Journal of Chemical Information and Modeling 最新文献

Graphs are one of the most natural and powerful representations available for molecules; natural because they have an intuitive correspondence to skeletal formulas, the language used by chemists worldwide, and powerful, because they are highly expressive both globally (molecular topology) and locally (atom and bond properties). Graph kernels are used to transform molecular graphs into fixed-length vectors, which, based on their capacity of measuring similarity, can be used as fingerprints for machine learning (ML). To date, graph kernels have mostly focused on the atomic nodes of the graph. In this work, we developed a graph kernel based on atom-atom, bond-bond, and bond-atom (AABBA) autocorrelations. The resulting vector representations were tested on regression ML tasks on a data set of transition metal complexes; a benchmark motivated by the higher complexity of these compounds relative to organic molecules. In particular, we tested different flavors of the AABBA kernel in the prediction of the energy barriers and bond distances of the Vaska's complex data set (Friederich et al., Chem. Sci., 2020, 11, 4584). For a variety of ML models, including neural networks, gradient boosting machines, and Gaussian processes, we showed that AABBA outperforms the baseline including only atom-atom autocorrelations. Dimensionality reduction studies also showed that the bond-bond and bond-atom autocorrelations yield many of the most relevant features. We believe that the AABBA graph kernel can accelerate the exploration of large chemical spaces and inspire novel molecular representations in which both atomic and bond properties play an important role.

Detecting doping agents in sports poses a significant challenge due to the continuous emergence of new prohibited substances and methods. Traditional detection methods primarily rely on targeted analysis, which is often labor-intensive and is susceptible to errors. In response, machine learning offers a transformative approach to enhancing doping screening and detection. With its powerful data analysis capabilities, machine learning enables the rapid identification of patterns and features in complex compound data, increasing both the efficiency and the accuracy of detection. Moreover, when integrated with nontargeted metabolomics, machine learning can predict unknown metabolites, aiding the discovery of long-lasting biomarkers of doping. It also excels in classifying novel compounds, thereby reducing false-negative rates. As instrumental analysis and machine learning technologies continue to advance, the development of rapid, scalable, and highly efficient doping detection methods becomes increasingly feasible, supporting the pursuit of fairness and integrity in sports competitions.

Protein-ligand binding affinity prediction is a crucial and challenging task in the field of drug discovery. However, traditional simulation-based computational approaches are often prohibitively time-consuming, limiting their practical utility. In this study, we introduce a novel deep learning method, CPIScore, which leverages the capabilities of Transformer and Graph Convolutional Networks (GCN) to enhance the prediction of protein-ligand binding affinity. CPIScore utilizes the Transformer architecture to capture comprehensive global contexts of protein and ligand sequences, while the GCN component effectively extracts local features from small molecular graphs. Our results demonstrate that CPIScore surpasses both traditional machine learning and other deep learning models in accuracy, achieving a Pearson's r of 0.74 on our test set. Furthermore, CPIScore has been validated across multiple targets, proving its ability to discern inhibitors from a diverse compound library with high enrichment rates. Notably, when applied to a generated focused library of compounds, CPIScore successfully identified six potent small-molecule inhibitors of ATR, which were tested experimentally and four small molecules exhibited inhibitory activity below ten nanomoles. These results highlight CPIScore's potential to significantly streamline and enhance the efficiency of drug discovery processes.

In this study, we introduced Matini-Net, which is a versatile framework for feature engineering and automated architecture design for materials informatics research using deep neural networks. Matini-Net provides the flexibility to design feature-based, graph-based, and combinations of these models, accommodating both single- and multimodal model architectures. For validation, we performed a performance evaluation on the MatBench benchmarking dataset of five properties, targeting five types of regression architectures that can be designed using Matini-Net. When applied to each of the five material property datasets, the best model performance for the various architectures exhibited R² > 0.84. This highlights the usefulness and flexibility of Matini-Net for accelerating materials discovery. Specifically, this framework was developed for researchers with limited experience in deep learning to easily apply it to research through automated feature engineering, hyperparameter tuning, and network construction. Moreover, Matini-Net improves the model interpretability by performing an importance analysis of the selected features. We believe that by employing Matini-Net, machine and deep learning can be applied more easily and effectively in various types of materials research.

With the advancement of deep learning (DL) methods in chemistry and materials science, the interpretability of DL models has become a critical issue in elucidating quantitative (molecular) structure-property relationships. Although attention mechanisms have been generally employed to explain the importance of molecular substructures that contribute to molecular properties, their interpretability remains limited. In this work, we introduce a versatile segmentation method and develop an interpretable subgraph attention (ISA) network with positive and negative streams (ISA-PN) to enhance the understanding of molecular structure-property relationships. The predictive performance of the ISA models was validated using data sets for aqueous solubility, lipophilicity, and melting temperature, with a particular focus on evaluating interpretability for the aqueous solubility data set. The ISA-PN model enables the quantification of the contributions of molecular substructures through positive and negative attention scores. Comparative analyses of the ISA, ISA-PN, and GC-Net (group contribution network) models demonstrate that the ISA-PN model significantly improves interpretability while maintaining similar accuracy levels. This study highlights the efficacy of the ISA-PN model in providing meaningful insights into the contributions of molecular substructures to molecular properties, thereby enhancing the interpretability of DL models in chemical applications.

Proteins are dynamic macromolecules. Knowledge of a protein's thermally accessible conformations is critical to determining important transitions and designing therapeutics. Accessible conformations are highly constrained by a protein's structure such that concerted structural changes due to external perturbations likely track intrinsic conformational transitions. These transitions can be thought of as paths through a conformational landscape. Crystallographic drug fragment screens are high-throughput perturbation experiments, in which thousands of crystals of a drug target are soaked with small-molecule drug precursors (fragments) and examined for fragment binding, mapping potential drug binding sites on the target protein. Here, we describe an open-source Python package, COnformational LAndscape Visualization (COLAV), to infer conformational landscapes from such large-scale crystallographic perturbation studies. We apply COLAV to drug fragment screens of two medically important systems: protein tyrosine phosphatase 1B (PTP1B), which regulates insulin signaling, and the SARS CoV-2 Main Protease (MPro). With enough fragment-bound structures, we find that such drug screens enable detailed mapping of proteins' conformational landscapes.

Free energy perturbation (FEP) methodologies have become commonplace methods for modeling potency in hit-to-lead and lead optimization stages of drug discovery. The conformational states of the initial poses of compounds for FEP+ calculations are often set up by alignment to a cocrystal structure ligand, but it is not clear if this method provides the best result for all proteins or all ligands. Not only are ligand conformational states potential variables in modeling compound potency in FEP but also the selection of crystallographic water molecules for inclusion in the FEP input structures can impact FEP models. Here, we report the results of FEP calculations using FEP+ from Schrödinger and starting from maximum common substructure alignment and docked poses generated with an array of docking methodologies. As a benchmark data set, we use monoacylglycerol lipase (MAGL), an important clinical drug target in cancer malignancy, neurological diseases, and metabolic disorders, and a set of 17 MAGL inhibitors. We found a large variation among FEP+ correlations to experimental IC₅₀ values depending on the method used to generate the input pose and that the inclusion of ligand-based information in the docking process, with some methods, increases the correlation between FEP+ free energies and IC₅₀ values. Upon analysis of the initial poses, we found that the differences in FEP+ correlations stemmed from rotation around a tertiary amide bond as well as translation of the compound toward the more hydrophobic side of the MAGL pocket. FEP+ estimation improved across all pose modeling methods when hydrogen bond constraint information was added. However, simple maximum common substructure alignment in the presence of all crystallographic water molecules outperformed all other methods in correlation between estimated and experimental IC₅₀ values. Taken together, these findings suggest that pose selection and crystallographic water inclusion greatly impact how well FEP+ estimated IC₅₀ values align with experimental IC₅₀ values and that modelers should benchmark a few different pose generation methodologies and different water inclusion strategies for their hit-to-lead and lead optimization drug discovery projects.

Identifying drug-target interactions (DTIs) is essential for drug discovery and development. Existing deep learning approaches to DTI prediction often employ powerful feature encoders to represent drugs and targets holistically, which usually cause significant redundancy and noise by neglecting the restricted binding regions. Furthermore, many previous DTI networks ignore or simplify the complex intermolecular interaction process involving diverse binding types, which significantly limits both predictive ability and interpretability. We propose ReduMixDTI, an end-to-end model that addresses feature redundancy and explicitly captures complex local interactions for DTI prediction. In this study, drug and target features are encoded by using graph neural networks and convolutional neural networks, respectively. These features are refined from channel and spatial perspectives to enhance the representations. The proposed attention mechanism explicitly models pairwise interactions between drug and target substructures, improving the model's understanding of binding processes. In extensive comparisons with seven state-of-the-art methods, ReduMixDTI demonstrates superior performance across three benchmark data sets and external test sets reflecting real-world scenarios. Additionally, we perform comprehensive ablation studies and visualize protein attention weights to enhance the interpretability. The results confirm that ReduMixDTI serves as a robust and interpretable model for reducing feature redundancy, contributing to advances in DTI prediction.

In medicinal chemistry, compound optimization relies on the generation of analogue series (AS) for exploring structure-activity relationships (SARs). Potency progression is a critical criterion for advancing AS. During optimization, a key question is which analogues to synthesize next. We introduce a new computational methodology for the extension of AS with potent compounds containing both core structure and substituent modifications at multiple sites, which has been reported for the first time. The approach combines a transformer chemical language model (CLM) with a SAR matrix (SARM) methodology that identifies and organizes structurally related AS. Therefore, the SARM approach was expanded to cover multisite AS. Consensus series extracted from SARMs representing a potency gradient served as input for CLM training to extend test AS with potent analogues. Different model variants were derived and investigated. Both general and fine-tuned models correctly predicted known potent analogues at high positions in probability-based compound rankings and chemically diversified AS through core structure modifications of the generated candidate compounds and substituent replacements at multiple sites.

The acid dissociation constant (pK_a), which quantifies the propensity for a solute to donate a proton to its solvent, is crucial for drug design and synthesis, environmental fate studies, chemical manufacturing, and many other fields. Unfortunately, the terminology used for describing acid-base phenomena is sometimes inconsistent, causing large potential for misinterpretation. In this work, we examine a systematic confusion underlying the definition of "acidic" and "basic" pK_a values for zwitterionic compounds. Due to this confusion, some pK_a data are misrepresented in data repositories, including the widely used and highly trusted ChEMBL database. Such datasets are frequently used to supply training data for pK_a prediction models, and hence, confusion and errors in the data make the model performance worse. Herein, we discuss the intricacies of this issue. We make suggestions for describing acid-base phenomena, training pK_a prediction models, and stewarding pK_a datasets, given the high potential for confusion and potentially high impact in downstream applications.