Molecular Informatics最新文献

The challenges of drug discovery from hit identification to clinical development sometimes involves addressing scaffold hopping issues, in order to optimise molecular biological activity or ADME properties, or mitigate toxicology concerns of a drug candidate. Docking is usually viewed as the method of choice for identification of isofunctional molecules, i. e. highly dissimilar molecules that share common binding modes with a protein target. However, the structure of the protein may not be suitable for docking because of a low resolution, or may even be unknown. This problem is frequently encountered in the case of membrane proteins, although they constitute an important category of the druggable proteome. In such cases, ligand-based approaches offer promise but are often inadequate to handle large-step scaffold hopping, because they usually rely on molecular structure. Therefore, we propose the Interaction Fingerprints Profile (IFPP), a molecular representation that captures molecules binding modes based on docking experiments against a panel of diverse high-quality proteins structures. Evaluation on the LH benchmark demonstrates the interest of IFPP for identification of isofunctional molecules. Nevertheless, computation of IFPPs is expensive, which limits its scalability for screening very large molecular libraries. We propose to overcome this limitation by leveraging Metric Learning approaches, allowing fast estimation of molecules IFPP similarities, thus providing an efficient pre-screening strategy that in applicable to very large molecular libraries. Overall, our results suggest that IFPP provides an interesting and complementary tool alongside existing methods, in order to address challenging scaffold hopping problems effectively in drug discovery.

Visualization of the combinatorial library chemical space provides a comprehensive overview of available compound classes, their diversity, and physicochemical property distribution - key factors in drug discovery. Typically, this visualization requires time- and resource-consuming compound enumeration, standardization, descriptor calculation, and dimensionality reduction. In this study, we present the Combinatorial Library Neural Network (CoLiNN) designed to predict the projection of compounds on a 2D chemical space map using only their building blocks and reaction information, thus eliminating the need for compound enumeration. Trained on 2.5 K virtual DNA-Encoded Libraries (DELs), CoLiNN demonstrated high predictive performance, accurately predicting the compound position on Generative Topographic Maps (GTMs). GTMs predicted by CoLiNN were found very similar to the maps built for enumerated structures. In the library comparison task, we compared the GTMs of DELs and the ChEMBL database. The similarity-based DELs/ChEMBL rankings obtained with "true" and CoLiNN predicted GTMs were consistent. Therefore, CoLiNN has the potential to become the go-to tool for combinatorial compound library design - it can explore the library design space more efficiently by skipping the compound enumeration.

In fragrance development, the framework development process is a bottleneck from the perspective of labor, cost, and human resource development. Odors vary greatly depending on the structure and functional groups of the molecule. Although odor has been predicted from only the structure of molecules, its practical application remains elusive. In this study, we developed a model for predicting the odor of molecules that have only small differences in structure. Focusing on the mechanism of human olfaction, we divided the mechanism into three levels and constructed three models: a classification model that predicts the presence or absence of binding between molecules and olfactory receptors, a regression model that predicts the strength of binding, and a classification model that predicts the presence or absence of odor based on the strength of binding. Olfactory receptors were used as descriptors to discriminate between similar molecular odors. Our models predicted odor differences between some similar molecules, including optical isomers.

Feature attribution methods from explainable artificial intelligence (XAI) provide explanations of machine learning models by quantifying feature importance for predictions of test instances. While features determining individual predictions have frequently been identified in machine learning applications, the consistency of feature importance-based explanations of machine learning models using different attribution methods has not been thoroughly investigated. We have systematically compared model explanations in molecular machine learning. Therefore, a test system of highly accurate compound activity predictions for different targets using different machine learning methods was generated. For these predictions, explanations were computed using methodological variants of the Shapley value formalism, a popular feature attribution approach in machine learning adapted from game theory. Predictions of each model were assessed using a model-agnostic and model-specific Shapley value-based method. The resulting feature importance distributions were characterized and compared by a global statistical analysis using diverse measures. Unexpectedly, methodological variants for Shapley value calculations yielded distinct feature importance distributions for highly accurate predictions. There was only little agreement between alternative model explanations. Our findings suggest that feature importance-based explanations of machine learning predictions should include an assessment of consistency using alternative methods.

Within a recent decade, graph neural network (GNN) has emerged as a powerful neural architecture for various graph-structured data modelling and task-driven representation learning problems. Recent studies have highlighted the remarkable capabilities of GNNs in handling complex graph representation learning tasks, achieving state-of-the-art results in node/graph classification, regression, and generation. However, most traditional GNN-based architectures like GCN and GraphSAGE still faced several challenges related to the capability of preserving the multi-scaled topological structures. These models primarily focus on capturing local neighborhood information, often failing to retain global structural features essential for graph-level representation and classification tasks. Furthermore, their expressiveness is limited when learning topological structures in complex molecular graph datasets. To overcome these limitations, in this paper, we proposed a novel graph neural architecture which is an integration between neuro-fuzzy network and topological graph learning approach, naming as: FTPG. Specifically, within our proposed FTPG model, we introduce a novel approach to molecular graph representation and property prediction by integrating multi-scaled topological graph learning with advanced neural components. The architecture employs separate graph neural learning modules to effectively capture both local graph-based structures as well as global topological features. Moreover, to further address feature uncertainty in the global-view representation, a multi-layered neuro-fuzzy network is incorporated within our model to enhance the robustness and expressiveness of the learned molecular graph embeddings. This combinatorial approach can assist to leverage the strengths of multi-view and multi-modal neural learning, enabling FTPG to deliver superior performance in molecular graph tasks. Extensive experiments on real-world/benchmark molecular datasets demonstrate the effectiveness of our proposed FTPG model. It consistently outperforms state-of-the-art GNN-based baselines categorized in different approaches, including canonical local proximity message passing based, graph transformer-based, and topology-driven approaches.

The human epidermal growth factor receptor 2 (HER2) is a critical oncogene implicated in the development of various aggressive cancers, particularly breast cancer. Discovering novel HER2 inhibitors is crucial for expanding therapeutic options for HER2-related malignancies. In this study, we present a computational workflow that focuses on generating pharmacophores derived from docked poses of a selected list of 15 diverse, potent HER2 inhibitors, utilizing flexible docking. The resulting pharmacophores, along with other physicochemical molecular descriptors, were then evaluated in a machine learning-quantitative structure-activity relationship (ML-QSAR) analysis against 1,272 HER2 inhibitors. Several machine learning methods were assessed, and a genetic function algorithm (GFA) was employed for feature selection. Ultimately, GFA combined with Bagging and J48Graft classifiers produced the best self-consistent and predictive models. These models highlighted the significance of two pharmacophores, Hypo_1 and Hypo_2, in distinguishing potent from less active inhibitors. The successful ML-QSAR models and their associated pharmacophores were used to screen the National Cancer Institute (NCI) database for novel HER2 inhibitors. Three promising anti-HER2 leads were identified, with the top-performing lead demonstrating an experimental anti-HER2 IC₅₀ value of 3.85 μM. Notably, the three inhibitors exhibited distinct chemical scaffolds compared to existing HER2 inhibitors, as indicated by principal component analysis.

Herein, we introduce MAYA (Multiple Activity Analyzer), a tool designed to automatically construct a chemical multiverse, generating multiple visualizations of chemical spaces of a compound data set described by structural descriptors of different nature such as Molecular ACCess Systems (MACCS) keys, extended connectivity fingerprints with different radius, molecular descriptors with pharmaceutical relevance, and bioactivity descriptors. These representations are integrated with various data visualization techniques for the automated analysis focused on structure - multiple activity/property relationships, enabling analysis for various problems set in user-friendly source software. The source code of MAYA is freely available on GitHub at https://github.com/IsrC11/MAYA.git.

Reactions commonly used in synthetic organic chemistry are named after their discoverers or developers. They are called the name reactions and generally consist of several elementary reactions. Quantum chemical calculations can optimize transition state (TS) structures of the elementary reactions. The geometrical feature of TS is called TS motif. We have constructed a database (QMRDB) with the TS motif information and have been continuing to accumulate them. In the present study, we extracted 102 elementary reactions from the QMRDB and attempted to classify them using the Kohonen self-organization map. As the results, all the TS motifs were clustered. By firing a target compound on a Kohonen map generated, we expect to be able to easily find the TS motifs most similar to the target.

Currently, numerous metrics allow chemists and computational chemists to refine and filter libraries of virtual molecules in order to prioritize their synthesis. Some of the most commonly used metrics and models are QSAR models, docking scores, diverse druggability metrics, and synthetic feasibility scores to name only a few. To our knowledge, among the known metrics, a function which estimates the price of a novel virtual molecule and which takes into account the availability and price of starting materials has not been considered before in literature. Being able to make such a prediction could improve and accelerate the decision-making process related to the cost-of-goods. Taking advantage of recent advances in the field of Computer Aided Synthetic Planning (CASP), we decided to investigate if the predicted retrosynthetic pathways of a given molecule and the prices of its associated starting materials could be good features to predict the price of that compound. In this work, we present a deep learning model, RetroPriceNet, that predicts the price of molecules using their predicted synthetic pathways. On a holdout test set, the model achieves better performance than the state-of-the-art model. The developed approach takes into account the synthetic feasibility of molecules and the availability and prices of the starting materials.

The widespread use of polymer solutions in the chemical industry poses a significant challenge in determining optimal dissolution conditions. Traditionally, researchers have relied on experimental methods to estimate the processing parameters needed to dissolve polymers, often requiring numerous iterations of testing different temperatures and pressures. This approach is both costly and time-consuming. In this study, for the first time, we present a machine learning-based approach to predict the minimum temperature and pressure required for polymer dissolution, correlating molecular weight and chemical structure of both the polymer and solvent and its weight percent. Using a dataset compiled from existing literature, which includes key factors influencing polymer dissolution, we also extracted chemical bond information from the molecular structures of polymer-solvent systems. Six different machine learning algorithms, including linear regression, k-nearest neighbors, regression trees, random forests, multilayer perceptron neural networks, and support vector regression, were employed to develop predictive models. Among these, the Random Forest model achieved the highest accuracy, with R² values of 0.931 and 0.942 for temperature and pressure predictions, respectively. This novel approach eliminates the need for repetitive experimental testing, offering a more efficient pathway to determining dissolution conditions.