Artificial intelligence chemistry最新文献

In this review, we explore the integration of intelligent algorithms in chemistry and materials science.We begin by delineating the core principles of Machine Learning, Deep Learning, and optimization algorithms, highlighting their bespoke adaptation to these scientific domains. The focus then shifts to the critical processes of data management, including collection, refinement, and feature engineering, alongside strategies for efficient data mining from targeted databases and literatures. Subsequently, we present a concise overview of the diverse applications of these algorithms, emphasizing their transformative impact in both fields. Finally, this review explores the future prospects and challenges of these emerging algorithms.

In computer-aided drug discovery, accurately determining the structure and properties of drug-like molecules is of utmost importance. This necessitates the use of precise and efficient electronic structure methods. Here, we developed two deep learning-based density functional methods, namely DeePHF and DeePKS, specifically tailored for drug-like molecules. Notably, DeePKS incorporates self-consistency into its framework. With a limited dataset labelled at the CCSD(T)/def2-TZVP level, both models have been able to achieve chemical accuracy in calculating molecular energies and have demonstrated excellent transferability. We anticipate that further advancements in this field will lead to the development of high-quality density functional methods designed specifically for drug discovery purposes. This research showcases the capabilities of deep learning approaches in simplifying the construction complexity associated with traditional DFT methods.

Developing full-dimensional machine-learned potentials with the current “gold-standard” coupled-cluster (CC) level is challenging for medium-sized molecules due to the high computational cost. Consequently, researchers are often bound to use lower-level electronic structure methods such as density functional theory or second-order Møller–Plesset perturbation theory (MP2). Here, we demonstrate on a representative example that gold-standard potentials can now be effectively constructed for molecules of 15 atoms using off-the-shelf hardware. This is achieved by accelerating the CCSD(T) computations via the accurate and cost-effective frozen natural orbital (FNO) approach. The Δ-machine learning (Δ-ML) approach is employed with the use of permutationally invariant polynomials to fit a full-dimensional potential energy surface of the acetylacetone molecule, but any other effective descriptor and ML approach can similarly benefit from the accelerated data generation proposed here. Our benchmarks for the global minima, H-transfer TS, and many high-lying configurations show the excellent agreement of FNO-CCSD(T) results with conventional CCSD(T) while achieving a significant time advantage of about a factor of 30–40. The obtained Δ-ML PES shows high fidelity from multiple perspectives including energetic, structural, and vibrational properties. We obtain the symmetric double well H-transfer barrier of 3.15 kcal/mol in excellent agreement with the direct FNO-CCSD(T) barrier of 3.11 kcal/mol as well as with the benchmark CCSD(F12*)(T+)/CBS value of 3.21 kcal/mol. Furthermore, the tunneling splitting due to H-atom transfer is calculated using a 1D double-well potential, providing improved estimates over previous ones obtained using an MP2-based PES. The methodology introduced here represents a significant advancement in the efficient and precise construction of potentials at the CCSD(T) level for molecules above the current limit of 15 atoms.

The effect of noise in the input data for learning potential energy surfaces (PESs) based on neural networks for chemical applications is assessed. Noise in energies and forces can result from aleatoric and epistemic errors in the quantum chemical reference calculations. Statistical (aleatoric) noise arises for example due to the need to set convergence thresholds in the self consistent field (SCF) iterations whereas systematic (epistemic) noise is due to, inter alia, particular choices of basis sets in the calculations. The two molecules considered here as proxies are H₂CO and HONO which are examples for single- and multi-reference problems, respectively, for geometries around the minimum energy structure. For H₂CO it is found that adding noise to energies and forces with magnitudes representative of single-point calculations does not deteriorate the quality of the final PESs whereas increasing the noise level commensurate with electronic structure calculations for more complicated, e.g. metal-containing, systems is expected to have a more notable effect. On the other hand, for HONO which requires a multi-reference treatment, a clear correlation between model quality and the degree of multi-reference character as measured by the T₁ amplitude is found. It is concluded that for chemically “simple” cases the effect of aleatoric and epistemic errors is manageable without evident deterioration of the trained model, but more care needs to be exercised for situations in which multi-reference effects are present.

Gaining insights into overarching trends in chemical reaction systems is crucial for refining reaction conditions and developing novel reactions. These knowledgements include preferences for certain reagents, solvents, and functional group tolerance rules. Traditionally, synthetic chemists have relied on extensive literature searching to acquire the knowledge, a process that is both time-consuming and laborious. To streamline this process, we construct a standardized dataset and knowledge graph on an emerging domain, transition-metal-free transformations with organoborons. The dataset, compiled from organic reaction literature, includes comprehensive details of reaction scopes and conditions. The subsequent construction of a knowledge graph offers a visual representation of the reactions and their interrelationships. Through knowledge graph-based hierarchical analysis and density functional theory (DFT) calculations, we revealed the currently most frequently used reactants, synthetic conditions, and functional group rules in this field. We anticipate this knowledge graph-based approach will accelerate the acquisition and transfer of chemical reaction knowledge, catalyzing the discovery of new reactions. This work provides an automatic and adaptive framework for extracting key insights from reaction datasets to inform the design of novel reactions.

This paper presents a new approach for protein generation based on one-shot learning and hybrid quantum neural networks. Given a single protein complex, the system learns how to predict the remaining unknown properties, without resorting to autoregression, from the physicochemical properties of the receptor and a prior on the physicochemical properties of the ligand. In contrast with other approaches, QuantumBound learns from a single instance, not from a large dataset, as is common in deep learning. By knowing half of the properties of the ligand, the system can predict the remaining half with an average relative error of 1.43% for a dataset consisting of one hundred and twenty Covid-19 spikes complexes. To the best of our knowledge, this is the first time that one-shot learning and hybrid quantum computing have been applied to protein generation.

The knowledge of chemical reactivity of substrates is a prerequisite to accurately design a chemical reaction; however, it has been a challenging task due to the slow trial-and-error experimental approaches and the high computational cost associated with in silico investigations. Artificial intelligence techniques could serve as an alternative to efficiently determine the relative reactivity of chemical entities. In the context of this research, we propose an artificial neural network model to predict the bond dissociation energies of hypervalent iodine reagents. An open-source cheminformatics package, namely, Mordred, was employed for calculating various 1D, 2D and topological descriptors. The approach utilizes a dataset of more than 1000 hypervalent iodine reagents, and the bond dissociation energies can be predicted with a remarkable accuracy, as suggested by an R² score of 0.97 and a mean absolute error of 1.96 kcal/mol. Owing to the low cost and high efficiency, this machine learning approach can provide an alternative to the theoretical/experimental approaches to rationally design a chemical reaction and without having to go through the hassle of high-throughput experimentation to reach the desired reaction outcome. In an effort to make the model interpretable, a feature importance algorithm was applied, which identified descriptors contributing most to the development of the model. Features describing electronegativity and polarizability are some of the important contributors to the model’s training.

Geometric information of molecules is closely related to their properties, and vibrational spectroscopy, as a common and powerful analytical tool for determining molecular structure, can assist in gaining precise geometric information. Traditional methods used to delineate spectrum-structure correlations are often expensive, time-consuming, and require extensive professional expertise. In this work, we used a machine learning protocol to construct a map from spectra to molecular geometric structures, and employed Grad-CAM, a convolutional network interpretation technology, to analyze which kinds of chemical information are important for determining our model’s results. The results obtained for six small molecules of differing structures demonstrate that the model is capable of (1) extracting the crucial spectral features that are vital to downstream tasks without necessitating any manual preprocessing, and (2) enabling retrieval of molecular structural information with high precision.

To accelerate the discovery of novel drug candidates for Coronavirus Disease 2019 (COVID-19) therapeutics, we reported a series of machine learning (ML)-based models to accurately predict the anti-SARS-CoV-2 activities of screening compounds. We explored 6 popular ML algorithms in combination with 15 molecular descriptors for molecular structures from 9 screening assays in the COVID-19 OpenData Portal hosted by NCATS. As a result, the models constructed by k-nearest neighbors (KNN) using the molecular descriptor GAFF+RDKit achieved the best overall performance with the highest average accuracy of 0.68 and relatively high average area under the receiver operating characteristic curve of 0.74, better than other ML algorithms. Meanwhile, The KNN model for all assays using GAFF+RDKit descriptor outperformed using other descriptors. The overall performance of our developed models was better than REDIAL-2020 (R). A web server (https://clickff.org/amberweb/covid-19-cp) was developed to enable users to predict anti-SARS-CoV-2 activities of arbitrary compounds using the COVID-19-CP (P) models. Besides the descriptor-based machine learning models, we also developed graph-based Attentive FP (A) models for the 9 assays. We found that the Attentive FP models achieved a comparable performance to that of COVID-19-CP and outperformed the REDIAL-2020 models. The consensus prediction utilizing both COVID-19-CP and Attentive FP can significantly boost the prediction accuracy as assessed by comparing its performance with other three individual models (R, P, A) utilizing the Wilcoxon signed-rank test, thus can ultimately improve the success rate of COVID-19 drug discovery.

The challenge of data-driven synthesis of advanced nanomaterials can be minimized by using machine learning algorithms to optimize synthesis parameters and expedite the innovation process. In this study, a high-throughput robotic platform was employed to synthesize over 1356 gold nanorods with varying aspect ratios via a seedless approach. The developed models guided us in synthesizing gold nanorods with customized morphology, resulting in highly repeatable morphological yield with quantifiable structure-modulating precursor adjustments. The study provides insight into the dynamic relationships between key structure-modulating precursors and the structural morphology of gold nanorods based on the expected aspect ratio. The high-throughput robotic platform-fabricated gold nanorods demonstrated precise aspect ratio control when spectrophotometrically investigated and further validated with the transmission electron microscopy characterization. These findings demonstrate the potential of high-throughput robot-assisted synthesis and machine learning in the synthesis optimization of gold nanorods and aided in the development of models that can aid such synthesis of as-desired gold nanorods.