Machine Learning Assisted Graphdiyne-Based Nanozyme Discovery

Abstract

Improving the catalytic activity and broadening the scope of nanozymes are prerequisites to supplement or even supersede natural enzymes. However, the discovery of nanozymes is mostly relied on serendipity with limited fine-tunings of chemical composition, which is often incomprehensive and fragmented. Machine learning (ML) is a promising solution to predict the nanozyme performance and thus accelerate the nanozyme development. Herein, a thorough investigation of the peroxidase (POD) mimic reaction catalyzed by nonmetal atom doped graphdiyne (GDY) is presented and two doped GDYs (B-GDY and N-GDY) with best performance were screened out. Specifically, the extreme gradient boosting (XGB) algorithm can mine the connection between the model parameters and maximum energy barrier (R 2 > 78%) or maximum energy consuming step (accuracy > 65%) from the data set constructed by all the nonmetal atom doped GDYs, which provides a method to effectively reduce 20% calculations. In addition, six nonmetal atom doped GDYs with different expected properties were synthesized, and their activity trends in three related experiments were consistent with the predicted results. This study demonstrates that ML can be a useful tool to assist density functional theory (DFT) computational screening and can serve as a guide for optimizing nanozyme performance, depending on doping strategy.

Theoretic Study of Sulfur-Doped Graphdiynes by X-ray Spectroscopy

“…As can be seen in Figure 4 and Table 2, the two stronger convolution peaks at 284.6 and 285.0 eV are designated as C−C(sp 2 ) and C−C(sp), respectively, which are consistent with the two peaks in the experiments. 16,20,40,41 In addition, the peak of the C−S single bond (shown as a purple spectrum in Figure 4) is convolved at 286.2 eV, which is in general agreement with the peaks assigned in the experimental report (285.9 and 286.1 eV). 16,20 The weak peak at 287.8 eV is attributed to the convolution of the C� S(O) double bond, which is slightly different from the experimental value.…”

“…In this paper, 10 different S-doped configurations were constructed based on the published literature on sulfur-doped graphdiyne. ,,,− In order to identify different doping configurations, the XPS and NEXAFS spectra of these 10 doping molecules and different bonding types of carbon atoms were simulated theoretically. The fine structure and spectral relationship of S-doped graphene were obtained by spectral analysis, which can help us to accurately identify the various sites of S-doped graphdiyne.…”

Clear‐Box Machine Learning for Virtual Screening of 2D Nanozymes to Target Tumor Hydrogen Peroxide

“…Recently, black-box machine learning models have been developed to predict the enzyme-like catalytic activities of NMs based on features of the materials and reaction conditions. [55,56] Differently, the present models are based on activity descriptors, and the machine learning algorithm is only used to replace DFT to quickly calculate the activity descriptors. Therefore, the present models take advantage of both interpretability of physics model and high efficiency of machine learning.…”

Rational Design Strategies for Nanozymes

“…Yu et al employed DFT to calculate the POD-like catalytic reaction paths of 168 GDY-based models containing different nonmetallic doping elements, sites, and concentrations, in order to establish a nonmetallic GDY data set. 47 By learning the nonmetallic GDY data set with the extreme gradient boosting (XGB) algorithm, two doped GDYs (B-GDY and N-GDY) with the best performance were screened out, the relationship between the model parameters and the maximum energy barrier (R 2 > 78%) or the maximum energy consumption step (accuracy >65%) was successfully mined, and 20% computa- tional cost reduction was achieved. In addition, feature importance and SHAP analysis disclosed that the type of doped element plays a key role in the POD-like activity.…”