avenue for accelerating the early drug discovery process.

The future of machine learning for small-molecule drug discovery will be driven by data

Many studies have prophesied that the integration of machine learning techniques into small-molecule therapeutics development will help to deliver a true leap forward in drug discovery. However, increasingly advanced algorithms and novel architectures have not always yielded substantial improvements in results. In this Perspective, we propose that a greater focus on the data for training and benchmarking these models is more likely to drive future improvement, and explore avenues for future research and strategies to address these data challenges.

Machine learning-assisted search for novel coagulants: When machine learning can be efficient even if data availability is low

Design of new drugs is a challenging process: a candidate molecule should satisfy multiple conditions to act properly and make the least side-effect-perfect candidates selectively attach to and influence only targets, leaving off-targets intact. The amount of experimental data about various properties of molecules constantly grows, promoting data-driven approaches. However, the applicability of typical predictive machine learning techniques can be substantially limited by a lack of experimental data about a particular target. For example, there are many known Thrombin inhibitors (acting as anticoagulants), but a very limited number of known Protein C inhibitors (coagulants). In this study, we present our approach to suggest new inhibitor candidates by building an effective representation of chemical space. For this aim, we developed a deep learning model-autoencoder, trained on a large set of molecules in the SMILES format to map the chemical space. Further, we applied different sampling strategies to generate novel coagulant candidates. Symmetrically, we tested our approach on anticoagulant candidates, where we were able to predict their inhibition towards Thrombin. We also compare our approach with MegaMolBART-another deep learning generative model, but exploiting similar principles of navigation in a chemical space.

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.

Machine learning accelerates pharmacophore-based virtual screening of MAO inhibitors

Nowadays, an efficient and robust virtual screening procedure is crucial in the drug discovery process, especially when performed on large and chemically diverse databases. Virtual screening methods, like molecular docking and classic QSAR models, are limited in their ability to handle vast numbers of compounds and to learn from scarce data, respectively. In this study, we introduce a universal methodology that uses a machine learning-based approach to predict docking scores without the need for time-consuming molecular docking procedures. The developed protocol yielded 1000 times faster binding energy predictions than classical docking-based screening. The proposed predictive model learns from docking results, allowing users to choose their preferred docking software without relying on insufficient and incoherent experimental activity data. The methodology described employs multiple types of molecular fingerprints and descriptors to construct an ensemble model that further reduces prediction errors and is capable of delivering highly precise docking score values for monoamine oxidase ligands, enabling faster identification of promising compounds. An extensive pharmacophore-constrained screening of the ZINC database resulted in a selection of 24 compounds that were synthesized and evaluated for their biological activity. A preliminary screen discovered weak inhibitors of MAO-A with a percentage efficiency index close to a known drug at the lowest tested concentration. The approach presented here can be successfully applied to other biological targets as target-specific knowledge is not incorporated at the screening phase.

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.

A flexible data-free framework for structure-based de novo drug design with reinforcement learning

Contemporary structure-based molecular generative methods have demonstrated their potential to model the geometric and energetic complementarity between ligands and receptors, thereby facilitating the design of molecules with favorable binding affinity and target specificity. Despite the introduction of deep generative models for molecular generation, the atom-wise generation paradigm that partially contradicts chemical intuition limits the validity and synthetic accessibility of the generated molecules. Additionally, the dependence of deep learning models on large-scale structural data has hindered their adaptability across different targets. To overcome these challenges, we present a novel search-based framework, 3D-MCTS, for structure-based de novo drug design. Distinct from prevailing atom-centric methods, 3D-MCTS employs a fragment-based molecular editing strategy. The fragments decomposed from small-molecule drugs are recombined under predefined retrosynthetic rules, offering improved drug-likeness and synthesizability, overcoming the inherent limitations of atom-based approaches. Leveraging multi-threaded parallel simulations combined with a real-time energy constraint-based pruning strategy, 3D-MCTS achieves remarkable efficiency. At a fixed computational cost, it outperforms other state-of-the-art (SOTA) methods by producing molecules with enhanced binding affinity. Furthermore, its fragment-based approach ensures the generation of more dependable binding conformations, exhibiting a success rate 43.6% higher than that of other SOTAs. This advantage becomes even more pronounced when handling targets that significantly deviate from the training dataset. 3D-MCTS is capable of achieving thirty times more hits with high binding affinity than traditional virtual screening methods, which demonstrates the superior ability of 3D-MCTS to explore chemical space. Moreover, the flexibility of our framework makes it easy to incorporate domain knowledge during the process, thereby enabling the generation of molecules with desirable pharmacophores and enhanced binding affinity. The adaptability of 3D-MCTS is further showcased in metalloprotein applications, highlighting its potential across various drug design scenarios.

Exploring the artificial intelligence and machine learning models in the context of drug design difficulties and future potential for the pharmaceutical sectors

Artificial intelligence (AI), particularly deep learning as a subcategory of AI, provides opportunities to accelerate and improve the process of discovering and developing new drugs. The use of AI in drug discovery is still in its early stages, but it has the potential to revolutionize the way new drugs are discovered and developed. As AI technology continues to evolve, it is likely that AI will play an even greater role in the future of drug discovery. AI is used to identify new drug targets, design new molecules, and predict the efficacy and safety of potential drugs. The inclusion of AI in drug discovery can screen millions of compounds in a matter of hours, identifying potential drug candidates that would have taken years to find using traditional methods. AI is highly utilized in the pharmaceutical industry by optimizing processes, reducing waste, and ensuring quality control. This review covers much-needed topics, including the different types of machine-learning techniques, their applications in drug discovery, and the challenges and limitations of using machine learning in this field. The state-of-the-art of AI-assisted pharmaceutical discovery is described, covering applications in structure and ligand-based virtual screening, de novo drug creation, prediction of physicochemical and pharmacokinetic properties, drug repurposing, and related topics. Finally, many obstacles and limits of present approaches are outlined, with an eye on potential future avenues for AI-assisted drug discovery and design.

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.

A Small Step Toward Generalizability: Training a Machine Learning Scoring Function for Structure-Based Virtual Screening

Over the past few years, many machine learning-based scoring functions for predicting the binding of small molecules to proteins have been developed. Their objective is to approximate the distribution which takes two molecules as input and outputs the energy of their interaction. Only a scoring function that accounts for the interatomic interactions involved in binding can accurately predict binding affinity on unseen molecules. However, many scoring functions make predictions based on data set biases rather than an understanding of the physics of binding. These scoring functions perform well when tested on similar targets to those in the training set but fail to generalize to dissimilar targets. To test what a machine learning-based scoring function has learned, input attribution, a technique for learning which features are important to a model when making a prediction on a particular data point, can be applied. If a model successfully learns something beyond data set biases, attribution should give insight into the important binding interactions that are taking place. We built a machine learning-based scoring function that aimed to avoid the influence of bias via thorough train and test data set filtering and show that it achieves comparable performance on the Comparative Assessment of Scoring Functions, 2016 (CASF-2016) benchmark to other leading methods. We then use the CASF-2016 test set to perform attribution and find that the bonds identified as important by PointVS, unlike those extracted from other scoring functions, have a high correlation with those found by a distance-based interaction profiler. We then show that attribution can be used to extract important binding pharmacophores from a given protein target when supplied with a number of bound structures. We use this information to perform fragment elaboration and see improvements in docking scores compared to using structural information from a traditional, data-based approach. This not only provides definitive proof that the scoring function has learned to identify some important binding interactions but also constitutes the first deep learning-based method for extracting structural information from a target for molecule design.