Journal of Cheminformatics最新文献

The prediction of drugs metabolism by in silico techniques is gaining a growing interest due to the possibility to process large datasets allowing the stability and safety of new drug candidates to be evaluated during the early stages of the drug discovery process. To date, in silico models for metabolism prediction mainly exploits the ligand-based (LB) properties of the training molecules to predict the occurrence of a given metabolic reaction and/or the reactive site involved in the biotransformation. However, recent reports highlighted that structure-based (SB) modeling can be conveniently integrated with LB methods for drug metabolism prediction purpose, with the advantages to predict if a given molecule can fit the enzyme active site and which moiety approaches the catalytic residues. Herein, we developed machine learning models for UDP-glucuronosyltransferase (UGT)-mediated metabolism by using both LB and SB methods. In particular, this study was focused on UGT2B7 and UGT2B15 isoforms which are involved in the clearance of many drugs as well as in clinically relevant drug-drug interactions. First, molecular dynamics (MD) and docking simulations were combined to explore the binding mechanism of cofactor and substrate within the catalytic pocket of the studied UGT isoforms exploiting their AlphaFold structures. The analysis of the MD trajectories allowed an appropriate conformation of both UGT isoforms to be identified for the development of binary classification models. For this purpose, Random Forest algorithm and the metabolic data extracted from the MetaQSAR database were used. SB models were trained on a set of scoring functions and protein–ligand interaction fingerprints derived from docking, while the LB models were built on a set of physicochemical and constitutional descriptors. When the single models were evaluated, the LB classifiers outperformed the SB models. However, the application of a consensus strategy led to an improvement of the prediction accuracy if compared to the individual models, highlighting that LB and SB approaches convey complementary information whose aggregation allowed us to achieve better predictions than the single models.

Reverse engineering in molecular design aims to identify optimal structures based on activities, or properties, computed through molecular descriptors like fingerprints. This task is known to be particularly difficult for the widely used Extended-Connectivity Fingerprints (ECFPs), due to significant loss of structural information during vectorization. While recent artificial intelligence-based works have raised awareness about the privacy risks associated with ECFP-based data sharing, we contribute a more conclusive demonstration by introducing a deterministic algorithm that reconstructs molecular structures from ECFPs. Using MetaNetX and eMolecules as databases of natural compounds and commercially available chemicals, the deterministic algorithm benchmarks a Transformer-based generative model trained to predict SMILES from ECFPs. The generative model achieves a top-ranked retrieval accuracy of 95.64% but struggles with exhaustive enumeration. Additionally, applying the deterministic method to a drug dataset reveals its potential for de novo drug design, as many of the reverse-engineered structures are found to be patented or supported by bioassay data.

Graphical Abstract

Repositioning drug‒disease relationships has always been a hot field of research. However, actual cases of biologically validated drug relocation remain very limited, and existing models have not yet fully utilized the structural information of the drug. Furthermore, most repositioning models are only used to complete the relationship matrix, and their practicality is poor when dealing with drug cold start problems. This paper proposes a structure-enhanced multimodal relationship prediction model (SMRP). SMPR is based on the SMILE structure of the drug, uses the MOL2VEC method to generate drug-embedded representations, and learns disease-embedded representations through heterogeneous network graph neural networks. Ultimately, a drug‒disease relationship matrix is constructed. In addition, to reduce the difficulty of use, SMPR also provides a cold start interface on the basis of structural similarity based on repositioning results to predict drug-related diseases simply and quickly. The repositioning ability and cold start capability of the model were verified from multiple perspectives. While the AUC and ACUPR scores of repositioning reached 99% and 61%, respectively, the AUC of the cold start method was 80%. In particular, the cold start recall indicator can reach more than 70%, which means that the SMPR is more sensitive to positive samples. Finally, case analysis was used to verify the practical value of the model, and visual analysis directly demonstrated the improvement in the structure of the model. For ease of use, we also provide local deployment of the model and packaged it into an executable program.

Scientific contribution

The SMPR model is a structure-enhanced multimodal learning model that focuses on the reference value of similar structures in practical docking and adds a drug embedding module. Considering that the predictive ability of the model is limited by the dataset, the model provides a cold start interface for the effective prediction of drugs that are not in the dataset. To further facilitate the use by pharmacology workers without programming knowledge and enhance its practical application value, the model was also encapsulated into a local executable program.

Structure-constrained molecular generation (SCMG) generates novel molecules that are structurally similar to a given molecule and have optimized properties. Deep learning solutions for SCMG are limited in that they are predisposed towards existing knowledge, and they suffer from a natural impedance mismatch problem due to the discrete nature of molecules, while deep learning methods for SCMG often operate in continuous space. Moreover, many task-specific evaluation metrics used during training often bias the model towards a particular metric -“metric-leakage”. To overcome these shortcomings, we propose Policy-guided Unbiased REpresentations (PURE) for SCMG that learns within a framework simulating molecular transformations for drug synthesis. PURE combines self-supervised learning with a policy-based reinforcement learning (RL) framework, thereby avoiding the need for external molecular metrics while learning high-quality representations that incorporate an inherent notion of similarity specific to the given task. Along with a semi-supervised training design, PURE utilizes template-based molecular simulations to better explore and navigate the discrete molecular search space. Despite the lack of metric biases, PURE achieves competitive or superior performance to state-of-the-art methods on multiple benchmarks. Our study emphasizes the importance of reevaluating current approaches for SCMG and developing strategies that naturally align with the problem. Finally, we illustrate how our methodology can be applied to combat drug resistance by identifying sorafenib-like compounds as a case study.

Aqueous solubility of a compound plays a crucial role throughout various stages of drug discovery and development. Despite numerous efforts using various machine learning models, accurately estimating aqueous solubility remains a challenge. One primary limitation is the absence of a single source, large dataset of druglike compounds for model training. Additionally, studies have highlighted the need for improvements in prediction algorithms and molecular representations. To address these challenges, the Johnson and Johnson (J&J) in-house solubility data was leveraged. Theoretical pH-solubility equations and in-house pKa prediction tools were utilized to calculate intrinsic solubility from J&J data. A multi-task graph transformer model was developed and trained on the calculated intrinsic solubility data of 13,306 compounds along with seven relevant physicochemical properties including solubility at pH 2/7, logP, and logD at three different pHs. When evaluated making use of high-quality test data, the developed model achieved a root mean square error (RMSE) of 0.61 and coefficient of determination (R²) of 0.60, demonstrating state-of-the-art performance in estimating intrinsic solubility for drug-like compounds.

Chemical space exploration has gained significant interest with the increasing availability of building blocks, enabling the creation of ultralarge virtual libraries containing billions or trillions of compounds. However, challenges remain in selecting the most suitable compounds for synthesis, especially in hit expansion. Thompson sampling, a probabilistic search method, has recently been proposed to improve efficiency by operating in reagent space rather than product space. Here, we address some of its limitations by introducing a roulette wheel selection method combined with a thermal cycling approach to balance greedy search and diversity-driven exploration. The effectiveness of this method is demonstrated through 109 queries against twenty distinct 1-million-compound libraries using ROCS.

Accurately classifying chemical structures is essential for cheminformatics and bioinformatics, including tasks such as identifying bioactive compounds of interest, screening molecules for toxicity to humans, finding non-organic compounds with desirable material properties, or organizing large chemical libraries for drug discovery or environmental monitoring. However, manual classification is labor-intensive and difficult to scale to large chemical databases. Existing automated approaches either rely on manually constructed classification rules, or are deep learning methods that lack explainability. This work presents an approach that uses generative artificial intelligence to automatically write chemical classifier programs for classes in the Chemical Entities of Biological Interest (ChEBI) database. These programs can be used for efficient deterministic run-time classification of SMILES structures, with natural language explanations. The programs themselves constitute an explainable computable ontological model of chemical class nomenclature, which we call the ChEBI Chemical Class Program Ontology (C3PO). We validated our approach against the ChEBI database, and compared our results against deep learning models and a naive SMARTS pattern based classifier. C3PO outperforms the naive classifier, but does not reach the performance of state of the art deep learning methods. However, C3PO has a number of strengths that complement deep learning methods, including explainability and reduced data dependence. C3PO can be used alongside deep learning classifiers to provide an explanation of the classification, where both methods agree. The programs can be used as part of the ontology development process, and iteratively refined by expert human curators.

Metabolomics data analysis is a multifaceted process often constrained by limited data sharing and a lack of transparency, which hinders reproducibility of results. While existing bioinformatics tools address some of these challenges, achieving greater simplicity and operational clarity remains essential for fully leveraging the potential of metabolomics. Here, we introduce TraceMetrix, a web-based platform designed for interactive traceability in metabolomics data analysis. TraceMetrix provides a flexible management system for both raw and derived data, enabling comprehensive tracking of file origins and destinations throughout the whole analysis pipeline. The platform documents the software and parameters used across four key modules, from raw data preprocessing, data cleaning, statistical analysis to functional analysis, enabling users to easily track critical factors influencing result accuracy. By mapping upstream and downstream relationships for nearly 19 analytical functions, TraceMetrix ensures end-to-end traceability, viewable interactively online or exportable as detailed reports. To address the limitations of single-machine environments in processing large-scale datasets, TraceMetrix is deployed on a high-performance computing cluster for efficient batch processing. Using a non-targeted metabolomics dataset, we demonstrated its traceability function to optimize parameter selection, successfully reproducing the analysis process and validating the original study's findings. TraceMetrix integrates traceability across data, software, and processes, significantly enhancing reproducibility in metabolomics research. The platform supports diverse applications and is freely available at https://www.biosino.org/tracemetrix.

Automatic extraction of molecules from scientific literature plays a crucial role in accelerating research across fields ranging from drug discovery to materials science. Patent documents, in particular, contain molecular information in visual form, which is often inaccessible through traditional text-based searches. In this work, we introduce SubGrapher, a method for the visual fingerprinting of molecule and Markush structure images. Unlike conventional Optical Chemical Structure Recognition (OCSR) models that attempt to reconstruct full molecular graphs, SubGrapher focuses on extracting fingerprints directly from images. Using learning-based instance segmentation, SubGrapher identifies functional groups and carbon backbones, constructing a substructure-based fingerprint that enables the retrieval of molecules and Markush structures. Our approach is evaluated against state-of-the-art OCSR and fingerprinting methods, demonstrating superior retrieval performance and robustness across diverse molecule and Markush structure depictions. The benchmark datasets, models, and inference code are publicly available..