Discovery of Surface‐Engineered Nanoparticles That Boost Enzyme Activity via High‐Throughput Screening and Machine Learning

Abstract

Nanoparticles (NPs) are known to enhance the activity of enzymes, but such findings remain largely empirical, lacking predictive design principles. Here, the first high‐throughput platform for the discovery of surface‐engineered nanoparticles (SENs) that modulate enzyme function is introduced. Guided by the hypothesis that surface ligands are primary drivers of activity enhancement, a library of 194 gold‐ and palladium‐based SENs functionalized with diverse peptide ligands is synthesized. These SENs are screened against three model enzymes: cytochrome c, lactoperoxidase (LPO), and lipase. Multiple SENs substantially increased enzymatic activity, with the most effective achieving ≈19‐fold increase. The resulting dataset enabled the training of a machine learning model that identified key ligand features associated with high‐performing SENs, establishing a predictive framework for designing activity‐enhancing NPs. Mechanistic studies confirm that the ligand shell plays a dominant role in driving enhancement, suggesting that effective ligands identified via this approach can be readily transferred across NP platforms. To demonstrate functional relevance, it is shown that an optimized SEN/LPO pair outperforms LPO in inhibiting the growth of multidrug‐resistant bacteria and disrupting biofilm formation. Collectively, this work offers a scalable and generalizable method to map and harness nanoscale structure‐function relationships at biointerfaces, with applications in biocatalysis, biosensing, and beyond.

A high‐throughput platform enables the discovery of surface‐engineered nanoparticles (SENs) that enhance enzyme activity. By focusing on ligand chemistry rather than core material, this work establishes predictive design principles for SENs, revealing their broad utility in biocatalysis and antimicrobial applications.

1. Introduction

Enzymes are essential catalysts in biotechnology, where they accelerate a wide range of chemical reactions across applications in biosensing, therapeutics, and industrial processing.^{[ 1} , ^{2 ]} To improve their stability and enable reuse, enzymes are often immobilized on solid supports.^{[ 3 ]} However, this strategy frequently leads to substantial losses in enzymatic activity, sometimes exceeding 90%.^{[ 4} , ⁵ , ⁶ , ^{7 ]}

Recent studies have shown that certain nanomaterials can not only preserve but, in some cases, significantly enhance enzyme activity. Enhanced catalytic performance has been reported for more than 40 enzymes when associated with nanomaterials such as metallic nanoparticles (NPs),^{[8, 9 ]} quantum dots,^{[ 10} , ^{11 ]} metal‐organic frameworks,^{[ 12} , ¹³ , ¹⁴ , ¹⁵ , ^{16 ]} carbon‐based materials,^{[ 17} , ¹⁸ , ^{19 ]} and DNA‐based nanostructures,^{[ 20} , ²¹ , ^{22 ]} among others.^{[ 23} , ²⁴ , ^{25 ]} Properties of these nanomaterials, including size, shape, and surface functionalization, have been shown to influence enzyme behavior.^{[ 23} , ²⁴ , ²⁵ , ^{26 ]} These enhancements are typically attributed to nanoscale effects such as conformational stabilization of the enzyme, altered substrate accessibility, or local environmental changes induced by the nanomaterial.^{[ 23} , ^{24 ]} While mechanistic explanations have been proposed on a case‐by‐case basis—often informed by systematic variation of nanomaterial features, most of the work remains phenomenological. As a result, there are still no generalizable or predictive rules to determine whether a given nanomaterial will enhance the activity of a particular enzyme.

In the absence of clear design principles, high‐throughput screening provides a powerful strategy for identifying effective NP–enzyme combinations. We reasoned that because interactions between NPs and enzymes occur primarily at their interface, the chemical nature of the surface‐displayed ligands, rather than the NP core, plays a dominant role in modulating enzymatic activity. To test this hypothesis, we developed the first scalable, high‐throughput platform for discovering surface‐engineered nanoparticles (SENs) that enhance enzyme function. We synthesized a library of 194 gold and palladium NPs functionalized with a chemically diverse set of peptidic ligands. These NPs were screened against three model enzymes, cytochrome c (CytC), lactoperoxidase (LPO), and lipase (LIP), using a 96‐well plate format that allowed the assay to be completed in under 1 h.

This high‐throughput approach enabled the rapid generation of a rich dataset quantifying how each NP affects enzyme activity, measured as fold‐changes in initial reaction rates. Leveraging this dataset, we trained a machine learning model that identified specific ligand features strongly associated with catalytic enhancement. Importantly, the resulting model provides a predictive framework for identifying promising ligand candidates, reducing the need for exhaustive experimental screening. Mechanistic studies further confirmed that ligands are a key driver of enzyme activity modulation, indicating that high‐performing ligands discovered through these screens can be readily transferred across different NP cores.

Remarkably, these screens identified a SEN that enhances LPO activity by up to ≈19‐fold compared to the native enzyme. We demonstrate that this enhancement can be harnessed to augment the antimicrobial properties of LPO. When paired with hydrogen peroxide and thiocyanate, the SEN/LPO complex exhibits strong bactericidal activity against multidrug‐resistant pathogens. This finding highlights the translational potential of designer SENs in addressing urgent biomedical challenges.

2. Results and Discussion

2.1. Design Strategy for Generating the SEN Library

We posited that a successful platform for studying how nanoscale surface chemistry influences enzyme activity must meet several key criteria: (1) it must be simple and highly modular, allowing easy customization of surface ligands; (2) it must be synthetically tractable to enable the creation of a large and diverse library; and (3) it should minimize barriers to substrate diffusion, avoiding potential confounding factors. Beyond these considerations, we chose to rely on noncovalent interactions between enzymes and NPs, reasoning that this would allow enzymes to adopt favorable binding orientations driven by the physicochemical properties of the surface ligands.^{[ 27 ]}

To meet these requirements, we developed a SEN platform using gold NPs (GNPs) functionalized with short peptides. GNPs were chosen due to their ease of synthesis, scalability, and compatibility with thiolated ligands that self‐assemble onto their surfaces.^{[ 28 ]} Select experiments were performed with palladium NPs (PdNPs) as cores (vide infra). Peptides served as ideal surface ligands due to their sequence programmability, which allows precise control over physicochemical properties such as charge, hydrophobicity, and polarity.^{[ 29 ]}

94 distinct peptides were used initially, and based on the experimental results, 6 additional peptides were included (Tables S1 and S2, Supporting Information). Each peptide was 7 amino acids in length—a design choice informed by previous studies showing this length offers sufficient chemical diversity while remaining synthetically accessible.^{[ 8 ]} All peptides included a cysteine (C) for gold‐thiol attachment, a tryptophan (W) for peptide quantification, and a glycine (G) spacer.^{[ 30 ]} The four remaining positions were varied to introduce broad chemical diversity.

To construct the peptide library, we selected amino acids representing key physicochemical categories. These included charged residues (e.g., E, D for negative; R, H for positive), hydrophobic residues (e.g., L, F, Y, W), and polar uncharged residues (e.g., S, Q). Peptides were designed with varying numbers and combinations of charged residues, including both single‐type (e.g., all negative or all positive) and mixed‐type charges. To investigate how sequence patterning influences enzyme interactions, a subset of peptides featured alternating patterns of charge and hydrophobicity or polarity (e.g., E‐L‐E‐L, R‐E‐R‐E), enabling us to explore the effects of sequence periodicity and spatial presentation. We also assessed the impact of residue position within otherwise similar sequences (e.g., EELLWGC vs LLEEWGC). Examples such as EEEEWGC and RRRRWGC represented high charge density, while SSEEWGC and QEQEWGC featured more moderate charge and polarity. In addition to linear sequences, we included branched architectures (e.g., EEEEWGC vs EEWCGEE) to assess how spatial configuration of functional groups influences enzyme activity.

The simplicity and versatility of the platform made it possible to generate a diverse library of 194 distinct SENs.

2.2. Synthesis and Characterization of the SEN Library

The SEN library was created by first synthesizing GNPs (Figure S1, Supporting Information) and subsequently functionalizing them with distinct peptides to yield peptide‐functionalized GNPs (PGNPs). 94 different peptides were used in our initial screen. Both GNP and PGNP synthesis followed previously established protocols,^{[ 8} , ^{31 ]} with modifications to enable parallel synthesis.

Briefly, GNPs were synthesized by reducing chloroauric acid with sodium borohydride to produce bare, unfunctionalized NPs (Figure S1, Supporting Information).^{[ 8} , ^{31 ]} To generate PGNPs, the as‐synthesized GNPs were aliquoted into a 96‐well plate, where 94 wells contained unique peptides. Two wells were left empty to serve as controls during subsequent enzyme activity screening. Each peptide was used at a 500‐fold molar excess relative to the GNPs.

Peptide conjugation was performed using a previously reported freeze–thaw method that enhances ligand packing on the NP surface.^{[ 8} , ^{32 ]} In this study, we adapted the method to a 96‐well plate format to enable high‐throughput, parallel synthesis of the PGNP library. Specifically, the GNP–peptide mixtures were flash‐frozen in liquid nitrogen for 2–3 min, then thawed to room temperature, resulting in PGNPs (Figure 1 ).

Figure 1.

Illustration of high‐throughput discovery of SENs that boost enzyme activity. A) Overview of the process: synthesis of a SEN library followed by evaluation of its effects on enzyme activity. B) Relative sizes of NP core (with GNP as an example), enzymes, and the PGNPs used in this study. Proteins shown in this figure correspond to crystal structures of CytC (PDB ID: 1HRC), LPO (PDB ID: 3GC1), and LIP (PDB ID: 1YS1). C) Schematic representation of the enzyme catalytic reactions used in this study.

The optical properties of GNPs before and after peptide functionalization were assessed using UV–vis spectroscopy. Unmodified GNPs exhibited a surface plasmon resonance (SPR) peak near 510 nm. This peak remained largely unchanged upon the addition of peptides with an isoelectric point (pI) below 6.9, indicating colloidal stability (Figures 2A,B; S2, Supporting Information). However, the addition of peptides with pI above 6.9 caused red shifts of up to 40 nm in the SPR peak, suggesting partial aggregation. This aggregation is consistent with electrostatic attraction between the positively charged peptides and the negatively charged surface of the as‐synthesized GNPs. Notably, these shifts occurred immediately upon peptide addition and did not change further after the freeze–thaw step, indicating that the peptide‐induced aggregation occurred prior to conjugation.

Figure 2.

Characterization of the PGNP Library. UV–vis absorbance spectra of representative PGNPs before A) and after B) freeze‐thaw treatment, the data is normalized by setting the maximum absorbance of the SPR peak to a value of 1. C) Comparative ζ‐potential of GNPs and representative PGNPs functionalized by different peptides (PGNP#1: EEEEWGC, pI = 3.58; PGNP#27: LLLLWGC, pI = 5.52; PGNP#55: RRWCGRR, pI = 12.00). Data represent the mean of three independent replicates, with error bars indicating standard deviation. TEM image of D) GNPs, E) PGNP#1, F) PGNP#27, and G) PGNP#55. The scale bar represents 20 nm.

Selected PGNPs were further characterized by electrophoretic light scattering and transmission electron microscopy (TEM). ζ‐Potential measurements supported successful peptide attachment (Figure 2C). Bare GNPs exhibited a ζ‐potential of −25.4 ± 3.2 mV. Functionalization with peptides with a low pI of 3.58 (i.e., EEEEWGC) shifted the ζ‐potential to −37.5 ± 0.3 mV. Functionalization with peptides with a moderate pI of 5.52 (i.e., LLLLWGC) and peptides with a high pI of 12.00 (i.e., RRWCGRR) shifted the ζ‐potential to ‐6.1 ± 0.2 mV and 21.7 ± 0.5 mV, respectively. TEM imaging showed that the unmodified GNPs were well‐dispersed, spherical particles with an average diameter of 4.1 ± 0.6 nm (Figures 2D; S3, Supporting Information). Peptide functionalization did not significantly alter particle size, although increased aggregation was observed following treatment with positively charged peptides (Figures 2E–G; S4–S6, Supporting Information).

2.3. High‐Throughput Screens to Assess Enzymatic Activity

Following the synthesis of the PGNP library, we performed enzyme activity screening to evaluate the impact of each PGNP on the catalytic performance of three model proteins: CytC and LPO, which exhibit heme‐dependent peroxidase activity, and LIP, which exhibits esterase activity. These enzymes were selected based on prior literature,^{[ 8} , ^{33 ]} their low cost, and the availability of chromogenic substrates that allow high‐throughput monitoring using a plate reader. Moreover, LPO and LIP are functionally important enzymes, with roles in antibacterial applications and industrial lipolysis, respectively.^{[ 33} , ³⁴ , ^{35 ]}

For peroxidase activity, we employed the well‐established 3,3′,5,5′‐tetramethylbenzidine (TMB) assay. In this reaction, TMB is oxidized by the heme‐containing protein in the presence of hydrogen peroxide (H₂O₂), producing a blue‐colored product with an absorbance peak at 652 nm.^{[ 36 ]} Lipase activity was assessed using p‐nitrophenyl acetate, which is hydrolyzed to form yellow‐colored p‐nitrophenol, detectable at 410 nm. Peroxidase activity was measured in acetate buffer at a pH of 5.0, whereas esterase activity was measured in PBS at a pH of 7.4.^{[ 37 ]}

Enzymes and substrates were added directly to the as‐synthesized PGNPs without further purification. This approach was justified by our observation that the peptides alone do not enhance enzyme activity (vide infra), eliminating the need to separate free peptides from the PGNPs prior to screening. Each 96‐well plate included two controls: one containing enzyme and substrates (positive control) and one containing only substrates (negative control).

Absorbance of the reaction products was monitored over a 30‐min period (Figure 3A–C), and the reaction velocities (V) were determined from the initial slopes of absorbance versus time curves. To assess the effect of each PGNP, we calculated the ratio of the V in the presence of PGNPs to that of the enzyme alone. This fold‐change in activity, referred to hereafter as “enhancement,” is shown as a heatmap in Figure 3D–F.

Figure 3.

High‐throughput screening of enzyme activity by PGNPs. Absorbance over time curves obtained using top‐performing PGNPs that enhanced the activity of A) CytC, B) LPO, and C) LIP, respectively. The absorbance values at t = 0 are non‐zero in the presence of PGNPs due to the intrinsic absorbance of the GNP core. Representative heat maps of high‐throughput screening results for D) CytC, E) LPO, and F) LIP. Each well corresponds to a unique PGNP. Well G12 represents the native enzyme control (activity set as 1), and well H12 is the substrate‐only control. Blue indicates enhanced activity, red indicates inhibited activity, and white represents activity comparable to the native enzyme.

As expected, PGNPs produced a range of effects on enzyme activity: some inhibited catalysis, others had no effect, and several significantly enhanced activities. Specifically, several PGNPs inhibited CytC, others enhanced it, and some had no effect (Figure 3D). In contrast, for LPO, most PGNPs enhanced activity to varying extents, with few showing no to minimal impact (Figures 3E; S7, Supporting Information). For LIP, most PGNPs had no effect, and only a select few enhanced activity (Figure 3F). These sparse hits underscore the importance of the high‐throughput screening method developed herein, as identifying rare but potent PGNP enhancers would have been unlikely without a systematic and scalable approach. The highest enhancements observed were ≈6.7‐fold for CytC, 19.0‐fold for LPO, and 1.8‐fold for LIP.

2.4. Machine Learning–Driven Identification of Key Peptide Features That Affect Enzyme Activity

The data generated from high‐throughput enzyme activity assays were analyzed using a machine learning–based framework to identify peptide features that significantly influence enzymatic function. Specifically, we employed supervised learning using random forest regression.^{[ 38 ]} Random forest was selected due to its robustness against overfitting, ability to capture nonlinear interactions, and built‐in feature importance estimation.^{[ 39 ]}

Because the observed activity changes are driven by the peptide ligands, the model focused on ligand‐specific characteristics. Each peptide was represented by 115 numerical features, capturing a range of biochemical and structural properties. These included global properties such as amino acid composition percentages (e.g., %C, %D, %E), pI, hydrophobicity, and molecular weight. To capture the importance of amino acids at specific positions in a sequence, one‐hot encoding was used.

The full dataset of 94 peptides was initially split into a training set and a test set using a 3:1 ratio, yielding 70 peptides for training and 24 for testing. K‐means clustering, followed by stratified sampling, was utilized to ensure that the test and training sets follow the same distribution. To assess reproducibility, this data‐splitting strategy was repeated six times to generate six different training–test set combinations.

Model performance was evaluated on the test set using mean squared error (MSE) and the Pearson correlation coefficient (r). r was used to quantify the strength of the relationship between predicted and observed enhancement values. We then extracted impurity‐based feature importance scores from the trained random forest model to identify the most predictive features (Figures S8 and S9, Supporting Information).

The model performed well for CytC and LPO, achieving average (across six training–test splits) MSE values of 0.37 ± 0.12 for CytC and 2.10 ± 0.60 for LPO, and Pearson r values of 0.89 ± 0.05 and 0.76 ± 0.04, respectively. Performance was weaker for LIP (MSE = 0.01 ± 0.00, r = 0.65 ± 0.08), likely due to the limited effect of most PGNPs on LIP activity. With fewer active hits, the model struggles to distinguish meaningful feature–response relationships.

For CytC, the top three predictive peptide features were pI, hydrophobicity score, and instability index. Notably, pI alone accounted for over 75% of the total feature importance score in the random forest model, and all three features were negatively correlated with enzyme enhancement. This aligns with previous findings that negatively charged PGNPs enhance CytC activity,^{[ 8 ]} likely due to favorable electrostatic interactions—CytC has a pI of ≈10–10.5 and is therefore positively charged at the assay pH of 5.^{[ 40 ]}

For LPO, the top features were again pI, molecular weight, and hydrophobicity score. As with CytC, pI was the dominant feature (contributing ≈55% of the feature importance score). However, here, pI was positively correlated with activity enhancement, while the other two features were negatively correlated. This was unexpected, as LPO also has a high pI (≈9.6) and should be positively charged in acetate buffer; thus, favorable interactions with positively charged PGNPs are counterintuitive. This prompted us to further investigate LPO's behavior and binding preferences (vide infra).

For LIP, all feature importance values were below 25%, with the top‐ranked being GRAVY score, hydrophobicity, and instability index. Given the minimal impact of the tested PGNPs on LIP activity, model predictions should be interpreted with caution, as the data likely lack a strong signal for learning meaningful structure–function relationships.

2.5. Mechanistic Investigation of Activity Enhancement by PGNPs

We investigated the mechanism behind PGNP‐mediated enzyme enhancement using LPO as a model system, given that it showed the strongest enhancement in our initial screen. To confirm that the observed enhancement was specifically due to interactions between LPO and the PGNP surface—rather than individual components—we performed a series of control experiments. Neither the peptides alone nor the PGNPs in the absence of enzyme exhibited any peroxidase activity (Figure S10 and S11, Supporting Information). Furthermore, simply mixing free peptides with LPO did not result in increased activity (Figure S12, Supporting Information). These findings indicate that the enhancement effect is not an additive property of the individual components but rather a unique consequence of the 3D display of peptides on the NP surface.

To evaluate whether the enhancement effect depends more on the peptide ligands than on the NP core, we synthesized 94 distinct peptide‐functionalized PdNPs (PPNPs) and repeated the high‐throughput screen with LPO. The resulting heatmap of enhancement values showed substantial similarity to that from the PGNP screen (Figure S13, Supporting Information). The correlation coefficient between enhancement values for matched PGNP and PPNP samples was r = 0.74, indicating a strong linear relationship. This suggests that peptide identity, rather than core composition, is the dominant factor in determining enhancement (Figure S13, Supporting Information).

We then performed structural and physicochemical characterization using the top‐performing PGNP identified in the screen (i.e., one functionalized with RRWCGRR). First, we confirmed that LPO and the PGNP interact by measuring the ζ‐potential of the mixture. Although the pI of LPO is ≈9.6, ζ‐potential measurements in 0.1× PBS (pH 7.4) indicated a slightly negative surface charge (≈ −0.5 mV), suggesting that free LPO is nearly electrically neutral under these conditions. This apparent discrepancy could be due to the ionic strength of the buffer compressing the electrical double layer, thereby reducing the apparent zeta potential, or due to interactions with phosphate ions in the buffer that influence surface charge. The PGNPs had a positive surface charge of 21.7 mV, which gradually decreased as increasing amounts of LPO were added, indicating that LPO adsorbs onto the PGNP surface (Figure 4A). The dissociation constant of the PGNP/LPO complex was quantitatively determined to be 0.03 nm (Figure S14 and Table S4, Supporting Information), which indicates that LPO is strongly bound to the PGNP.

Figure 4.

Mechanistic investigation of activity enhancement by PGNPs. A) ζ‐potential measurements of GNPs, PGNP#55, and PGNP#55 mixed with increasing concentrations of LPO, showing the gradual change in surface charge upon enzyme binding. Data represent the mean of three independent replicates, with error bars indicating standard deviation. B) CD spectra of LPO before and after incubation with PGNP#55 for 1 and 30 min, respectively. Data represent the average of three independent measurements, collected in 0.1 × acetate buffer. C). Quantification of LPO secondary structure components using the BeStSel algorithm based on CD data.^{[ 41 ]} D) 3D structure of LPO (PDB ID: 3GC1), highlighting the asymmetric distribution of surface charge. E) Structure of the heme cavity in LPO. Residues involved in the substrate diffusion channel are shown in green, substrate binding residues in pink, and the heme moiety in blue. Yellow highlights H351, which is located in the proximal heme cavity. F) Heat map of enzyme activity screening for PGNP#95–#100 and additional PGNPs with similar peptide sequences. Each well corresponds to a distinct PGNP. Wells C4 and D4 represent the native enzyme and substrate‐only controls, respectively. In the heap map, blue indicates enhanced activity, red indicates inhibition, and white indicates activity similar to the native enzyme.

Circular dichroism (CD) spectroscopy was used to study whether binding to PGNPs alters the structure of LPO (Figure 4B; Figure S15, Supporting Information). Characteristic CD signals at 190, 208, and 230 nm were significantly diminished following PGNP treatment, indicating substantial alterations to the protein's secondary structure. Typically, α‐helical proteins exhibit negative ellipticity near 208 nm and a positive signal ≈190 nm. The observed spectral changes suggest a shift in the balance of α‐helical and β‐sheet content. Analysis using the BeStSel tool (Figure 4C) revealed a dramatic reduction in α‐helical content—from ≈20% in the native state to ≈0% post‐treatment—accompanied by a marked increase in antiparallel β‐sheet content.^{[ 41 ]} This structural reorganization may reflect a conformational change that improves substrate access to the heme active site, potentially contributing to the observed enhancement in enzymatic activity. We note similar structural changes have been previously reported in other heme‐containing proteins, such as CytC, when they show activity enhancement in the presence of nanomaterials.^{[ 8} , ⁴² , ^{43 ]}

Kinetic analysis (Figure S16 andTable S5, Supporting Information) revealed that PGNPs increased the catalytic turnover number (k_cat ) of LPO by ≈8‐fold, while also increasing the apparent Michaelis constant (K_M ). This suggests that while the NP surface facilitates catalysis, it also reduces the binding affinity between the enzyme and the substrate. This decrease in affinity may result from nonspecific interactions either between the substrate and the PGNP surface or between the enzyme and the PGNP, both of which could interfere with the substrate binding effectively to the enzyme.

To gain deeper mechanistic insight, we analyzed the top 20 peptide sequences from the PGNP screen. A majority (75%) contained basic residues such as histidine (H) or arginine (R), consistent with machine learning predictions that identified a positive correlation between peptide pI and enhancement. Glutamic acid (E) and leucine (L) appeared in 25% of top peptides, while G—beyond the conserved spacer shared by all peptides—and serine (S) each appeared in two. Threonine (T) and glutamine (Q) were observed only once.

Interestingly, examination of the LPO crystal structure revealed that all of these residues, exceptS and T—are located within 3 Å of the heme center, the substrate‐binding pocket, or the substrate diffusion channel (Figures 4D,E; S17, Supporting Information).^{[ 44} , ^{45 ]} This suggests that PGNPs bearing such residues may mimic key features of the enzyme's active site, thereby promoting local substrate enrichment and enhancing catalysis. However, if substrates bind too tightly to the NP surface, their effective availability to the enzyme could be reduced, a scenario supported by the observed increase in K_M . While a detailed mechanistic investigation is beyond the scope of this study, these observations provide a rationale for the sequence‐dependent enhancement effect.

Given the apparent importance of H and R, as well as the high frequency of phenylalanine (F) residues in the native LPO active site, we synthesized 5 additional peptides incorporating these residues in various combinations to assess whether activity could be further improved (Tables S1 and S3, Supporting Information). One additional peptide was synthesized to evaluate the effect of replacingW withG (Tables S1 and S3, Supporting Information). We functionalized PGNPs with these 6 new peptides and performed a high‐throughput screen alongside 8 PGNPs bearing similar sequences, chosen from our initial library. Sequences with high H and R content consistently emerged as the strongest enhancers. However, none of the newly tested sequences led to a substantial improvement over the original top‐performing PGNP (Figure 4F).

2.6. Potential Application of the PGNP/LPO System as Bactericidal Agents

To demonstrate the broader applicability of the PGNP platform, we conducted a proof‐of‐concept study evaluating its potential antimicrobial applications.

LPO is a well‐known antimicrobial enzyme that functions via the oxidation of thiocyanate ions in the presence of hydrogen peroxide.^{[ 34} , ^{46 ]} This reaction produces hypothiocyanite, a reactive oxygen species that disrupts bacterial metabolism by targeting sulfhydryl groups in bacterial proteins and membranes (Figure 5A).^{[ 47 ]} Unlike traditional antibiotics, this mechanism targets metabolic pathways that are less prone to developing resistance, making LPO an attractive candidate for antimicrobial strategies.^{[ 48 ]} Given the strong enhancement of LPO activity by PGNPs, we hypothesized that this effect could translate into improved bactericidal efficacy, particularly valuable in addressing antibiotic‐resistant bacteria.

Figure 5.

Evaluation of the PGNP/LPO system as a bactericidal agent. A) Schematic showing the antibacterial mechanism of the LPO system. B) Representative photographs showing the turbidity of kanamycin‐resistant E. coli cultures under different treatment conditions, 24 h after incubation. C) Representative photographs showing colony formation resulting from bacterial plating from suspensions in (B). D) Growth curves of kanamycin‐resistant E. coli monitored over 24 h under different treatment conditions, measured by OD₆₀₀. E) Quantification of biofilm formation by multidrug‐resistant E. coli under different treatment conditions. Biofilm formation was evaluated using the crystal violet (CV) assay. Biofilms were formed over 24 h, exposed to different treatments after an additional 12 h, and then stained with 0.1% CV for 15 min. The absorbance was measured at 570 nm. (D,E): Data represent the mean of three independent replicates, with error bars indicating standard deviation. Statistical analysis was performed using the one‐tailed Student's t test to evaluate a directional hypothesis. A Bonferroni correction was applied for three comparisons, adjusting the significance threshold to p < 0.0167 to maintain an overall 95% confidence level (^* p < 0.0167). F) Photographs corresponding to different treatment groups in (E). G) SEM images showing morphological changes in bacteria following various treatments.

For antibacterial evaluation, Escherichia coli (E. coli) was selected as the model organism due to its well‐characterized genome and common use in antimicrobial studies. We first examined kanamycin‐resistant E. coli.

Bacterial cultures were grown in Luria Broth (LB) at an initial concentration of $\sim 1\times {10}^6$ CFU/mL. Turbidity was used as a qualitative indicator of bacterial growth. After 24 h, untreated E. coli cultures became visibly turbid, while those treated with 100 nm LPO showed reduced turbidity (Figure 5B). Remarkably, cultures treated with both 100 nm LPO and 5 nm PGNPs exhibited minimal turbidity, suggesting strong bacterial inhibition. All treated cultures also contained the LPO substrates hydrogen peroxide and potassium thiocyanate. To assess viable bacteria, 5 µL of each suspension was then plated onto LB‐agar plates and incubated overnight. Colonies were visible the following day (Figure 5C). Untreated plates showed numerous, slightly overgrown colonies. Plates from LPO‐treated cultures displayed a substantially smaller number of colonies, while those treated with the PGNP/LPO combination showed very few, indicating enhanced antimicrobial activity.

To quantitatively assess bacterial viability, we measured the optical density (OD₆₀₀) of bacterial suspensions with and without treatment over a 24‐h period. The untreated control exhibited rapid growth, reaching OD₆₀₀ ≈ 0.3 within 6 h. LPO alone delayed bacterial growth, with OD₆₀₀ reaching the same value after 24 h. In contrast, the PGNP/LPO‐treated group showed no appreciable increase in OD₆₀₀ throughout the entire 24‐h period (Figures 5D; S18, Supporting Information).

Encouraged by these results, we extended the study to a multidrug‐resistant E. coli strain that is resistant to both kanamycin and β‐lactam antibiotics such as penicillin. This strain expresses β‐lactamase, an enzyme that hydrolyses the β‐lactam ring, thereby inactivating β‐lactam drugs and contributing to its antibiotic resistance profile.^{[ 49} , ^{50 ]} As before, treated and untreated bacterial suspensions were plated on LB‐agar, and colony formation was visualized (Figure S19, Supporting Information). The PGNP/LPO system again outperformed native LPO, demonstrating its enhanced efficacy against multidrug‐resistant strains.

LPO is also known to disrupt biofilm formation.^{[ 51 ]} Therefore, we next investigated whether the PGNPs can also enhance this function of LPO. Biofilms were allowed to form over 24 h by inoculating multidrug‐resistant E. coli in 96‐well plates containing LB broth at an initial concentration of $\sim 1\times {10}^7$ CFU mL⁻¹. We then exposed the biofilms to different treatment conditions. The PGNP/LPO system showed significantly improved disruption of biofilms compared to the LPO alone and all control groups (Figures 5E,F; S20, Supporting Information).

Further insight into the antibacterial mechanism was obtained via scanning electron microscopy (SEM). Untreated E. coli cells displayed smooth surfaces and a characteristic rod‐shaped morphology. In contrast, cells treated with the PGNP/LPO system exhibited extensive membrane deformation, including shrinkage, surface collapse, and structural disintegration, consistent with irreversible cell damage (Figures 5G; S21–S24, Supporting Information). Cells treated with LPO alone also exhibited signs of membrane disruption, but the damage was noticeably less severe compared to the PGNP/LPO‐treated group.

Together, these findings suggest that PGNP‐enhanced LPO retains potent antibacterial activity, even against multidrug‐resistant bacteria, and may offer a promising strategy for developing NP‐assisted antimicrobial therapies.

3. Conclusion

This study introduces a powerful, high‐throughput strategy for identifying SENs that enhance enzymatic activity, addressing a longstanding challenge in enzyme immobilization and nanobiocatalysis. By functionalizing GNPs with a library of 100 chemically diverse, sequence‐defined peptides, we developed a platform capable of assessing enzyme activity rapidly (in <1 h) in a 96‐well format. This is made possible by the synthetic accessibility and parallelizability of peptide‐modified GNPs, which also minimize mass transport limitations and enable efficient exploration of nanoscale structure–function relationships. While the current implementation uses a 96‐well format, the approach is readily scalable to higher‐throughput formats such as 384‐ or 1536‐well plates.

Our results show that the ligands on the NP surface are the primary determinants of enzyme modulation. These ligands mediate strong interactions with enzymes (with dissociation constants as low as K_d = 0.03 nm) and are responsible for the observed catalytic enhancement. Critically, these enhancements require no modifications to the enzyme sequence or structure, making the approach broadly compatible with native, unaltered proteins. These results indicate that once beneficial ligands are identified, they can be readily transferred across different nanoparticle platforms, allowing the NP core to be selected based on the specific requirements of the intended application, whether biomedical, catalytic, or sensing‐related.

Systematic screening of three model enzymes, CytC, LPO, and LIP, revealed that SENs modulate activity in an enzyme‐ and ligand‐specific manner. Notably, we observed up to 19‐fold enhancement in LPO activity. While several SENs enhanced the activity of CytC and LPO, hits were sparse for LIP, underscoring the value of high‐throughput screens to uncover otherwise unpredictable interactions.

Importantly, the resulting dataset enabled the development of predictive machine learning models. Using random forest regression, we identified key peptide features that correlate with enhancement, though the specific features varied across enzymes. This highlights the complex and context‐dependent nature of nano‐bio interactions. While the current model provides useful insights, its predictive power could be further improved through hyperparameter tuning, advanced feature engineering, and by expanding the size and chemical diversity of the peptide training set. Additionally, alternative modeling approaches such as transformer‐based architectures may offer enhanced performance as the dataset grows.^{[ 52 ]}

As a proof‐of‐concept demonstration of practical utility, we showed that SEN‐enhanced LPO improves functional performance in biological settings. The PGNP/LPO complex exhibited potent antibacterial activity against both kanamycin‐resistant and multidrug‐resistant E. coli, reducing bacterial growth, colony formation, and biofilm integrity—suggesting that catalytic enhancements at the nano–bio interface can be translated into therapeutic outcomes.

Altogether, this work establishes a generalizable framework for engineering nanoparticle–enzyme interfaces through high‐throughput screening and machine learning. Beyond offering new insights into nano–bio interactions, this platform lays the groundwork for the rational design of programmable nanomaterials for applications in biocatalysis, biosensing, and antimicrobial therapy.

4. Experimental Section

Materials

All chemicals were available commercially and used without further purification. Chemical reagents include: Sodium borohydride (NaBH₄, Sigma–Aldrich, Cat. No. 452 882), gold(III) chloride trihydrate (HAuCl₄⋅3H₂O, Sigma–Aldrich, Cat. No. 520 918), hydrogen peroxide (H₂O₂, Cat. No. H1009), and palladium(II) nitrate hydrate (Sigma–Aldrich, Cat. No. 205 761). Enzymes include: Cytochrome C from equine heart (CytC, Sigma–Aldrich, Cat. No. C2506), lactoperoxidase from bovine milk (LPO, Sigma–Aldrich, Cat. No. L2005‐), and amano lipase PS from Burkholderia cepacian (LIP, Sigma–Aldrich, Cat. No.534641). Biochemical reagents include: 3,3′,5,5′‐tetramethylbenzidine dihydrochloride hydrate (TMB, Sigma–Aldrich, Cat. No. 87 750), 2,2′‐Azino‐bis(3‐ethylbenzothiazoline‐6‐sulfonic acid) diammonium salt (ABTS, Sigma–Aldrich, Cat. No. A1888), 4‐Nitrophenyl acetate (pNPA, Sigma–Aldrich, Cat. No. N8130), potassium thiocyanate (KSCN, Sigma–Aldrich, Cat. No. AA 1 431 822), and Alexa Fluor 488 NHS ester (Thermo Fisher, Cat. No. A20000). Reagents used for bacterial experiments include: Luria Broth (LB, Sigma–Aldrich, Cat. No. 28 713), LB broth with agar (Sigma–Aldrich, Cat. No. L3147), Crystal violet (Thermo Fisher, Cat. No. B21932.22), kanamycin disulfate salt (Sigma–Aldrich, Cat. No. K1876), and penicillin G sodium salt (Sigma–Aldrich, Cat. No. P3032). BLA plasmid was purchased from Genscript. Peptides were ordered from Genscript. TEM grids with ultrathin carbon film on a lacey carbon support (400 mesh, Ted Pella, Cat. No. 1824) were used for TEM sample preparation. Additional materials include: Amicon Ultra‐0.5 spin filter (Millipore), 96‐well polystyrene plates (Fisherbrand, Cat. No. 12‐565‐501), petri dishes (Fisherbrand, 100 mm × 15 mm, slippable lid, Cat. No. FB0875713), and NAP‐5 desalting columns (Fisher Scientific, Cat. No. 45 000 151).

Synthesis of SEN Library

GNPs were synthesized following a previously reported protocol. Briefly, in a 50 mL centrifuge tube, 10 mL of deionized water was mixed with 14.0 µL of 100.0 mm HAuCl₄ under vortex agitation. Subsequently, 140.0 µL of freshly prepared NaBH₄ solution was added rapidly under vigorous stirring. The reaction mixture was stirred for an additional 5 min, then stored in the dark at room temperature for 30 min before use.

PGNPs were synthesized using a freeze‐directed approach. In a 96‐well plate, 100.0 µL of the prepared GNP solution was added to each well using a multichannel pipette. Subsequently, 10.0 µL of 100.0 µm peptide solution (corresponding to a 500‐fold molar excess relative to GNP) was added to each well. The plate was sealed with an adhesive cover and immersed in liquid nitrogen until all wells were completely frozen (2‐3 min). The frozen plate was then incubated at 37 °C to allow complete thawing (15 min). PdNPs and PPNPs were synthesized using a similar procedure. Details were provided in the supporting information.

ζ‐Potential Measurements

ζ‐Potential was measured (using Malvern Zetasizer Nano instrument) to characterize the surface charge of GNPs, representative PGNPs with distinct net charges, and PGNP/LPO complexes. All selected PGNPs were prepared at a final concentration of 20.0 nm in deionized water. The selected peptides included RRWCGRR (peptide #55, positively charged), EEEEWGC (peptide #1, negatively charged), and LLLLWGC (peptide #27, near‐neutral at the assay pH of 5.0). To investigate the interaction between positively charged PGNPs (peptide #55) and LPO, ζ‐potential was measured at increasing concentrations of LPO ranging from 25.0 to 150.0 nm. All measurements were conducted in triplicate, and the averaged values were reported.

TEM

TEM was employed to investigate the morphology of GNPs and representative PGNPs. For sample preparation, 2.0 µL of each NP solution was dropped onto Lacey carbon‐coated copper grids (400 mesh) and allowed to dry overnight at room temperature. Imaging was performed (JEOL NEOARM Low kV STEM Corrected) at an accelerating voltage of 200 kV.

Enzyme Activity Assays

All high‐throughput enzyme activity assays were performed in 96‐well plates. Each plate contained 94 different PGNP samples prepared in the same batch, along with one well containing the native enzyme alone (positive control) and one well containing only the substrate (negative control). Buffer, enzyme, and substrates were added to each well, containing 110.0 µL of the as‐synthesized PGNP, to reach a final volume of 210.0 µL. Buffer was used in place of the enzyme for the positive control. Water was used in place of PGNPs for both the positive and negative controls.

1. CytC and LPO Activity Assays. The 3,3′,5,5′‐tetramethylbenzidine (TMB) oxidation assay was used to evaluate the peroxidase‐like activity of CytC and LPO in the presence of PGNPs. In the presence of H₂O₂, TMB was oxidized by CytC or LPO to form a blue‐colored product with a characteristic absorbance peak at 652 nm. All reactions were conducted in 1X acetate buffer (0.1 m, pH 5.0) at room temperature. The CytC assay was carried out with 125.0 nm CytC, 1.5 mm TMB, and 2.0 mm H₂O₂; the LPO assay used 3.0 nm LPO, 1.5 mm TMB, and 0.1 mm H₂O₂. All concentrations refer to final values in the reaction mixture. Reaction components were added sequentially to the buffer. The formation of oxidized TMB was monitored at 652 nm every minute for 30 min using a microplate reader.

2. Lipase Activity Assay. Lipase high‐throughput activity screening was determined using a colorimetric assay based on the enzymatic hydrolysis of para‐nitrophenyl acetate, which releases para‐nitrophenol, a chromogenic product measurable at 410 nm. Reactions were carried out in 1X PBS containing 0.5 mm para‐nitrophenyl acetate and 34.8 nm lipase. Absorbance at 410 nm was recorded every minute for 30 min using a microplate reader to monitor enzyme activity in real time.

CD Spectroscopy

CD spectroscopy (Jasco CD J‐815 instrument) was used to assess the conformational changes of peptides before and after conjugation to GNPs, as well as the structural changes in LPO upon binding with PGNPs. CD spectra were acquired in 0.1 × acetate buffer (pH 5.0) over a wavelength range of 190–260 nm, using baseline‐corrected data averaged from three independent measurements. For CD measurements, PGNPs were prepared using peptide #55 (RRWCGRR), and the final sample concentrations were 100.0 nm GNP/PGNP and 50.0 µm peptide (500‐fold excess relative to GNP). To investigate enzyme–NP interactions, PGNPs were mixed with 5.0 µm LPO and incubated for either 1 min or 30 min prior to measurement.

Antibacterial Activity of PGNP/LPO System on Kanamycin‐Resistant Bacteria

The antibacterial effect of the PGNP/LPO system against kanamycin‐resistant E. coli was evaluated by monitoring bacterial growth (OD₆₀₀) in a 96‐well plate. Each well contained 150.0 µL LB broth and 50.0 µL reaction mixture (1X acetate buffer, premixed for 30 min), with a final bacterial concentration of 1 × 10⁷CFU mL⁻¹ and a total volume of 200.0 µL. A range of treatment conditions, including different combinations of PGNP, LPO, KSCN, and H₂O₂, were tested. The working concentrations were: PGNP (2.5 nm), LPO (0.05 µm), KSCN (1.0 mm), and H₂O₂ (0.5 mm). Plates were incubated at 37 °C, and OD₆₀₀ was recorded hourly for 24 h. Antibacterial activity was further assessed by plating 5.0 µL of each sample on LB agar, followed by incubation at 37 °C for 16 h. The plates were then photographed. Detailed protocols were provided in the supporting information.

Antibacterial Activity of PGNP/LPO System on Multidrug‐Resistant Bacteria

The antibacterial activity of the PGNP/LPO system against multidrug‐resistant E. coli was assessed using CFU plating. Bacteria (finalconcentration: 1 × 10⁷CFU°mL⁻¹)were treated in 96‐well plates containing 150.0 µL of LB broth and 50.0 µL of reaction mixture (prepared in 1X acetate buffer, premixed for 30 min), for a total volume of 200.0 µL per well. A range of treatment conditions, including different combinations of PGNP, LPO, KSCN, and H₂O₂, were tested. The working concentrations were: PGNP (2.5 nm), LPO (0.05 µm), KSCN (1.0 mm), and H₂O₂ (0.5 mm). Control groups received equal volumes of sterile water in place of reagents. All treatments were performed in triplicate. After 12 h, 5.0 µL from each well was plated on LB agar and incubated at 37 °C for an additional 16 h. The plates were then photographed. Detailed protocols were provided in the supporting information.

Study of Biofilm Disruption of Multidrug‐Resistant Bacteria

Biofilm disruption was assessed using a crystal violet assay. Multidrug‐resistant E. coli (OD₆₀₀ = 0.05) was cultured in 96‐well plates at 37 °C for 24 h to form biofilms. Wells were washed to remove planktonic cells, then treated with various reagent combinations. A range of treatment conditions, including different combinations of PGNP, LPO, KSCN, and H₂O₂, were tested. The working concentrations were: PGNP (2.5 nm), LPO (0.05 µm), KSCN (1.0 mm), and H₂O₂ (0.5 mm). After 12 h incubation, wells were washed, stained with 0.5% crystal violet, and the dye was solubilized with 33% acetic acid. Absorbance at 570 nm quantified biofilm biomass. Detailed protocols were provided in the supporting information.

Morphological Investigation of Bacteria by SEM

SEM was used to assess morphological changes in E. coli (OD₆₀₀ = 0.4) after 3 h treatment at 37 °C with one of four conditions: (1) untreated, (2) PGNP/LPO system, (3) LPO system alone, or (4) H₂O₂ alone. Working concentrations were PGNP (2.5 nm), LPO (0.1 µm), KSCN (2.0 mm), and H₂O₂ (0.5 mm). Cells were fixed in 2% glutaraldehyde, dehydrated through graded ethanol and HMDS series, air‐dried, sputter‐coated with platinum, and imaged by SEM (Apreo 2C LoVac SEM). Detailed protocols were provided in the supporting information.

Data Analysis

Unless stated otherwise, all data represent the mean of three independent measurements. Error bars indicate the standard deviation of these replicates. Normalized absorbance was calculated by normalizing the highest absorbance value in a data set to 100%. The fold change in enzyme activity was calculated based on the initial reaction rates of the enzymatic assay. Specifically, the reaction rate of the free enzyme was set as the baseline with a value of 1, and fold changes for other samples were determined relative to this baseline. Values were corrected for substrate autohydrolysis or oxidation. These fold change values were subsequently used to generate heat maps. Statistical analysis was performed using the one‐tailed Student's t test to evaluate a directional hypothesis. A Bonferroni correction was applied for three comparisons, adjusting the significance threshold to p < 0.0167 to maintain an overall 95% confidence level (^* p < 0.0167). The statistical tests were conducted using GraphPad Prism version 10. Detailed methods were outlined in the Supporting Information.

Conflict of Interest

The authors declare no conflict of interest.

Supporting information

Supporting Information

Sun Y., Pandit S., Satish N., Gilman G., Snider D. M., and Samanta D., “Discovery of Surface‐Engineered Nanoparticles That Boost Enzyme Activity via High‐Throughput Screening and Machine Learning.” Small 21, no. 40 (2025): e07126. 10.1002/smll.202507126

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.