Mining Toxicity Information from Large Amounts of Toxicity Data

MDNN-DTA: a multimodal deep neural network for drug-target affinity prediction

Determining drug-target affinity (DTA) is a pivotal step in drug discovery, where in silico methods can significantly improve efficiency and reduce costs. Artificial intelligence (AI), especially deep learning models, can automatically extract high-dimensional features from the biological sequences of drug molecules and target proteins. This technology demonstrates lower complexity in DTA prediction compared to traditional experimental methods, particularly when handling large-scale data. In this study, we introduce a multimodal deep neural network model for DTA prediction, referred to as MDNN-DTA. This model employs Graph Convolutional Networks (GCN) and Convolutional Neural Networks (CNN) to extract features from the drug and protein sequences, respectively. One notable strength of our method is its ability to accurately predict DTA directly from the sequences of the target proteins, obviating the need for protein 3D structures, which are frequently unavailable in drug discovery. To comprehensively extract features from the protein sequence, we leverage an ESM pre-trained model for extracting biochemical features and design a specific Protein Feature Extraction (PFE) block for capturing both global and local features of the protein sequence. Furthermore, a Protein Feature Fusion (PFF) Block is engineered to augment the integration of multi-scale protein features derived from the abovementioned techniques. We then compare MDNN-DTA with other models on the same dataset, conducting a series of ablation experiments to assess the performance and efficacy of each component. The results highlight the advantages and effectiveness of the MDNN-DTA method.

Single-step retrosynthesis prediction via multitask graph representation learning

Inferring appropriate synthesis reaction (i.e., retrosynthesis) routes for newly designed molecules is vital. Recently, computational methods have produced promising single-step retrosynthesis predictions. However, template-based methods are limited by the known synthesis templates; template-free methods are weakly interpretable; and semi template-based methods are deficient with regard to utilizing the associations between chemical entities. To address these issues, this paper leverages the intra-associations between synthons, the inter-associations between synthons and leaving groups (LGs), and the intra-associations between LGs. It develops a multitask graph representation learning model for single-step retrosynthesis prediction (Retro-MTGR) to solve reaction centre deduction and LG identification simultaneously. A comparison with 16 state-of-the-art methods first demonstrates the superiority of Retro-MTGR. Then, its robustness and scalability and the contributions of its crucial components are validated. More importantly, it can determine whether a bond can be a reaction centre and what LGs are appropriate for a given synthon, respectively. The answers reflect underlying chemical synthesis rules, especially opposite electrical properties between chemical entities (e.g., reaction sites, synthons, and LGs). Finally, case studies demonstrate that the retrosynthesis routes inferred by Retro-MTGR are promising for single-step synthesis reactions. The code and data of this study are freely available at https://doi.org/10.5281/zenodo.14346324 .

Construction of interpretable ensemble learning models for predicting bioaccumulation parameters of organic chemicals in fish

Accurate prediction of bioaccumulation parameters is essential for assessing exposure, hazards, and risks of chemicals. However, the majority of prediction models on bioaccumulation parameters are individual models based on a single algorithm and lack model interpretation, resulting in unsatisfactory prediction accuracy due to inherent constraints of the algorithm and weak interpretability. Ensemble learning (EL) that combine multiple algorithms, coupled with SHapley Additive exPlanation (SHAP) method, may overcome the limitations. Herein, EL models were constructed for three bioaccumulation parameters using datasets covering 2496 chemicals. The EL models demonstrated superior prediction accuracy compared to both individual models developed in this study and those from previous research, achieving a coefficient of determination of up to 0.861 on the validation sets. Applicability domains were characterized using a structure-activity landscape-based (abbreviated as ADSAL) methodology. The optimal EL models, together with the ADSAL, were successfully used to predict bioaccumulation parameters for 4374 chemicals included in the Inventory of Existing Chemical Substances of China. Model interpretation using the SHAP method offered insight into key features influencing bioaccumulation potential, including hydrophobicity, water solubility, polarizability, ionization potential, weight, and volume of molecules. Overall, the study provides data and models to support the sound management and risk assessment of chemicals.

GCLmf: A Novel Molecular Graph Contrastive Learning Framework Based on Hard Negatives and Application in Toxicity Prediction

In silico methods for prediction of chemical toxicity can decrease the cost and increase the efficiency in the early stage of drug discovery. However, due to low accessibility of sufficient and reliable toxicity data, constructing robust and accurate prediction models is challenging. Contrastive learning, a type of self-supervised learning, leverages large unlabeled data to obtain more expressive molecular representations, which can boost the prediction performance on downstream tasks. While molecular graph contrastive learning has gathered growing attentions, current models neglect the quality of negative data set. Here, we proposed a self-supervised pretraining deep learning framework named GCLmf. We first utilized molecular fragments that meet specific conditions as hard negative samples to boost the quality of the negative set and thus increase the difficulty of the proxy tasks during pre-training to learn informative representations. GCLmf has shown excellent predictive power on various molecular property benchmarks and demonstrates high performance in 33 toxicity tasks in comparison with multiple baselines. In addition, we further investigated the necessity of introducing hard negatives in model building and the impact of the proportion of hard negatives on the model.

Advanced AI-Driven Prediction of Pregnancy-Related Adverse Drug Reactions

Ensuring drug safety during pregnancy is critical due to the potential risks to both the mother and fetus. However, the exclusion of pregnant women from clinical trials complicates the assessment of adverse drug reactions (ADRs) in this population. This study aimed to develop and validate risk prediction models for pregnancy-related ADRs of drugs using advanced Machine Learning (ML) and Deep Learning (DL) techniques, leveraging real-world data from the FDA Adverse Event Reporting System. We explored three methods─Information Component, Reporting Odds Ratio, and 95% confidence interval of ROR─for classifying drugs into high-risk and low-risk categories. DL models, including Directed Message Passing Neural Networks (DMPNN), Graph Neural Networks, and Graph Convolutional Networks, were developed and compared to traditional ML models like Random Forest, Support Vector Machines, and XGBoost. Among these, the DMPNN model, which integrated molecular graph information and molecular descriptors, exhibited the highest predictive performance, particularly at the preferred term level. The model was validated against external data sets from SIDER and DailyMed, demonstrating strong generalizability. Additionally, the model was applied to assess the risk of 22 oral hypoglycemic drugs, and potential substructure alerts for pregnancy-related ADRs were identified. These findings suggest that the DMPNN model is a valuable tool for predicting ADRs in pregnant women, offering significant advancement in drug safety assessment and providing crucial insights for safer medication use during pregnancy.

Prediction and Interpretation Microglia Cytotoxicity by Machine Learning

Ameliorating microglia-mediated neuroinflammation is a crucial strategy in developing new drugs for neurodegenerative diseases. Plant compounds are an important screening target for the discovery of drugs for the treatment of neurodegenerative diseases. However, due to the spatial complexity of phytochemicals, it becomes particularly important to evaluate the effectiveness of compounds while avoiding the mixing of cytotoxic substances in the early stages of compound screening. Traditional high-throughput screening methods suffer from high cost and low efficiency. A computational model based on machine learning provides a novel avenue for cytotoxicity determination. In this study, a microglia cytotoxicity classifier was developed using a machine learning approach. First, we proposed a data splitting strategy based on the molecule murcko generic scaffold, under this condition, three machine learning approaches were coupled with three kinds of molecular representation methods to construct microglia cytotoxicity classifier, which were then compared and assessed by the predictive accuracy, balanced accuracy, F1-score, and Matthews Correlation Coefficient. Then, the recursive feature elimination integrated with support vector machine (RFE-SVC) dimension reduction method was introduced to molecular fingerprints with high dimensions to further improve the model performance. Among all the microglial cytotoxicity classifiers, the SVM coupled with ECFP4 fingerprint after feature selection (ECFP4-RFE-SVM) obtained the most accurate classification for the test set (ACC of 0.99, BA of 0.99, F1-score of 0.99, MCC of 0.97). Finally, the Shapley additive explanations (SHAP) method was used in interpreting the microglia cytotoxicity classifier and key substructure smart identified as structural alerts. Experimental results show that ECFP4-RFE-SVM have reliable classification capability for microglia cytotoxicity, and SHAP can not only provide a rational explanation for microglia cytotoxicity predictions, but also offer a guideline for subsequent molecular cytotoxicity modifications.

FGTN: Fragment-based graph transformer network for predicting reproductive toxicity

Reproductive toxicity is one of the important issues in chemical safety. Traditional laboratory testing methods are costly and time-consuming with raised ethical issues. Only a few in silico models have been reported to predict human reproductive toxicity, but none of them make full use of the topological information of compounds. In addition, most existing atom-based graph neural network methods focus on attributing model predictions to individual nodes or edges rather than chemically meaningful fragments or substructures. In current studies, we develop a novel fragment-based graph transformer network (FGTN) approach to generate the QSAR model of human reproductive toxicity by considering internal topological structure information of compounds. In the FGTN model, the compound is represented by a graph architecture using fragments to be nodes and bonds linking two fragments to be edges. A super molecule-level node is further proposed to connect all fragment nodes by undirected edges, obtaining global molecular features from fragment embeddings. The FGTN model achieved an accuracy (ACC) of 0.861 and an area under the receiver operating characteristic curve (AUC) value of 0.914 on nonredundant blind tests, outperforming traditional fingerprint-based machine learning models and atom-based GCN model. The FGTN model can attribute toxic predictions to fragments, generating specific structural alerts for the positive compound. Moreover, FGTN may also have the capability to distinguish various chemical isomers. We believe that FGTN can be used as a reliable and effective tool for human reproductive toxicity prediction in contribution to the advancement of chemical safety assessment.

Molecular designing of potential environmentally friendly PFAS based on deep learning and generative models

Perfluoroalkyl and polyfluoroalkyl substances (PFAS) are widely used across a spectrum of industrial and consumer goods. Nonetheless, their persistent nature and tendency to accumulate in biological systems pose substantial environmental and health threats. Consequently, striking a balance between maximizing product efficiency and minimizing environmental and health risks by tailoring the molecular structure of PFAS has become a pivotal challenge in the fields of environmental chemistry and sustainable development. To address this issue, a computational workflow was proposed for designing an environmentally friendly PFAS by incorporating deep learning (DL) and molecular generative models. The hybrid DL architecture MolHGT+ based on heterogeneous graph neural network with transformer-like attention was applied to predict the surface tension, bioaccumulation, and hepatotoxicity of the molecules. Through virtual screening of the PFAS master database using MolHGT+, the findings indicate that incorporating the siloxane group and betaine fragment can effectively decrease both the bioaccumulation and hepatotoxicity of PFAS while preserving low surface tension. In addition, molecular generative models were employed to create a structurally diverse pool of novel PFASs with the aforementioned hit molecules serving as the initial template structures. Overall, our study presents a promising AI-driven method for advancing the development of environmentally friendly PFAS.

DigFrag as a digital fragmentation method used for artificial intelligence-based drug design

Fragment-Based Drug Design (FBDD) plays a pivotal role in the field of drug discovery and development. The construction of high-quality fragment libraries is a critical step in FBDD. Conventional fragmentation approaches often rely on rigid rules and chemical intuition, limiting their adaptability to diverse molecular structures. The rapid development of Artificial Intelligence (AI) technology offers a transformative opportunity to rethink traditional methods. Here, we present DigFrag, a digital fragmentation method that highlights important substructures by focusing locally within the molecular graph. In addition, we feed the fragments segmented by machine intelligence and human expertise into the deep generative model to compare the preference for data from different sources. Experimental results show that the structural diversity of fragments segmented by DigFrag is higher, and more desirable compounds are generated based on these fragments. These results also demonstrate that data generated based on AI methods may be more suitable for AI models. Moreover, a user-friendly platform called MolFrag ( https://dpai.ccnu.edu.cn/MolFrag/ ) is developed based on various fragmentation techniques to support molecular segmentation.

Multimodal Representation Learning via Graph Isomorphism Network for Toxicity Multitask Learning

Toxicity is paramount for comprehending compound properties, particularly in the early stages of drug design. Due to the diversity and complexity of toxic effects, it became a challenge to compute compound toxicity tasks. To address this issue, we propose a multimodal representation learning model, termed multimodal graph isomorphism network (MMGIN), to address this challenge for compound toxicity multitask learning. Based on fingerprints and molecular graphs of compounds, our MMGIN model incorporates a multimodal representation learning model to acquire a comprehensive compound representation. This model adopts a two-channel structure to independently learn fingerprint representation and molecular graph representation. Subsequently, two feedforward neural networks utilize the learned multimodal compound representation to perform multitask learning, encompassing compound toxicity classification and multiple compound category classification simultaneously. To test the effectiveness of our model, we constructed a novel data set, termed the compound toxicity multitask learning (CTMTL) data set, derived from the TOXRIC data set. We compare our MMGIN model with other representative machine learning and deep learning models on the CTMTL and Tox21 data sets. The experimental results demonstrate significant advancements achieved by our MMGIN model. Furthermore, the ablation study underscores the effectiveness of the introduced fingerprints, molecular graphs, the multimodal representation learning model, and the multitask learning model, showcasing the model's superior predictive capability and robustness.

Data-driven toxicity prediction in drug discovery: Current status and future directions

Early toxicity assessment plays a vital role in the drug discovery process on account of its significant influence on the attrition rate of candidates. Recently, constant upgrading of information technology has greatly promoted the continuous development of toxicity prediction. To give an overview of the current state of data-driven toxicity prediction, we reviewed relevant studies and summarized them in three main respects: the features and difficulties of toxicity prediction, the evolution of modeling approaches, and the available tools for toxicity prediction. For each part, we expound the research status, existing challenges, and feasible solutions. Finally, several new directions and suggestions for toxicity prediction are also put forward.

Attentive graph neural network models for the prediction of blood brain barrier permeability

The blood brain barrier's (BBB) unique endothelial cells and tight junctions selectively regulate passage of molecules to the central nervous system (CNS) to prevent pathogen entry and maintain neural homeostasis. Various neurological conditions and neurodegenerative diseases benefit from small molecules capable of BBB penetration (BBBP) to elicit a therapeutic effect. Predicting BBBP often involves in silico assessment of molecular properties such as lipophilicity (log P ) and polar surface area (PSA) using the CNS multiparameter optimization (MPO) method. This study curated an open-source dataset to benchmark rigorously machine learning (ML) and neural network (NN) models with each other and with MPO for predicting BBBP. Our analysis demonstrated that AI models, especially attentive NNs using stereochemical features, significantly outperform MPO in predicting BBBP. An attentive graph neural network (GNN), we refer to as CANDID-CNS™, achieved a 0.23-0.26 higher AUROC score than MPO on full test sets, and a 0.17-0.19 higher score on stereoisomers filtered subsets. Regarding stereoisomers that differ in BBBP, which MPO cannot distinguish, attentive GNNs correctly classify these with AUROC and MCC metrics comparable to or better than MPO's AUROC and MCC on less difficult test molecules. These findings suggest that integrating attentive GNN models into pharmaceutical drug discovery processes can substantially improve prediction rates, and thereby reduce the timeline, cost, and increase probability of success of designing brain penetrant therapeutics for the treatment of a wide variety of neurological and neurodegenerative diseases.

Multi-task pretrained language model with novel application domains enables more comprehensive health and ecological toxicity prediction

In silico models for screening substances of healthy and ecological concern are essential for effective chemical management. However, current data-driven toxicity prediction models confront formidable challenges related to expressive capacity, data scarcity, and reliability issues. Thus, this study introduces TOX-BERT, a SMILES-based pretrained model for screening health and ecological toxicity. Results show that masked atom recovery pretraining and multi-task learning offer promising solutions to enhance model capacity and address data scarcity issues. Two novel application domain (AD) parameters, termed PCA-AD and LDS, were proposed to improve prediction reliability of TOX-BERT with accuracy surpassing 90 % and mean absolute error (MAE) below 0.52. TOX-BERT was applied to 18,905 IECSC chemicals, revealing distinct toxicity relationships that align with experimental studies such as those between cardiotoxicity and acute ecotoxicity. In addition to previous PBT screening, 156 potential high-risk chemicals for specific endpoint were identified covering 7 categories. Furthermore, a SMILES-based toxicity site detection approach was developed for structural toxicity analysis. These advancements carry profound implications to address challenges faced by current data-driven toxicity prediction models. TOX-BERT emerges as a valuable tool for more comprehensive, reliable, and applicable predictions of health and ecological toxicity in chemical risk assessment and management.

Integrated Transfer Learning and Multitask Learning Strategies to Construct Graph Neural Network Models for Predicting Bioaccumulation Parameters of Chemicals

Accurate prediction of parameters related to the environmental exposure of chemicals is crucial for the sound management of chemicals. However, the lack of large data sets for training models may result in poor prediction accuracy and robustness. Herein, integrated transfer learning (TL) and multitask learning (MTL) was proposed for constructing a graph neural network (GNN) model (abbreviated as TL-MTL-GNN model) using n-octanol/water partition coefficients as a source domain. The TL-MTL-GNN model was trained to predict three bioaccumulation parameters based on enlarged data sets that cover 2496 compounds with at least one bioaccumulation parameter. Results show that the TL-MTL-GNN model outperformed single-task GNN models with and without the TL, as well as conventional machine learning models trained with molecular descriptors or fingerprints. Applicability domains were characterized by a state-of-the-art structure-activity landscape-based (abbreviated as ADSAL) methodology. The TL-MTL-GNN model coupled with the optimal ADSAL was employed to predict bioaccumulation parameters for around 60,000 chemicals, with more than 13,000 compounds identified as bioaccumulative chemicals. The high predictive accuracy and robustness of the TL-MTL-GNN model demonstrate the feasibility of integrating the TL and MTL strategy in modeling small-sized data sets. The strategy holds significant potential for addressing small data challenges in modeling environmental chemicals.

MultiCBlo: Enhancing predictions of compound-induced inhibition of cardiac ion channels with advanced multimodal learning

Predicting compound-induced inhibition of cardiac ion channels is crucial and challenging, significantly impacting cardiac drug efficacy and safety assessments. Despite the development of various computational methods for compound-induced inhibition prediction in cardiac ion channels, their performance remains limited. Most methods struggle to fuse multi-source data, relying solely on specific dataset training, leading to poor accuracy and generalization. We introduce MultiCBlo, a model that fuses multimodal information through a progressive learning approach, designed to predict compound-induced inhibition of cardiac ion channels with high accuracy. MultiCBlo employs progressive multimodal information fusion technology to integrate the compound's SMILES sequence, graph structure, and fingerprint, enhancing its representation. This is the first application of progressive multimodal learning for predicting compound-induced inhibition of cardiac ion channels, to our knowledge. The objective of this study was to predict the compound-induced inhibition of three major cardiac ion channels: hERG, Cav1.2, and Nav1.5. The results indicate that MultiCBlo significantly outperforms current models in predicting compound-induced inhibition of cardiac ion channels. We hope that MultiCBlo will facilitate cardiac drug development and reduce compound toxicity risks. Code and data are accessible at: https://github.com/taowang11/MultiCBlo. The online prediction platform is freely accessible at: https://huggingface.co/spaces/wtttt/PCICB.

Deep active learning with high structural discriminability for molecular mutagenicity prediction

The assessment of mutagenicity is essential in drug discovery, as it may lead to cancer and germ cells damage. Although in silico methods have been proposed for mutagenicity prediction, their performance is hindered by the scarcity of labeled molecules. However, experimental mutagenicity testing can be time-consuming and costly. One solution to reduce the annotation cost is active learning, where the algorithm actively selects the most valuable molecules from a vast chemical space and presents them to the oracle (e.g., a human expert) for annotation, thereby rapidly improving the model's predictive performance with a smaller annotation cost. In this paper, we propose muTOX-AL, a deep active learning framework, which can actively explore the chemical space and identify the most valuable molecules, resulting in competitive performance with a small number of labeled samples. The experimental results show that, compared to the random sampling strategy, muTOX-AL can reduce the number of training molecules by about 57%. Additionally, muTOX-AL exhibits outstanding molecular structural discriminability, allowing it to pick molecules with high structural similarity but opposite properties.

DeepRA: A novel deep learning-read-across framework and its application in non-sugar sweeteners mutagenicity prediction

Non-sugar sweeteners (NSSs) or artificial sweeteners have long been used as food chemicals since World War II. NSSs, however, also raise a concern about their mutagenicity. Evaluating the mutagenic ability of NSSs is crucial for food safety; this step is needed for every new chemical registration in the food and pharmaceutical industries. A computational assessment provides less time, money, and involved animals than the in vivo experiments; thus, this study developed a novel computational method from an ensemble convolutional deep neural network and read-across algorithms, called DeepRA, to classify the mutagenicity of chemicals. The mutagenicity data were obtained from the curated Ames test data set. The DeepRA model was developed using both molecular descriptors and molecular fingerprints. The obtained DeepRA model provides accurate and reliable mutagenicity classification through an independent test set. This model was then used to examine the NSSs-related chemicals, enabling the evaluation of mutagenicity from the NSSs-like substances. Finally, this model was publicly available at https://github.com/taraponglab/deepra for further use in chemical regulation and risk assessment.

DrugFlow: An AI-Driven One-Stop Platform for Innovative Drug Discovery

Artificial intelligence (AI)-aided drug design has demonstrated unprecedented effects on modern drug discovery, but there is still an urgent need for user-friendly interfaces that bridge the gap between these sophisticated tools and scientists, particularly those who are less computer savvy. Herein, we present DrugFlow, an AI-driven one-stop platform that offers a clean, convenient, and cloud-based interface to streamline early drug discovery workflows. By seamlessly integrating a range of innovative AI algorithms, covering molecular docking, quantitative structure-activity relationship modeling, molecular generation, ADMET (absorption, distribution, metabolism, excretion and toxicity) prediction, and virtual screening, DrugFlow can offer effective AI solutions for almost all crucial stages in early drug discovery, including hit identification and hit/lead optimization. We hope that the platform can provide sufficiently valuable guidance to aid real-word drug design and discovery. The platform is available at https://drugflow.com.

ChemFH: an integrated tool for screening frequent false positives in chemical biology and drug discovery

High-throughput screening rapidly tests an extensive array of chemical compounds to identify hit compounds for specific biological targets in drug discovery. However, false-positive results disrupt hit compound screening, leading to wastage of time and resources. To address this, we propose ChemFH, an integrated online platform facilitating rapid virtual evaluation of potential false positives, including colloidal aggregators, spectroscopic interference compounds, firefly luciferase inhibitors, chemical reactive compounds, promiscuous compounds, and other assay interferences. By leveraging a dataset containing 823 391 compounds, we constructed high-quality prediction models using multi-task directed message-passing network (DMPNN) architectures combining uncertainty estimation, yielding an average AUC value of 0.91. Furthermore, ChemFH incorporated 1441 representative alert substructures derived from the collected data and ten commonly used frequent hitter screening rules. ChemFH was validated with an external set of 75 compounds. Subsequently, the virtual screening capability of ChemFH was successfully confirmed through its application to five virtual screening libraries. Furthermore, ChemFH underwent additional validation on two natural products and FDA-approved drugs, yielding reliable and accurate results. ChemFH is a comprehensive, reliable, and computationally efficient screening pipeline that facilitates the identification of true positive results in assays, contributing to enhanced efficiency and success rates in drug discovery. ChemFH is freely available via https://chemfh.scbdd.com/.

ToxMPNN: A deep learning model for small molecule toxicity prediction

Machine learning (ML) has shown a great promise in predicting toxicity of small molecules. However, the availability of data for such predictions is often limited. Because of the unsatisfactory performance of models trained on a single toxicity endpoint, we collected toxic small molecules with multiple toxicity endpoints from previous study. The dataset comprises 27 toxic endpoints categorized into seven toxicity classes, namely, carcinogenicity and mutagenicity, acute oral toxicity, respiratory toxicity, irritation and corrosion, cardiotoxicity, CYP450, and endocrine disruption. In addition, a binary classification Common-Toxicity task was added based on the aforementioned dataset. To improve the performance of the models, we added marketed drugs as negative samples. This study presents a toxicity predictive model, ToxMPNN, based on the message passing neural network (MPNN) architecture, aiming to predict the toxicity of small molecules. The results demonstrate that ToxMPNN outperforms other models in capturing toxic features within the molecular structure, resulting in more precise predictions with the ROC_AUC testing score of 0.886 for the Toxicity_drug dataset. Furthermore, it was observed that adding marketed drugs as negative samples not only improves the predictive performance of the binary classification Common-Toxicity task but also enhances the stability of the model prediction. It shows that the graph-based deep learning (DL) algorithms in this study can be used as a trustworthy and effective tool to assess small molecule toxicity in the development of new drugs.

A review of deep learning methods for ligand based drug virtual screening

Drug discovery is costly and time consuming, and modern drug discovery endeavors are progressively reliant on computational methodologies, aiming to mitigate temporal and financial expenditures associated with the process. In particular, the time required for vaccine and drug discovery is prolonged during emergency situations such as the coronavirus 2019 pandemic. Recently, the performance of deep learning methods in drug virtual screening has been particularly prominent. It has become a concern for researchers how to summarize the existing deep learning in drug virtual screening, select different models for different drug screening problems, exploit the advantages of deep learning models, and further improve the capability of deep learning in drug virtual screening. This review first introduces the basic concepts of drug virtual screening, common datasets, and data representation methods. Then, large numbers of common deep learning methods for drug virtual screening are compared and analyzed. In addition, a dataset of different sizes is constructed independently to evaluate the performance of each deep learning model for the difficult problem of large-scale ligand virtual screening. Finally, the existing challenges and future directions in the field of virtual screening are presented.

Enhancing Molecular Property Prediction through Task-Oriented Transfer Learning: Integrating Universal Structural Insights and Domain-Specific Knowledge

Precisely predicting molecular properties is crucial in drug discovery, but the scarcity of labeled data poses a challenge for applying deep learning methods. While large-scale self-supervised pretraining has proven an effective solution, it often neglects domain-specific knowledge. To tackle this issue, we introduce Task-Oriented Multilevel Learning based on BERT (TOML-BERT), a dual-level pretraining framework that considers both structural patterns and domain knowledge of molecules. TOML-BERT achieved state-of-the-art prediction performance on 10 pharmaceutical datasets. It has the capability to mine contextual information within molecular structures and extract domain knowledge from massive pseudo-labeled data. The dual-level pretraining accomplished significant positive transfer, with its two components making complementary contributions. Interpretive analysis elucidated that the effectiveness of the dual-level pretraining lies in the prior learning of a task-related molecular representation. Overall, TOML-BERT demonstrates the potential of combining multiple pretraining tasks to extract task-oriented knowledge, advancing molecular property prediction in drug discovery.

Predicting the Activity of Unidentified Chemicals in Complementary Bioassays from the HRMS Data to Pinpoint Potential Endocrine Disruptors

The majority of chemicals detected via nontarget liquid chromatography high-resolution mass spectrometry (HRMS) in environmental samples remain unidentified, challenging the capability of existing machine learning models to pinpoint potential endocrine disruptors (EDs). Here, we predict the activity of unidentified chemicals across 12 bioassays related to EDs within the Tox21 10K dataset. Single- and multi-output models, utilizing various machine learning algorithms and molecular fingerprint features as an input, were trained for this purpose. To evaluate the models under near real-world conditions, Monte Carlo sampling was implemented for the first time. This technique enables the use of probabilistic fingerprint features derived from the experimental HRMS data with SIRIUS+CSI:FingerID as an input for models trained on true binary fingerprint features. Depending on the bioassay, the lowest false-positive rate at 90% recall ranged from 0.251 (sr.mmp, mitochondrial membrane potential) to 0.824 (nr.ar, androgen receptor), which is consistent with the trends observed in the models' performances submitted for the Tox21 Data Challenge. These findings underscore the informativeness of fingerprint features that can be compiled from HRMS in predicting the endocrine-disrupting activity. Moreover, an in-depth SHapley Additive exPlanations analysis unveiled the models' ability to pinpoint structural patterns linked to the modes of action of active chemicals. Despite the superior performance of the single-output models compared to that of the multi-output models, the latter's potential cannot be disregarded for similar tasks in the field of in silico toxicology. This study presents a significant advancement in identifying potentially toxic chemicals within complex mixtures without unambiguous identification and effectively reducing the workload for postprocessing by up to 75% in nontarget HRMS.

Expanding Predictive Capacities in Toxicology: Insights from Hackathon-Enhanced Data and Model Aggregation

In the realm of predictive toxicology for small molecules, the applicability domain of QSAR models is often limited by the coverage of the chemical space in the training set. Consequently, classical models fail to provide reliable predictions for wide classes of molecules. However, the emergence of innovative data collection methods such as intensive hackathons have promise to quickly expand the available chemical space for model construction. Combined with algorithmic refinement methods, these tools can address the challenges of toxicity prediction, enhancing both the robustness and applicability of the corresponding models. This study aimed to investigate the roles of gradient boosting and strategic data aggregation in enhancing the predictivity ability of models for the toxicity of small organic molecules. We focused on evaluating the impact of incorporating fragment features and expanding the chemical space, facilitated by a comprehensive dataset procured in an open hackathon. We used gradient boosting techniques, accounting for critical features such as the structural fragments or functional groups often associated with manifestations of toxicity.

CatNet: Sequence-based deep learning with cross-attention mechanism for identifying endocrine-disrupting chemicals

Endocrine-disrupting chemicals (EDCs) pose significant environmental and health risks due to their potential to interfere with nuclear receptors (NRs), key regulators of physiological processes. Despite the evident risks, the majority of existing research narrows its focus on the interaction between compounds and the individual NR target, neglecting a comprehensive assessment across the entire NR family. In response, this study assembled a comprehensive human NR dataset, capturing 49,244 interactions between 35,467 unique compounds and 42 NRs. We introduced a cross-attention network framework, "CatNet", innovatively integrating compound and protein representations through cross-attention mechanisms. The results showed that CatNet model achieved excellent performance with an area under the receiver operating characteristic curve (AUCROC) = 0.916 on the test set, and exhibited reliable generalization on unseen compound-NR pairs. A distinguishing feature of our research is its capacity to expand to novel targets. Beyond its predictive accuracy, CatNet offers a valuable mechanistic perspective on compound-NR interactions through feature visualization. Augmenting the utility of our research, we have also developed a graphical user interface, empowering researchers to predict chemical binding to diverse NRs. Our model enables the prediction of human NR-related EDCs and shows the potential to identify EDCs related to other targets.

Genetic Algorithm-Based Receptor Ligand: A Genetic Algorithm-Guided Generative Model to Boost the Novelty and Drug-Likeness of Molecules in a Sampling Chemical Space

Deep learning-based de novo molecular design has recently gained significant attention. While numerous DL-based generative models have been successfully developed for designing novel compounds, the majority of the generated molecules lack sufficiently novel scaffolds or high drug-like profiles. The aforementioned issues may not be fully captured by commonly used metrics for the assessment of molecular generative models, such as novelty, diversity, and quantitative estimation of the drug-likeness score. To address these limitations, we proposed a genetic algorithm-guided generative model called GARel (genetic algorithm-based receptor-ligand interaction generator), a novel framework for training a DL-based generative model to produce drug-like molecules with novel scaffolds. To efficiently train the GARel model, we utilized dense net to update the parameters based on molecules with novel scaffolds and drug-like features. To demonstrate the capability of the GARel model, we used it to design inhibitors for three targets: AA2AR, EGFR, and SARS-Cov2. The results indicate that GARel-generated molecules feature more diverse and novel scaffolds and possess more desirable physicochemical properties and favorable docking scores. Compared with other generative models, GARel makes significant progress in balancing novelty and drug-likeness, providing a promising direction for the further development of DL-based de novo design methodology with potential impacts on drug discovery.

AttentionMGT-DTA: A multi-modal drug-target affinity prediction using graph transformer and attention mechanism

The accurate prediction of drug-target affinity (DTA) is a crucial step in drug discovery and design. Traditional experiments are very expensive and time-consuming. Recently, deep learning methods have achieved notable performance improvements in DTA prediction. However, one challenge for deep learning-based models is appropriate and accurate representations of drugs and targets, especially the lack of effective exploration of target representations. Another challenge is how to comprehensively capture the interaction information between different instances, which is also important for predicting DTA. In this study, we propose AttentionMGT-DTA, a multi-modal attention-based model for DTA prediction. AttentionMGT-DTA represents drugs and targets by a molecular graph and binding pocket graph, respectively. Two attention mechanisms are adopted to integrate and interact information between different protein modalities and drug-target pairs. The experimental results showed that our proposed model outperformed state-of-the-art baselines on two benchmark datasets. In addition, AttentionMGT-DTA also had high interpretability by modeling the interaction strength between drug atoms and protein residues. Our code is available at https://github.com/JK-Liu7/AttentionMGT-DTA.

Predicting drug-induced liver injury using graph attention mechanism and molecular fingerprints

Drug-induced liver injury (DILI) is a significant issue in drug development and clinical treatment due to its potential to cause liver dysfunction or damage, which, in severe cases, can lead to liver failure or even fatality. DILI has numerous pathogenic factors, many of which remain incompletely understood. Consequently, it is imperative to devise methodologies and tools for anticipatory assessment of DILI risk in the initial phases of drug development. In this study, we present DMFPGA, a novel deep learning predictive model designed to predict DILI. To provide a comprehensive description of molecular properties, we employ a multi-head graph attention mechanism to extract features from the molecular graphs, representing characteristics at the level of compound nodes. Additionally, we combine multiple fingerprints of molecules to capture features at the molecular level of compounds. The fusion of molecular fingerprints and graph features can more fully express the properties of compounds. Subsequently, we employ a fully connected neural network to classify compounds as either DILI-positive or DILI-negative. To rigorously evaluate DMFPGA's performance, we conduct a 5-fold cross-validation experiment. The obtained results demonstrate the superiority of our method over four existing state-of-the-art computational approaches, exhibiting an average AUC of 0.935 and an average ACC of 0.934. We believe that DMFPGA is helpful for early-stage DILI prediction and assessment in drug development.

EDC-Predictor: A Novel Strategy for Prediction of Endocrine-Disrupting Chemicals by Integrating Pharmacological and Toxicological Profiles

Identification of endocrine-disrupting chemicals (EDCs) is crucial in the reduction of human health risks. However, it is hard to do so because of the complex mechanisms of the EDCs. In this study, we propose a novel strategy named EDC-Predictor to integrate pharmacological and toxicological profiles for the prediction of EDCs. Different from conventional methods that only focus on a few nuclear receptors (NRs), EDC-Predictor considers more targets. It uses computational target profiles from network-based and machine learning-based methods to characterize compounds, including both EDCs and non-EDCs. The best model constructed by these target profiles outperformed those models by molecular fingerprints. In a case study to predict NR-related EDCs, EDC-Predictor showed a wider applicability domain and higher accuracy than four previous tools. Another case study further demonstrated that EDC-Predictor could predict EDCs targeting other proteins rather than NRs. Finally, a free web server was developed to make EDC prediction easier (http://lmmd.ecust.edu.cn/edcpred/). In summary, EDC-Predictor would be a powerful tool in EDC prediction and drug safety assessment.

ADMET property prediction via multi-task graph learning under adaptive auxiliary task selection

It is a critical step in lead optimization to evaluate the absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties of drug-like compounds. Classical single-task learning (STL) has effectively predicted individual ADMET endpoints with abundant labels. Conversely, multi-task learning (MTL) can predict multiple ADMET endpoints with fewer labels, but ensuring task synergy and highlighting key molecular substructures remain challenges. To tackle these issues, this work elaborates a multi-task graph learning framework for predicting multiple ADMET properties of drug-like small molecules (MTGL-ADMET) by holding a new paradigm of MTL, "one primary, multiple auxiliaries." It first adeptly combines status theory with maximum flow for auxiliary task selection. The subsequent phase introduces a primary-task-centric MTL model with integrated modules. MTGL-ADMET not only outstrips existing STL and MTL methods but also offers a transparent lens into crucial molecular substructures. It is anticipated that this work can promote lead compound finding and optimization in drug discovery.

Chemistry-intuitive explanation of graph neural networks for molecular property prediction with substructure masking

Graph neural networks (GNNs) have been widely used in molecular property prediction, but explaining their black-box predictions is still a challenge. Most existing explanation methods for GNNs in chemistry focus on attributing model predictions to individual nodes, edges or fragments that are not necessarily derived from a chemically meaningful segmentation of molecules. To address this challenge, we propose a method named substructure mask explanation (SME). SME is based on well-established molecular segmentation methods and provides an interpretation that aligns with the understanding of chemists. We apply SME to elucidate how GNNs learn to predict aqueous solubility, genotoxicity, cardiotoxicity and blood-brain barrier permeation for small molecules. SME provides interpretation that is consistent with the understanding of chemists, alerts them to unreliable performance, and guides them in structural optimization for target properties. Hence, we believe that SME empowers chemists to confidently mine structure-activity relationship (SAR) from reliable GNNs through a transparent inspection on how GNNs pick up useful signals when learning from data.

Learning with uncertainty to accelerate the discovery of histone lysine-specific demethylase 1A (KDM1A/LSD1) inhibitors

Machine learning including modern deep learning models has been extensively used in drug design and screening. However, reliable prediction of molecular properties is still challenging when exploring out-of-domain regimes, even for deep neural networks. Therefore, it is important to understand the uncertainty of model predictions, especially when the predictions are used to guide further experiments. In this study, we explored the utility and effectiveness of evidential uncertainty in compound screening. The evidential Graphormer model was proposed for uncertainty-guided discovery of KDM1A/LSD1 inhibitors. The benchmarking results illustrated that (i) Graphormer exhibited comparative predictive power to state-of-the-art models, and (ii) evidential regression enabled well-ranked uncertainty estimates and calibrated predictions. Subsequently, we leveraged time-splitting on the curated KDM1A/LSD1 dataset to simulate out-of-distribution predictions. The retrospective virtual screening showed that the evidential uncertainties helped reduce false positives among the top-acquired compounds and thus enabled higher experimental validation rates. The trained model was then used to virtually screen an independent in-house compound set. The top 50 compounds ranked by two different ranking strategies were experimentally validated, respectively. In general, our study highlighted the importance to understand the uncertainty in prediction, which can be recognized as an interpretable dimension to model predictions.

TOXRIC: a comprehensive database of toxicological data and benchmarks

The toxic effects of compounds on environment, humans, and other organisms have been a major focus of many research areas, including drug discovery and ecological research. Identifying the potential toxicity in the early stage of compound/drug discovery is critical. The rapid development of computational methods for evaluating various toxicity categories has increased the need for comprehensive and system-level collection of toxicological data, associated attributes, and benchmarks. To contribute toward this goal, we proposed TOXRIC (https://toxric.bioinforai.tech/), a database with comprehensive toxicological data, standardized attribute data, practical benchmarks, informative visualization of molecular representations, and an intuitive function interface. The data stored in TOXRIC contains 113 372 compounds, 13 toxicity categories, 1474 toxicity endpoints covering in vivo/in vitro endpoints and 39 feature types, covering structural, target, transcriptome, metabolic data, and other descriptors. All the curated datasets of endpoints and features can be retrieved, downloaded and directly used as output or input to Machine Learning (ML)-based prediction models. In addition to serving as a data repository, TOXRIC also provides visualization of benchmarks and molecular representations for all endpoint datasets. Based on these results, researchers can better understand and select optimal feature types, molecular representations, and baseline algorithms for each endpoint prediction task. We believe that the rich information on compound toxicology, ML-ready datasets, benchmarks and molecular representation distribution can greatly facilitate toxicological investigations, interpretation of toxicological mechanisms, compound/drug discovery and the development of computational methods.

Kinome-wide polypharmacology profiling of small molecules by multi-task graph isomorphism network approach

Prediction of the interactions between small molecules and their targets play important roles in various applications of drug development, such as lead discovery, drug repurposing and elucidation of potential drug side effects. Therefore, a variety of machine learning-based models have been developed to predict these interactions. In this study, a model called auxiliary multi-task graph isomorphism network with uncertainty weighting (AMGU) was developed to predict the inhibitory activities of small molecules against 204 different kinases based on the multi-task Graph Isomorphism Network (MT-GIN) with the auxiliary learning and uncertainty weighting strategy. The calculation results illustrate that the AMGU model outperformed the descriptor-based models and state-of-the-art graph neural networks (GNN) models on the internal test set. Furthermore, it also exhibited much better performance on two external test sets, suggesting that the AMGU model has enhanced generalizability due to its great transfer learning capacity. Then, a naïve model-agnostic interpretable method for GNN called edges masking was devised to explain the underlying predictive mechanisms, and the consistency of the interpretability results for 5 typical epidermal growth factor receptor (EGFR) inhibitors with their structure‒activity relationships could be observed. Finally, a free online web server called KIP was developed to predict the kinome-wide polypharmacology effects of small molecules (http://cadd.zju.edu.cn/kip).

Revealing the impact of lifestyle stressors on the risk of adverse pregnancy outcomes with multitask machine learning

Psychosocial and stress-related factors (PSFs), defined as internal or external stimuli that induce biological changes, are potentially modifiable factors and accessible targets for interventions that are associated with adverse pregnancy outcomes (APOs). Although individual APOs have been shown to be connected to PSFs, they are biologically interconnected, relatively infrequent, and therefore challenging to model. In this context, multi-task machine learning (MML) is an ideal tool for exploring the interconnectedness of APOs on the one hand and building on joint combinatorial outcomes to increase predictive power on the other hand. Additionally, by integrating single cell immunological profiling of underlying biological processes, the effects of stress-based therapeutics may be measurable, facilitating the development of precision medicine approaches.

Objectives

The primary objectives were to jointly model multiple APOs and their connection to stress early in pregnancy, and to explore the underlying biology to guide development of accessible and measurable interventions.

Materials and Methods

In a prospective cohort study, PSFs were assessed during the first trimester with an extensive self-filled questionnaire for 200 women. We used MML to simultaneously model, and predict APOs (severe preeclampsia, superimposed preeclampsia, gestational diabetes and early gestational age) as well as several risk factors (BMI, diabetes, hypertension) for these patients based on PSFs. Strongly interrelated stressors were categorized to identify potential therapeutic targets. Furthermore, for a subset of 14 women, we modeled the connection of PSFs to the maternal immune system to APOs by building corresponding ML models based on an extensive single cell immune dataset generated by mass cytometry time of flight (CyTOF).

Results

Jointly modeling APOs in a MML setting significantly increased modeling capabilities and yielded a highly predictive integrated model of APOs underscoring their interconnectedness. Most APOs were associated with mental health, life stress, and perceived health risks. Biologically, stressors were associated with specific immune characteristics revolving around CD4/CD8 T cells. Immune characteristics predicted based on stress were in turn found to be associated with APOs.

Conclusions

Elucidating connections among stress, multiple APOs simultaneously, and immune characteristics has the potential to facilitate the implementation of ML-based, individualized, integrative models of pregnancy in clinical decision making. The modifiable nature of stressors may enable the development of accessible interventions, with success tracked through immune characteristics.

ALipSol: An Attention-Driven Mixture-of-Experts Model for Lipophilicity and Solubility Prediction

Lipophilicity (logD) and aqueous solubility (logS_w) play a central role in drug development. The accurate prediction of these properties remains to be solved due to data scarcity. Current methodologies neglect the intrinsic relationships between physicochemical properties and usually ignore the ionization effects. Here, we propose an attention-driven mixture-of-experts (MoE) model named ALipSol, which explicitly reproduces the hierarchy of task relationships. We adopt the principle of divide-and-conquer by breaking down the complex end point (logD or logS_w) into simpler ones (acidic pK_a, basic pK_a, and logP) and allocating a specific expert network for each subproblem. Subsequently, we implement transfer learning to extract knowledge from related tasks, thus alleviating the dilemma of limited data. Additionally, we substitute the gating network with an attention mechanism to better capture the dynamic task relationships on a per-example basis. We adopt local fine-tuning and consensus prediction to further boost model performance. Extensive evaluation experiments verify the success of the ALipSol model, which achieves RMSE improvement of 8.04%, 2.49%, 8.57%, 12.8%, and 8.60% on the Lipop, ESOL, AqSolDB, external logD, and external logS data sets, respectively, compared with Attentive FP and the state-of-the-art in silico tools. In particular, our model yields more significant advantages (Welch's t-test) for small training data, implying its high robustness and generalizability. The interpretability analysis proves that the atom contributions learned by ALipSol are more reasonable compared with the vanilla Attentive FP, and the substitution effects in benzene derivatives agreed well with empirical constants, revealing the potential of our model to extract useful patterns from data and provide guidance for lead optimization.

GSAML-DTA: An interpretable drug-target binding affinity prediction model based on graph neural networks with self-attention mechanism and mutual information

Identifying drug-target affinity (DTA) has great practical importance in the process of designing efficacious drugs for known diseases. Recently, numerous deep learning-based computational methods have been developed to predict drug-target affinity and achieved impressive performance. However, most of them construct the molecule (drug or target) encoder without considering the weights of features of each node (atom or residue). Besides, they generally combine drug and target representations directly, which may contain irrelevant-task information. In this study, we develop GSAML-DTA, an interpretable deep learning framework for DTA prediction. GSAML-DTA integrates a self-attention mechanism and graph neural networks (GNNs) to build representations of drugs and target proteins from the structural information. In addition, mutual information is introduced to filter out redundant information and retain relevant information in the combined representations of drugs and targets. Extensive experimental results demonstrate that GSAML-DTA outperforms state-of-the-art methods for DTA prediction on two benchmark datasets. Furthermore, GSAML-DTA has the interpretation ability to analyze binding atoms and residues, which may be conducive to chemical biology studies from data. Overall, GSAML-DTA can serve as a powerful and interpretable tool suitable for DTA modelling.

Deep learning methods for molecular representation and property prediction

With advances in artificial intelligence (AI) methods, computer-aided drug design (CADD) has developed rapidly in recent years. Effective molecular representation and accurate property prediction are crucial tasks in CADD workflows. In this review, we summarize contemporary applications of deep learning (DL) methods for molecular representation and property prediction. We categorize DL methods according to the format of molecular data (1D, 2D, and 3D). In addition, we discuss some common DL models, such as ensemble learning and transfer learning, and analyze the interpretability methods for these models. We also highlight the challenges and opportunities of DL methods for molecular representation and property prediction.

Profiling prediction of nuclear receptor modulators with multi-task deep learning methods: toward the virtual screening

Nuclear receptors (NRs) are ligand-activated transcription factors, which constitute one of the most important targets for drug discovery. Current computational strategies mainly focus on a single target, and the transfer of learned knowledge among NRs was not considered yet. Herein we proposed a novel computational framework named NR-Profiler for prediction of potential NR modulators with high affinity and specificity. First, we built a comprehensive NR data set including 42 684 interactions to connect 42 NRs and 31 033 compounds. Then, we used multi-task deep neural network and multi-task graph convolutional neural network architectures to construct multi-task multi-classification models. To improve the predictive capability and robustness, we built a consensus model with an area under the receiver operating characteristic curve (AUC) = 0.883. Compared with conventional machine learning and structure-based approaches, the consensus model showed better performance in external validation. Using this consensus model, we demonstrated the practical value of NR-Profiler in virtual screening for NRs. In addition, we designed a selectivity score to quantitatively measure the specificity of NR modulators. Finally, we developed a freely available standalone software for users to make profiling predictions for their compounds of interest. In summary, our NR-Profiler provides a useful tool for NR-profiling prediction and is expected to facilitate NR-based drug discovery.

Boosting Protein-Ligand Binding Pose Prediction and Virtual Screening Based on Residue-Atom Distance Likelihood Potential and Graph Transformer

The past few years have witnessed enormous progress toward applying machine learning approaches to the development of protein-ligand scoring functions. However, the robust performance and wide applicability of scoring functions remain a big challenge for increasing the success rate of docking-based virtual screening. Herein, a novel scoring function named RTMScore was developed by introducing a tailored residue-based graph representation strategy and several graph transformer layers for the learning of protein and ligand representations, followed by a mixture density network to obtain residue-atom distance likelihood potential. Our approach was resolutely validated on the CASF-2016 benchmark, and the results indicate that RTMScore can outperform almost all of the other state-of-the-art methods in terms of both the docking and screening powers. Further evaluation confirms the robustness of our approach that can not only retain its docking power on cross-docked poses but also achieve improved performance as a rescoring tool in larger-scale virtual screening.

Distribution- and Metabolism-Based Drug Discovery: A Potassium-Competitive Acid Blocker as a Proof of Concept

Conventional methods of drug design require compromise in the form of side effects to achieve sufficient efficacy because targeting drugs to specific organs remains challenging. Thus, new strategies to design organ-specific drugs that induce little toxicity are needed. Based on characteristic tissue niche-mediated drug distribution (TNMDD) and patterns of drug metabolism into specific intermediates, we propose a strategy of distribution- and metabolism-based drug design (DMBDD); through a physicochemical property-driven distribution optimization cooperated with a well-designed metabolism pathway, SH-337, a candidate potassium-competitive acid blocker (P-CAB), was designed. SH-337 showed specific distribution in the stomach in the long term and was rapidly cleared from the systemic compartment. Therefore, SH-337 exerted a comparable pharmacological effect but a 3.3-fold higher no observed adverse effect level (NOAEL) compared with FDA-approved vonoprazan. This study contributes a proof-of-concept demonstration of DMBDD and provides a new perspective for the development of highly efficient, organ-specific drugs with low toxicity.

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.

Exploring Low-Toxicity Chemical Space with Deep Learning for Molecular Generation

Creating a wide range of new compounds that not only have ideal pharmacological properties but also easily pass long-term toxicity evaluation is still a challenging task in current drug discovery. In this study, we developed a conditional generative model by combining a semisupervised variational autoencoder (SSVAE) with an MGA toxicity predictor. Our aim is to generate molecules with low toxicity, good drug-like properties, and structural diversity. For multiobjective optimization, we have developed a method with hierarchical constraints on the toxicity space of small molecules to generate drug-like small molecules, which can also minimize the effect on the diversity of generated results. The evaluation results of the metrics indicate that the developed model has good effectiveness, novelty, and diversity. The generated molecules by this model are mainly distributed in low-toxicity regions, which suggests that our model can efficiently constrain the generation of toxic structures. In contrast to simply filtering toxic ones after generation, the low-toxicity molecular generative model can generate molecules with structural diversity. Our strategy can be used in target-based drug discovery to improve the quality of generated molecules with low-toxicity, drug-like, and highly active properties.

RELATION: A Deep Generative Model for Structure-Based De Novo Drug Design

Deep learning (DL)-based de novo molecular design has recently gained considerable traction. Many DL-based generative models have been successfully developed to design novel molecules, but most of them are ligand-centric and the role of the 3D geometries of target binding pockets in molecular generation has not been well-exploited. Here, we proposed a new 3D-based generative model called RELATION. In the RELATION model, the BiTL algorithm was specifically designed to extract and transfer the desired geometric features of the protein-ligand complexes to a latent space for generation. The pharmacophore conditioning and docking-based Bayesian sampling were applied to efficiently navigate the vast chemical space for the design of molecules with desired geometric properties and pharmacophore features. As a proof of concept, the RELATION model was used to design inhibitors for two targets, AKT1 and CDK2. The calculation results demonstrated that the RELATION model could efficiently generate novel molecules with favorable binding affinity and pharmacophore features.

BayeshERG: a robust, reliable and interpretable deep learning model for predicting hERG channel blockers

Unintended inhibition of the human ether-à-go-go-related gene (hERG) ion channel by small molecules leads to severe cardiotoxicity. Thus, hERG channel blockage is a significant concern in the development of new drugs. Several computational models have been developed to predict hERG channel blockage, including deep learning models; however, they lack robustness, reliability and interpretability. Here, we developed a graph-based Bayesian deep learning model for hERG channel blocker prediction, named BayeshERG, which has robust predictive power, high reliability and high resolution of interpretability. First, we applied transfer learning with 300 000 large data in initial pre-training to increase the predictive performance. Second, we implemented a Bayesian neural network with Monte Carlo dropout to calibrate the uncertainty of the prediction. Third, we utilized global multihead attentive pooling to augment the high resolution of structural interpretability for the hERG channel blockers and nonblockers. We conducted both internal and external validations for stringent evaluation; in particular, we benchmarked most of the publicly available hERG channel blocker prediction models. We showed that our proposed model outperformed predictive performance and uncertainty calibration performance. Furthermore, we found that our model learned to focus on the essential substructures of hERG channel blockers via an attention mechanism. Finally, we validated the prediction results of our model by conducting in vitro experiments and confirmed its high validity. In summary, BayeshERG could serve as a versatile tool for discovering hERG channel blockers and helping maximize the possibility of successful drug discovery. The data and source code are available at our GitHub repository (https://github.com/GIST-CSBL/BayeshERG).

In Silico Prediction of Chemical Aquatic Toxicity by Multiple Machine Learning and Deep Learning Approaches

Fish is one of the model animals used to evaluate the adverse effects of a chemical exposed to the ecosystem. However, its low throughput and relevantly high expense make it impossible to test all new chemicals in manufacture. Hence using in silico models to prioritize compounds to be tested has been widely applied in environmental risk assessment and drug discovery. In this study, we constructed the local predictive models for four fish species, including bluegill sunfish, rainbow trout, fathead minnow, and sheepshead minnow, and the global models with all four fish data. A total of 1874 unique compounds with their labels, i.e. toxic (LC₅₀ < 10 ppm) or nontoxic were collected from ECOTOX and literature. Both conventional machine learning methods and the deep learning architecture, graph convolutional network (GCN), were used to build predictive models. The classification accuracy of the best local model for each fish species was higher than 0.83. For the global models, two strategies including consistency prediction and probability threshold were adopted to improve the predictive capability at the cost of limiting applicability domain. For 63% of compounds in domain, the accuracy was around 0.97. By comparison of the deep learning and machine learning methods, we found that the single-task GCN showed specific advantages in performance and multi-task GCN showed no advantages over the conventional machine learning methods. The data and models are available on GitHub (https://github.com/ChemPredict/ChemicalAquaticToxicity).

IDL-PPBopt: A Strategy for Prediction and Optimization of Human Plasma Protein Binding of Compounds via an Interpretable Deep Learning Method

The prediction and optimization of pharmacokinetic properties are essential in lead optimization. Traditional strategies mainly depend on the empirical chemical rules from medicinal chemists. However, with the rising amount of data, it is getting more difficult to manually extract useful medicinal chemistry knowledge. To this end, we introduced IDL-PPBopt, a computational strategy for predicting and optimizing the plasma protein binding (PPB) property based on an interpretable deep learning method. At first, a curated PPB data set was used to construct an interpretable deep learning model, which showed excellent predictive performance with a root mean squared error of 0.112 for the entire test set. Then, we designed a detection protocol based on the model and Wilcoxon test to identify the PPB-related substructures (named privileged substructures, PSubs) for each molecule. In total, 22 general privileged substructures (GPSubs) were identified, which shared some common features such as nitrogen-containing groups, diamines with two carbon units, and azetidine. Furthermore, a series of second-level chemical rules for each GPSub were derived through a statistical test and then summarized into substructure pairs. We demonstrated that these substructure pairs were equally applicable outside the training set and accordingly customized the structural modification schemes for each GPSub, which provided alternatives for the optimization of the PPB property. Therefore, IDL-PPBopt provides a promising scheme for the prediction and optimization of the PPB property and would be helpful for lead optimization of other pharmacokinetic properties.

An Algorithm Framework for Drug-Induced Liver Injury Prediction Based on Genetic Algorithm and Ensemble Learning

In the process of drug discovery, drug-induced liver injury (DILI) is still an active research field and is one of the most common and important issues in toxicity evaluation research. It directly leads to the high wear attrition of the drug. At present, there are a variety of computer algorithms based on molecular representations to predict DILI. It is found that a single molecular representation method is insufficient to complete the task of toxicity prediction, and multiple molecular fingerprint fusion methods have been used as model input. In order to solve the problem of high dimensional and unbalanced DILI prediction data, this paper integrates existing datasets and designs a new algorithm framework, Rotation-Ensemble-GA (R-E-GA). The main idea is to find a feature subset with better predictive performance after rotating the fusion vector of high-dimensional molecular representation in the feature space. Then, an Adaboost-type ensemble learning method is integrated into R-E-GA to improve the prediction accuracy. The experimental results show that the performance of R-E-GA is better than other state-of-art algorithms including ensemble learning-based and graph neural network-based methods. Through five-fold cross-validation, the R-E-GA obtains an ACC of 0.77, an F1 score of 0.769, and an AUC of 0.842.

Knowledge-based BERT: a method to extract molecular features such as computational chemists

Molecular property prediction models based on machine learning algorithms have become important tools to triage unpromising lead molecules in the early stages of drug discovery. Compared with the mainstream descriptor- and graph-based methods for molecular property predictions, SMILES-based methods can directly extract molecular features from SMILES without human expert knowledge, but they require more powerful algorithms for feature extraction and a larger amount of data for training, which makes SMILES-based methods less popular. Here, we show the great potential of pre-training in promoting the predictions of important pharmaceutical properties. By utilizing three pre-training tasks based on atom feature prediction, molecular feature prediction and contrastive learning, a new pre-training method K-BERT, which can extract chemical information from SMILES like chemists, was developed. The calculation results on 15 pharmaceutical datasets show that K-BERT outperforms well-established descriptor-based (XGBoost) and graph-based (Attentive FP and HRGCN+) models. In addition, we found that the contrastive learning pre-training task enables K-BERT to 'understand' SMILES not limited to canonical SMILES. Moreover, the general fingerprints K-BERT-FP generated by K-BERT exhibit comparative predictive power to MACCS on 15 pharmaceutical datasets and can also capture molecular size and chirality information that traditional binary fingerprints cannot capture. Our results illustrate the great potential of K-BERT in the practical applications of molecular property predictions in drug discovery.