Cytotoxicity Induction by the Oxidative Reactivity of Nanoparticles Revealed by a Combinatorial GNP Library with Diverse Redox Properties

It is crucial to establish relationship between nanoparticle structures (or properties) and nanotoxicity. Previous investigations have shown that a nanoparticle’s size, shape, surface and core materials all impact its toxicity. However, the relationship between the redox property of nanoparticles and their toxicity has not been established when all other nanoparticle properties are identical. Here, by synthesizing an 80-membered combinatorial gold nanoparticle (GNP) library with diverse redox properties, we systematically explored this causal relationship. The compelling results revealed that the oxidative reactivity of GNPs, rather than their other physicochemical properties, directly caused cytotoxicity via induction of cellular oxidative stress. Our results show that the redox diversity of nanoparticles is regulated by GNPs modified with redox reactive ligands.


Introduction
The rapid increase in the application of nanomaterials in consumer products, biomedical materials, and industries, is a result of the vast diversity of their structural properties. Gold nanoparticles (GNPs) have been the object of investigations for usages as contrast agents for computed tomography, as drug delivery vehicles and as agents for photothermal cancer therapy, despite having no FDA approval yet. This presents a major challenge for effective and timely screening for their potential adverse effects in biological and environmental systems [1][2][3][4]. The inherent properties of nanoparticles, including size, shape, chemical composition, surface functionalization, solubility, hydrophobicity, surface charge and agglomeration state may all affect their interactions with biomolecules and cells [5][6][7]. Therefore, these physicochemical factors can determine the bioactivity and toxicity of nanoparticles [8][9][10][11].
Previous reports have focused on the influences of certain physicochemical properties of nanoparticles, such as size, surface chemistry and shape or core composition, on cytotoxicity [12][13][14]. More solid conclusions were achieved for the effects of nanoparticle size and surface chemistry. The effects of core composition on cellular perturbations were recently reported by us [15], and studies on the influence of shape have also been published [16][17][18]. However, as an intrinsic property of all nanoparticles, the effects of the redox reactivity of nanoparticles on cellular perturbations have not been systematically studied. Attribution of physicochemical determinants of nanoparticles to their toxicity is complex and requires systematical investigations because all these factors may participate in biological regulation simultaneously [19][20][21]. Therefore, it is difficult to solve this problem without systematic and isolated studies because of the lack of comparability. Facing this challenge, we developed a nano-combinatorial nanoparticle library approach and solved issues related to surface chemistry, size and core material. These systematic studies provided more convincing results [22][23][24].
In this study, we explored the relationship between the redox reactivity of nanoparticles and cellular perturbations using a combinatorial GNP library with diverse redox properties. In this systematic work, the redox-diversified library was first synthesized by incorporating a ferrocene moiety and organic ligand molecules with diversified structures. Cellular studies then demonstrated that the redox reactivity of GNPs, rather than other physicochemical properties, causes cytotoxicity via induction of cellular oxidative stress.

Combinatorial GNP Library with Diversified Redox Properties
In order to model the naturally diverse redox properties of nanoparticles, we designed a combinatorial GNP library. A ferrocene group was introduced into each ligand molecule on the surface of GNPs. Ferrocene can undergo redox reactions repeatedly with oxidative or reductive species in human cells [25][26][27]. These reactions are through the reversible transformation of Fe(II) and Fe(III) with the loss and gain of electrons. Moreover, the redox activity of the ferrocene group is probably affected by the ligand structure so that redox activities of GNPs may be modulated. In line with these designs, we initiated the synthesis of a combinatorial GNP library following Scheme 1. in biological regulation simultaneously [19][20][21]. Therefore, it is difficult to solve this problem without systematic and isolated studies because of the lack of comparability. Facing this challenge, we developed a nano-combinatorial nanoparticle library approach and solved issues related to surface chemistry, size and core material. These systematic studies provided more convincing results [22][23][24].
In this study, we explored the relationship between the redox reactivity of nanoparticles and cellular perturbations using a combinatorial GNP library with diverse redox properties. In this systematic work, the redox-diversified library was first synthesized by incorporating a ferrocene moiety and organic ligand molecules with diversified structures. Cellular studies then demonstrated that the redox reactivity of GNPs, rather than other physicochemical properties, causes cytotoxicity via induction of cellular oxidative stress.

Combinatorial GNP Library with Diversified Redox Properties
In order to model the naturally diverse redox properties of nanoparticles, we designed a combinatorial GNP library. A ferrocene group was introduced into each ligand molecule on the surface of GNPs. Ferrocene can undergo redox reactions repeatedly with oxidative or reductive species in human cells [25][26][27]. These reactions are through the reversible transformation of Fe(II) and Fe(III) with the loss and gain of electrons. Moreover the redox activity of the ferrocene group is probably affected by the ligand structure so that redox activities of GNPs may be modulated. In line with these designs, we initiated the synthesis of a combinatorial GNP library following Scheme 1.

Scheme 1.
Synthesis of the combinatorial GNP library. Using the Ugi multicomponent reaction, ligand molecules with a ferrocene group and various molecules structures were synthesized before a combinatorial GNP library with various redox activities was constructed using these ligands.
The key to success in this study was to first design diverse ligand molecules that could modulate the ferrocene group and therefore the redox activity of the GNPs. With this in mind, we introduced linear and cyclic substituent structures to diversify the influence of molecule geometry and introduced aromatic ring structures with different substituents to reflect the influence of π electron density. The introduction of heteroatoms reflected the effect of hydrophilicity and hydrogen bond acceptors, the amino and carboxyl groups reflected the influence of charge and the hydroxyl groups reflected the influence of hydrogen bond donors ( Figure 1). Scheme 1. Synthesis of the combinatorial GNP library. Using the Ugi multicomponent reaction, ligand molecules with a ferrocene group and various molecules structures were synthesized before a combinatorial GNP library with various redox activities was constructed using these ligands.
The key to success in this study was to first design diverse ligand molecules that could modulate the ferrocene group and therefore the redox activity of the GNPs. With this in mind, we introduced linear and cyclic substituent structures to diversify the influence of molecule geometry and introduced aromatic ring structures with different substituents to reflect the influence of π electron density. The introduction of heteroatoms reflected the effect of hydrophilicity and hydrogen bond acceptors, the amino and carboxyl groups reflected the influence of charge and the hydroxyl groups reflected the influence of hydrogen bond donors (Figure 1).

Characterization of Physiochemical Properties of the GNP Library
The GNP library was fully characterized for the GNPs' physicochemical properties. Morphologies of GNPs were determined by TEM (Figure 2a and Figure S1). The GNPs had a spherical shape and uniform diameter. The average diameter of the entire library was 5.6 ± 0.5 nm (Figure 2b). In solution, GNPs were well dispersed with a slight aggregation tendency. Electrostatic and electrodynamic properties of GNPs were assessed by their zeta potential, which was on average −24.8 ± 9.1 mV (Figure 2d and Figure S2b).

Characterization of Physiochemical Properties of the GNP Library
The GNP library was fully characterized for the GNPs' physicochemical properties. Morphologies of GNPs were determined by TEM (Figures 2a and S1). The GNPs had a spherical shape and uniform diameter. The average diameter of the entire library was 5.6 ± 0.5 nm (Figure 2b). In solution, GNPs were well dispersed with a slight aggregation tendency. Electrostatic and electrodynamic properties of GNPs were assessed by their zeta potential, which was on average −24.8 ± 9.1 mV (Figures 2d and S2b). The hydrophobicity of a GNP determines its ability to penetrate cell membranes or bind to important biomolecules. This reactivity is characterized by GNPs' water/octanol participation coefficient or log p value. The log p values of all GNPs were determined with a modified "shaking flask" method [28][29][30]. Log p values of the 80 GNPs in this library were between −1.0 and 1.0 and the average was 0.59 ± 0.24 (Figures 2e and S2a).

Direct In Situ Quantification of the GNP Surface Modifications
GNP properties, in large part, are determined by the surface ligand molecules or surface chemistry. However, determination of the number of surface ligands is not an easy The hydrophobicity of a GNP determines its ability to penetrate cell membranes or bind to important biomolecules. This reactivity is characterized by GNPs' water/octanol participation coefficient or log p value. The log p values of all GNPs were determined with a modified "shaking flask" method [28][29][30]. Log p values of the 80 GNPs in this library were between −1.0 and 1.0 and the average was 0.59 ± 0.24 ( Figure 2e and Figure S2a).

Direct In Situ Quantification of the GNP Surface Modifications
GNP properties, in large part, are determined by the surface ligand molecules or surface chemistry. However, determination of the number of surface ligands is not an easy task, especially in high-throughput screening. In most previous papers, this was not reported. In others, various analytical methods were applied, such as elemental analysis [31], thermogravimetric analyses [32] or cleavage and analysis methods [33]. For a large nanoparticle library or ligand molecules with multiple common elements, the applications of the above methods are limited. In this work, we established a direct, quick and accurate method to obtain both the number of ligands on each GNP and the number of GNPs in solution. By incorporating the ferrocene group in every ligand, we achieved both of the above goals using ICP-MS analysis (Scheme 2).
olecules 2021, 26, x FOR PEER REVIEW 5 of 1 task, especially in high-throughput screening. In most previous papers, this was not re ported. In others, various analytical methods were applied, such as elemental analysi [31], thermogravimetric analyses [32] or cleavage and analysis methods [33]. For a large nanoparticle library or ligand molecules with multiple common elements, the application of the above methods are limited. In this work, we established a direct, quick and accurate method to obtain both the number of ligands on each GNP and the number of GNPs in solution. By incorporating the ferrocene group in every ligand, we achieved both of the above goals using ICP-MS analysis (Scheme 2). In our method, the concentrations of Au and Fe were simultaneously determined by ICP-MS in one step analysis, and the number of GNPs and number of ligands were calcu lated accordingly. The number of loaded ligands per GNP was the ratio of the ligand number to GNP number. For the GNP library, the number of ligands per GNP was in the range of 240~359 (Figures 2f and S2C). Compared with the previous methods, our method involves fewer samples, simpler operation, fewer steps, reduces the possibility of erro and is suitable for high-throughput characterization of a large number of nanoparticle samples.

Diverse Redox Activities of the GNP Library and Underlying Mechanisms
The ferrocene group offers GNPs a unique electron exchange ability. These GNPs can thus be expected to have electrochemical reactivities. Diverse molecule structures of lig ands create different microenvironments, so the electrochemical reactivities of GNPs may be highly diverse. Such reactivities of GNPs were determined by cyclic voltammetry using the electrocatalytic reduction of H2O2. Different GNPs exhibited different intensities o reduction peak currents produced by H2O2 because of their different H2O2 reducing abili ties. The ferrocene group surrounded by each individual molecular environment deter mined the redox activity of GNPs ( Figure S3). The peak current is defined as the curren density at −0.20 V (Ag/AgCl) as shown in Figure 3a. Cyclic voltammetry and the peak currents of all GNPs were experimentally determined (Figures 3b and S4). The peak cur rent values of the GNP library varied, ranging from −0.04 to −0.23 mA, suggesting thei different H2O2 reducing abilities [34][35][36][37]. After uptake by cells, GNPs with a stronger re ducing ability will convert more H2O2 into OH -, which is relatively nontoxic [38][39][40][41]. In this process, the levels of ROS in cells are inhibited better, thus avoiding the occurrence o oxidative stress and cell damage. In our method, the concentrations of Au and Fe were simultaneously determined by ICP-MS in one step analysis, and the number of GNPs and number of ligands were calculated accordingly. The number of loaded ligands per GNP was the ratio of the ligand number to GNP number. For the GNP library, the number of ligands per GNP was in the range of 240~359 (Figure 2f and Figure S2C). Compared with the previous methods, our method involves fewer samples, simpler operation, fewer steps, reduces the possibility of error and is suitable for high-throughput characterization of a large number of nanoparticle samples.

Diverse Redox Activities of the GNP Library and Underlying Mechanisms
The ferrocene group offers GNPs a unique electron exchange ability. These GNPs can thus be expected to have electrochemical reactivities. Diverse molecule structures of ligands create different microenvironments, so the electrochemical reactivities of GNPs may be highly diverse. Such reactivities of GNPs were determined by cyclic voltammetry using the electrocatalytic reduction of H 2 O 2 . Different GNPs exhibited different intensities of reduction peak currents produced by H 2 O 2 because of their different H 2 O 2 reducing abilities. The ferrocene group surrounded by each individual molecular environment determined the redox activity of GNPs ( Figure S3). The peak current is defined as the current density at −0.20 V (Ag/AgCl) as shown in Figure 3a. Cyclic voltammetry and the peak currents of all GNPs were experimentally determined (Figure 3b and Figure S4). The peak current values of the GNP library varied, ranging from −0.04 to −0.23 mA, suggesting their different H 2 O 2 reducing abilities [34][35][36][37]. After uptake by cells, GNPs with a stronger reducing ability will convert more H 2 O 2 into OH -, which is relatively nontoxic [38][39][40][41]. In this process, the levels of ROS in cells are inhibited better, thus avoiding the occurrence of oxidative stress and cell damage.  Furthermore, we explored how the structural features of surface ligands actually regulated the reduction activity of GNPs by analyzing the resulting predictive model. Based on the coefficients of each variable in the model, the top seven ranked descriptors were selected and are shown in Figure 3d. The high ranking of a descriptor use indicated its critical contribution to the final modeling and the reduction activity of GNPs. Therefore, the reduction activity of GNPs was mostly relevant to the mean atomic electronegativity and mean atomic polarizability. Electronegativity is the tendency of an atom to attract shared electrons to itself, and polarizability is defined as the dipole moment induced in the atom in response to the application of an electric field. Obviously, the distribution of electrons or electron clouds inside atoms greatly affects the reduction activity of GNPs, while these properties are significantly enhanced by ligand structural diversity. The properties (physical and chemical properties and reactivity) of molecules are determined by their molecular orbitals and the arrangement of electrons in the molecular orbitals. The structural diversity of ligands inevitably affects the composition of molecular orbitals and the energy levels of the HOMO and LUMO, modulating the physical and chemical properties and reactivity of molecules and GNPs. Previous studies have also shown that the electronegativity and polarizability of atoms are closely related to the reduction activity Furthermore, we explored how the structural features of surface ligands actually regulated the reduction activity of GNPs by analyzing the resulting predictive model. Based on the coefficients of each variable in the model, the top seven ranked descriptors were selected and are shown in Figure 3d. The high ranking of a descriptor use indicated its critical contribution to the final modeling and the reduction activity of GNPs. Therefore, the reduction activity of GNPs was mostly relevant to the mean atomic electronegativity and mean atomic polarizability. Electronegativity is the tendency of an atom to attract shared electrons to itself, and polarizability is defined as the dipole moment induced in the atom in response to the application of an electric field. Obviously, the distribution of electrons or electron clouds inside atoms greatly affects the reduction activity of GNPs, while these properties are significantly enhanced by ligand structural diversity. The properties (physical and chemical properties and reactivity) of molecules are determined by their molecular orbitals and the arrangement of electrons in the molecular orbitals. The structural diversity of ligands inevitably affects the composition of molecular orbitals and the energy levels of the HOMO and LUMO, modulating the physical and chemical properties and reactivity of molecules and GNPs. Previous studies have also shown that the electronegativity and polarizability of atoms are closely related to the reduction activity of materials [42,43]. The model constructed here and the corresponding mechanism analysis can help us design nanomaterials with the required reduction activity in the future. Based on the theoretical calculation of the quantum chemistry, the weight of many influencing factors can be summarized and analyzed so as to realize the effect prediction of nanomaterials and the auxiliary design of new nanomaterials.

Redox Activities of GNPs Induce Cytotoxicity via Regulating the Cellular Oxidative Stress
In order to examine whether the redox activity and other physicochemical properties of GNPs play a role in inducing cellular perturbations, we determined cellular uptake (Figure 4b and Figure S2d), cellular oxidative stress ( Figure S2e) and cytotoxicity along with their relationships with GNP properties. The GNP-induced changes in relative fluorescence intensity of the DCFH-DA probe reflected the level of cellular oxidative stress. There was a clear trend that cellular oxidative stress was positively correlated with the oxidative properties of GNPs (Figure 4e). GNP-induced cytotoxicity was determined with the CellTiter-Glo ® assay (Figure 4d and Figure S5). GNP concentration that induced 50% cell death (EC50) was used to reflect cytotoxicity. The redox activity of the GNP library was correlated with cytotoxicity, with more oxidative GNPs inducing more cytotoxicity. It is well-known that H 2 O 2 is one of the important ROS species in cells [44][45][46]. Combining the above results, GNP-induced cellular oxidative stress was positively correlated with cytotoxicity (the lower the EC50 values, the higher the cytotoxicity). Excessive ROS in cells may result in random oxidation of nucleic acids, proteins and lipids, which may lead to the loss of their normal functions by changing the structure of these biomolecules [47][48][49][50][51]. When this kind of damage exceeds the self-healing ability of cells, it induces programmed cell death.
To examine whether other basic physicochemical properties of GNPs play a role in GNP-induced oxidative stress in cells, the effects of hydrophobic properties, zeta potential, cellular uptake and surface ligand loading were analyzed. Our results showed that the number of surface ligands ( Figure S6), hydrophobic properties (Figure 4a and Figure S7), zeta potential (Figure 4c and Figure S7) and cellular uptake (Figure 4b) did not have any correlation with the induction of cellular oxidative stress or cytotoxicity in A549 cells. Therefore, the diverse redox activities of GNPs were the only reason for cytotoxicity via induction of cellular oxidative stress.
The 80 GNP library members were all modified by organic ligand molecules. Although ligands were diverse, they hardly altered some of the basic physical properties, such as the hydrophobicity and the zeta potential. However, these structural diversities were enough to change the redox reactivities of the GNPs and, therefore, the induction of cellular oxidative stress and cytotoxicity. The GNP library approach has been used previously to develop cancer cell-targeting nanoparticles [22] and cellular oxidative stress-enhancing nanoparticles [52].  To examine whether other basic physicochemical properties of GNPs play a role in GNP-induced oxidative stress in cells, the effects of hydrophobic properties, zeta potential, cellular uptake and surface ligand loading were analyzed. Our results showed that the number of surface ligands ( Figure S6), hydrophobic properties (Figures 4a and S7), zeta potential (Figures 4c and S7) and cellular uptake (Figure 4b) did not have any correlation with the induction of cellular oxidative stress or cytotoxicity in A549 cells. Therefore, the diverse redox activities of GNPs were the only reason for cytotoxicity via induction of cellular oxidative stress.
The 80 GNP library members were all modified by organic ligand molecules. Although ligands were diverse, they hardly altered some of the basic physical properties, such as the hydrophobicity and the zeta potential. However, these structural diversities were enough to change the redox reactivities of the GNPs and, therefore, the induction of

Synthesis of Ligands
In an ice bath, 11.5 g (0.05 mol) of ferrocenecarboxylic acid was mixed with 100 mL of dichloromethane (DCM). Under vigorous stirring, 7.0 g (0.06 mol) of N-hydroxysuccinimide (NHS) and 11.5 g (0.06 mol) of 3-ethylcarbodiimide hydrochloride (EDC·HCl) were added to the reaction system. The reaction was stirred for 4-6 h in an ice bath and monitored by TLC. After reaction, filtration was carried out and a DCM solution of (1) was obtained.
In an ice bath, 8.5 g (0.075 mol) of triethylenediamine was added to 50 mL of DCM. Then, 24.6 mL of 1,4-cyclohexanediamine was added and stirred until the reaction system was uniformly stabilized. The DCM solution (1) was added dropwise to the reaction system within 2 to 3 h. The reaction was maintained continuously for 24 h. The reaction system gradually became viscous and the reaction was monitored by TLC. After reaction, the solvent was removed by rotary evaporation, and the crude product was washed repeatedly with water and purified by column chromatography. An orange-red solid product (2) was obtained.
Intermediate (2) (0.75 g, 5 mmol), benzaldehyde (0.50 mL, 5 mmol), cyclohexyl isocyanide (0.60 mL, 5 mmol) and lipoic acid (1.03 g, 5 mmol) were added to 10 mL of methanol and reacted at 320 K for 48 h. The reaction was monitored by TLC. After reaction, the solvent was removed by rotary evaporation, the crude product was purified by column chromatography and ligand 1 was obtained.

Synthesis of the GNP Library
We mixed an orange solution containing 0.0032 mmol of ligand (Ligand 1) in 20 mL of DMF, 12.5 mg of HAuCl 4 ·3H 2 O, 0.6 mg of ammonium chloride and 0.95 mg of sodium citrate in 2.25 mL of water at room temperature. Then, in an ice bath with vigorous stirring, 15 mg of sodium borohydride in 15 mL of water was slowly added to the reaction system through peristaltic pumping over 60 min. The solution color became dark and then red gradually. GNPs were named from GNP-1 to GNP-80 according to the types of modified ligands. After the Ugi reaction, 80 different ligands were obtained. Each ligand molecule had only one ferrocene group (one iron atom) with the "labeling" of the ligands undertaken by using the Fe atom.

Measurements of Zeta Potentials of GNPs
GNPs were suspended in aqueous solution with sonication. The zeta potentials of GNPs were measured at room temperature using a Malvern Zetasizer instrument (Nano ZS90, Malvern Instruments Ltd., Worcestershire, UK). All samples were measured at the same concentration. Each sample was measured in triplicate. The samples were diluted with Milli-Q water with a pH value of 6.83 when measuring zeta potential. Amine groups in the starting materials were converted to amides after the reactions so that there were no amine groups in the final products.

Determination of Loaded Ligands on GNPs
GNPs were suspended in aqueous solution with sonication. Then, 50 µL of GNP solution was mixed with 1.0 mL of aqua regia. The sample was completely digested and diluted with Milli-Q. The contents of gold and iron were determined by ICP-MS (Varian 820-MS). The GNP concentration was determined by the Au concentration, and the ligand concentration was determined by the Fe concentration.
GNPs can be approximately regarded as spheres. The gold atom number in a spherical GNP with a specific particle size can be calculated (m Au ·N A = M Au , m Au · n Au = m GNP = ρ Au ·V GNP ). Therefore, the results of ICP-MS can reflect the number of GNPs in a specific volume of GNP solution. In addition, the number of iron atoms is equal to the number of ligands (n Fe = n Ligand ).

Number of ligands = (C
An example of such analysis can be shown for GNP-5: the volume of GNP-5 was 87 nm 3 (V = 4/3 · π · r 3 ) since the average particle diameter was 5.5 nm (Figure 2a). ρ Au = 19.32 kg/m 3 . Based on the ICP-MS results, the Au content was 146.51 ppb and the Fe content was 2.73 ppb (all results after deducting the background value). Using Equation (1), we were able to determine that the number of ligands loaded on GNP-5 was 337.

TEM Images Characterization
TEM images of the GNPs were taken using a JEOL-1011 transmission electron microscope (JEM, Tokyo, Japan) at 100 KV. The images were acquired using an AMT 2k CCD camera. Image data were analyzed by ImageJ.

Hydrophobicity Measurements of GNPs
An equal volume of deionized water and n-octanol were mixed and stirred for 24 h. The mixture solution was allowed to stand quietly. After the separation, the upper layer was the water-saturated n-octanol, and the lower layer was the n-octanol-saturated aqueous solution. The two phases were collected and used in the next step.
Hydrophobicity of GNPs was measured according a previously reported method. Firstly, 0.3 mg of GNPs was dispersed with 1.0 mL of n-octanol-saturated aqueous solution. Secondly, 1.0 mL of water-saturated n-octanol was added. Thirdly, the mixtures were vigorously shaken on a shaker for 24 h. After that, the mixtures were left to stand quietly until the two phases were completely separated. Each phase was carefully collected and dried at 120 • C under vacuum for 4 h. Finally, 500 µL of fresh aqua regia was added to digest GNPs over 12 h. After diluting, gold concentrations in each phase were measured by ICP-MS. Log p values of GNPs were calculated according to the gold concentrations in the two phases.

Preparation of GNP Solution for Cell-Based Experiments
The aqueous dispersion of the gold nanoparticles was sonicated for 15 min. Sterile water was then added to dilute GNPs to a concentration of 1.0 mg/mL, and uniformly dispersed stock solutions were obtained. After sterilization by batch sterilization, stock solutions were stored at 4 • C before use. For cell experiments, after sonication, stock solutions were diluted with cell culture medium to desired concentrations and vortexed before adding to plate wells.

Cytotoxicity Test
A549 cells were cultured with RPMI 1640 medium and, respectively, supplemented with 10% fetal bovine serum, 100 U/mL penicillin and 100 µg/mL streptomycin. Then, 100 µL of A549 cells in the logarithmic growth phase at a concentration of 6 × 10 4 cells/mL was seeded to each well of 96-well plates. After incubation for 24 h, the medium was removed, and fresh, complete medium containing different concentrations of GNPs (i.e., 12.5 µg/mL, 25 µg/mL, 50 µg/mL, 100 µg/mL, 200 µg/mL, and 400 µg/mL) was added to each well. The cells were then incubated with GNPs for 48 h. After removing the medium and washing with PBS twice, 50 µL of the complete medium and 50 µL of the CellTiter-Glo ® assay working fluid were added in sequence. After gently shaking for 2 min, the plates were incubated for 10 min at room temperature in the dark. Then, 70 µL of the mixture solutions in each well were transferred to another set of white, opaque 96-well plates, and the fluorescence signal intensity was detected by a plate reader (VICTORTM X2, Perkin Elmer, Waltham, MA, USA). Deducting the background signals, cell viability was calculated. Each experiment was done in parallel three times.

Cellular Oxidative Stress Test
A549 cells were seeded in 24-well plates (6 × 10 4 cells/well). After 24 h incubation, the medium in each well was removed. GNPs in cell culture medium (500 µL, 50 µg/mL) were added to each well and incubated with cells for 24 h. The cells were then washed with PBS twice and 2,7-dichlorodihydro-fluoresceindiacetate (DCFH-DA, 10 µM) in medium were added. After incubation for 30 min in the darkness, cells were washed with PBS three times, treated with trypsin, and collected for measurements with a fluorescence microplate reader. Groups with only cell culture medium added and H 2 O 2 (500 nM) were set as negative and positive controls, respectively. Cellular oxidative stress was determined using the DCFH-DA method. In the absence of GNPs, cellular oxidative stress values with and without H 2 O 2 were used as positive and negative controls.

Cell Uptake Experiment
A549 cells in the logarithmic growth phase were seeded in 24-well plates with a density of 4 ×10 5 cells/well. After 24 h incubation at 37 • C, each well was washed twice with 500 µL of pre-warmed, serum-free RPMI1640 medium. Then, 500 µL of RPMI1640 complete medium containing 50 µg/mL GNPs was added to each well.
After another 24 h incubation at 37 • C, each well was washed three times with cold PBS and 200 µL of trypsin was added to each well. After 4 min digestion, 200 µL of the RPMI1640 complete medium was added to stop this process. Cells in each well were counted and 200 µL of the cells was collected. Fresh aqua regia was then added to completely digest the GNPs overnight at 37 • C. After diluting, gold contents in each well were measured using ICP-MS. Each experiment was done in parallel three times.

Determination of Redox Reactivity of GNPs
The hydrogenation ability of GNPs was evaluated by electrocatalysis of H 2 O 2 to OH − using an electrochemical workstation (CHI660c, CH Instruments, Inc.). A glassy carbon electrode was used as the working electrode, the Pt electrode as the counter electrode and the Ag/AgCl electrode as the reference electrode. Each GNP from the library with the same concentration (1.0 mg/mL, 10 µL) was added to the polished glassy carbon electrodes. The electrochemical reactions were performed in a 0.01 M PBS solution with 0.3%, 0.6% and 0.9% H 2 O 2 at a scan rate of 20 mV·s −1 . The initial potential, the maximum potential and the minimum potential were set at 0 V, 0.6 V and −1.4 V, respectively.

Correlation Analysis
Based on the definition of the Pearson correlation coefficient, the correlation between different characteristics (physical and chemical properties, biological effects) of GNPs (quotient of covariance and standard deviation between different characteristics) was calculated, and the correlation degree between these characteristics was analyzed.

Structure Activity Relationship
To explore the quantitative relationships between the nanostructures and the redox properties of GNPs, we used the structural information (i.e., Dragon descriptors) of the surface ligands to develop a quantitative structure-activity relationship (QSAR) model to predicting their redox reactivities. Here, we used a multiple linear regression method, called the least absolute shrinkage and selection operator (LASSO) model, to build the QSAR model. As a classic machine learning method, LASSO can enhance the prediction accuracy and interpretability of the resulting models through performing both variable selection and regularization. In the present LASSO model, the regularization parameter was set to 0.0001 and the maximum number of iterations set to 100,000. As a result, a total of 58 descriptors were selected to fit the model. Both model construction and prediction were performed by the machine learning library in Python, scikit-learn v0.19.2. The correlation between experimental and predicted reduction activity values is shown in Figure 3c. The high determination coefficient (R2 = 0.82) showed that the machine learning model correctly predicted GNPs' redox activities.

Conclusions
In this study, we explored the induction of cytotoxicity by a combinatorial GNP library with diverse redox activity. In this systematic work, we established that the cytotoxicity was directly caused by the oxidative reactivity, rather than other properties, of GNPs. There are two important implications of this study. First, a simple electrochemical characterization of nanoparticles should in the future be used to predict the basic toxicity risk of nanoparticles. Second, screening of a GNP library yields an array of nanoprobes, which can be embedded into human cells to probe cellular mechanisms or generate cell models with various oxidative stresses. Both of these findings will accelerate nanomedicine and nanotoxicity studies significantly. Through regulation of redox activity, GNPs may regulate cells to the appropriate level of oxidative stress to enforce a synergistic effect, which may further enhance the applications of GNPs in more biomedical fields, such as imaging, drug delivery, diagnosis and clinical treatment.
Supplementary Materials: The following are available online. Figure S1: TEM images of 80 GNPs in the whole library. Figure S2: The characterization of physicochemical properties and biological effects of 80 GNPs in the whole library. Figure S3: When the ligand does not contain ferrocene substituents, GNPs do not show obvious redox activity. Figure S4: The redox activities of 80 GNPs in the whole library were characterized by cyclic voltammetry. Figure S5: The dose response curves of 80 GNPs in the whole library. Figure S6: The relationship between the redox activity of 80 GNPs and their hydrodynamic diameter or number of ligand uploads. Figure S7: The relationship between the cellular effects of 80 GNPs and hydrophobic property or surface charge.