Search

Abstract Building reliable and robust quantitative structure–property relationship (QSPR) models is a challenging task. First, the experimental data needs to be obtained, analyzed and curated. Second, the number of available methods is continuously growing and evaluating different algorithms and methodologies can be arduous. Finally, the last hurdle that researchers face is to ensure the reproducibility of their models and facilitate their transferability into practice. In this work, we introduce QSPRpred, a toolkit for analysis of bioactivity data sets and QSPR modelling, which attempts to address the aforementioned challenges. QSPRpred’s modular Python API enables users to intuitively describe different parts of a modelling workflow using a plethora of pre-implemented components, but also integrates customized implementations in a “plug-and-play” manner. QSPRpred data sets and models are directly serializable, which means they can be readily reproduced and put into operation after training as the models are saved with all required data pre-processing steps to make predictions on new compounds directly from SMILES strings. The general-purpose character of QSPRpred is also demonstrated by inclusion of support for multi-task and proteochemometric modelling. The package is extensively documented and comes with a large collection of tutorials to help new users. In this paper, we describe all of QSPRpred’s functionalities and also conduct a small benchmarking case study to illustrate how different components can be leveraged to compare a diverse set of models. QSPRpred is fully open-source and available at https://github.com/CDDLeiden/QSPRpred . Scientific Contribution QSPRpred aims to provide a complex, but comprehensive Python API to conduct all tasks encountered in QSPR modelling from data preparation and analysis to model creation and model deployment. In contrast to similar packages, QSPRpred offers a wider and more exhaustive range of capabilities and integrations with many popular packages that also go beyond QSPR modelling. A significant contribution of QSPRpred is also in its automated and highly standardized serialization scheme, which significantly improves reproducibility and transferability of models.

Abstract The accurate identification of protein–ligand binding sites is of critical importance in understanding and modulating protein function. Accordingly, ligand binding site prediction has remained a research focus for over three decades with over 50 methods developed and a change of paradigm from geometry-based to machine learning. In this work, we collate 13 ligand binding site predictors, spanning 30 years, focusing on the latest machine learning-based methods such as VN-EGNN, IF-SitePred, GrASP, PUResNet, and DeepPocket and compare them to the established P2Rank, PRANK and fpocket and earlier methods like PocketFinder, Ligsite and Surfnet. We benchmark the methods against the human subset of our new curated reference dataset, LIGYSIS. LIGYSIS is a comprehensive protein–ligand complex dataset comprising 30,000 proteins with bound ligands which aggregates biologically relevant unique protein–ligand interfaces across biological units of multiple structures from the same protein. LIGYSIS is an improvement for testing methods over earlier datasets like sc-PDB, PDBbind, binding MOAD, COACH420 and HOLO4K which either include 1:1 protein–ligand complexes or consider asymmetric units. Re-scoring of fpocket predictions by PRANK and DeepPocket display the highest recall (60%) whilst IF-SitePred presents the lowest recall (39%). We demonstrate the detrimental effect that redundant prediction of binding sites has on performance as well as the beneficial impact of stronger pocket scoring schemes, with improvements up to 14% in recall (IF-SitePred) and 30% in precision (Surfnet). Finally, we propose top-N+2 recall as the universal benchmark metric for ligand binding site prediction and urge authors to share not only the source code of their methods, but also of their benchmark. Scientific contributions This study conducts the largest benchmark of ligand binding site prediction methods to date, comparing 13 original methods and 15 variants using 10 informative metrics. The LIGYSIS dataset is introduced, which aggregates biologically relevant protein–ligand interfaces across multiple structures of the same protein. The study highlights the detrimental effect of redundant binding site prediction and demonstrates significant improvement in recall and precision through stronger scoring schemes. Finally, top-N+2 recall is proposed as a universal benchmark metric for ligand binding site prediction, with a recommendation for open-source sharing of both methods and benchmarks.

Abstract Background Target identification and hit identification can be transformed through the application of biomedical knowledge analysis, AI-driven virtual screening and robotic cloud lab systems. However there are few prospective studies that evaluate the efficacy of such integrated approaches. Results We synergistically integrate our in-house-developed target evaluation (SpectraView) and deep-learning-driven virtual screening (HydraScreen) tools with an automated robotic cloud lab designed explicitly for ultra-high-throughput screening, enabling us to validate these platforms experimentally. By employing our target evaluation tool to select IRAK1 as the focal point of our investigation, we prospectively validate our structure-based deep learning model. We can identify 23.8% of all IRAK1 hits within the top 1% of ranked compounds. The model outperforms traditional virtual screening techniques and offers advanced features such as ligand pose confidence scoring. Simultaneously, we identify three potent (nanomolar) scaffolds from our compound library, 2 of which represent novel candidates for IRAK1 and hold promise for future development. Conclusion This study provides compelling evidence for SpectraView and HydraScreen to provide a significant acceleration in the processes of target identification and hit discovery. By leveraging Ro5’s HydraScreen and Strateos’ automated labs in hit identification for IRAK1, we show how AI-driven virtual screening with HydraScreen could offer high hit discovery rates and reduce experimental costs. Scientific contribution We present an innovative platform that leverages Knowledge graph-based biomedical data analytics and AI-driven virtual screening integrated with robotic cloud labs. Through an unbiased, prospective evaluation we show the reliability and robustness of HydraScreen in virtual and high-throughput screening for hit identification in IRAK1. Our platforms and innovative tools can expedite the early stages of drug discovery.

Abstract Predicting protein-small molecule binding sites, the initial step in structure-guided drug design, remains challenging for proteins lacking experimentally derived ligand-bound structures. Here, we propose CLAPE-SMB, which integrates a pre-trained protein language model with contrastive learning to provide high accuracy predictions of small molecule binding sites that can accommodate proteins without a published crystal structure. We trained and tested CLAPE-SMB on the SJC dataset, a non-redundant dataset based on sc-PDB, JOINED, and COACH420, and achieved an MCC of 0.529. We also compiled the UniProtSMB dataset, which merges sites from similar proteins based on raw data from UniProtKB database, and achieved an MCC of 0.699 on the test set. In addition, CLAPE-SMB achieved an MCC of 0.815 on our intrinsically disordered protein (IDP) dataset that contains 336 non-redundant sequences. Case studies of DAPK1, RebH, and Nep1 support the potential of this binding site prediction tool to aid in drug design. The code and datasets are freely available at https://github.com/JueWangTHU/CLAPE-SMB . Scientific contribution CLAPE-SMB combines a pre-trained protein language model with contrastive learning to accurately predict protein-small molecule binding sites, especially for proteins without experimental structures, such as IDPs. Trained across various datasets, this model shows strong adaptability, making it a valuable tool for advancing drug design and understanding protein-small molecule interactions.

Abstract We introduce an advanced model for predicting protein–ligand interactions. Our approach combines the strengths of graph neural networks with physics-based scoring methods. Existing structure-based machine-learning models for protein–ligand binding prediction often fall short in practical virtual screening scenarios, hindered by the intricacies of binding poses, the chemical diversity of drug-like molecules, and the scarcity of crystallographic data for protein–ligand complexes. To overcome the limitations of existing machine learning-based prediction models, we propose a novel approach that fuses three independent neural network models. One classification model is designed to perform binary prediction of a given protein–ligand complex pose. The other two regression models are trained to predict the binding affinity and root-mean-square deviation of a ligand conformation from an input complex structure. We trained the model to account for both deviations in experimental and predicted binding affinities and pose prediction uncertainties. By effectively integrating the outputs of the triplet neural networks with a physics-based scoring function, our model showed a significantly improved performance in hit identification. The benchmark results with three independent decoy sets demonstrate that our model outperformed existing models in forward screening. Our model achieved top 1% enrichment factors of 32.7 and 23.1 with the CASF2016 and DUD-E benchmark sets, respectively. The benchmark results using the LIT-PCBA set further confirmed its higher average enrichment factors, emphasizing the model’s efficiency and generalizability. The model’s efficiency was further validated by identifying 23 active compounds from 63 candidates in experimental screening for autotaxin inhibitors, demonstrating its practical applicability in hit discovery. Scientific contribution Our work introduces a novel training strategy for a protein–ligand binding affinity prediction model by integrating the outputs of three independent sub-models and utilizing expertly crafted decoy sets. The model showcases exceptional performance across multiple benchmarks. The high enrichment factors in the LIT-PCBA benchmark demonstrate its potential to accelerate hit discovery.

Abstract Molecular dynamics simulations serve as a prevalent approach for investigating the dynamic behaviour of proteins and protein–ligand complexes. Due to its versatility and speed, GROMACS stands out as a commonly utilized software platform for executing molecular dynamics simulations. However, its effective utilization requires substantial expertise in configuring, executing, and interpreting molecular dynamics trajectories. Existing automation tools are constrained in their capability to conduct simulations for large sets of compounds with minimal user intervention, or in their ability to distribute simulations across multiple servers. To address these challenges, we developed a Python-based tool that streamlines all phases of molecular dynamics simulations, encompassing preparation, execution, and analysis. This tool minimizes the required knowledge for users engaging in molecular dynamics simulations and can efficiently operate across multiple servers within a network or a cluster. Notably, the tool not only automates trajectory simulation but also facilitates the computation of free binding energies for protein–ligand complexes and generates interaction fingerprints across the trajectory. Our study demonstrated the applicability of this tool on several benchmark datasets. Additionally, we provided recommendations for end-users to effectively utilize the tool. Scientific contribution The developed tool, StreaMD, is applicable to different systems (proteins, ligands and their complexes including co-factors) and requires a little user knowledge to setup and run molecular dynamics simulations. Other features of StreaMD are seamless integration with calculation of MM-GBSA/PBSA binding free energies and protein-ligand interaction fingerprints, and running of simulations within distributed environments. All these will facilitate routine and massive molecular dynamics simulations.

Abstract The adverse outcome pathway (AOP) concept has gained attention as a way to explore the mechanism of chemical toxicity. In this study, quantitative structure–activity relationship (QSAR) models were developed to predict compound activity toward protein targets relevant to molecular initiating events (MIE) upstream of organ-specific toxicities, namely liver steatosis, cholestasis, nephrotoxicity, neural tube closure defects, and cognitive functional defects. Utilizing bioactivity data from the ChEMBL 33 database, various machine learning algorithms, chemical features and methods to assess prediction reliability were compared and applied to develop robust models to predict compound activity. The results demonstrate high predictive performance across multiple targets, with balanced accuracy exceeding 0.80 for the majority of models. Furthermore, stability checks confirmed the consistency of predictive performance across multiple training-test splits. The results obtained by using QSAR predictions to identify known markers of adversities highlighted the utility of the models for risk assessment and for prioritizing compounds for further experimental evaluation. Scientific contribution The work describes the development of QSAR models as tools for screening chemicals with potential systemic toxicity, thus contributing to resource savings and providing indications for further better-targeted testing. This study provides advances in the field of computational modeling of MIEs and information from AOP which is still relatively young and unexplored. The comprehensive modeling procedure is highly generalizable, and offers a robust framework for predicting a wide range of toxicological endpoints.

Abstract Over the past ~ 25 years, chemoinformatics has evolved as a scientific discipline, with a strong foundation in pharmaceutical research and scientific roots that can be traced back to the late 1950s. It covers a wide methodological spectrum and is perhaps best positioned in the greater context of chemical information science. Herein, the chemoinformatics discipline is delineated, characteristic (and partly problematic) features are discussed, and a global view of the field is provided, emphasizing key developments.

Abstract State‑of‑the‑art medical studies proved that predicting CYP450 enzyme inhibitors is beneficial in the early stage of drug discovery. However, accurate machine learning-based (ML) in silico methods for predicting CYP450 inhibitors remains challenging. Here, we introduce GTransCYPs, an improved graph neural network (GNN) with a transformer mechanism for predicting CYP450 inhibitors. This model significantly enhances the discrimination between inhibitors and non-inhibitors for five major CYP450 isozymes: 1A2, 2C9, 2C19, 2D6, and 3A4. GTransCYPs learns information patterns from molecular graphs by aggregating node and edge representations using a transformer. The GTransCYPs model utilizes transformer convolution layers to process features, followed by a global attention-pooling technique to synthesize the graph-level information. This information is then fed through successive linear layers for final output generation. Experimental results demonstrate that the GTransCYPs model achieved high performance, outperforming other state-of-the-art methods in CYP450 prediction. Scientific contribution The prediction of CYP450 inhibition via computational techniques utilizing biological information has emerged as a cost-effective and highly efficient approach. Here, we presented a deep learning (DL) architecture based on GNN with transformer mechanism and attention pooling (GTransCYPs) to predict CYP450 inhibitors. Four GTransCYPs of different pooling technique were tested on an experimental tasks on the CYP450 prediction problem for the first time. Graph transformer with attention pooling algorithm achieved the best performances. Comparative and ablation experiments provide evidence of the efficacy of our proposed method in predicting CYP450 inhibitors. The source code is publicly available at https://github.com/zonwoo/GTransCYPs .

Abstract The evaluation of compound-target interactions (CTIs) is at the heart of drug discovery efforts. Given the substantial time and monetary costs of classical experimental screening, significant efforts have been dedicated to develop deep learning-based models that can accurately predict CTIs. A comprehensive comparison of these models on a large, curated CTI dataset is, however, still lacking. Here, we perform an in-depth comparison of 12 state-of-the-art deep learning architectures that use different protein and compound representations. The models were selected for their reported performance and architectures. To reliably compare model performance, we curated over 300 thousand binding and non-binding CTIs and established several gold-standard datasets of varying size and information. Based on our findings, DeepConv-DTI consistently outperforms other models in CTI prediction performance across the majority of datasets. It achieves an MCC of 0.6 or higher for most of the datasets and is one of the fastest models in training and inference. These results indicate that utilizing convolutional-based windows as in DeepConv-DTI to traverse trainable embeddings is a highly effective approach for capturing informative protein features. We also observed that physicochemical embeddings of targets increased model performance. We therefore modified DeepConv-DTI to include normalized physicochemical properties, which resulted in the overall best performing model Phys-DeepConv-DTI. This work highlights how the systematic evaluation of input features of compounds and targets, as well as their corresponding neural network architectures, can serve as a roadmap for the future development of improved CTI models. Scientific contribution This work features comprehensive CTI datasets to allow for the objective comparison and benchmarking of CTI prediction algorithms. Based on this dataset, we gained insights into which embeddings of compounds and targets and which deep learning-based algorithms perform best, providing a blueprint for the future development of CTI algorithms. Using the insights gained from this screen, we provide a novel CTI algorithm with state-of-the-art performance.

Abstract Drug solubility is an important parameter in the drug development process, yet it is often tedious and challenging to measure, especially for expensive drugs or those available in small quantities. To alleviate these challenges, machine learning (ML) has been applied to predict drug solubility as an alternative approach. However, the majority of existing ML research has focused on the predictions of aqueous solubility and/or solubility at specific temperatures, which restricts the model applicability in pharmaceutical development. To bridge this gap, we compiled a dataset of 27,000 solubility datapoints, including solubility of small molecules measured in a range of binary solvent mixtures under various temperatures. Next, a panel of ML models were trained on this dataset with their hyperparameters tuned using Bayesian optimization. The resulting top-performing models, both gradient boosted decision trees (light gradient boosting machine and extreme gradient boosting), achieved mean absolute errors (MAE) of 0.33 for LogS (S in g/100 g) on the holdout set. These models were further validated through a prospective study, wherein the solubility of four drug molecules were predicted by the models and then validated with in-house solubility experiments. This prospective study demonstrated that the models accurately predicted the solubility of solutes in specific binary solvent mixtures under different temperatures, especially for drugs whose features closely align within the solutes in the dataset (MAE

Abstract A crucial mechanism for controlling the actions of proteins is allostery. Allosteric modulators have the potential to provide many benefits compared to orthosteric ligands, such as increased selectivity and saturability of their effect. The identification of new allosteric sites presents prospects for the creation of innovative medications and enhances our comprehension of fundamental biological mechanisms. Allosteric sites are increasingly found in different protein families through various techniques, such as machine learning applications, which opens up possibilities for creating completely novel medications with a diverse variety of chemical structures. Machine learning methods, such as PASSer, exhibit limited efficacy in accurately finding allosteric binding sites when relying solely on 3D structural information. Scientific Contribution Prior to conducting feature selection for allosteric binding site identification, integration of supporting amino-acid–based information to 3D structural knowledge is advantageous. This approach can enhance performance by ensuring accuracy and robustness. Therefore, we have developed an accurate and robust model called Multimodel Ensemble Feature Selection for Allosteric Site Identification (MEF-AlloSite) after collecting 9460 relevant and diverse features from the literature to characterise pockets. The model employs an accurate and robust multimodal feature selection technique for the small training set size of only 90 proteins to improve predictive performance. This state-of-the-art technique increased the performance in allosteric binding site identification by selecting promising features from 9460 features. Also, the relationship between selected features and allosteric binding sites enlightened the understanding of complex allostery for proteins by analysing selected features. MEF-AlloSite and state-of-the-art allosteric site identification methods such as PASSer2.0 and PASSerRank have been tested on three test cases 51 times with a different split of the training set. The Student’s t test and Cohen’s D value have been used to evaluate the average precision and ROC AUC score distribution. On three test cases, most of the p-values ( $$< 0.05$$ < 0.05 ) and the majority of Cohen’s D values ( $$> 0.5$$ > 0.5 ) showed that MEF-AlloSite’s 1–6% higher mean of average precision and ROC AUC than state-of-the-art allosteric site identification methods are statistically significant.

Abstract Neural processes (NPs) are models for meta-learning which output uncertainty estimates. So far, most studies of NPs have focused on low-dimensional datasets of highly-correlated tasks. While these homogeneous datasets are useful for benchmarking, they may not be representative of realistic transfer learning. In particular, applications in scientific research may prove especially challenging due to the potential novelty of meta-testing tasks. Molecular property prediction is one such research area that is characterized by sparse datasets of many functions on a shared molecular space. In this paper, we study the application of graph NPs to molecular property prediction with DOCKSTRING, a diverse dataset of docking scores. Graph NPs show competitive performance in few-shot learning tasks relative to supervised learning baselines common in chemoinformatics, as well as alternative techniques for transfer learning and meta-learning. In order to increase meta-generalization to divergent test functions, we propose fine-tuning strategies that adapt the parameters of NPs. We find that adaptation can substantially increase NPs' regression performance while maintaining good calibration of uncertainty estimates. Finally, we present a Bayesian optimization experiment which showcases the potential advantages of NPs over Gaussian processes in iterative screening. Overall, our results suggest that NPs on molecular graphs hold great potential for molecular property prediction in the low-data setting. Scientific contribution Neural processes are a family of meta-learning algorithms which deal with data scarcity by transferring information across tasks and making probabilistic predictions. We evaluate their performance on regression and optimization molecular tasks using docking scores, finding them to outperform classical single-task and transfer-learning models. We examine the issue of generalization to divergent test tasks, which is a general concern of meta-learning algorithms in science, and propose strategies to alleviate it.

Abstract Tunnels in enzymes with buried active sites are key structural features allowing the entry of substrates and the release of products, thus contributing to the catalytic efficiency. Targeting the bottlenecks of protein tunnels is also a powerful protein engineering strategy. However, the identification of functional tunnels in multiple protein structures is a non-trivial task that can only be addressed computationally. We present a pipeline integrating automated structural analysis with an in-house machine-learning predictor for the annotation of protein pockets, followed by the calculation of the energetics of ligand transport via biochemically relevant tunnels. A thorough validation using eight distinct molecular systems revealed that CaverDock analysis of ligand un/binding is on par with time-consuming molecular dynamics simulations, but much faster. The optimized and validated pipeline was applied to annotate more than 17,000 cognate enzyme–ligand complexes. Analysis of ligand un/binding energetics indicates that the top priority tunnel has the most favourable energies in 75% of cases. Moreover, energy profiles of cognate ligands revealed that a simple geometry analysis can correctly identify tunnel bottlenecks only in 50% of cases. Our study provides essential information for the interpretation of results from tunnel calculation and energy profiling in mechanistic enzymology and protein engineering. We formulated several simple rules allowing identification of biochemically relevant tunnels based on the binding pockets, tunnel geometry, and ligand transport energy profiles. Scientific contributions The pipeline introduced in this work allows for the detailed analysis of a large set of protein–ligand complexes, focusing on transport pathways. We are introducing a novel predictor for determining the relevance of binding pockets for tunnel calculation. For the first time in the field, we present a high-throughput energetic analysis of ligand binding and unbinding, showing that approximate methods for these simulations can identify additional mutagenesis hotspots in enzymes compared to purely geometrical methods. The predictor is included in the supplementary material and can also be accessed at https://github.com/Faranehhad/Large-Scale-Pocket-Tunnel-Annotation.git . The tunnel data calculated in this study has been made publicly available as part of the ChannelsDB 2.0 database, accessible at https://channelsdb2.biodata.ceitec.cz/ .

Abstract Focused screening on target-prioritized compound sets can be an efficient alternative to high throughput screening (HTS). For most biomolecular targets, compound prioritization models depend on prior screening data or a target structure. For phenotypic or multi-protein pathway targets, it may not be clear which public assay records provide relevant data. The question also arises as to whether data collected from disparate assays might be usefully consolidated. Here, we report on the development and application of a data mining pipeline to examine these issues. To illustrate, we focus on identifying inhibitors of oxidative phosphorylation, a druggable metabolic process in epithelial ovarian tumors. The pipeline compiled 8415 available OXPHOS-related bioassays in the PubChem data repository involving 312,093 unique compound records. Application of PubChem assay activity annotations, PAINS (Pan Assay Interference Compounds), and Lipinski-like bioavailability filters yields 1852 putative OXPHOS-active compounds that fall into 464 clusters. These chemotypes are diverse but have relatively high hydrophobicity and molecular weight but lower complexity and drug-likeness. These chemotypes show a high abundance of bicyclic ring systems and oxygen containing functional groups including ketones, allylic oxides (alpha/beta unsaturated carbonyls), hydroxyl groups, and ethers. In contrast, amide and primary amine functional groups have a notably lower than random prevalence. UMAP representation of the chemical space shows strong divergence in the regions occupied by OXPHOS-inactive and -active compounds. Of the six compounds selected for biological testing, 4 showed statistically significant inhibition of electron transport in bioenergetics assays. Two of these four compounds, lacidipine and esbiothrin, increased in intracellular oxygen radicals (a major hallmark of most OXPHOS inhibitors) and decreased the viability of two ovarian cancer cell lines, ID8 and OVCAR5. Finally, data from the pipeline were used to train random forest and support vector classifiers that effectively prioritized OXPHOS inhibitory compounds within a held-out test set (ROCAUC 0.962 and 0.927, respectively) and on another set containing 44 documented OXPHOS inhibitors outside of the training set (ROCAUC 0.900 and 0.823). This prototype pipeline is extensible and could be adapted for focus screening on other phenotypic targets for which sufficient public data are available. Scientific contribution Here, we describe and apply an assay data mining pipeline to compile, process, filter, and mine public bioassay data. We believe the procedure may be more broadly applied to guide compound selection in early-stage hit finding on novel multi-protein mechanistic or phenotypic targets. To demonstrate the utility of our approach, we apply a data mining strategy on a large set of public assay data to find drug-like molecules that inhibit oxidative phosphorylation (OXPHOS) as candidates for ovarian cancer therapies. Graphical Abstract

Abstract In untargeted metabolomics, structures of small molecules are annotated using liquid chromatography-mass spectrometry by leveraging information from the molecular retention time (RT) in the chromatogram and m/z (formerly called ''mass-to-charge ratio'') in the mass spectrum. However, correct identification of metabolites is challenging due to the vast array of small molecules. Therefore, various in silico tools for mass spectrometry peak alignment and compound prediction have been developed; however, the list of candidate compounds remains extensive. Accurate RT prediction is important to exclude false candidates and facilitate metabolite annotation. Recent advancements in artificial intelligence (AI) have led to significant breakthroughs in the use of deep learning models in various fields. Release of a large RT dataset has mitigated the bottlenecks limiting the application of deep learning models, thereby improving their application in RT prediction tasks. This review lists the databases that can be used to expand training datasets and concerns the issue about molecular representation inconsistencies in datasets. It also discusses the application of AI technology for RT prediction, particularly in the 5 years following the release of the METLIN small molecule RT dataset. This review provides a comprehensive overview of the AI applications used for RT prediction, highlighting the progress and remaining challenges. Scientific contribution This article focuses on the advancements in small molecule retention time prediction in computational metabolomics over the past five years, with a particular emphasis on the application of AI technologies in this field. It reviews the publicly available datasets for small molecule retention time, the molecular representation methods, the AI algorithms applied in recent studies. Furthermore, it discusses the effectiveness of these models in assisting with the annotation of small molecule structures and the challenges that must be addressed to achieve practical applications.

Abstract Bitter taste is an unpleasant taste modality that affects food consumption. Bitter peptides are generated during enzymatic processes that produce functional, bioactive protein hydrolysates or during the aging process of fermented products such as cheese, soybean protein, and wine. Understanding the underlying peptide sequences responsible for bitter taste can pave the way for more efficient identification of these peptides. This paper presents BitterPep-GCN, a feature-agnostic graph convolution network for bitter peptide prediction. The graph-based model learns the embedding of amino acids in the bitter peptide sequences and uses mixed pooling for bitter classification. BitterPep-GCN was benchmarked using BTP640, a publicly available bitter peptide dataset. The latent peptide embeddings generated by the trained model were used to analyze the activity of sequence motifs responsible for the bitter taste of the peptides. Particularly, we calculated the activity for individual amino acids and dipeptide, tripeptide, and tetrapeptide sequence motifs present in the peptides. Our analyses pinpoint specific amino acids, such as F, G, P, and R, as well as sequence motifs, notably tripeptide and tetrapeptide motifs containing FF, as key bitter signatures in peptides. This work not only provides a new predictor of bitter taste for a more efficient identification of bitter peptides in various food products but also gives a hint into the molecular basis of bitterness. Scientific Contribution Our work provides the first application of Graph Neural Networks for the prediction of peptide bitter taste. The best-developed model, BitterPep-GCN, learns the embedding of amino acids in the bitter peptide sequences and uses mixed pooling for bitter classification. The embeddings were used to analyze the sequence motifs responsible for the bitter taste.

Abstract Nuclear receptors (NRs) play a crucial role as biological targets in drug discovery. However, determining which compounds can act as endocrine disruptors and modulate the function of NRs with a reduced amount of candidate drugs is a challenging task. Moreover, the computational methods for NR-binding activity prediction mostly focus on a single receptor at a time, which may limit their effectiveness. Hence, the transfer of learned knowledge among multiple NRs can improve the performance of molecular predictors and lead to the development of more effective drugs. In this research, we integrate graph neural networks (GNNs) and Transformers to introduce a few-shot GNN-Transformer, Meta-GTNRP to predict the binding activity of compounds using the combined information of different NRs and identify potential NR-modulators with limited data. The Meta-GTNRP model captures the local information in graph-structured data and preserves the global-semantic structure of molecular graph embeddings for NR-binding activity prediction. Furthermore, a few-shot meta-learning approach is proposed to optimize model parameters for different NR-binding tasks and leverage the complementarity among multiple NR-specific tasks to predict binding activity of compounds for each NR with just a few labeled molecules. Experiments with a compound database containing annotations on the binding activity for 11 NRs shows that Meta-GTNRP outperforms other graph-based approaches. The data and code are available at: https://github.com/ltorres97/Meta-GTNRP . Scientific contribution The proposed few-shot GNN-Transformer model, Meta-GTNRP captures the local structure of molecular graphs and preserves the global-semantic information of graph embeddings to predict the NR-binding activity of compounds with limited available data; A few-shot meta-learning framework adapts model parameters across NR-specific tasks for different NRs in a joint learning procedure to predict the binding activity of compounds for each NR with just a few labeled molecules in highly imbalanced data scenarios; Meta-GTNRP is a data-efficient approach that combines the strengths of GNNs and Transformers to predict the NR-binding properties of compounds through an optimized meta-learning procedure and deliver robust results valuable to identify potential NR-based drug candidates.

Abstract This paper proposes a novel multi-view ensemble predictor model that is designed to address the challenge of determining synergistic drug combinations by predicting both the synergy score value values and synergy class label of drug combinations with cancer cell lines. The proposed methodology involves representing drug features through four distinct views: Simplified Molecular-Input Line-Entry System (SMILES) features, molecular graph features, fingerprint features, and drug-target features. On the other hand, cell line features are captured through four views: gene expression features, copy number features, mutation features, and proteomics features. To prevent overfitting of the model, two techniques are employed. First, each view feature of a drug is paired with each corresponding cell line view and input into a multi-task attention deep learning model. This multi-task model is trained to simultaneously predict both the synergy score value and synergy class label. This process results in sixteen input view features being fed into the multi-task model, producing sixteen prediction values. Subsequently, these prediction values are utilized as inputs for an ensemble model, which outputs the final prediction value. The ‘MVME’ model is assessed using the O’Neil dataset, which includes 38 distinct drugs combined across 39 distinct cancer cell lines to output 22,737 drug combination pairs. For the synergy score value, the proposed model scores a mean square error (MSE) of 206.57, a root mean square error (RMSE) of 14.30, and a Pearson score of 0.76. For the synergy class label, the model scores 0.90 for accuracy, 0.96 for precision, 0.57 for kappa, 0.96 for the area under the ROC curve (ROC-AUC), and 0.88 for the area under the precision-recall curve (PR-AUC).

Abstract The identification, establishment, and exploration of potential pharmacological drug targets are major steps of the drug development pipeline. Target validation requires diverse chemical tools that come with a spectrum of functionality, e.g., inhibitors, activators, and other modulators. Particularly tools with rare modes-of-action allow for a proper kinetic and functional characterization of the targets-of-interest (e.g., channels, enzymes, receptors, or transporters). Despite, functional innovation is a prime criterion for patentability and commercial exploitation, which may lead to therapeutic benefit. Unfortunately, data on new, and thus, undruggable or barely druggable targets are scarce and mostly available for mainstream modes-of-action only (e.g., inhibition). Here we present a novel cheminformatic workflow—computer-aided pattern scoring (C@PS)—which was specifically designed to project its prediction capabilities into an uncharted domain of applicability.

Abstract Natural products are molecules that fulfil a range of important ecological functions. Many natural products have been exploited for pharmaceutical and agricultural applications. In contrast to many other specialised metabolites, the products of modular nonribosomal peptide synthetase (NRPS) and polyketide synthase (PKS) systems can often (partially) be predicted from the DNA sequence of the biosynthetic gene clusters. This is because the biosynthetic pathways of NRPS and PKS systems adhere to consistent rulesets. These universal biosynthetic rules can be leveraged to generate biosynthetic models of biosynthetic pathways. While these principles have been largely deciphered, software that leverages these rules to automatically generate visualisations of biosynthetic models has not yet been developed. To enable high-quality automated visualisations of natural product biosynthetic pathways, we developed RAIChU (Reaction Analysis through Illustrating Chemical Units), which produces depictions of biosynthetic transformations of PKS, NRPS, and hybrid PKS/NRPS systems from predicted or experimentally verified module architectures and domain substrate specificities. RAIChU also boasts a library of functions to perform and visualise reactions and pathways whose specifics (e.g., regioselectivity, stereoselectivity) are still difficult to predict, including terpenes, ribosomally synthesised and posttranslationally modified peptides and alkaloids. Additionally, RAIChU includes 34 prevalent tailoring reactions to enable the visualisation of biosynthetic pathways of fully maturated natural products. RAIChU can be integrated into Python pipelines, allowing users to upload and edit results from antiSMASH, a widely used BGC detection and annotation tool, or to build biosynthetic PKS/NRPS systems from scratch. RAIChU’s cluster drawing correctness (100%) and drawing readability (97.66%) were validated on 5000 randomly generated PKS/NRPS systems, and on the MIBiG database. The automated visualisation of these pathways accelerates the generation of biosynthetic models, facilitates the analysis of large (meta-) genomic datasets and reduces human error. RAIChU is available at https://github.com/BTheDragonMaster/RAIChU and https://pypi.org/project/raichu . Scientific contribution RAIChU is the first software package capable of automating high-quality visualisations of natural product biosynthetic pathways. By leveraging universal biosynthetic rules, RAIChU enables the depiction of complex biosynthetic transformations for PKS, NRPS, ribosomally synthesised and posttranslationally modified peptide (RiPP), terpene and alkaloid systems, enhancing predictive and analytical capabilities. This innovation not only streamlines the creation of biosynthetic models, making the analysis of large genomic datasets more efficient and accurate, but also bridges a crucial gap in predicting and visualising the complexities of natural product biosynthesis.

Abstract Despite recent advancement in 3D molecule conformation generation driven by diffusion models, its high computational cost in iterative diffusion/denoising process limits its application. Here, an equivariant consistency model (EC-Conf) was proposed as a fast diffusion method for low-energy conformation generation. In EC-Conf, a modified SE (3)-equivariant transformer model was directly used to encode the Cartesian molecular conformations and a highly efficient consistency diffusion process was carried out to generate molecular conformations. It was demonstrated that, with only one sampling step, it can already achieve comparable quality to other diffusion-based models running with thousands denoising steps. Its performance can be further improved with a few more sampling iterations. The performance of EC-Conf is evaluated on both GEOM-QM9 and GEOM-Drugs sets. Our results demonstrate that the efficiency of EC-Conf for learning the distribution of low energy molecular conformation is at least two magnitudes higher than current SOTA diffusion models and could potentially become a useful tool for conformation generation and sampling. Scientific Contributions In this work, we proposed an equivariant consistency model that significantly improves the efficiency of conformation generation in diffusion-based models while maintaining high structural quality. This method serves as a general framework and can be further extended to more complex structure generation and prediction tasks, including those involving proteins, in future steps.

Abstract This article highlights research from the last century that has provided the basis for the searching techniques that are used in present-day cheminformatics systems, and thus provides an acknowledgement of the contributions made by early pioneers in the field.

Abstract In recent years computational methods for molecular modeling have become a prime focus of computational biology and cheminformatics. Many dedicated systems exist for modeling specific classes of molecules such as proteins or small drug-like ligands. These are often heavily tailored toward the automated generation of molecular structures based on some meta-input by the user and are not intended for expert-driven structure assembly. Dedicated manual or semi-automated assembly software tools exist for a variety of molecule classes but are limited in the scope of structures they can produce. In this work we present BuildAMol, a highly flexible and extendable, general-purpose fragment-based molecular assembly toolkit. Written in Python and featuring a well-documented, user-friendly API, BuildAMol empowers researchers with a framework for detailed manual or semi-automated construction of diverse molecular models. Unlike specialized software, BuildAMol caters to a broad range of applications. We demonstrate its versatility across various use cases, encompassing generating metal complexes or the modeling of dendrimers or integrated into a drug discovery pipeline. By providing a robust foundation for expert-driven model building, BuildAMol holds promise as a valuable tool for the continuous integration and advancement of powerful deep learning techniques. Scientific contribution BuildAMol introduces a cutting-edge framework for molecular modeling that seamlessly blends versatility with user-friendly accessibility. This innovative toolkit integrates modeling, modification, optimization, and visualization functions within a unified API, and facilitates collaboration with other cheminformatics libraries. BuildAMol, with its shallow learning curve, serves as a versatile tool for various molecular applications while also laying the groundwork for the development of specialized software tools, contributing to the progress of molecular research and innovation.

Abstract Ion Mobility coupled with Mass Spectrometry (IM-MS) is a promising analytical technique that enhances molecular characterization by measuring collision cross-section (CCS) values, which are indicative of the molecular size and shape. However, the effective application of CCS values in structural analysis is still constrained by the limited availability of experimental data, necessitating the development of accurate machine learning (ML) models for in silico predictions. In this study, we evaluated state-of-the-art Graph Neural Networks (GNNs), trained to predict CCS values using the largest publicly available dataset to date. Although our results confirm the high accuracy of these models within chemical spaces similar to their training environments, their performance significantly declines when applied to structurally novel regions. This discrepancy raises concerns about the reliability of in silico CCS predictions and underscores the need for releasing further publicly available CCS datasets. To mitigate this, we introduce Mol2CCS which demonstrates how generalization can be partially improved by extending models to account for additional features such as molecular fingerprints, descriptors, and the molecule types. Lastly, we also show how confidence models can support by enhancing the reliability of the CCS estimates. Scientific contribution We have benchmarked state-of-the-art graph neural networks for predicting collision cross section. Our work highlights the accuracy of these models when trained and predicted in similar chemical spaces, but also how their accuracy drops when evaluated in structurally novel regions. Lastly, we conclude by presenting potential approaches to mitigate this issue.

Abstract Molecular fragmentation is an effective suite of approaches to reduce the formal computational complexity of quantum chemistry calculations while enhancing their algorithmic parallelisability. However, the practical applicability of fragmentation techniques remains hindered by a dearth of automation and effective metrics to assess the quality of a fragmentation scheme. In this article, we present the Quick Fragmentation via Automated Genetic Search (QFRAGS), a novel automated fragmentation algorithm that uses a genetic optimisation procedure to generate molecular fragments that yield low energy errors when adopted in Many Body Expansions (MBEs). Benchmark testing of QFRAGS on protein systems with less than 500 atoms, using two-body (MBE2) and three-body (MBE3) MBE calculations at the HF/6-31G* level, reveals mean absolute energy errors (MAEE) of 20.6 and 2.2 kJ $$\hbox {mol}^{-1}$$ mol - 1 , respectively. For larger protein systems exceeding 500 atoms, MAEEs are 181.5 kJ $$\hbox {mol}^{-1}$$ mol - 1 for MBE2 and 24.3 kJ $$\hbox {mol}^{-1}$$ mol - 1 for MBE3. Furthermore, when compared to three manual fragmentation schemes on a 40-protein dataset, using both MBE and Fragment Molecular Orbital techniques, QFRAGS achieves comparable or often lower MAEEs. When applied to a 10-lipoglycan/glycolipid dataset, MAEs of 7.9 and 0.3 kJ $$\hbox {mol}^{-1}$$ mol - 1 were observed at the MBE2 and MBE3 levels, respectively. Scientific Contribution This Article presents the Quick Fragmentation via Automated Genetic Search (QFRAGS), an innovative molecular fragmentation algorithm that significantly improves upon existing molecular fragmentation approaches by specifically addressing their lack of automation and effective fragmentation quality metrics. With an evolutionary optimisation strategy, QFRAGS actively pursues high quality fragments, generating fragmentation schemes that exhibit minimal energy errors on systems with hundreds to thousands of atoms. The advent of QFRAGS represents a significant advancement in molecular fragmentation, greatly improving the accessibility and computational feasibility of accurate quantum chemistry calculations.

Abstract Motivation Computational techniques for drug-disease prediction are essential in enhancing drug discovery and repositioning. While many methods utilize multimodal networks from various biological databases, few integrate comprehensive multi-omics data, including transcriptomes, proteomes, and metabolomes. We introduce STRGNN, a novel graph deep learning approach that predicts drug-disease relationships using extensive multimodal networks comprising proteins, RNAs, metabolites, and compounds. We have constructed a detailed dataset incorporating multi-omics data and developed a learning algorithm with topological regularization. This algorithm selectively leverages informative modalities while filtering out redundancies. Results STRGNN demonstrates superior accuracy compared to existing methods and has identified several novel drug effects, corroborating existing literature. STRGNN emerges as a powerful tool for drug prediction and discovery. The source code for STRGNN, along with the dataset for performance evaluation, is available at https://github.com/yuto-ohnuki/STRGNN.git .

Abstract One challenge that current de novo drug design models face is a disparity between the user’s expectations and the actual output of the model in practical applications. Tailoring models to better align with chemists’ implicit knowledge, expectation and preferences is key to overcoming this obstacle effectively. While interest in preference-based and human-in-the-loop machine learning in chemistry is continuously increasing, no tool currently exists that enables the collection of standardized and chemistry-specific feedback. Metis is a Python-based open-source graphical user interface (GUI), designed to solve this and enable the collection of chemists’ detailed feedback on molecular structures. The GUI enables chemists to explore and evaluate molecules, offering a user-friendly interface for annotating preferences and specifying desired or undesired structural features. By providing chemists the opportunity to give detailed feedback, allows researchers to capture more efficiently the chemist’s implicit knowledge and preferences. This knowledge is crucial to align the chemist’s idea with the de novo design agents. The GUI aims to enhance this collaboration between the human and the “machine” by providing an intuitive platform where chemists can interactively provide feedback on molecular structures, aiding in preference learning and refining de novo design strategies. Metis integrates with the existing de novo framework REINVENT, creating a closed-loop system where human expertise can continuously inform and refine the generative models. Scientific contribution We introduce a novel Graphical User Interface, that allows chemists/researchers to give detailed feedback on substructures and properties of small molecules. This tool can be used to learn the preferences of chemists in order to align de novo drug design models with the chemist’s ideas. The GUI can be customized to fit different needs and projects and enables direct integration into de novo REINVENT runs. We believe that Metis can facilitate the discussion and development of novel ways to integrate human feedback that goes beyond binary decisions of liking or disliking a molecule.

Abstract With the increased availability of chemical data in public databases, innovative techniques and algorithms have emerged for the analysis, exploration, visualization, and extraction of information from these data. One such technique is chemical grouping, where chemicals with common characteristics are categorized into distinct groups based on physicochemical properties, use, biological activity, or a combination. However, existing tools for chemical grouping often require specialized programming skills or the use of commercial software packages. To address these challenges, we developed a user-friendly chemical grouping workflow implemented in KNIME, a free, open-source, low/no-code, data analytics platform. The workflow serves as an all-encompassing tool, expertly incorporating a range of processes such as molecular descriptor calculation, feature selection, dimensionality reduction, hyperparameter search, and supervised and unsupervised machine learning methods, enabling effective chemical grouping and visualization of results. Furthermore, we implemented tools for interpretation, identifying key molecular descriptors for the chemical groups, and using natural language summaries to clarify the rationale behind these groupings. The workflow was designed to run seamlessly in both the KNIME local desktop version and KNIME Server WebPortal as a web application. It incorporates interactive interfaces and guides to assist users in a step-by-step manner. We demonstrate the utility of this workflow through a case study using an eye irritation and corrosion dataset. Scientific contributions This work presents a novel, comprehensive chemical grouping workflow in KNIME, enhancing accessibility by integrating a user-friendly graphical interface that eliminates the need for extensive programming skills. This workflow uniquely combines several features such as automated molecular descriptor calculation, feature selection, dimensionality reduction, and machine learning algorithms (both supervised and unsupervised), with hyperparameter optimization to refine chemical grouping accuracy. Moreover, we have introduced an innovative interpretative step and natural language summaries to elucidate the underlying reasons for chemical groupings, significantly advancing the usability of the tool and interpretability of the results.

Abstract Chemical engineers heavily rely on precise knowledge of physicochemical properties to model chemical processes. Despite the growing popularity of deep learning, it is only rarely applied for property prediction due to data scarcity and limited accuracy for compounds in industrially-relevant areas of the chemical space. Herein, we present a geometric deep learning framework for predicting gas- and liquid-phase properties based on novel quantum chemical datasets comprising 124,000 molecules. Our findings reveal that the necessity for quantum-chemical information in deep learning models varies significantly depending on the modeled physicochemical property. Specifically, our top-performing geometric model meets the most stringent criteria for “chemically accurate” thermochemistry predictions. We also show that by carefully selecting the appropriate model featurization and evaluating prediction uncertainties, the reliability of the predictions can be strongly enhanced. These insights represent a crucial step towards establishing deep learning as the standard property prediction workflow in both industry and academia. Scientific contribution We propose a flexible property prediction tool that can handle two-dimensional and three-dimensional molecular information. A thermochemistry prediction methodology that achieves high-level quantum chemistry accuracy for a broad application range is presented. Trained deep learning models and large novel molecular databases of real-world molecules are provided to offer a directly usable and fast property prediction solution to practitioners.

Abstract The exponential growth of data is challenging for humans because their ability to analyze data is limited. Especially in chemistry, there is a demand for tools that can visualize molecular datasets in a convenient graphical way. We propose a new, ready-to-use, multi-tool, and open-source framework for visualizing and navigating chemical space. This framework adheres to the low-code/no-code (LCNC) paradigm, providing a KNIME node, a web-based tool, and a Python package, making it accessible to a broad cheminformatics community. The core technique of the MolCompass framework employs a pre-trained parametric t-SNE model. We demonstrate how this framework can be adapted for the visualisation of chemical space and visual validation of binary classification QSAR/QSPR models, revealing their weaknesses and identifying model cliffs. All parts of the framework are publicly available on GitHub, providing accessibility to the broad scientific community. Scientific contribution We provide an open-source, ready-to-use set of tools for the visualization of chemical space. These tools can be insightful for chemists to analyze compound datasets and for the visual validation of QSAR/QSPR models.

Abstract In recent years, significant advancements have been made in molecular generation algorithms aimed at facilitating drug development, and molecular diversity holds paramount importance within the realm of molecular generation. Nonetheless, the effective quantification of molecular diversity remains an elusive challenge, as extant metrics exemplified by Richness and Internal Diversity fall short in concurrently encapsulating the two main aspects of such diversity: quantity and dissimilarity. To address this quandary, we propose Hamiltonian diversity, a novel molecular diversity metric predicated upon the shortest Hamiltonian circuit. This metric embodies both aspects of molecular diversity in principle, and we implement its calculation with high efficiency and accuracy. Furthermore, through empirical experiments we demonstrate the high consistency of Hamiltonian diversity with real-world chemical diversity, and substantiate its effects in promoting diversity of molecular generation algorithms. Our implementation of Hamiltonian diversity in Python is available at: https://github.com/HXYfighter/HamDiv . Scientific contribution We propose a more rational molecular diversity metric for the community of cheminformatics and drug development. This metric can be applied to evaluation of existing molecular generation methods and enhancing drug design algorithms.

Abstract The performance of molecular docking can be improved by comparing the shape similarity of the flexibly sampled poses against the target proteins’ inverted binding cavities. The effectiveness of these pseudo-ligands or negative image-based models in docking rescoring is boosted further by performing enrichment-driven optimization. Here, we introduce a novel shape-focused pharmacophore modeling algorithm O-LAP that generates a new class of cavity-filling models by clumping together overlapping atomic content via pairwise distance graph clustering. Top-ranked poses of flexibly docked active ligands were used as the modeling input and multiple alternative clustering settings were benchmark-tested thoroughly with five demanding drug targets using random training/test divisions. In docking rescoring, the O-LAP modeling typically improved massively on the default docking enrichment; furthermore, the results indicate that the clustered models work well in rigid docking. The C+ +/Qt5-based algorithm O-LAP is released under the GNU General Public License v3.0 via GitHub ( https://github.com/jvlehtonen/overlap-toolkit ). Scientific contribution This study introduces O-LAP, a C++/Qt5-based graph clustering software for generating new type of shape-focused pharmacophore models. In the O-LAP modeling, the target protein cavity is filled with flexibly docked active ligands, the overlapping ligand atoms are clustered, and the shape/electrostatic potential of the resulting model is compared against the flexibly sampled molecular docking poses. The O-LAP modeling is shown to ensure high enrichment in both docking rescoring and rigid docking based on comprehensive benchmark-testing. Graphical Abstract

Abstract Designing compounds with a range of desirable properties is a fundamental challenge in drug discovery. In pre-clinical early drug discovery, novel compounds are often designed based on an already existing promising starting compound through structural modifications for further property optimization. Recently, transformer-based deep learning models have been explored for the task of molecular optimization by training on pairs of similar molecules. This provides a starting point for generating similar molecules to a given input molecule, but has limited flexibility regarding user-defined property profiles. Here, we evaluate the effect of reinforcement learning on transformer-based molecular generative models. The generative model can be considered as a pre-trained model with knowledge of the chemical space close to an input compound, while reinforcement learning can be viewed as a tuning phase, steering the model towards chemical space with user-specific desirable properties. The evaluation of two distinct tasks—molecular optimization and scaffold discovery—suggest that reinforcement learning could guide the transformer-based generative model towards the generation of more compounds of interest. Additionally, the impact of pre-trained models, learning steps and learning rates are investigated. Scientific contribution Our study investigates the effect of reinforcement learning on a transformer-based generative model initially trained for generating molecules similar to starting molecules. The reinforcement learning framework is applied to facilitate multiparameter optimisation of starting molecules. This approach allows for more flexibility for optimizing user-specific property profiles and helps finding more ideas of interest.

Abstract enviPath is a widely used database and prediction system for microbial biotransformation pathways of primarily xenobiotic compounds. Data and prediction system are freely available both via a web interface and a public REST API. Since its initial release in 2016, we extended the data available in enviPath and improved the performance of the prediction system and usability of the overall system. We now provide three diverse data sets, covering microbial biotransformation in different environments and under different experimental conditions. This also enabled developing a pathway prediction model that is applicable to a more diverse set of chemicals. In the prediction engine, we implemented a new evaluation tailored towards pathway prediction, which returns a more honest and holistic view on the performance. We also implemented a novel applicability domain algorithm, which allows the user to estimate how well the model will perform on their data. Finally, we improved the implementation to speed up the overall system and provide new functionality via a plugin system. Scientific contribution The main scientific contributions are the development of a pathway prediction model applicable to diverse chemicals, a specialized evaluation method for holistic performance assessment, and a novel applicability domain algorithm for user-specific performance estimation. The introduction of two new data sets, and the creation of links to EC classes make enviPath a unique resource in microbial biotransformation research.

Abstract An automated pipeline for comprehensive calculation of intermolecular interaction energies based on molecular force-fields using the Tinker molecular modelling package is presented. Starting with non-optimized chemically intuitive monomer structures, the pipeline allows the approximation of global minimum energy monomers and dimers, configuration sampling for various monomer–monomer distances, estimation of coordination numbers by molecular dynamics simulations, and the evaluation of differential pair interaction energies. The latter are used to derive Flory–Huggins parameters and isotropic particle–particle repulsions for Dissipative Particle Dynamics (DPD). The computational results for force fields MM3, MMFF94, OPLS-AA and AMOEBA09 are analyzed with Density Functional Theory (DFT) calculations and DPD simulations for a mixture of the non-ionic polyoxyethylene alkyl ether surfactant C10E4 with water to demonstrate the usefulness of the approach. Scientific Contribution To our knowledge, there is currently no open computational pipeline for differential pair interaction energies at all. This work aims to contribute an (at least academically available, open) approach based on molecular force fields that provides a robust and efficient computational scheme for their automated calculation for small to medium-sized (organic) molecular dimers. The usefulness of the proposed new calculation scheme is demonstrated for the generation of mesoscopic particles with their mutual repulsive interactions.

Abstract Data scarcity is one of the most critical issues impeding the development of prediction models for chemical effects. Multitask learning algorithms leveraging knowledge from relevant tasks showed potential for dealing with tasks with limited data. However, current multitask methods mainly focus on learning from datasets whose task labels are available for most of the training samples. Since datasets were generated for different purposes with distinct chemical spaces, the conventional multitask learning methods may not be suitable. This study presents a novel multitask learning method MTForestNet that can deal with data scarcity problems and learn from tasks with distinct chemical space. The MTForestNet consists of nodes of random forest classifiers organized in the form of a progressive network, where each node represents a random forest model learned from a specific task. To demonstrate the effectiveness of the MTForestNet, 48 zebrafish toxicity datasets were collected and utilized as an example. Among them, two tasks are very different from other tasks with only 1.3% common chemicals shared with other tasks. In an independent test, MTForestNet with a high area under the receiver operating characteristic curve (AUC) value of 0.911 provided superior performance over compared single-task and multitask methods. The overall toxicity derived from the developed models of zebrafish toxicity is well correlated with the experimentally determined overall toxicity. In addition, the outputs from the developed models of zebrafish toxicity can be utilized as features to boost the prediction of developmental toxicity. The developed models are effective for predicting zebrafish toxicity and the proposed MTForestNet is expected to be useful for tasks with distinct chemical space that can be applied in other tasks. Scieific contribution A novel multitask learning algorithm MTForestNet was proposed to address the challenges of developing models using datasets with distinct chemical space that is a common issue of cheminformatics tasks. As an example, zebrafish toxicity prediction models were developed using the proposed MTForestNet which provide superior performance over conventional single-task and multitask learning methods. In addition, the developed zebrafish toxicity prediction models can reduce animal testing.

Abstract Protein language models (PLMs) play a dominant role in protein representation learning. Most existing PLMs regard proteins as sequences of 20 natural amino acids. The problem with this representation method is that it simply divides the protein sequence into sequences of individual amino acids, ignoring the fact that certain residues often occur together. Therefore, it is inappropriate to view amino acids as isolated tokens. Instead, the PLMs should recognize the frequently occurring combinations of amino acids as a single token. In this study, we use the byte-pair-encoding algorithm and unigram to construct advanced residue vocabularies for protein sequence tokenization, and we have shown that PLMs pre-trained using these advanced vocabularies exhibit superior performance on downstream tasks when compared to those trained with simple vocabularies. Furthermore, we introduce PETA, a comprehensive benchmark for systematically evaluating PLMs. We find that vocabularies comprising 50 and 200 elements achieve optimal performance. Our code, model weights, and datasets are available at https://github.com/ginnm/ProteinPretraining . Scientific contribution This study introduces advanced protein sequence tokenization analysis, leveraging the byte-pair-encoding algorithm and unigram. By recognizing frequently occurring combinations of amino acids as single tokens, our proposed method enhances the performance of PLMs on downstream tasks. Additionally, we present PETA, a new comprehensive benchmark for the systematic evaluation of PLMs, demonstrating that vocabularies of 50 and 200 elements offer optimal performance. Graphical Abstract

Abstract Here, we present a new method for evaluating questions on chemical reactions in the context of remote education. This method can be used when binary grading is not sufficient as some tolerance may be acceptable. In order to determine a grade, the developed workflow uses the pairwise similarity assessment of two considered reactions, each encoded by a single molecular graph with the help of the Condensed Graph of Reaction (CGR) approach. This workflow is part of the ChemMoodle project and is implemented as a Moodle Plugin. It uses the Chemdoodle engine for reaction drawing and visualization and communicates with a REST server calculating the similarity score using ISIDA fragment descriptors. The plugin is open-source, accessible in GitHub ( https://github.com/Laboratoire-de-Chemoinformatique/moodle-qtype_reacsimilarity ) and on the Moodle plugin store ( https://moodle.org/plugins/qtype_reacsimilarity?lang=en ). Both similarity measures and fragmentation can be configured. Scientific contribution This work introduces an open-source method for evaluating chemical reaction questions within Moodle using the CGR approach. Our contribution provides a nuanced grading mechanism that accommodates acceptable tolerances in reaction assessments, enhancing the accuracy and flexibility of the grading process.

Abstract Machine learning is becoming a preferred method for the virtual screening of organic materials due to its cost-effectiveness over traditional computationally demanding techniques. However, the scarcity of labeled data for organic materials poses a significant challenge for training advanced machine learning models. This study showcases the potential of utilizing databases of drug-like small molecules and chemical reactions to pretrain the BERT model, enhancing its performance in the virtual screening of organic materials. By fine-tuning the BERT models with data from five virtual screening tasks, the version pretrained with the USPTO–SMILES dataset achieved R2 scores exceeding 0.94 for three tasks and over 0.81 for two others. This performance surpasses that of models pretrained on the small molecule or organic materials databases and outperforms three traditional machine learning models trained directly on virtual screening data. The success of the USPTO–SMILES pretrained BERT model can be attributed to the diverse array of organic building blocks in the USPTO database, offering a broader exploration of the chemical space. The study further suggests that accessing a reaction database with a wider range of reactions than the USPTO could further enhance model performance. Overall, this research validates the feasibility of applying transfer learning across different chemical domains for the efficient virtual screening of organic materials. Scientific contribution This study verifies the feasibility of applying transfer learning to large language models in different chemical fields to help organic materials perform virtual screening. Through the comparison of transfer learning from different chemical fields to a variety of organic material molecules, the high precision virtual screening of organic materials is realized.

Abstract Every year, more than 19 million cancer cases are diagnosed, and this number continues to increase annually. Since standard treatment options have varying success rates for different types of cancer, understanding the biology of an individual's tumour becomes crucial, especially for cases that are difficult to treat. Personalised high-throughput profiling, using next-generation sequencing, allows for a comprehensive examination of biopsy specimens. Furthermore, the widespread use of this technology has generated a wealth of information on cancer-specific gene alterations. However, there exists a significant gap between identified alterations and their proven impact on protein function. Here, we present a bioinformatics pipeline that enables fast analysis of a missense mutation’s effect on stability and function in known oncogenic proteins. This pipeline is coupled with a predictor that summarises the outputs of different tools used throughout the pipeline, providing a single probability score, achieving a balanced accuracy above 86%. The pipeline incorporates a virtual screening method to suggest potential FDA/EMA-approved drugs to be considered for treatment. We showcase three case studies to demonstrate the timely utility of this pipeline. To facilitate access and analysis of cancer-related mutations, we have packaged the pipeline as a web server, which is freely available at https://loschmidt.chemi.muni.cz/predictonco/ . Scientific contribution This work presents a novel bioinformatics pipeline that integrates multiple computational tools to predict the effects of missense mutations on proteins of oncological interest. The pipeline uniquely combines fast protein modelling, stability prediction, and evolutionary analysis with virtual drug screening, while offering actionable insights for precision oncology. This comprehensive approach surpasses existing tools by automating the interpretation of mutations and suggesting potential treatments, thereby striving to bridge the gap between sequencing data and clinical application.

Abstract Mass spectral libraries have proven to be essential for mass spectrum annotation, both for library matching and training new machine learning algorithms. A key step in training machine learning models is the availability of high-quality training data. Public libraries of mass spectrometry data that are open to user submission often suffer from limited metadata curation and harmonization. The resulting variability in data quality makes training of machine learning models challenging. Here we present a library cleaning pipeline designed for cleaning tandem mass spectrometry library data. The pipeline is designed with ease of use, flexibility, and reproducibility as leading principles. Scientific contribution This pipeline will result in cleaner public mass spectral libraries that will improve library searching and the quality of machine-learning training datasets in mass spectrometry. This pipeline builds on previous work by adding new functionality for curating and correcting annotated libraries, by validating structure annotations. Due to the high quality of our software, the reproducibility, and improved logging, we think our new pipeline has the potential to become the standard in the field for cleaning tandem mass spectrometry libraries. Graphical Abstract

Abstract Motivation Chemical space embedding methods are widely utilized in various research settings for dimensional reduction, clustering and effective visualization. The maps generated by the embedding process can provide valuable insight to medicinal chemists in terms of the relationships between structural, physicochemical and biological properties of compounds. However, these maps are known to be difficult to interpret, and the ‘‘landscape’’ on the map is prone to ‘‘rearrangement’’ when embedding different sets of compounds. Results In this study we present the Hilbert-Curve Assisted Space Embedding (HCASE) method which was designed to create maps by organizing structures according to a logic familiar to medicinal chemists. First, a chemical space is created with the help of a set of ‘‘reference scaffolds’’. These scaffolds are sorted according to the medicinal chemistry inspired Scaffold-Key algorithm found in prior art. Next, the ordered scaffolds are mapped to a line which is folded into a higher dimensional (here: 2D) space. The intricately folded line is referred to as a pseudo-Hilbert-Curve. The embedding of a compound happens by locating its most similar reference scaffold in the pseudo-Hilbert-Curve and assuming the respective position. Through a series of experiments, we demonstrate the properties of the maps generated by the HCASE method. Subjects of embeddings were compounds of the DrugBank and CANVASS libraries, and the chemical spaces were defined by scaffolds extracted from the ChEMBL database. Scientific contribution The novelty of HCASE method lies in generating robust and intuitive chemical space embeddings that are reflective of a medicinal chemist’s reasoning, and the precedential use of space filling (Hilbert) curve in the process. Availability https://github.com/ncats/hcase Graphical Abstract

Abstract Synthetic accessibility prediction is a task to estimate how easily a given molecule might be synthesizable in the laboratory, playing a crucial role in computer-aided molecular design. Although synthesis planning programs can determine synthesis routes, their slow processing times make them impractical for large-scale molecule screening. On the other hand, existing rapid synthesis accessibility estimation methods offer speed but typically lack integration with actual synthesis routes and building block information. In this work, we introduce BR-SAScore, an enhanced version of SAScore that integrates the available building block information (B) and reaction knowledge (R) from synthesis planning programs into the scoring process. In particular, we differentiate fragments inherent in building blocks and fragments to be derived from synthesis (reactions) when scoring synthetic accessibility. Compared to existing methods, our experimental findings demonstrate that BR-SAScore offers more accurate and precise identification of a molecule's synthetic accessibility by the synthesis planning program with a fast calculation time. Moreover, we illustrate how BR-SAScore provides chemically interpretable results, aligning with the capability of the synthesis planning program embedded with the same reaction knowledge and available building blocks. Scientific contribution We introduce BR-SAScore, an extension of SAScore, to estimate the synthetic accessibility of molecules by leveraging known building-block and reactivity information. In our experiments, BR-SAScore shows superior prediction performance on predicting molecule synthetic accessibility compared to previous methods, including SAScore and deep-learning models, while requiring significantly less computation time. In addition, we show that BR-SAScore is able to precisely identify the chemical fragment contributing to the synthetic infeasibility, holding great potential for future molecule synthesizability optimization.

Abstract Pretrained Graph Neural Networks have been widely adopted for various molecular property prediction tasks. Despite their ability to encode structural and relational features of molecules, traditional fine-tuning of such pretrained GNNs on the target task can lead to poor generalization. To address this, we explore the adaptation of pretrained GNNs to the target task by jointly training them with multiple auxiliary tasks. This could enable the GNNs to learn both general and task-specific features, which may benefit the target task. However, a major challenge is to determine the relatedness of auxiliary tasks with the target task. To address this, we investigate multiple strategies to measure the relevance of auxiliary tasks and integrate such tasks by adaptively combining task gradients or by learning task weights via bi-level optimization. Additionally, we propose a novel gradient surgery-based approach, Rotation of Conflicting Gradients ( $$\mathop {\texttt{RCGrad}}\limits$$ RCGrad ), that learns to align conflicting auxiliary task gradients through rotation. Our experiments with state-of-the-art pretrained GNNs demonstrate the efficacy of our proposed methods, with improvements of up to 7.7% over fine-tuning. This suggests that incorporating auxiliary tasks along with target task fine-tuning can be an effective way to improve the generalizability of pretrained GNNs for molecular property prediction. Scientific contribution We introduce a novel framework for adapting pretrained GNNs to molecular tasks using auxiliary learning to address the critical issue of negative transfer. Leveraging novel gradient surgery techniques such as $$\mathop {\texttt{RCGrad}}\limits$$ RCGrad , the proposed adaptation framework represents a significant departure from the dominant pretraining fine-tuning approach for molecular GNNs. Our contributions are significant for drug discovery research, especially for tasks with limited data, filling a notable gap in the efficient adaptation of pretrained models for molecular GNNs.

Abstract It is well-accepted that knowledge of a small molecule’s target can accelerate optimization. Although chemogenomic databases are helpful resources for predicting or finding compound interaction partners, they tend to be limited and poorly annotated. Furthermore, unlike genes, compound identifiers are often not standardized, and many synonyms may exist, especially in the biological literature, making batch analysis of compounds difficult. Here, we constructed an open-source annotation and target hypothesis prediction tool that explores some of the largest chemical and biological databases, mining these for both common name, synonyms, and structurally similar molecules. We used this Chemical Analysis and Clustering for Target Identification (CACTI) tool to analyze the Pathogen Box collection, an open-source set of 400 drug-like compounds active against a variety of microbial pathogens. Our analysis resulted in 4,315 new synonyms, 35,963 pieces of new information and target prediction hints for 58 members. Scientific contributions With the employment of this tool, a comprehensive report with known evidence, close analogs and drug-target prediction can be obtained for large-scale chemical libraries that will facilitate their evaluation and future target validation and optimization efforts.

Abstract While drug combination therapies are of great importance, particularly in cancer treatment, identifying novel synergistic drug combinations has been a challenging venture. Computational methods have emerged in this context as a promising tool for prioritizing drug combinations for further evaluation, though they have presented limited performance, utility, and interpretability. Here, we propose a novel predictive tool, piscesCSM, that leverages graph-based representations to model small molecule chemical structures to accurately predict drug combinations with favourable anticancer synergistic effects against one or multiple cancer cell lines. Leveraging these insights, we developed a general supervised machine learning model to guide the prediction of anticancer synergistic drug combinations in over 30 cell lines. It achieved an area under the receiver operating characteristic curve (AUROC) of up to 0.89 on independent non-redundant blind tests, outperforming state-of-the-art approaches on both large-scale oncology screening data and an independent test set generated by AstraZeneca (with more than a 16% improvement in predictive accuracy). Moreover, by exploring the interpretability of our approach, we found that simple physicochemical properties and graph-based signatures are predictive of chemotherapy synergism. To provide a simple and integrated platform to rapidly screen potential candidate pairs with favourable synergistic anticancer effects, we made piscesCSM freely available online at https://biosig.lab.uq.edu.au/piscescsm/ as a web server and API. We believe that our predictive tool will provide a valuable resource for optimizing and augmenting combinatorial screening libraries to identify effective and safe synergistic anticancer drug combinations. Scientific contribution This work proposes piscesCSM, a machine-learning-based framework that relies on well-established graph-based representations of small molecules to identify and provide better predictive accuracy of syngenetic drug combinations. Our model, piscesCSM, shows that combining physiochemical properties with graph-based signatures can outperform current architectures on classification prediction tasks. Furthermore, implementing our tool as a web server offers a user-friendly platform for researchers to screen for potential synergistic drug combinations with favorable anticancer effects against one or multiple cancer cell lines.

Abstract Purpose Reaction databases are a key resource for a wide variety of applications in computational chemistry and biochemistry, including Computer-aided Synthesis Planning (CASP) and the large-scale analysis of metabolic networks. The full potential of these resources can only be realized if datasets are accurate and complete. Missing co-reactants and co-products, i.e., unbalanced reactions, however, are the rule rather than the exception. The curation and correction of such incomplete entries is thus an urgent need. Methods The SynRBL framework addresses this issue with a dual-strategy: a rule-based method for non-carbon compounds, using atomic symbols and counts for prediction, alongside a Maximum Common Subgraph (MCS)-based technique for carbon compounds, aimed at aligning reactants and products to infer missing entities. Results The rule-based method exceeded 99% accuracy, while MCS-based accuracy varied from 81.19 to 99.33%, depending on reaction properties. Furthermore, an applicability domain and a machine learning scoring function were devised to quantify prediction confidence. The overall efficacy of this framework was delineated through its success rate and accuracy metrics, which spanned from 89.83 to 99.75% and 90.85 to 99.05%, respectively. Conclusion The SynRBL framework offers a novel solution for recalibrating chemical reactions, significantly enhancing reaction completeness. With rigorous validation, it achieved groundbreaking accuracy in reaction rebalancing. This sets the stage for future improvement in particular of atom-atom mapping techniques as well as of downstream tasks such as automated synthesis planning. Scientific Contribution SynRBL features a novel computational approach to correcting unbalanced entries in chemical reaction databases. By combining heuristic rules for inferring non-carbon compounds and common subgraph searches to address carbon unbalance, SynRBL successfully addresses most instances of this problem, which affects the majority of data in most large-scale resources. Compared to alternative solutions, SynRBL achieves a dramatic increase in both success rate and accurary, and provides the first freely available open source solution for this problem.

Abstract Motivation Retrosynthesis planning poses a formidable challenge in the organic chemical industry, particularly in pharmaceuticals. Single-step retrosynthesis prediction, a crucial step in the planning process, has witnessed a surge in interest in recent years due to advancements in AI for science. Various deep learning-based methods have been proposed for this task in recent years, incorporating diverse levels of additional chemical knowledge dependency. Results This paper introduces UAlign, a template-free graph-to-sequence pipeline for retrosynthesis prediction. By combining graph neural networks and Transformers, our method can more effectively leverage the inherent graph structure of molecules. Based on the fact that the majority of molecule structures remain unchanged during a chemical reaction, we propose a simple yet effective SMILES alignment technique to facilitate the reuse of unchanged structures for reactant generation. Extensive experiments show that our method substantially outperforms state-of-the-art template-free and semi-template-based approaches. Importantly, our template-free method achieves effectiveness comparable to, or even surpasses, established powerful template-based methods. Scientific contribution We present a novel graph-to-sequence template-free retrosynthesis prediction pipeline that overcomes the limitations of Transformer-based methods in molecular representation learning and insufficient utilization of chemical information. We propose an unsupervised learning mechanism for establishing product-atom correspondence with reactant SMILES tokens, achieving even better results than supervised SMILES alignment methods. Extensive experiments demonstrate that UAlign significantly outperforms state-of-the-art template-free methods and rivals or surpasses template-based approaches, with up to 5% (top-5) and 5.4% (top-10) increased accuracy over the strongest baseline.

Abstract Background Previous deep learning methods for predicting protein binding pockets mainly employed 3D convolution, yet an abundance of convolution operations may lead the model to excessively prioritize local information, thus overlooking global information. Moreover, it is essential for us to account for the influence of diverse protein folding structural classes. Because proteins classified differently structurally exhibit varying biological functions, whereas those within the same structural class share similar functional attributes. Results We proposed LVPocket, a novel method that synergistically captures both local and global information of protein structure through the integration of Transformer encoders, which help the model achieve better performance in binding pockets prediction. And then we tailored prediction models for data of four distinct structural classes of proteins using the transfer learning. The four fine-tuned models were trained on the baseline LVPocket model which was trained on the sc-PDB dataset. LVPocket exhibits superior performance on three independent datasets compared to current state-of-the-art methods. Additionally, the fine-tuned model outperforms the baseline model in terms of performance. Scientific contribution We present a novel model structure for predicting protein binding pockets that provides a solution for relying on extensive convolutional computation while neglecting global information about protein structures. Furthermore, we tackle the impact of different protein folding structures on binding pocket prediction tasks through the application of transfer learning methods. Graphical Abstract