Identification of Polymeric Nanoparticles Using Strategic Peptide Sensor Configurations and Machine Learning

Abstract

Environmental pollution by miniaturized plastics such as micro- and nanoplastics continues to escalate, posing serious risks to ecosystems and human health. Therefore, there is an urgent need to detect or identify the plastics. Although the techniques for microplastics have been advanced, those for nanoplastics remain challenging owing to the difficulty of sample collection and sensing reliability. In this study, the identification of polymeric nanoparticles dispersed in water was demonstrated using peptide sensors with a microenvironment-sensitive fluorophore. The fluorescence spectra obtained from peptide sensors were different depending on the polymer species of polymeric nanoparticles. Supervised and unsupervised machine learning on the signal patterns of fluorescence intensities obtained from the spectra successfully identified polymeric nanoparticles with slightly different chemical structures. Systematic evaluation revealed the critical role of both the number and combination of peptide sensors in achieving the precise identification of polymeric nanoparticles. Our approach offers new and foundational insights into the forthcoming identification of nanoplastics dispersed in water.

The annual release of plastics into the environment is escalating, exacerbating global pollution concerns. , Because of their low degradability, plastics that enter terrestrial and marine ecosystems fragment into smaller particles, referred to as microplastics and nanoplastics, through physical degradation and fragmentation such as mechanical abrasion, UV radiation, and microbiological degradation. − Recent studies have highlighted that microplastics and nanoplastics can pose significant health risks to humans, − plants, − and other organisms, , primarily through bioaccumulation and adsorption of chemical contaminants on their surfaces. Consequently, the development of effective methods for identifying the chemical species of micro/nanoplastics or detecting subtle variations in their surface conditions, such as partial degradation by oxidation, is becoming increasingly critical. For the identification of microplastics, Fourier transform infrared (FT-IR) − or Raman − spectroscopy has typically been employed to dry samples collected physically from water. However, because of their significantly small size, the collection of nanoplastics from water presents considerable challenges. Moreover, the signals of nanoplastics from FT-IR or Raman spectroscopy are often too weak for reliable identification. To address these issues, surface-enhanced Raman scattering (SERS) has been utilized for nanoplastics immobilized on suitable substrates, − and there has been progress in combining microscopic Raman spectroscopy or SERS with machine learning techniques. − Including these nanoplastics identification, machine learning techniques have recently garnered attention across various scientific and technological fields. − Despite these advances, methods for identifying nanoplastics remain limited and require further development. Importantly, to the best of our knowledge, nanoplastics dispersed in water have not yet been identified.

In human taste , and olfactory systems, multiple receptor proteins interact with taste and odor molecules in a combinatorial manner. These interactions generate characteristic signal patterns that enable the discrimination of an enormous variety of tastes and odors. The system performs remarkably well despite the total number of receptor proteins being smaller than the vast number of taste and odorant molecules it can detect. Inspired by these systems, the approach of identifying target analytes through the recognition of signal patterns from multiple chemical sensors using machine learning techniques has garnered significant interest. This method relies on the fact that each chemical sensor interacts with a wide variety of target analytes in different degrees for providing discriminative signals from the analytes, and therefore contrasts with traditional chemical sensors based on a ‘lock and key’ model with high specificity and selectivity. Various chemical sensors, including those made from nanoparticles, − synthetic polymers, − supramolecules, − and peptides, have been developed to detect proteins, metal ions, and cells. Among these, peptide sensors stand out because of their ability to provide discriminative signal patterns for identifying target analytes, as minor variations in amino acid sequences can finely tune their interaction with target analytes. In fact, we have demonstrated the successful identification of water-soluble polymers using machine learning on fluorescence signal patterns obtained by using short peptide sensors with a microenvironment-sensitive fluorophore, where the peptide moiety physically interacted with water-soluble polymers to provide fluorescence signals. , These promising results with peptide sensors have motivated their application in identifying water-dispersed polymeric nanoparticles, serving as a potential model for nanoplastics.

In this study, we present a new concept demonstrating that peptide sensors with a microenvironment-sensitive fluorophore can be used to precisely identify water-dispersed polymeric nanoparticles composed of homopolymers and copolymers with different chemical structures (Figure ). Four peptide sensors with a microenvironment-sensitive 1-anilinonaphythalene (AN) group, previously designed to identify water-soluble polymers, have also been used for polymeric nanoparticles. The peptide sensor named P1 comprised the His-Gln-Ile-Ala-His-Lys-Ala-Glu-His-Arg-Leu-Arg sequence (Figure A). Our previous study has demonstrated that P1 can interact moderately with various water-soluble polymers, potentially through mechanisms such as hydrogen bonding and/or hydrophobic interactions. This interaction resulted in the generation of diverse fluorescence signals. Based on these findings, we anticipated that P1 would similarly produce diverse fluorescent signals by adequately interacting with polymeric nanoparticles of different chemical structures. For the other peptide sensors, named P2, P3, and P4, each His residue from the N-terminus of P1 was replaced with an Ala residue, respectively (Figure A). Replacing His residues with Ala residues is expected to alter the hydrogen bonding and/or hydrophobic interactions between the original peptide (P1) and polymeric nanoparticles. Indeed, our previous study has shown that even the substitution of a single residue (His to Ala) can lead to diverse interactions with water-soluble polymers, resulting in varied fluorescent signals. Based on this, we employed these substituted peptide sensors in the present study, anticipating that their interactions with polymeric nanoparticles would similarly vary, leading to the generation of diverse fluorescent signals (Figure B). We systematically evaluated the performance of a single-peptide sensor and up to four peptide sensors in providing fluorescence signal patterns discernible through machine learning (Figure C). These findings underscore the critical role of both the number and specific combinations of peptide sensors in achieving precise identification of structurally similar nanoparticles. In contrast to conventional approaches such as SERS and Raman spectroscopy, the present method eliminates the requirement for sample drying during preparation. This crucial feature allows for the direct examination of polymeric nanoparticles in their aqueous state, thereby suggesting the system’s potential for analyzing nanoplastics directly in their native aquatic environments. Our approach leveraged the unique interactions between short peptides and diverse polymeric nanoparticles, providing novel and foundational knowledge for detecting nanoplastics in water.

1.

Identification of polymeric nanoparticles using peptide sensors. (A) Chemical structures of peptide sensors. (B) Measurements of fluorescence spectra of each peptide sensor in the presence of water-dispersed polymeric nanoparticles. (C) Machine learning on fluorescence signal patterns obtained from peptide sensors for polymeric nanoparticles.

Results

Fluorescence Spectra of Peptide Sensors in the Presence of Homopolymeric Nanoparticles

To evaluate whether peptide sensors can provide fluorescence signals from polymeric nanoparticles, the fluorescence spectra of P1 at a concentration of 1 μM in the presence or absence of different concentrations of poly­(butyl acrylate) (PBA) nanoparticles (Figure A) with an average particle size of 74 nm were measured at an excitation wavelength of 350 nm in Britton–Robinson (BR) buffer (pH 7.0) at 25 °C (Figure B). The fluorescence intensity of the spectra increased with increasing nanoparticle concentration, suggesting that the microenvironment (typically, hydrophobicity) − of the AN group of P1 was changed by the interaction with the PBA nanoparticle, and that the amount of P1 interacting with the nanoparticle increased with increasing nanoparticle concentration. The increase in the fluorescence intensity at 425 nm showed a saturation curve (Figure C). The saturation curve fitted well with the Langmuir isotherm, with a coefficient of determination of 0.99. Therefore, the apparent association constant was estimated to be 2.41 × 10¹⁰ M^–1. It is noted that the molar concentration of the PBA nanoparticle was estimated from the weight concentration and particle size, assuming the density of the PBA nanoparticle to be 1 g mL^–1. The apparent association constant suggested that peptide sensors could potentially detect polymeric nanoparticles at several tens of femtomolar concentrations for their identification.

2.

Fluorescence spectra of peptide sensors in the presence of homopolymeric nanoparticles for PCA-based extraction of discriminative fluorescence signals. (A) Structural formula of constituent polymers for homopolymeric nanoparticles. (B) Fluorescence spectra of P1 (1 μM) in the presence or absence of different concentrations of the PBA nanoparticle with an excitation wavelength of 350 nm in BR buffer (pH 7.0) at 25 °C. (C) Fluorescence intensities at 425 nm for P1 against the concentration of the PBA nanoparticle. (D) Normalized fluorescence spectra of peptide sensors (1 μM) in the presence of homopolymeric nanoparticles (10 μg mL^–1) with an excitation wavelength of 350 nm in BR buffer (pH 7.0) at 25 °C. (E) Principal component loadings derived from PCA of the spectra.

Since P1 successfully provided fluorescence signals from the PBA nanoparticle, the fluorescence spectra of each peptide sensor (P1, P2, P3, and P4) were measured in the presence of eight homopolymeric nanoparticles composed of acrylate- and methacrylate-type homopolymers such as poly­(methyl acrylate) (PMA), poly­(ethyl acrylate) (PEA), PBA, poly­(2-ethylhexyl acrylate) (PEHA), poly­(methyl methacrylate) (PMMA), poly­(butyl methacrylate) (PBMA), poly­(cyclohexyl methacrylate) (PCHMA), and poly­(benzyl methacrylate) (PBzMA) (Figure A) at sensor and nanoparticle concentrations of 1 μM and 10 μg mL^–1, respectively (Figure S1). Since the peptide sensors are considerably smaller than the polymeric nanoparticles, the fluorescence spectra may be affected by variations in the nanoparticles’ specific surface area. To mitigate these potential biases and acquire spectra comparable across samples with differing surface areas, the spectra were normalized to the specific surface area of each homopolymeric nanoparticle. Namely, each fluorescence spectrum was multiplied by the average particle size of the corresponding nanoparticle with round shapes (Table S1 and Figure S2) for normalization, assuming that the density of the nanoparticles was 1 g mL^–1 (Figure D). The maximum intensity and its wavelength of the fluorescence spectra for each peptide sensor were different for the nanoparticles, possibly suggesting that each peptide sensor provided discriminative signals for identifying the nanoparticles.

Extraction of Discriminative Fluorescence Intensities from Fluorescence Spectra

To extract discriminative fluorescence intensities at appropriate wavelengths for the identification of homopolymeric nanoparticles, normalized spectra were applied to principal component analysis (PCA), an unsupervised machine learning algorithm for compressing high-dimensional data to low dimensions. PCA creates new Cartesian coordinate axes (principal components) to maximize the variance of the data and rearrange the data along these axes. This process reduces the dimensionality of the original data with as little loss of features as possible, thereby facilitating data visualization and analysis. The contribution or weight of the original variable to the principal component is expressed by the principal component loading. The higher the absolute value of the loading weight of the principal component loading, the stronger the contribution of the variable to the principal component. The principal component loading derived from PCA of the normalized fluorescence spectra for the four peptide sensors exhibited peaks at 425 and 475 nm for the first and second principal components, respectively (Figure E). The absolute values of the loading weights for the former peaks were maximal for all sensors; therefore, the fluorescence intensity at 425 nm was used as one of the potentially discriminative signals for the identification of homopolymeric nanoparticles. On the other hand, the wavelength of the latter peaks was significantly different from that of the former ones; therefore, the fluorescence intensity at 475 nm was used as another signal. Although the absolute loading weight for the second principal component was greater at 400 nm than at 475 nm, the 475 nm wavelength was selected. This choice was made to ensure sufficient spectral separation from the 425 nm wavelength (selected based on the first principal component) and thereby maximize the discriminative power. The extraction of fluorescence intensities at the two specific wavelengths as discriminative signals for each peptide sensor enables the evaluation of the potential of each peptide sensor for the identification of polymeric nanoparticles through machine learning (see below).

Classification and Identification of Homopolymeric Nanoparticles

To visualize the fluorescence intensities at the two wavelengths, the data sets obtained from ten-time experiments are shown in the form of a heatmap (Figure A). Each peptide sensor appeared to reproducibly provide different signals from homopolymeric nanoparticles, producing a fluorescence signal pattern discernible through machine learning. To reveal the potential of a single-peptide sensor for the classification of homopolymeric nanoparticles, the fluorescence intensities obtained from each peptide sensor at the two wavelengths for different homopolymeric nanoparticles were applied to linear discriminant analysis (LDA), which is widely recognized as a supervised machine learning algorithm for data classification and dimensionality reduction. As the present data sets were two-dimensional, the two-dimensional LDA score plot for each peptide sensor (Figure B) provided more information on the discrimination between polymeric nanoparticles without dimensionality reduction. The two discriminant scores in the resulting discriminant function accounted for 100% of the total variance, indicating that all information in the data sets was involved in the plot. The ten scores for each homopolymeric nanoparticle in the plot were enclosed in an ellipse with a 95% confidence level to classify the scores into each cluster of the corresponding homopolymeric nanoparticle. Except for P4, the eight ellipses obtained from P1, P2, and P3 for the homopolymeric nanoparticles were separated without overwrapping, indicating that there were statistically significant differences in the clusters. This result strongly suggested that the three peptide sensors could potentially discriminate between the eight homopolymeric nanoparticles.

3.

Classification of homopolymeric nanoparticles through machine learning. (A) Heatmap visualization of the fluorescence intensities obtained from peptide sensors for homopolymeric nanoparticles. As shown on the right, one block contains ten data sets obtained by ten-time experiments for each homopolymeric nanoparticle. (B) Two-dimensional LDA score plots obtained from the fluorescence intensities of each peptide sensor. The ten scores for each homopolymeric nanoparticle in the plot were enclosed in an ellipse with a 95% confidence level. (C) HCA dendrograms for the data sets obtained from each peptide sensor for homopolymeric nanoparticles.

Two cross-validation techniques were employed for the data sets of each peptide sensor to evaluate the generalization performance of the learning models for the identification of homopolymeric nanoparticles. The first technique was leave-one-out cross-validation (LOOCV), where one data set for a specific homopolymeric nanoparticle was excluded from the ten data sets and used as the test data set. Following generation of the LDA score plot from the remaining all data sets and ellipses with a 95% confidence level, the test data set was applied to the discriminant function. Then, the resulting plot of the test data set was assigned to an ellipse with the shortest Mahalanobis distance. This procedure was repeated for every data set (i.e., 80 data sets) to classify of each test data set into one of the homopolymeric nanoparticles. Significantly, all the test data sets for P1, P2, and P3 were accurately classified into the corresponding homopolymeric nanoparticles, resulting in overall classification accuracy of 100% (Tables S2–S5). However, in the case of P4, the three data sets for the PEA nanoparticle were unsuccessfully classified into the PBzMA nanoparticle, resulting in an overall classification accuracy of 96%. The second technique was stratified k-fold cross-validation with 5-fold configurations (S5CV), where one-fifth of the ten data sets for certain homopolymeric nanoparticle and the remaining all data sets were used as the test and training data sets, respectively. This procedure was repeated for every data set combination to classify the test data sets into one of the homopolymeric nanoparticles, similarly to LOOCV. The 100% classification accuracy was achieved by P1, P2, and P3, while the classification accuracy for P4 was 96%. Consequently, the two cross-validation results revealed a high confidence in the data sets obtained from P1, P2, and P3 for the identification of homopolymeric nanoparticles.

Hierarchical cluster analysis (HCA) was employed to classify the same data sets based on a different machine-learning algorithm (Figure C). HCA is a well-known technique that classifies distinct groups within a data set according to their spatial distances, such as Euclidean distances, in an unsupervised manner (but not in a supervised manner like LDA). Shorter distances along the branches of the dendrogram represented more closely related classes. Remarkably, the ten data sets for each homopolymeric nanoparticle were closely clustered in the cases of P1, P2, and P3; the exceptions were some data sets obtained from P4 for the PEA and PBzMA nanoparticle. This result indicated that P1, P2, and P3 successfully provided the discriminative signals even for an unsupervised machine learning algorithm such as HCA, as well as for a supervised LDA. Remarkably, only in the dendrogram obtained from P1, the four acrylate- and four methacrylate-type nanoparticles were clustered separately, possibly suggesting that P1 provided discriminative signals for the two different types of homopolymeric nanoparticles (namely, acrylate versus methacrylate). Furthermore, the PBA and PMA nanoparticles clustered closely to the PEHA and PEA nanoparticles, respectively, suggesting that P1 provided discriminative signals for the critical difference in alkyl chain lengths. Additionally, the PCHMA and PBzMA nanoparticles were classified in close proximity, suggesting that P1 provided similar signals for the cyclohexyl and benzyl groups. These results suggest that P1 provides the fluorescence signals that reflect the structural similarities between the polymer species of the homopolymeric nanoparticles. In contrast, in the cases of other peptide sensors (i.e., P2, P3, and P4), the distances between the clusters for homopolymeric nanoparticles did not show a clear dependence on the chemical structures of homopolymeric nanoparticles.

Assembling all results, the peptide sensors (particularly P1, P2, and P3) had a high potential to provide discriminative signals for the identification of homopolymeric nanoparticles through supervised and unsupervised machine learning.

Application to Classification and Identification of Copolymeric Nanoparticles

To broaden the potential applicability of the present strategy, the identification of polymeric nanoparticles composed of copolymers with slightly different chemical structures was evaluated. For copolymeric nanoparticles, butyl acrylate monomers were copolymerized with small amounts of the seven comonomers (Figure A and Table S6). The composition ratios changed only for the glycidyl methacrylate comonomer. Identification methods were essentially the same as those for homopolymeric nanoparticles. The fluorescence spectra of the four peptide sensors in the presence of copolymeric nanoparticles were obtained (Figure S3), which were normalized by the specific surface area of each copolymeric nanoparticle with round shapes (Figure B, Table S6, and Figure S4). The spectra of each peptide sensor were significantly different between the nanoparticles, even though the comonomer ratios were small (i.e., 2–15 wt %). The spectra were then applied to PCA. The peak wavelengths of the first and second principal components of the principal component loadings were obtained for each peptide sensor (P1: 430 and 445 nm; P2: 425 and 475 nm; P3: 425 and 520 nm; P4: 425 and 450 nm, respectively) as wavelengths for the potentially discriminative signals (Figure C). To visualize the fluorescence intensities at these wavelengths, the data sets obtained from ten-time experiments are shown in the form of a heatmap (Figure D). Each peptide sensor seemed to provide different signals from the copolymeric nanoparticles, producing the fluorescence signal pattern discernible through machine learning.

4.

Fluorescence spectra of peptide sensors in the presence of the PBA and copolymeric nanoparticles for PCA-based extraction of discriminative fluorescence signals. (A) Structural formula of constituent copolymers for copolymeric nanoparticles. (B) Normalized fluorescence spectra of peptide sensors (1 μM) in the presence of the PBA and copolymeric nanoparticles (10 μg mL^–1) with an excitation wavelength of 350 nm in BR buffer (pH 7.0) at 25 °C. (C) Principal component loadings derived from PCA of the spectra. (D) Heatmap visualization of the fluorescence intensities obtained from peptide sensors for the PBA and copolymeric nanoparticles.

To reveal the potential of a single-peptide sensor for the classification of copolymeric nanoparticles and homopolymeric PBA nanoparticle, the fluorescence intensities obtained from each peptide sensor at the two wavelengths were applied to the LDA (Figure S5). Importantly, although the clusters (i.e., ellipses with a 95% confidence level) for the P­(BA-GMM4) and P­(BA-GMA15) nanoparticles in the two-dimensional LDA score plots were separated from other clusters for every peptide sensor, other clusters were overwrapped with one or more of the clusters in the case of at least one peptide sensor. This result indicates that each peptide sensor hardly discriminated between copolymeric and PBA nanoparticles. In fact, the classification accuracies for LOOCV were obtained to be 95%, 98%, 95%, and 77% for P1, P2, P3, and P4, respectively (Tables S7–S10), indicating the poor generalization performance of each learning model. Therefore, each peptide sensor could not provide the discriminative signals for the identification of copolymeric nanoparticles. These results differed from the successful identification of homopolymeric nanoparticles using a single-peptide sensor. The differences in the surface structures of the copolymeric nanoparticles seemed to be smaller than those of homopolymeric nanoparticles. The signals only from each peptide sensor were insufficient for the identification of copolymeric nanoparticles.

To identify of copolymeric nanoparticles, the number and combination of peptide sensors were systematically changed. Two-dimensional LDA score plots were obtained for every sensor combination (Figures S6, S7, and A). The data sets obtained from two, three, and four peptide sensors were comprised of four-, six-, and eight-dimensional data, respectively. The Calinski–Harabasz (CH) index, which represents the separation performance of clusters by calculating intracluster cohesion and intercluster dispersion, was evaluated. The CH index cannot be interpreted in absolute terms because its value depends on the number of clusters and samples. However, in this study, the CH index was used solely for relative evaluation, as the number of clusters and samples remained consistent across all analyses. The CH index tended to increase with an increasing number of peptide sensors (Table ), indicating that the clusters were more dispersed and the plots of each cluster were more cohered with increasing number of peptide sensors. Indeed, the overlap of the ellipses tended to decrease when two or more peptide sensors were used. These results suggest that an increased number of peptide sensors potentially promotes the identification of copolymeric nanoparticles.

5.

Classification of the PBA and copolymeric nanoparticles through machine learning. (A) Two- and (B) three-dimensional LDA score plots obtained from the fluorescence intensities of four peptide sensors. The ten scores for each polymeric nanoparticle in the plot (A) were enclosed in an ellipse with a 95% confidence level. (C) HCA dendrogram for the data sets obtained from four peptide sensors for the PBA and copolymeric nanoparticles.

1. CH Index and Classification Accuracies of Two Cross-Validations for the PBA and Copolymeric Nanoparticles.

	peptide sensor(s) used
name of peptide sensor(s) and their combinations	P1	P2	P3	P4	CH index	LOOCV accuracy (%)	S5CV accuracy (%)
P1	●				1184	95.5	96.4
P2		●			2297	98.2	97.3
P3			●		2780	95.5	94.5
P4				●	2739	77.3	80.9
P1/P2	●	●			2936	99.1	99.1
P1/P3	●		●		2956	98.2	99.1
P1/P4	●			●	3050	100.0	100.0
P2/P3		●	●		4797	99.1	99.1
P2/P4		●		●	4017	100.0	100.0
P3/P4			●	●	4080	100.0	100.0
P1/P2/P3	●	●	●		4917	99.1	99.1
P1/P2/P4	●	●		●	4210	100.0	100.0
P1/P3/P4	●		●	●	4142	100.0	100.0
P2/P3/P4		●	●	●	5231	100.0	100.0
P1/P2/P3/P4	●	●	●	●	5313	100.0	100.0

The generalization performance of the learning models was evaluated using LOOCV and S5CV (Tables and S11–S14). In the case of two-sensor combinations, half of the combinations (P1/P4, P2/P4, and P3/P4) achieved 100% classification accuracy for both validations. Even though the single use of P4 provided poorer discriminative signals than the others (see the smallest LOOCV and S5CV accuracies compared with other sensors), the combination of P4 with other sensors clearly provided discriminative signals. From the viewpoint of practical application, a smaller number of peptide sensors might be useful. The successful classification of two-sensor combinations clearly demonstrates the high potential of peptide sensors. In the case of the three-sensor combinations, three of the four combinations achieved 100% classification accuracy for both validations, suggesting that the ratio of better learning models increased with an increasing number of peptide sensors. Considering the best score of the CH index and the 100% classification accuracy for both validations, the use of the four peptide sensors seemed to be a better choice for the identification of copolymeric nanoparticles. In fact, the three-dimensional LDA score plot for the four peptide sensors clearly visualized the separation of each cluster composed of ten plots (Figure B).

HCA was applied to the data sets obtained from seven promising combinations of peptide sensors, which achieved 100% classification accuracy for both validations (Figures S8 and C). Remarkably, the ten data sets obtained from four (P1/P4, P2/P3/P4, P1/P3/P4, and P1/P2/P3/P4) of the seven combinations were successfully clustered into the same nanoparticles, indicating that the four combinations provided robust signals for both supervised and unsupervised machine learning to identify the copolymeric nanoparticles.

Evaluation of the Relationship between Fluorescent Signals and Chemical Structure of Poly­(butyl acrylate) and Copolymeric Nanoparticles

To further evaluate the potential relationship between the fluorescent signals and chemical structures of copolymeric and PBA nanoparticles, canonical correlation analysis (CCA) was performed. CCA is a statistical technique used to explore relationships between two multivariate data sets by calculating the linear transformations of variables that maximize their correlation. For CCA, the chemical structures of the copolymeric nanoparticles and PBA were converted into SMILES representations and described using MACCS keys. The copolymeric nanoparticles were represented as fingerprints weighted by their composition ratios (mol %) and used as the objective variable, while the signals extracted from the fluorescence spectra of the four peptide sensors were used as the explanatory variable. By applying CCA to these variables, linear transformations that maximized the correlation between the objective and explanatory variables were obtained. Two-dimensional score plots were created by plotting the explanatory variables against the linear transformations (Figure ).

6.

CCA score plot of the PBA and copolymeric nanoparticles.

The first and second canonical correlation coefficients were exceptionally high (the first canonical correlation: 0.999; the second canonical correlation: 0.995), suggesting that the fluorescence signals contain information highly correlated with the chemical structures of the polymeric nanoparticles. In other words, the fluorescent signals obtained from the peptide sensors contain information that reflects the chemical structures of polymeric nanoparticles. Moreover, the score plots showed clear separation even when structural information was included, with distinct plots of the same type of polymeric nanoparticles clustering. This observation reinforces the notion that the fluorescent signals differ significantly across nanoparticle types. In general, excessively high canonical correlation coefficients may indicate potential overfitting in the model. To assess this possibility, two cross-validation methods (namely, LOOCV and S5CV) were applied to the CCA model. The classification accuracies of both the methods exceeded 98%, indicating that the model could not be overfitted. These findings demonstrate that the fluorescent signals obtained from peptide sensors not only strongly correlate with the chemical structures of polymeric nanoparticles, but also provide highly reliable differentiation between nanoparticle types, underscoring the potential of peptide sensors as robust analytical tools for characterizing the surface of polymeric nanoparticles.

Discussion

Among the sensor substances potentially used for water-dispersed polymeric nanoparticles, this study focused on short peptides conjugated with a microenvironment-sensitive fluorophore, because of their structural designability for functional modulation. Peptide sensors seemed to physically interact with polymeric nanoparticles in the aqueous phase to effectively provide fluorescence spectra that are distinctive for polymeric nanoparticles. Remarkably, the slight substitution of amino acids involving the substitution of a single amino acid (namely, each His residue of P1) with another one (namely, Ala for P2, P3, and P4) effectively modulated the interaction with polymeric nanoparticles, providing variations in the fluorescence spectra. Machine learning of signal patterns composed of fluorescence intensities at appropriate wavelengths was useful for identifying a wide variety of polymeric nanoparticles. The use of multiple peptide sensors (i.e., more signals for machine learning) was beneficial to identify structurally similar polymeric nanoparticles. Meanwhile, the appropriate combination of peptide sensors was also significant for polymeric nanoparticles identification. Although the species and position of a microenvironment-sensitive fluorophore (namely, the AN group) did not vary in this study, such modulation would provide further variations in the fluorescence spectra. Considering the structural designability of peptide sensors, there would be no limit to the identification of polymeric nanoparticles. In fact, a quantum computing-based method for predicting peptide sequences with high affinity for microplastics has recently been reported, providing a promising approach for further optimization of peptide sequences as sensors. These findings suggest the high potential of peptide sensors for future applications to nanoplastics sensing.

Many efforts have been made to develop technologies for nanoplastics identification; most existing studies rely on the use of as-synthesized polymeric nanoparticles, such as polystyrene or PMMA nanoparticles. ,,,, However, actual nanoplastics in the environment are far more complex than as-synthesized polymeric nanoparticles because of variations in morphologies and changes in chemical structures (i.e., partial degradation, including oxidation) induced by various environmental factors. Nevertheless, the present results obtained from as-synthesized polymeric nanoparticles with subtle differences in chemical structures (i.e., polymeric nanoparticles composed of different homopolymers and copolymers) would be significant to demonstrate the potential of peptide sensors that discriminate the slight structural differences of polymeric nanoparticles. Although recent efforts have focused on artificially developing standard samples for nanoplastics, − there are currently no standard samples that consider the morphological changes that nanoplastics undergo in the environment. To fully assess the practical usefulness of our system, more realistic or standard samples will be applied in the future.

The machine learning techniques employed in this study (namely, PCA, LDA, and CCA) are relatively simple linear algorithms that offer the advantages of low computational cost and effectiveness with smaller data sets. However, these techniques have inherent limitations when dealing with complex nonlinear relationships and high-dimensional data, where the performance may degrade. Although the identification of polymeric nanoparticles was successfully demonstrated using these algorithms, the accuracy was likely to decrease as the number of target samples increased. Several studies have explored the application of deep-learning algorithms to detect and identify polymeric nanoparticles from complex signals. − These approaches leverage the power of deep learning algorithms to uncover intricate patterns within large data sets, offering enhanced accuracy in detecting and/or distinguishing polymeric nanoparticles even when dealing with highly complex and noisy data. As deep learning algorithms evolve, their ability to handle complicated environmental samples will open a promising avenue in this field.

Conclusion

In summary, inspired by the recognition mechanism for human taste and olfactory systems, we succeeded in the identification of polymeric nanoparticles with slightly different chemical structures through machine learning of fluorescence signal patterns obtained from peptide sensors. Our systematic evaluation revealed the critical role of both the number and combination of peptide sensors in achieving precise identification of polymeric nanoparticles. Furthermore, the fluorescent signals generated by the peptide sensors likely reflect information about the chemical structure of the polymers composing the polymeric nanoparticles. This suggests that the proposed identification system potentially represents a robust analytical approach not only for identifying polymeric nanoparticles but also for assessing their surface characteristics. Unlike SERS and Raman spectroscopy, which have been proposed as analytical methods for nanoplastics, the method presented here does not require a drying process for sample preparation. This crucial feature allows for the direct analysis of nanoplastics in their solution state, a capability particularly valuable for investigating nanoplastics in their native state as they exist in aquatic environments like rivers and oceans. Furthermore, this approach is expected to enable more detailed analysis of surface properties and facilitate toxicity assessments of nanoplastics in aqueous environments in the near future.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acssensors.5c01000.

• Experimental details, raw fluorescence spectra, characterizations of polymeric nanoparticles, LOOCV classification accuracies, LDA score plots, and HCA dendrograms (PDF)

S.H.: Methodology, Investigation, and Writingoriginal draft. T.S.: Methodology and Writingoriginal draft. Y.K.: Methodology, Investigation, and Writingoriginal draft. M.N.: Methodology, Investigation, and Writingreview and editing. T.K.: Methodology, Supervision, and Writingreview and editing. T.S.: Conceptualization, Methodology, Supervision, and Writingreview and editing.

This study was partially supported by a Grant-in-Aid for Challenging Research Exploratory (JP22K19844) from the Japan Society for the Promotion of Science (JSPS) for T.S., and The Japan Science and Technology Agency SPRING (JPMJSP2106 and JPMJSP2180) for S.H.

The authors declare no competing financial interest.