Chemical Space Exploration with Active Learning and Alchemical Free Energies

Active learning approaches in molecule pKi prediction

During the early stages of drug design, identifying compounds with suitable bioactivities is crucial. Given the vast array of potential drug databases, it's feasible to assay only a limited subset of candidates. The optimal method for selecting the candidates, aiming to minimize the overall number of assays, involves an active learning (AL) approach. In this work, we benchmarked a range of AL strategies with two main objectives: (1) to identify a strategy that ensures high model performance and (2) to select molecules with desired properties using minimal assays. To evaluate the different AL strategies, we employed the simulated AL workflow based on "virtual" experiments. These experiments leveraged ChEMBL datasets, which come with known biological activity values for the molecules. Furthermore, for classification tasks, we proposed the hybrid selection strategy that unified both exploration and exploitation AL strategies into a single acquisition function, defined by parameters n and c. We have also shown that popular minimal margin and maximal variance selection approaches for exploration selection correspond to minimization of the hybrid acquisition function with n=1 and 2 respectively. The balance between the exploration and exploitation strategies can be adjusted using a coefficient (c), making the optimal strategy selection straightforward. The primary strength of the hybrid selection method lies in its adaptability; it offers the flexibility to adjust the criteria for molecule selection based on the specific task by modifying the value of the contribution coefficient. Our analysis revealed that, in regression tasks, AL strategies didn't succeed at ensuring high model performance, however, they were successful in selecting molecules with desired properties using minimal number of tests. In analogous experiments in classification tasks, exploration strategy and the hybrid selection function with a constant c<1 (for n=1) and c≤0.2 (for n=2) were effective in achieving the goal of constructing a high-performance predictive model using minimal data. When searching for molecules with desired properties, exploitation, and the hybrid function with c≥1 (n=1) and c≥0.7 (n=2) demonstrated efficiency identifying molecules in fewer iterations compared to random selection method. Notably, when the hybrid function was set to an intermediate coefficient value (c=0.7), it successfully addressed both tasks simultaneously.

Structure-aware dual-target drug design through collaborative learning of pharmacophore combination and molecular simulation

Dual-target drug design has gained significant attention in the treatment of complex diseases, such as cancers and autoimmune disorders. A widely employed design strategy is combining pharmacophores to leverage the knowledge of structure-activity relationships of both targets. Unfortunately, pharmacophore combination often struggles with long and expensive trial and error, because the protein pockets of the two targets impose complex structural constraints. In this study, we propose AIxFuse, a structure-aware dual-target drug design method that learns pharmacophore fusion patterns to satisfy the dual-target structural constraints simulated by molecular docking. AIxFuse employs two self-play reinforcement learning (RL) agents to learn pharmacophore selection and fusion by comprehensive feedback including dual-target molecular docking scores. Collaboratively, the molecular docking scores are learned by active learning (AL). Through collaborative RL and AL, AIxFuse learns to generate molecules with multiple desired properties. AIxFuse is shown to outperform state-of-the-art methods in generating dual-target drugs against glycogen synthase kinase-3 beta (GSK3β) and c-Jun N-terminal kinase 3 (JNK3). When applied to another task against retinoic acid receptor-related orphan receptor γ-t (RORγt) and dihydroorotate dehydrogenase (DHODH), AIxFuse exhibits consistent performance while compared methods suffer from performance drops, leading to a 5 times higher performance in success rate. Docking studies demonstrate that AIxFuse can generate molecules concurrently satisfying the binding mode required by both targets. Further free energy perturbation calculation indicates that the generated candidates have promising binding free energies against both targets.

The present state and challenges of active learning in drug discovery

Active learning (AL) is an iterative feedback process that efficiently identifies valuable data within vast chemical space, even with limited labeled data. This characteristic renders it a valuable approach to tackle the ongoing challenges faced in drug discovery, such as the ever-expanding explore space and the limitations of labeled data. Consequently, AL is increasingly gaining prominence in the field of drug development. In this paper, we comprehensively review the application of AL at all stages of drug discovery, including compounds-target interaction prediction, virtual screening, molecular generation and optimization, as well as molecular properties prediction. Additionally, we discuss the challenges and prospects associated with the current applications of AL in drug discovery.

FEP Augmentation as a Means to Solve Data Paucity Problems for Machine Learning in Chemical Biology

In the realm of medicinal chemistry, the primary objective is to swiftly optimize a multitude of chemical properties of a set of compounds to yield a clinical candidate poised for clinical trials. In recent years, two computational techniques, machine learning (ML) and physics-based methods, have evolved substantially and are now frequently incorporated into the medicinal chemist's toolbox to enhance the efficiency of both hit optimization and candidate design. Both computational methods come with their own set of limitations, and they are often used independently of each other. ML's capability to screen extensive compound libraries expediently is tempered by its reliance on quality data, which can be scarce especially during early-stage optimization. Contrarily, physics-based approaches like free energy perturbation (FEP) are frequently constrained by low throughput and high cost by comparison; however, physics-based methods are capable of making highly accurate binding affinity predictions. In this study, we harnessed the strength of FEP to overcome data paucity in ML by generating virtual activity data sets which then inform the training of algorithms. Here, we show that ML algorithms trained with an FEP-augmented data set could achieve comparable predictive accuracy to data sets trained on experimental data from biological assays. Throughout the paper, we emphasize key mechanistic considerations that must be taken into account when aiming to augment data sets and lay the groundwork for successful implementation. Ultimately, the study advocates for the synergy of physics-based methods and ML to expedite the lead optimization process. We believe that the physics-based augmentation of ML will significantly benefit drug discovery, as these techniques continue to evolve.

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.

Benchmarking Active Learning Protocols for Ligand-Binding Affinity Prediction

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

The SAMPL9 host-guest blind challenge: an overview of binding free energy predictive accuracy

We report the results of the SAMPL9 host-guest blind challenge for predicting binding free energies. The challenge focused on macrocycles from pillar[n]-arene and cyclodextrin host families, including WP6, and bCD and HbCD. A variety of methods were used by participants to submit binding free energy predictions. A machine learning approach based on molecular descriptors achieved the highest accuracy (RMSE of 2.04 kcal mol-1) among the ranked methods in the WP6 dataset. Interestingly, predictions for WP6 obtained via docking tended to outperform all methods (RMSE of 1.70 kcal mol-1), most of which are MD based and computationally more expensive. In general, methods applying force fields achieved better correlation with experiments for WP6 opposed to the machine learning and docking models. In the cyclodextrin-phenothiazine challenge, the ATM approach emerged as the top performing method with RMSE less than 1.86 kcal mol-1. Correlation metrics of ranked methods in this dataset were relatively poor compared to WP6. We also highlight several lessons learned to guide future work and help improve studies on the systems discussed. For example, WP6 may be present in other microstates other than its -12 state in the presence of certain guests. Machine learning approaches can be used to fine tune or help train force fields for certain chemistry (i.e. WP6-G4). Certain phenothiazines occupy distinct primary and secondary orientations, some of which were considered individually for accurate binding free energies. The accuracy of predictions from certain methods while starting from a single binding pose/orientation demonstrates the sensitivity of calculated binding free energies to the orientation, and in some cases the likely dominant orientation for the system. Computational and experimental results suggest that guest phenothiazine core traverses both the secondary and primary faces of the cyclodextrin hosts, a bulky cationic side chain will primarily occupy the primary face, and the phenothiazine core substituent resides at the larger secondary face.

Exploring Chemical Space Using Ab Initio Hyperreactor Dynamics

In recent years, first-principles exploration of chemical reaction space has provided valuable insights into intricate reaction networks. Here, we introduce ab initio hyperreactor dynamics, which enables rapid screening of the accessible chemical space from a given set of initial molecular species, predicting new synthetic routes that can potentially guide subsequent experimental studies. For this purpose, different hyperdynamics derived bias potentials are applied along with pressure-inducing spherical confinement of the molecular system in ab initio molecular dynamics simulations to efficiently enhance reactivity under mild conditions. To showcase the advantages and flexibility of the hyperreactor approach, we present a systematic study of the method's parameters on a HCN toy model and apply it to a recently introduced experimental model for the prebiotic formation of glycinal and acetamide in interstellar ices, which yields results in line with experimental findings. In addition, we show how the developed framework enables the study of complicated transitions like the first step of a nonenzymatic DNA nucleoside synthesis in an aqueous environment, where the molecular fragmentation problem of earlier nanoreactor approaches is avoided.

ChemSpaceAL: An Efficient Active Learning Methodology Applied to Protein-Specific Molecular Generation

The incredible capabilities of generative artificial intelligence models have inevitably led to their application in the domain of drug discovery. Within this domain, the vastness of chemical space motivates the development of more efficient methods for identifying regions with molecules that exhibit desired characteristics. In this work, we present a computationally efficient active learning methodology and demonstrate its applicability to targeted molecular generation. When applied to c-Abl kinase, a protein with FDA-approved small-molecule inhibitors, the model learns to generate molecules similar to the inhibitors without prior knowledge of their existence and even reproduces two of them exactly. We also show that the methodology is effective for a protein without any commercially available small-molecule inhibitors, the HNH domain of the CRISPR-associated protein 9 (Cas9) enzyme. To facilitate implementation and reproducibility, we made all of our software available through the open-source ChemSpaceAL Python package.

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.

Alchemical Free Energy Calculations on Membrane-Associated Proteins

Membrane proteins have diverse functions within cells and are well-established drug targets. The advances in membrane protein structural biology have revealed drug and lipid binding sites on membrane proteins, while computational methods such as molecular simulations can resolve the thermodynamic basis of these interactions. Particularly, alchemical free energy calculations have shown promise in the calculation of reliable and reproducible binding free energies of protein-ligand and protein-lipid complexes in membrane-associated systems. In this review, we present an overview of representative alchemical free energy studies on G-protein-coupled receptors, ion channels, transporters as well as protein-lipid interactions, with emphasis on best practices and critical aspects of running these simulations. Additionally, we analyze challenges and successes when running alchemical free energy calculations on membrane-associated proteins. Finally, we highlight the value of alchemical free energy calculations calculations in drug discovery and their applicability in the pharmaceutical industry.

Deep Learning with Geometry-Enhanced Molecular Representation for Augmentation of Large-Scale Docking-Based Virtual Screening

Structure-based virtual screening has been a crucial tool in drug discovery for decades. However, as the chemical space expands, the existing structure-based virtual screening techniques based on molecular docking and scoring struggle to handle billion-entry ultralarge libraries due to the high computational cost. To address this challenge, people have resorted to machine learning techniques to enhance structure-based virtual screening for efficiently exploring the vast chemical space. In those cases, compounds are usually treated as sequential strings or two-dimensional topology graphs, limiting their ability to incorporate three-dimensional structural information for downstream tasks. We herein propose a novel deep learning protocol, GEM-Screen, which utilizes the geometry-enhanced molecular representation of the compounds docking to a specific target and is trained on docking scores of a small fraction of a library through an active learning strategy to approximate the docking outcome for yet nontraining entries. This protocol is applied to virtual screening campaigns against the AmpC and D4 targets, demonstrating that GEM-Screen enriches more than 90% of the hit scaffolds for AmpC in the top 4% of model predictions and more than 80% of the hit scaffolds for D4 in the same top-ranking size of library. GEM-Screen can be used in conjunction with traditional docking programs for docking of only the top-ranked compounds to avoid the exhaustive docking of the whole library, thus allowing for discovering top-scoring compounds from billion-entry libraries in a rapid yet accurate fashion.

Efficient exploration of transition-metal decorated MXene for carbon monoxide sensing using integrated active learning and density functional theory

The urgent demand for chemical safety necessitates the real-time detection of carbon monoxide (CO), a highly toxic gas. MXene, a 2D material, has shown potential for gas sensing applications (e.g., NH3, NO, SO2, CO2) due to its high surface accessibility, electrical conductivity, stability, and flexibility in surface functionalization. However, the pristine MXene generally exhibits poor interaction with CO; still, transition metal decoration can strengthen the interaction between CO and MXene. This study presents a high-throughput screening of 450 combinations of transition-metal (TM) decorated MXene (TM@MXene) for CO sensing applications using an integrated active learning (AL) and density functional theory (DFT) screening pipeline. Our AL pipeline, adopting a crystal graph convolutional neural network (CGCNN) as a surrogate model, successfully accelerates the screening of CO sensor candidates with minimal computational resources. This study identifies Sc@Zr3C2O2 and Y@Zr3C2O2 as the optimal TM@MXene candidates with promising CO sensing performance regarding the screening criteria of recovery time, surface stability, charge transfer, and sensitivity to CO. The proposed AL framework can be extended for property finetuning in the combinatorial chemical space.

Using machine learning to go beyond potential energy surface benchmarking for chemical reactivity

We train an equivariant machine learning (ML) model to predict energies and forces for hydrogen combustion under conditions of finite temperature and pressure. This challenging case for reactive chemistry illustrates that ML potential energy surfaces are difficult to make complete, due to overreliance on chemical intuition of what data are important for training. Instead, a 'negative design' data acquisition strategy using metadynamics as part of an active learning workflow helps to create a ML model that avoids unforeseen high-energy or unphysical energy configurations. This strategy more rapidly converges the potential energy surfaces such that it is now more efficient to make calls to the external ab initio source when query-by-committee models disagree to further molecular dynamics in time without need for ML retraining. With the hybrid ML-physics model we realize two orders of magnitude reduction in cost, allowing for prediction of the free-energy change in the transition-state mechanism for several hydrogen combustion reaction channels.

A Simple Way to Incorporate Target Structural Information in Molecular Generative Models

Deep learning generative models are now being applied in various fields including drug discovery. In this work, we propose a novel approach to include target 3D structural information in molecular generative models for structure-based drug design. The method combines a message-passing neural network model that predicts docking scores with a generative neural network model as its reward function to navigate the chemical space searching for molecules that bind favorably with a specific target. A key feature of the method is the construction of target-specific molecular sets for training, designed to overcome potential transferability issues of surrogate docking models through a two-round training process. Consequently, this enables accurate guided exploration of the chemical space without reliance on the collection of prior knowledge about active and inactive compounds for the specific target. Tests on eight target proteins showed a 100-fold increase in hit generation compared to conventional docking calculations and the ability to generate molecules similar to approved drugs or known active ligands for specific targets without prior knowledge. This method provides a general and highly efficient solution for structure-based molecular generation.

ZINC-22─A Free Multi-Billion-Scale Database of Tangible Compounds for Ligand Discovery

Purchasable chemical space has grown rapidly into the tens of billions of molecules, providing unprecedented opportunities for ligand discovery but straining the tools that might exploit these molecules at scale. We have therefore developed ZINC-22, a database of commercially accessible small molecules derived from multi-billion-scale make-on-demand libraries. The new database and tools enable analog searching in this vast new space via a facile GUI, CartBlanche, drawing on similarity methods that scale sublinearly in the number of molecules. The new library also uses data organization methods, enabling rapid lookup of molecules and their physical properties, including conformations, partial atomic charges, c Log P values, and solvation energies, all crucial for molecule docking, which had become slow with older database organizations in previous versions of ZINC. As the libraries have continued to grow, we have been interested in finding whether molecular diversity has suffered, for instance, because certain scaffolds have come to dominate via easy analoging. This has not occurred thus far, and chemical diversity continues to grow with database size, with a log increase in Bemis-Murcko scaffolds for every two-log unit increase in database size. Most new scaffolds come from compounds with the highest heavy atom count. Finally, we consider the implications for databases like ZINC as the libraries grow toward and beyond the trillion-molecule range. ZINC is freely available to everyone and may be accessed at cartblanche22.docking.org, via Globus, and in the Amazon AWS and Oracle OCI clouds.