Journal of Cheminformatics最新文献

Over the last few decades the pharmaceutical industry has generated a vast corpus of knowledge on the safety and efficacy of drugs. Much of this information is contained in toxicology reports, which summarise the results of animal studies designed to analyse the effects of the tested compound, including unintended pharmacological and toxic effects, known as treatment-related findings. Despite the potential of this knowledge, the fact that most of this relevant information is only available as unstructured text with variable degrees of digitisation has hampered its systematic access, use and exploitation. Text mining technologies have the ability to automatically extract, analyse and aggregate such information, providing valuable new insights into the drug discovery and development process. In the context of the eTRANSAFE project, we present PretoxTM (Preclinical Toxicology Text Mining), the first system specifically designed to detect, extract, organise and visualise treatment-related findings from toxicology reports. The PretoxTM tool comprises three main components: PretoxTM Corpus, PretoxTM Pipeline and PretoxTM Web App. The PretoxTM Corpus is a gold standard corpus of preclinical treatment-related findings annotated by toxicology experts. This corpus was used to develop, train and validate the PretoxTM Pipeline, which extracts treatment-related findings from preclinical study reports. The extracted information is then presented for expert visualisation and validation in the PretoxTM Web App.

Scientific Contribution

While text mining solutions have been widely used in the clinical domain to identify adverse drug reactions from various sources, no similar systems exist for identifying adverse events in animal models during preclinical testing. PretoxTM fills this gap by efficiently extracting treatment-related findings from preclinical toxicology reports. This provides a valuable resource for toxicology research, enhancing the efficiency of safety evaluations, saving time, and leading to more effective decision-making in the drug development process.

More sophisticated representations of compounds attempt to incorporate not only information on the structure and physicochemical properties of molecules, but also knowledge about their biological traits, leading to the so-called bioactivity profile. The bioactive profiling of air pollutants is challenging and crucial, as their biological activity and toxicological effects have not been deeply investigated yet, and further exploration could shed light on the impact of air pollution on complex disorders. Therefore, a biological signature that simultaneously captures the chemistry and the biology of small molecules may be beneficial in predicting the behaviour of such ligands towards a protein target. Moreover, the interactivity between biological entities can be represented through combined feature vectors that can be given as input to a machine learning (ML) model to capture the underlying interaction. To this end, we propose a chemogenomic approach, called Air Pollutant Bioactivity (APBIO), which integrates compound bioactivity signatures and target sequence descriptors to train ML classifiers subsequently used to predict potential compound-target interactions (CTIs). We report the performances of the proposed methodology and, via external validation sets, demonstrate its outperformance compared to existing molecular representations in terms of model generalizability. We have also developed a publicly available Streamlit application for APBIO at ap-bio.streamlit.app, allowing users to predict associations between investigated compounds and protein targets.

Scientific contribution

We derived ex novo bioactivity signatures for air pollutant molecules to capture their biological behaviour and associations with protein targets. The proposed chemogenomic methodology enables the prediction of novel CTIs for known or similar compounds and targets through well-established and efficient ML models, deepening our insight into the molecular interactions and mechanisms that may have a deleterious impact on human biological systems.

MLinvitroTox is an automated Python pipeline developed for high-throughput hazard-driven prioritization of toxicologically relevant signals detected in complex environmental samples through high-resolution tandem mass spectrometry (HRMS/MS). MLinvitroTox is a machine learning (ML) framework comprising 490 independent XGBoost classifiers trained on molecular fingerprints from chemical structures and target-specific endpoints from the ToxCast/Tox21 invitroDBv4.1 database. For each analyzed HRMS feature, MLinvitroTox generates a 490-bit bioactivity fingerprint used as a basis for prioritization, focusing the time-consuming molecular identification efforts on features most likely to cause adverse effects. The practical advantages of MLinvitroTox are demonstrated for groundwater HRMS data. Among the 874 features for which molecular fingerprints were derived from spectra, including 630 nontargets, 185 spectral matches, and 59 targets, around 4% of the feature/endpoint relationship pairs were predicted to be active. Cross-checking the predictions for targets and spectral matches with invitroDB data confirmed the bioactivity of 120 active and 6791 nonactive pairs while mislabeling 88 active and 56 non-active relationships. By filtering according to bioactivity probability, endpoint scores, and similarity to the training data, the number of potentially toxic features was reduced by at least one order of magnitude. This refinement makes the analytical confirmation of the toxicologically most relevant features feasible, offering significant benefits for cost-efficient chemical risk assessment.

Scientific Contribution:

In contrast to the classical ML-based approaches for toxicity prediction, MLinvitroTox predicts bioactivity for HRMS features (i.e., distinct m/z signals) based on MS2 fragmentation spectra rather than the chemical structures from the identified features. While the original proof of concept study was accompanied by the release of a MLinvitroTox v1 KNIME workflow, in this study, we release a Python MLinvitroTox v2 package, which, in addition to automation, expands functionality to include predicting toxicity from structures, cleaning up and generating chemical fingerprints, customizing models, and retraining on custom data. Furthermore, as a result of improvements in bioactivity data processing, realized in the concurrently released pytcpl Python package for the custom processing of invitroDBv4.1 input data used for training MLinvitroTox, the current release introduces enhancements in model accuracy, coverage of biological mechanistic targets, and overall interpretability.

The human ether-a-go-go-related gene (hERG) channel plays a critical role in the electrical activity of the heart, and its blockers can cause serious cardiotoxic effects. Thus, screening for hERG channel blockers is a crucial step in the drug development process. Many in silico models have been developed to predict hERG blockers, which can efficiently save time and resources. However, previous methods have found it hard to achieve high performance and to interpret the predictive results. To overcome these challenges, we have proposed hERGAT, a graph neural network model with an attention mechanism, to consider compound interactions on atomic and molecular levels. In the atom-level interaction analysis, we applied a graph attention mechanism (GAT) that integrates information from neighboring nodes and their extended connections. The hERGAT employs a gated recurrent unit (GRU) with the GAT to learn information between more distant atoms. To confirm this, we performed clustering analysis and visualized a correlation heatmap, verifying the interactions between distant atoms were considered during the training process. In the molecule-level interaction analysis, the attention mechanism enables the target node to focus on the most relevant information, highlighting the molecular substructures that play crucial roles in predicting hERG blockers. Through a literature review, we confirmed that highlighted substructures have a significant role in determining the chemical and biological characteristics related to hERG activity. Furthermore, we integrated physicochemical properties into our hERGAT model to improve the performance. Our model achieved an area under the receiver operating characteristic of 0.907 and an area under the precision-recall of 0.904, demonstrating its effectiveness in modeling hERG activity and offering a reliable framework for optimizing drug safety in early development stages.

Scientific contribution:

hERGAT is a deep learning model for predicting hERG blockers by combining GAT and GRU, enabling it to capture complex interactions at atomic and molecular levels. We improve the model's interpretability by analyzing the highlighted molecular substructures, providing valuable insights into their roles in determining hERG activity. The model achieves high predictive performance, confirming its potential as a preliminary tool for early cardiotoxicity assessment and enhancing the reliability of the results.

Accurate prediction of ligand-receptor binding affinity is crucial in structure-based drug design, significantly impacting the development of effective drugs. Recent advances in machine learning (ML)–based scoring functions have improved these predictions, yet challenges remain in modeling complex molecular interactions. This study introduces the AGL-EAT-Score, a scoring function that integrates extended atom-type multiscale weighted colored subgraphs with algebraic graph theory. This approach leverages the eigenvalues and eigenvectors of graph Laplacian and adjacency matrices to capture high-level details of specific atom pairwise interactions. Evaluated against benchmark datasets such as CASF-2016, CASF-2013, and the Cathepsin S dataset, the AGL-EAT-Score demonstrates notable accuracy, outperforming existing traditional and ML-based methods. The model’s strength lies in its comprehensive similarity analysis, examining protein sequence, ligand structure, and binding site similarities, thus ensuring minimal bias and over-representation in the training sets. The use of extended atom types in graph coloring enhances the model’s capability to capture the intricacies of protein-ligand interactions. The AGL-EAT-Score marks a significant advancement in drug design, offering a tool that could potentially refine and accelerate the drug discovery process.

Scientific Contribution

The AGL-EAT-Score presents an algebraic graph-based framework that predicts ligand-receptor binding affinity by constructing multiscale weighted colored subgraphs from the 3D structure of protein-ligand complexes. It improves prediction accuracy by modeling interactions between extended atom types, addressing challenges like dataset bias and over-representation. Benchmark evaluations demonstrate that AGL-EAT-Score outperforms existing methods, offering a robust and systematic tool for structure-based drug design.

This article introduces StreamChol, a software for developing and applying mechanistic models to predict cholestasis. StreamChol is a Streamlit application, usable as a desktop application or web-accessible software when installed on a server using a docker container.

StreamChol allows a seamless integration of pharmacokinetic analyses with Machine Learning models. This integration not only enables cholestasis prediction but also opens avenues for predicting other toxicological endpoints requiring similar integrations. StreamChol's Docker containerization also streamlines deployment across diverse environments, addressing potential compatibility issues. StreamChol is distributed as open-source under GNU GPL v3, reflecting our commitment to open science. Through StreamChol, researchers are offered a potent tool for predictive modelling in toxicology, harnessing its strengths within an intuitive and user-friendly interface, without the need for any programming knowledge.

Scientific contribution This work offers a user-friendly web-based tool for cholestasis prediction and a complete workflow for creating web platforms that require the combination of both programming languages, R and Python.

Machine learning models for chemistry require large datasets, often compiled by combining data from multiple assays. However, combining data without careful curation can introduce significant noise. While absolute values from different assays are rarely comparable, trends or differences between compounds are often assumed to be consistent. This study evaluates that assumption by analyzing potency differences between matched compound pairs across assays and assessing the impact of assay metadata curation on error reduction. We find that potency differences between matched pairs exhibit less variability than individual compound measurements, suggesting systematic assay differences may partially cancel out in paired data. Metadata curation further improves inter-assay agreement, albeit at the cost of dataset size. For minimally curated compound pairs, agreement within 0.3 pChEMBL units was found to be 44–46% for K_i and IC₅₀ values respectively, which improved to 66–79% after curation. Similarly, the percentage of pairs with differences exceeding 1 pChEMBL unit dropped from 12 to 15% to 6–8% with extensive curation. These results establish a benchmark for expected noise in matched molecular pair data from the ChEMBL database, offering practical metrics for data quality assessment.

Chemistry has diversified from a basic understanding of the elements to studying millions of highly diverse molecules and materials, which together are conceptualized as the chemical space. A map of this chemical space where distances represent similarities between compounds can represent the mutual relationships between different subfields of chemistry and help the discipline to be viewed and understood globally.

Traditional best practices for quantitative structure activity relationship (QSAR) modeling recommend dataset balancing and balanced accuracy (BA) as the key desired objective of model development. This study explores the value of the conventional norms in the context of using QSAR models for virtual screening of modern large and ultra-large chemical libraries. For this increasingly common task, we now recommend the use of models with the highest positive predictive value (PPV) built on imbalanced training sets as preferred virtual screening tools. This recommendation stems from practical considerations of how the results of virtual screening are used in experimental laboratories where only a small fraction of virtually screened molecules can be tested using standard well plates. As a proof of concept, we have developed QSAR models for five expansive datasets with different ratios of active and inactive molecules and compared model performance in virtual screening using BA, PPV, and other metrics. We show that training on imbalanced datasets achieves a hit rate at least 30% higher than using balanced datasets, and that the PPV metric captured this difference of performance with no parameter tuning. Importantly, hit rates were estimated for top scoring compounds organized in batches of the size of plates (for instance, 128 molecules) used in the experimental high throughput screening. Based on the results of our studies, we posit that QSAR models trained on imbalanced datasets with the highest PPV should be relied upon to identify and test hit compounds in early drug discovery studies.