Artificial intelligence chemistry最新文献

Machine Learning (ML) stands as a disruptive technology, finding application across a diverse array of scientific disciplines. When applied to homogeneous catalysis, this technology accelerates catalyst discovery through virtual screening, which not only reduces experimental iterations but also yields significant savings in time, resources, and waste generation. ML algorithms, often integrated with cheminformatic tools and quantum mechanics featurization, excel in predicting reaction outcomes that guide the engineering of catalysts for desired reactivity and selectivity. This minireview presents recent studies regarding databases as well as supervised and unsupervised problems, offering a general yet insightful perspective on the current ML-driven progress in homogeneous catalysis.

Photocatalytic reactions, achieving chemical synthesis in a more sustainable manner than thermal reactions, have been demonstrated to become more efficient, greener and easier to scale up when combined with continuous microflow technology. Nevertheless, the report on the kinetics measurement for photocatalytic reactions in continuous microflow, especially in a fully automated way, is very rare. In this work, two challenging parameters, i.e., the reaction order with respect to oxygen (2.48) and photoreaction activation energy (-16.83 kJ/mol) of the photocatalytic oxidation of 9,10-diphenylanthracene, were acquired in an automated continuous flow platform using the Steady-state Method. Moreover, the Ramping Method was also successfully implemented in the automated continuous flow photoreaction platform, exhibiting a predictive accuracy of 4.42 %, with 64.3 % less time and 58.0 % less material consumption than the Steady-state Method. And it was found that the improvement in the residence time distribution of the microreactor could improve the accuracy of the Ramping Method. The automated continuous flow process developed in this work could offer an efficient and accurate way to attain the reaction kinetics information for homogeneous photocatalytic reactions.

Metal oxides (MOs) are a class of electrocatalysts which could be the low-cost alternatives to precious metals. However, many MOs suffer from poor stability under electrochemical operating conditions. The Materials Project stands out as one of the largest computational materials databases to date, where the bulk Pourbaix diagrams are essential in assessing the aqueous stability of potential electrocatalysts. Herein, we performed data mining from the Materials Project database to identify potentially stable MOs for industrially important electrocatalytic reactions including oxygen reduction reaction (ORR), oxygen evolution reaction (OER), chlorine evolution reaction (CER), hydrogen evolution reaction (HER), and nitrogen reduction reaction (NRR). We found that many MOs can be potentially stable under electrocatalytic conditions, especially in neutral and alkaline medium. Finally, we summarized those MOs that had been previously experimentally synthesized but haven’t been explored as electrocatalysts. This comprehensive assessment effectively narrows down the exploration scope and facilitates the evaluation of material stability.

Aqueous Organic Redox Flow Batteries (AORFBs) are considered as one of the most appealing technologies for large-scale energy storage due to their electroactive organic materials, which are abundant, easy to produce, and recyclable. A prevailing challenge for the redox chemistries applied in AORFBs is to achieve high power and energy density. The chemical design and molecular engineering of the electroactive compounds is an effective approach for the optimization of their physicochemical properties. Among them, the reaction energy of redox couples is often used as a proxy for the measured potentials. In this study, we present RedPred, a machine learning (ML) model that predicts the one-step two-electron two-proton redox reaction energy of redox-active molecule pairs. RedPred comprises an ensemble of Artificial Neural Networks, Random Forests, and Graph Convolutional Networks, trained using the RedDB database, which contains over 15,000 reactant-product pairs for AORFBs. We evaluated RedPred’s performance using six different molecular encoders and five prominent ML algorithms applied in chemical science. The predictive capability of RedPred was tested on both its training chemical space and the chemical space outside its training domain using two separate test datasets. We released a user-friendly web tool with open-source code to promote software sustainability and broad use.

Artificial intelligence (AI) has huge potential to accelerate drug discovery, but challenges remain in implementing AI algorithms that can be used by the broad scientific community. Identification of molecular features and their subsequent use in training of machine learning models may permit prediction of new molecules with enhanced properties. Predictive modeling is particularly applicable to analysis of structure-activity relationships (SARs) and would be a useful tool in the hands of laboratory medicinal chemists. This requires a software platform that is chemically intuitive while providing the user with access to AI methods. The KNIME platform provides such an environment through inclusion of broad chemical toolsets and a user-friendly approach for utilization of machine learning for analysis of SAR data. Here, we illustrate use of KNIME for this purpose, with a focus on discovery of features of highly potent tau inhibitors from a series of structurally diverse polyphenols. Workflows are described that enable implementation of AI tools in KNIME for diverse SAR projects.

Fe-based Fischer-Tropsch Synthesis (FTS) enables the selective conversion of syngas into long-chain hydrocarbons, which can be further refined to produce highly demanded liquid fuels and high-value chemical products. However, developing novel heterogeneous catalysts for FTS with desirable performance characteristics is a challenging task, as their performance depends on various factors such as precursor, support material, promoters, pretreatment conditions and the catalyst structures. Thus, it remains difficult to understand the structure-performance relationship of FTS and to optimize the catalyst formulations and operating conditions rationally. By integrating traditional chemistry with machine learning, we herein establish intrinsic correlations among reduction, reaction conditions, phase information and the methane selectivity of Fe-based FTS, using high quality experimental data. The content of the iron phases in the post-reaction phase, particularly χ-Fe₅C₂, significantly influences the methane selectivity of the catalyst. Four types of additives K, Cu, SiO₂, and Ca could effectively suppress the methane selectivity, most likely by promoting or stabilizing the iron carbide phases, indicated by their strong correlation. The machine learned structure-performance relationships offers new insights into the design of Fe-based FTS catalysts, and could guide the further optimization of the preprocessing conditions and various parameter factors to minimize the methane selectivity of FTS.

The application of artificial intelligence to chemistry usually focuses on the identification of good correlations between descriptors and a given property of interest. The descriptors often come from arbitrary sets, with the implicit assumption that the evaluation of a sufficiently wide range of descriptors will lead to a satisfactory choice. Recent work in our group has focused on applying statistical analysis to large amounts of DFT results with the goal of finding optimal descriptor sets for a given property, which we label as hidden descriptors. This article briefly discusses this treatment and the chemical knowledge that has been gained through its application in two different domains: metal-ligand bond strength in transition metal complexes, and energy barriers in bimolecular nucleophilic substitution reactions.

Drug repurposing is an approach to identifying new uses for existing drugs, where advanced computational methods, such as text and graph embedding techniques, are playing an ever-increasing role. This review provides a timely overview of these embedding methods for drug repurposing and discusses their integration with machine learning. Text embedding techniques, such as Word2Vec, FastText, BERT, and Doc2Vec, enable the analysis of biomedical literature and clinical data to discover potential drug-disease relationships. These methods convert textual data into numerical representations, allowing for similarity calculations and predictive modeling. Several successful applications of text embedding for drug repurposing are highlighted. In addition, graph embedding methods, such as Node2Vec and GraphSAGE, are being employed to convert complex biological knowledge graphs into vector representations. These representations facilitate various network analysis tasks, including predicting drug-target interactions and identifying hidden associations between drugs and diseases. Case studies in both technologies demonstrate their effectiveness in drug repurposing. The advantages and limitations of both text and graph embedding technologies, and their complementarity with traditional structure-based approaches have been discussed. Finally, text and graph embedding methods can be employed in conjunction with traditional approaches of computational methods, which can offer a promising path to identifying novel drug repurposing opportunities, particularly for rare diseases.

The choice of a proper machine learning (ML) algorithm for constructing potential energy surface (PES) models has become a crucial tool in the fields of quantum chemistry and computational modeling. These algorithms offer the ability to make reliable and accurate predictions at a reasonable computational cost, and thus they can be then used in various molecular dynamics and spectroscopic studies. For that, it is not surprising that much of the current research focuses on the development of software that generates machine learning models using precalculated ab initio data points. This study is primarily dedicated to the application and assessment of various automated learning models. These models are trained and tested using datasets derived from CCSD(T)/CBS[56] calculations, aiming to represent intermolecular interactions in small molecules, such as the NgH${}_2^{+}$ complexes, where Ng represents helium (He), neon (Ne), and argon (Ar) atoms. These noble gas-containing molecules have gained increasing significance in the field of molecular astrochemistry, due to the recent discovery of HeH⁺ and ArH⁺ molecular cations in the interstellar medium (ISM), thereby opening up a wide range of possibilities in this scientific area. Consequently, the ML-generated PESs are employed to compute vibrational bound states for these molecular cations, with the goal of characterizing all their known isotopologues. Furthermore, the results are compared with spectroscopic data, when available, from previous studies in the literature. Our findings have the potential to provide valuable guidance for future ML-PES development and benchmarking studies involving noble gas-containing cations of astrophysical importance.

We present SOmicsFusion, a software toolbox for ’fusing’ spatial omics with classical biomedical imaging modalities, capitalizing on their inherent correspondences and complementarity when characterizing the same subject. By augmenting radiological and histological images with spatially resolved molecular profiling, this fusion offers a panoramic characterization of the biochemical perturbations underlying pathological conditions, thereby advancing our understanding of diseases like brain disorders and cancers. The cornerstone of SOmicsFusion is a coregistration tool that leverages an innovative two-stage machine learning pipeline to tackle the longstanding challenge of spatially aligning data from fundamentally different modalities, priming them for subsequent fusion analysis that often requires precise pixel-wise correspondence between the datasets. Specifically, the pipeline utilizes an original dimension reduction algorithm for representational domain alignment, followed by a Deep Learning-based method for spatial domain alignment. SOmicsFusion is demonstrated using mass spectrometry imaging (MSI)-mediated spatial metabolomics and four other modalities: magnetic resonance imaging (MRI), microscopy, brain atlas, and spatial transcriptomics. By reducing coregistration errors by 38–69% compared to existing pipelines, SOmicsFusion enhances the precision of associating molecule distribution with anatomy and pathology features, ultimately leading to more statistically robust findings. Furthermore, SOmicsFusion incorporates various downstream analysis tools, including overlay visualization, spatial correlation/co-expression analysis, pansharpening, and automated anatomy annotation. These tools facilitate the extraction of biological insights that would be unattainable through individual modalities alone. For instance, the coregistration and correlation between MSI and in vivo MRI datasets unveil that the spatial heterogeneity in metabolites stems from the temporal heterogeneity in the development of cerebral ischemia-reperfusion injury.