Machine learning–enhanced surface-enhanced spectroscopic detection of polycyclic aromatic hydrocarbons in the human placenta

Significance

Employing advanced techniques for the detection of environmental toxins in pregnancy can potentially result in improving public health policy and clinical management, leading to more effective and widespread screening practices for toxins with known associations with adverse birth outcomes. Here, we combine light-based chemical identification methods enhanced by machine learning to clearly identify the presence of harmful polycyclic aromatic hydrocarbons (PAHs) and polycyclic aromatic compounds (PACs) in the human placenta of mothers who are smokers, clearly distinct from nonsmokers. This approach demonstrates significant potential for the development of streamlined, rapid methods for detecting human exposures to multiple environmental toxins across various tissue and fluid samples, as well as understanding their impact on human health.

Abstract

The detection and identification of polycyclic aromatic hydrocarbons (PAHs) and their derivatives, polycyclic aromatic compounds (PACs), are essential for environmental and health monitoring, for assessing toxicological exposure and their associated health risks. PAHs/PACs are the most dangerous chemicals found in tobacco smoke, and cigarette use during pregnancy can convey these molecules to the developing fetus through the placenta. This exposure is associated with many negative health outcomes, from premature birth to sudden infant death syndrome and adverse neurodevelopmental disorders. This study demonstrates the use of surface-enhanced Raman and surface-enhanced infrared absorption spectroscopies for direct detection of PAHs/PACs in human placental tissue. We applied two spectroscopy-informed machine learning algorithms, Characteristic Peak Extraction (CaPE) and Characteristic Peak Similarity (CaPSim), to identify the specific PAHs and PACs present in the placenta of women who smoked tobacco cigarettes in pregnancy compared to spectra of the placenta from self-reported nonsmokers. CaPE and CaPSim analysis enabled a clear distinction between these two groups. Independent verification was accomplished by detecting PAH-DNA and PAC-DNA adducts in the smoking group by means of a ³²P-postlabeling assay. These findings highlight the effectiveness of combining surface-enhanced spectroscopies with informed ML analysis for the streamlined detection of hazardous environmental compounds in human tissues, suggesting broader applications in clinical diagnostics and public health surveillance.

Polycyclic aromatic hydrocarbons (PAHs) are well known as a significant environmental health concern due to their known toxicity and carcinogenicity (1–3). These compounds are predominantly produced through the incomplete combustion of organic materials and can enter the human body via inhalation, ingestion, or dermal absorption. Once inside, these hydrophobic molecules can diffuse across cell membranes, bind to receptors, and activate signal pathways that alter gene expression (3, 4). The metabolic path of PAHs involves the activation of cytochrome P450 (CYP) response genes, including CYP1A1 and CYP1B1, enzymes that biotransform PAHs into DNA-reactive metabolites as well as hydroxyl-radical cation intermediates. These intermediates are subsequently processed by Phase II detoxifying enzymes, facilitating their excretion from the body (5). However, the reactivity of these intermediates also allows for their covalent binding to DNA, forming adducts that disrupt cellular functions such as cell proliferation, apoptosis, and DNA repair (6–9). PAH exposure has been linked to a range of health issues, including childhood asthma, adult respiratory disorders, cardiovascular diseases, and reproductive problems (1, 2, 10, 11). PAHs can undergo environmental or metabolic transformations to form polycyclic aromatic compounds (PACs) by adding oxygen and nitrogen atoms into their structures (12). These functional groups facilitate the passive diffusion of these molecules into cells, potentially leading to oxidative stress, cytotoxicity, and chronic disease development (13–16). These multiple serious impacts of PAHs and PACs on human health highlight the importance of their timely and accurate detection in biological and clinical samples.

Studies of PAH exposures during pregnancy are important for public health. Since PAHs can cross the placenta, smoking during pregnancy exposes the developing fetus to these hazardous chemicals. Smoking during pregnancy is associated with increased risk of miscarriage, prematurity, stillbirth, low birth weight, perinatal morbidity, sudden infant death syndrome, and adverse neurodevelopmental disorders such as Attention-Deficit Hyperactivity, anxiety, and depression (17–22). PAHs can also disrupt placental function. In vitro studies using placental cell culture (i.e., trophoblast cells) demonstrate that exposure to PAHs disrupts the natural apoptosis pathway, alters endocrine signaling, and changes proliferative potential (23–26). Our studies, along with findings from others, indicate that PAH exposures are associated with PAH-induced DNA bulky adducts in the placenta, revealing mutagenic DNA damage to placental cells as exposure levels increase.

PAHs are among the most harmful substances found in tobacco smoke (27, 28). A single cigarette can produce between 474.17 to 1,607.2 ng of PAHs, depending on brand (29, 30). Smoking during pregnancy leads to changes in the placental transcriptome and epigenome and reveals an increase in markers of oxidative damage (31–35). We have also reported increased expression of CYP1A1 in the placenta from smokers (35). To effectively monitor these compounds, various analytical methods have been developed, such as high-performance liquid chromatography coupled with fluorescence detection, gas chromatography-mass spectrometry (GC-MS), time-of-flight secondary ion mass spectrometry (ToF-SIMS), and ultraviolet-visible spectroscopy (36–41). While each method offers unique advantages and collectively contributes to the comprehensive study and management of PAH and PAC exposure, most require significant preparation and operational time. In contrast, surface-enhanced Raman spectroscopy (SERS) and surface-enhanced infrared absorption (SEIRA) leverage the unique vibrational fingerprints of molecules, enabling the detection of compounds with similar structures that pose challenges for GC-MS analysis. Integrating machine learning (ML) with these approaches further enhances their capability to detect and classify unknown or novel compounds by recognizing patterns in the spectral data that may not be apparent through GC-MS analysis alone. Therefore, developing streamlined approaches like SERS and SEIRA represents a significant step toward reducing analysis time and cost, thereby making toxicity screenings more accessible and efficient, particularly in addressing environmental toxins with significant public health implications.

Recent advancements have integrated SERS with SEIRA and ML to enhance the detection capabilities of PAHs/Cs (19, 20, 37, 42). Combining SERS with SEIRA provides the complementary detection of both Raman-active and IR-active vibrational modes, providing sufficient information for the chemical identification of unknown molecules not available with either Raman or IR spectroscopies alone (43, 44). Furthermore, combining surface-enhanced spectroscopies with the analytic power of ML enables more efficient and precise analysis of complex mixtures, including studies of exposures within complex media where the specific compounds and their concentrations may be unknown (42). In addition, ML can be used to connect Raman spectroscopy and SERS, where there are frequent nuisance variations and spectral discrepancies between the two methods, facilitating the direct use of Raman library data to identify molecules using SERS (45). This SERS-ML tandem approach is essential for rapidly assessing human exposures following events such as wildfires, which release significant amounts of PAHs. Implementing automated platforms for SERS and SEIRA data acquisition can further streamline this process, enhancing spectral library expansion and facilitating simultaneous analysis of multiple samples. By leveraging ML-trained algorithms, the interpretation of complex spectral data is significantly accelerated, reducing human intervention while improving the accuracy and comprehensiveness of environmental and biological analyses. In this study, we combine SERS and SEIRA to detect PAHs and PACs in placental tissues from smokers, using nonsmokers as controls. We utilized specialized, spectroscopy-informed ML algorithms, CaPE and CaPSim, developed for characteristic peak selection of the spectral features of SERS with the capability of being particularly robust to the inherent nuisance variations of this spectroscopy (45). Using this approach, we can clearly distinguish a distinct chemical signature of smokers vs. nonsmokers due to the presence of PAHs and PACs in the smokers’ placenta. We then analyzed for the presence of PAH-DNA adducts in the placenta using a ³²P-postlabeling assay, which showed clear, independent evidence of these adducts in the placenta of smokers vs. nonsmokers in the subject pool. Not only does this set of experiments highlight the clear ability to detect known toxins and carcinogens being passed from mother to fetus via the placenta, but it also shows that surface-enhanced spectroscopic monitoring of placental tissue is capable of providing this crucial screening information to doctors and patients.

Results

A schematic representation of the sample preparation and data collection, along with the substrate design specifically tailored for PAH/PAC detection using both SERS and SEIRA, is shown in Fig. 1A. The substrate contains core-shell (SiO2-Au) nanostructures called Au nanoshells (NS) with a diameter of 168 ± 10 nm (see Materials and Methods section for details). The NS plasmon resonance mode at 800 nm aligns closely with the 785 nm laser excitation wavelength used in SERS, which enhances surface plasmon resonance and increases detection sensitivity (46). In the case of SEIRA, the substrate boosts mid-IR electromagnetic fields in the molecular fingerprinting region by optimizing the SiO₂’ core diameter (D = 180 nm), Au shell thickness of 10 nm, and interparticle spacing. This mid-IR (MIR) enhancement specifically amplifies the vibrational modes of molecules in the NS junctions (44). A high-angle annular dark field (HAADF) scanning transmission electron microscopy (STEM) image shows the core-shell spherical geometry of Au NSs (Fig. 1B). Energy-dispersive X-ray spectroscopy (EDS) elemental mapping of Au NS confirms the spatial distribution of elements within the nanoparticles. A dark field image of the nanoparticles (NS) highlighting their morphology and an inset of the particle size distribution is shown (Fig. 1C). The intensity profile (Fig. 1D) from a representative cross-sectional area generated along a line drawn across a particle highlights the geometric composition of the SiO2 core (red and cyan) and the Au shell (yellow). Additionally, Energy Dispersive X-ray Spectroscopy (EDX) analysis (Fig. 1E) provides detailed insights into the elemental composition of these nanoparticles, confirming the presence of Si, O, Cu, and Au atoms (where the Cu signal is attributed to the copper samples supporting the TEM grid).

Fig. 1.

Schematic diagram of sample preparation, data collection, and Au NS elemental analysis. (A) Schematic diagram illustrating the PAH extraction from human placental tissue and the Au NS-based SERS-SEIRA substrate used for PAHs/PACs detection. (B) A HAADF microscopy image displays a core-shell structure consisting of a 120 nm silica core surrounded by an approximately 15 nm thick gold shell. Elemental mapping images showing the distribution of gold (yellow), silica (cyan), and oxygen (red), along with a composite image integrating all elements. (C) Darkfield image of the NS, with an Inset showing the size distribution, (D) elemental map distribution, and (E) EDX analysis.

Smoking during pregnancy is a major source of PAH exposure, with smokers inhaling approximately 0.26 µg of benzo[a]pyrene (B[a]P) per pack of 20 cigarettes, along with other harmful compounds (47). Previous research reported increased levels of B[a]P, benzo(b)fluoranthene (B[b]F), and dibenz[a,h]anthracene (DB[a,h]A) in the placentae of preterm deliveries compared to term deliveries, with PAHs quantified by GC-MS (26). Here we analyzed placental samples collected from women who self-reported the use of tobacco cigarettes in pregnancy and compared them with samples from self-reported nonsmokers (Fig. 2). Representative SERS and SEIRA spectra of placental samples collected from smokers and nonsmoker patients are shown in Fig. 2 A and B. By immediate visual analysis, the smoker vs. nonsmoker spectra are nearly indistinguishable. Additionally, reference Raman and FTIR spectra of B[a]P, B[b]F, DB[a,h]A, and TQE are presented for comparison. The characteristic SERS peaks (Fig. 2A) corresponding to various compounds were identified exclusively in the placenta of smokers. For B[a]P, peaks were observed at 527 cm⁻¹ (CC, CH in-plane bending, and CCCC torsion), 613 cm⁻¹ (CC and CH in-plane bending), 898 cm⁻¹ (CC and CH out-of-plane bending), and 1,443 cm⁻¹ (CC stretching, CH in-plane bending) (48). B[b]F exhibited peaks at 689 cm⁻¹ (CCCC stretching), 1,004 cm⁻¹ (CCCH torsion), and 1,443 cm⁻¹ (CC stretching) (49). DB[a,h]A showed peaks at 1,296 cm⁻¹ (CC stretching and CH bending) and 1,562 cm⁻¹ (CC stretching) (49). For TQE, peaks were identified at 450 cm⁻¹ (in-plane skeletal distortion), 612 cm⁻¹ (out-of-plane skeletal deformation), and 1,200 cm⁻¹ (in-plane skeletal distortion), which were exclusively present in the placenta of smokers. These peaks, highlighted with light blue dash lines, were exclusive to the placental tissue of smokers while absent in nonsmoker tissue.

Fig. 2.

SERS and SEIRA spectra of PAH/C in the human placenta. Left: (A) SERS and (B) SEIRA averaged spectra from placental samples collected from smoker and nonsmoker individuals. Dashed lines highlight the anticipated peak locations.

SEIRA peaks (Fig. 2B) identified the presence of TQE with vibrational peaks 1,165 cm⁻¹ (CH bending), 1,209 cm⁻¹ (CO stretching, OH in-plane deformation), and 964 cm⁻¹ (COC stretching) (37). B[a]P exhibited peaks at 1,176 cm⁻¹ (CH in-plane bending) and 829 cm⁻¹ (CH out-of-plane bending). DB[a,h]A displayed peaks at 759 cm⁻¹ and 818 cm⁻¹ (CH out-of-plane bending), while B[b]F presented peaks at 1,220 cm⁻¹ (CH in-plane bending). These SEIRA peaks, highlighted with light green dash lines, were found only in the placental tissue of smokers while absent in nonsmoker tissue. Both SERS and SEIRA spectroscopies confirmed the presence of PAHs in the placentas of smokers.

SERS spectra of placental samples from smokers and nonsmokers and the corresponding reference extracted peaks from normal Raman spectra of targeted PAH/Cs are shown in Fig. 3 A and B. The Characteristic Peak Extraction (CaPE) algorithm was utilized to extract important and reliable peaks from the SERS spectra (42), while the characteristic peak similarity (CaPSim) algorithm (45) was employed to compute the similarity between the extracted SERS peaks from the placental samples and the Raman reference PAH and PAC spectra. The peaks in the SERS spectra aligned with the CaPE-extracted features from the normal Raman spectra (Fig. 3 A and B), providing a clear visual comparison of how closely the sample peaks match with those known for TQE, DB[ah]A, B[b]F, B[a]P, pyrene (PYR), and anthracene (ANTH).

Fig. 3.

CaPE SERS analysis of PAH/C in the human placenta. SERS-ML spectral analysis of placenta samples from (A) smokers and (B) nonsmokers. Top spectra: 25 superimposed spectra from (A) smokers and (B) nonsmokers. Lower spectra: CaPE-extracted peaks from the normal Raman spectra of TQE, DB[ah]A, B[b]F, and B[a]P, respectively.

A detailed analysis of the similarity value distributions across six specific PAHs/PACs (B[a]P, B[b]F, DB[a,h]A, PYR, ANTH, TQE) is displayed in Fig. 4 A–D. Two distinct similarity measurement approaches, CaPSim and NormSim, are utilized. CaPSim measures the similarity between the extracted peaks of a PAH/PAC SERS spectrum using CaPE and those of a placenta sample SERS spectrum, while NormSim measures the similarity between the full spectra. The distribution of NormSim shows significant overlap between smokers and nonsmokers for all analyzed PAHs/PACs, indicating its limited capability in differentiating between these groups. In contrast, CaPSim distributions reveal a more distinct separation between smokers and nonsmokers for B[a]P, B[b]F, D[a,h]A, and TQE. A closer examination of TQE (last row in Fig. 4B) shows that, although the similarity distribution for smokers and nonsmokers are not fully separated in either case, CaPSim captures different modalities: a simple Gaussian for smokers and a mixture of two Gaussians for nonsmokers. NormSim, however, models both distributions as Gaussian, suggesting that CaPSim still provides a larger gap for TQE discrimination, as confirmed in Fig. 4E. It is important to note that PACs such as TQE are generally better distinguished by IR and SEIRA spectroscopy due to the presence of their dipole-active, heteroatomic functional groups (in this case, the quinoid ring of TQE). For PYR and ANTH, the CaPSim similarity values distribution shows a substantial overlap between the smoker and nonsmoker groups, indicating the absence of these compounds in the placental samples.

Fig. 4.

Similarity values for PAHs/PACs in placental samples from smokers vs. nonsmokers. Distribution of similarity values for smokers and nonsmokers in placental samples across different PAHs/PACs using (A) CaPSim and (B) NormSim algorithms. In (A), red histograms represent the similarity values for smokers, and blue histograms represent the values for nonsmokers. In (B), green histograms represent the similarity values for smokers, and orange histograms represent the values for nonsmokers. (C) CaPSim similarity values for Pyrene and Anthracene. (D) NormSim similarity values for Pyrene and Anthracene. (E) Wasserstein distance, a metric for measuring the dissimilarity between probability distributions (see details in the Materials and Methods section) between smokers and nonsmokers, is calculated for each PAH/C and similarity type (CaPSim and NormSim).

To further quantify the separation between smoker and nonsmoker distributions observed in Fig. 4 A–D, we calculated the Wasserstein distance (50) (see details in the Materials and Methods section) between these distributions for each PAH/PAC and similarity method, as shown in Fig. 4E. The results confirm CaPSim’s superior differentiation ability, which produces much larger distances for B[a]P, B[b]F, DB[a,h]A and TQE, compared to NormSim. The smaller distances for PYR and ANTH corroborate their absence in both groups. This analysis shows that combining SERS with ML-based peak extraction techniques provides a promising approach to accurately and robustly identifying the impact of smoking on placental tissues.

In determining the detection capabilities of our system, we analyzed the signal-to-noise ratio (SNR) of our spectral measurements, as this is a critical parameter for calculating the Limit of Detection (LOD). The SNR was determined by the ratio of the peak-to-peak amplitude (signal strength) to the SD of the spectral signal (noise level). Raw SERS spectra from placental samples (SI Appendix, Fig. S1A) exhibited robust SNR values ranging from 5.4 to 9.2, indicating strong signal quality even before processing. After applying baseline removal and Savitzky–Golay smoothing, the SNR values improved significantly (SI Appendix, Fig. S1B), ranging from 9.2 to 18.3, particularly in smokers’ samples, where SNR increased from 7.5 to 9.2 to 10.0 to 18.3. Furthermore, our implementation of the CaPE algorithm and CaPSim analysis provides additional improvement in signal quality by specifically isolating and analyzing characteristic spectral features while reducing the influence of background noise. Using our combined SERS-ML method, we achieved LOD values of 0.027 ppm for B[a]P and B[b]F and 0.03 ppm for DB[a,h]A, with a limit of quantification (LOQ) of 89.57 ppm for B[a]P and B[b]F and 98.81 ppm for DB[a,h]A (details in SI Appendix). Similarly, the SEIRA method demonstrated a LOD of = 1.96 ppm and a LOQ of 6.54 ppm for TEQ, emphasizing the high sensitivity of these approaches for detecting harmful substances transmitted from mother to fetus.

The placenta of smokers has been shown to also exhibit up-regulated CYP enzymes, including CYP1A1, and shows increased markers of oxidative damage, likely due to PAH exposures from tobacco smoke (34, 35). These by-products can bind to cellular DNA and proteins, causing cellular damage, inflammation, and an increased risk of cancer development (2, 51). This is consistent with other studies that have shown that smokers have increased levels of placental DNA adducts (52). To further investigate the presence of PAH-induced DNA adducts in the placenta, we used an ultrasensitive assay called ³²P-postlabeling (53, 54). This was performed using placenta from smokers (N = 6) and nonsmoking controls (N = 6). Smoking status was self-reported and not determined through urinary cotinine levels due to the lack of sample availability through the biobank. PAH-induced DNA adducts are formed at very low levels and serve as a biomarker of exposure (55). Using this technique, we obtained qualitative and quantitative data on placental DNA adducts. Placental DNA adducts in smokers showed significantly higher intensities (Fig. 5B) compared to nonsmokers (Fig. 5A), and quantitative analyses confirmed a significant elevation of cigarette smoking-related DNA adducts (Fig. 5C) in placental tissue from smokers. The levels of chromatographically distinct placental DNA adducts were higher in smoker samples compared to nonsmokers, as indicated by the circles and numbers in Fig. 5 A and B. These adducts are likely derived from PAHs, consistent with previous findings in humans and mouse placentas using ³²P-postlabeling. In a landmark study, Randerath et al. first reported that cigarette smoking was linked to covalent DNA adducts in the human placenta (56). Additionally, Lu et al. demonstrated that several carcinogens, including B[a]P and its metabolites, could cross the placenta and form DNA adducts in fetal tissues, with potential implications for transplacental carcinogenesis (57). Our DNA adduct data strongly support the use of SERS and SEIRA, highlighting their potential in detecting such alterations.

Fig. 5.

Comparison of bulky DNA adducts from placentae. An autoradiograph of multidirectional axion-exchange thin-layer chromatography shows the profiles of DNA covalent adducts from placenta tissues collected from (A) nonsmokers and (B) smokers. Intensifying screen-enhanced autoradiography was performed at −80 °C for 92 h. (C) Comparison of PAH-induced adducts in human placentae tissues collected from (blue) nonsmokers and (red) smokers. Values are means of measurements from N = 6 placental samples/group, and error bars represent the SEM (* indicates P < 0.001).

Discussion

In this study, we observe that surface-enhanced vibrational spectroscopies such as SERS and SEIRA can detect PAHs within human tissue samples. Analysis of placental tissue from smokers and nonsmoker patients reveals specific spectral peak signatures characteristic of PAHs/Cs contaminants only in the smoker’s tissue. While qualitative analysis of the SERS and SEIRA spectra alone may not be sufficient to definitively confirm or deny the presence of PAHs/PACs in smoking pregnant individuals, integrating SERS with ML analysis demonstrates a higher likelihood of detection of the contaminants in smokers. 32P-postlabeling assays provide independent evidence of elevated levels of PAHs in placental tissues from smokers compared to nonsmokers through the detection of PAH-DNA adducts, which correlate positively with levels of PAH exposure. Overall, ML-enhanced SERS+SEIRA offers an accessible and efficient alternative to standard methods for clinical and toxicological assessments. While self-reported smokers during pregnancy provided this study with a known source of PAH exposure, this type of streamlined detection methodology could also be useful to determine PAH exposure following natural disasters, such as wildfires, as well as man-made disasters, such as industrial fires or battlefield burn pits (58). Ultrasensitive detection leveraging SERS and SEIRA can provide information on complex mixtures of PAHs that will help inform public health measures. Further studies are warranted to determine the potential for SERS to detect multiple other chemicals with known harmful effects on human health.

Materials and Methods

Materials.

(3-aminopropyl) triethoxysilane (ES, 99%), tetra chloroauric acid (HAuCl4·3H2O), tetrakis hydroxymethyl phosphonium chloride (THPC), poly-L-lysine hydrobromide (MW 150,000 to 300,000) poly-L-Lysine (PLL) pyrene, and anthracene were purchased from Sigma-Aldrich. Formaldehyde (37%), sulfuric acid (H2SO4, 100%), hydrogen peroxide (H2O2, 30%), potassium dihydrogen phosphate (KH2PO4), and 200-proof ethanol were obtained from Fisher Scientific. All chemicals were used as received without further purification. Water was deionized and filtered by a Milli-Q water system (18.2 MΩ cm at 25 °C, Millipore). Quartz slides were obtained for substrate fabrication (Fisher Scientific). The 120 nm diameter aminated SiO2 cores in ethanol (10 mg/mL) were purchased from nanoComposix, Inc., San Diego, CA.

Synthesis of Au NS.

The Au NSs were synthesized using a previously reported procedure (59). Small Au colloids (1 to 3 nm diameters) were synthesized by reducing chloroauric acid with a THPC reducing agent. Then, 300 μL aminated SiO₂ cores were incubated with 20 mL THPC-Au solution for 24 h. to enable the attachment of the Au colloids onto the SiO₂ surface, followed by multiple cleaning cycles to remove the excess reaction chemicals. Following this step, an electroless plating process reduced Au onto the SiO₂ surface, forming a continuous Au shell. The electroless deposition was achieved by reducing Au from a 3 mL solution of 1.8 mM potassium carbonate and 0.4 μM chloroauric acids with 15 μL formaldehyde. The Au NSs were fabricated with the core and shell radii [r1, r2] = [63, 75] nm, corresponding to an Au NS plasmon resonance of 785 nm (Raman laser wavelength) in H₂O.

SERS Substrate Preparation.

Cleaned quartz slides (from Thermo Scientific™ Quartz microscope slide, fused, 76.2 × 25.4 mm) were immersed in 0.01% w/v aqueous solution of PLL (MW 1,50,000 to 3,00,000) for 5 h to facilitate subsequent attachment of nanoparticles on the quartz surface. The surfaces were cleaned with ethanol and dried in a flow of N₂. Cut silicone isolators (Grace Bio-Labs Press-To-Seal silicone isolator, No PSA, 9 mm diameter) were placed on the quartz-PLL substrate, followed by a dry drop deposition of 100 μL 1010NSs/mL Au NS or NPs aqueous solution into the isolator well. Then, 20 μL of biosamples were directly dried on the Au NSs SERS substrates. Before acquiring the SERS spectra, the substrates were fully immersed in Milli-Q water.

SEIRA Substrate Preparation.

A 525 μm thick silicon wafer (undoped,<100>, FZ, University Wafers) was cut into small square supports and cleaned through sonication, first in acetone and then in isopropanol. Subsequently, the silicon supports were blow-dried with dry nitrogen. Next, 100 μL of the NSs was drop-cast onto silicon support and allowed to dry under ambient conditions. Following this, 20 μL of the biosample was directly drop-cast onto the substrate and left to dry. The NSs utilized for the SEIRA experiments were fabricated with the core and shell radii [r1, r2] = [100, 115] nm, resulting in a dipole plasmon resonance of 965 nm and a quadrupole plasmon resonance of 685 nm in H2O.

Placenta Samples.

Placenta tissue was obtained through our PeriBank biobank, as previously described (60). Subjects were enrolled by trained research personnel at the time of delivery. Subjects provided informed consent at the time of enrollment for biobanking, as well as future research use with the biobanked samples. All samples were assigned a unique code at the time of enrollment. Sample collection protocols were approved by the Baylor College of Medicine Institutional Review Board (H-26364). Following delivery of the placenta, cuboidal sections of the placenta were uniformly excised from a location approximately 4 cm from the cord insertion site. Decidua and chorion were excised and discarded, and the sample was snap-frozen “as is” at −80 °C until future use. All samples were collected from subjects who self-reported smoking during the third trimester (N = 6) and from nonsmoking controls (N = 6), all of whom had term deliveries without any pregnancy complications.

Extraction Procedure.

Approximately 100 mg of each placenta tissue sample was homogenized in 1,000 µL of hexane using a Janke & Kunkel Ultra-Turrax T25 mixer at 20,000 RPM for 30 s. Following initial homogenization, 1,000 µL of chloroform was added, and the mixture was vortexed. The homogenate mixture was then allowed to stand at room temperature for 30 min to facilitate phase separation. Subsequently, the samples were centrifuged at 5,000 rpm for 10 min. The supernatant was carefully transferred to a new container and dried under a stream of nitrogen at room temperature to ensure complete solvent removal. Finally, the dried residues were reconstituted in 500 µL of acetone vortex to ensure homogeneity and used for further assay analysis.

DNA Bulky Adduct Labeling.

The nuclease P1-enhanced bisphosphate version of the ³²P-postlabeling assay was performed as previously described (26, 61, 62). Briefly, placental DNA was enzymatically degraded to normal (Np) and adducted (Xp) deoxyribonucleoside 3′-monophosphates with a mixture of micrococcal nuclease and spleen phosphodiesterase. Adducted nucleotides are converted to 5′−³²P-labeled deoxyribonucleoside 3′,5′-bisphosphates (pXp) by incubation with carrier-free [γ−³²P]ATP and polynucleotide kinase. Radioactively labeled products were purified and partially resolved by one-dimensional development overnight with solvent 2.3 M sodium phosphate, pH 5.7 (D1). Labeled DNA adducts retained in the lower (2.8 × 1.0 cm) part of the D1 chromatogram were separated with 3.82 M lithium formate, 6.75 M urea, pH 3.35 and 0.72 M sodium phosphate, 0.45 M Tris–HCl, 7.65 M urea, pH 8.2 in the first and second dimensions, respectively (63). ³²P-labeled DNA adducts were detected with the aid of an Instant Imager (64).

Measurements.

SERS measurements were acquired with a Renishaw inVia Raman microscope (Renishaw) with a 785-nm excitation wavelength and 55 μW laser power at the samples. Backscattered light was collected using a 63× water immersion objective lens (Leica) NA = 0.9 with a 20-s exposure time. SEIRA spectra were recorded in a Bruker VERTEX 80v FTIR spectrophotometer in transmission mode at normal incidence. The instrument has a mercury cadmium telluride detector, a KBr beam splitter, and a SiC Globar light source. The mirror velocity was 20 kHz, and the spectral resolution was 4 cm⁻¹. Spectra were measured under a rough vacuum (∼1.0 hPa) with 256 scans and the background of the empty chambers. For baseline subtraction, a standard IR spectrum of silicone substrate was employed. All SEIRA data were processed using baseline subtraction. Extinction measurements were performed on a Cary 5000 UV/Vis/NIR Varian spectrophotometer. HAADF STEM images and EDX spectrum images were recorded using an aberration-corrected Titan Themis3 300 kV equipped with a monochromator and a Super-X energy-dispersive X-ray spectrometer. HAADF results and EDX spectral images were acquired using Velox software. SEM measurements were performed using an FEI Quanta 400 field emission SEM at a 20 kV scanning electron microscope acceleration voltage. Homogenization of the biosamples was achieved with a T25 ULTRA-TURRAX® instrument.

CaPSim.

CaPSim is a similarity metric designed for spectral recognition, particularly focusing on the characteristic peaks (CPs) of SERS spectra (45). Initially, an algorithm is utilized to identify potential peak locations, followed by extracting CPs from query and reference spectra. Since the CPs correspond to specific SERS vibrational modes, serving as molecular fingerprints, spectral recognition with CaPSim is more robust against noise and background peak variations. Ju et al. employ the CaPSim metric to identify chemical substances by comparing unknown SERS spectra against a database of known Raman spectra (45). We adapt this approach to detect specific PAHs/PACs in organic samples with the following procedure:

Preprocessing.

We first resample all spectra using the “interpolate” function from SciPy, ensuring consistent wavenumbers across spectra (65). The process involves linear interpolation for intensity values within the established data range, while values outside this range are addressed through extrapolation, ensuring uniformity across all spectra. To enhance the clarity of spectra from organic samples, we employ the Whitaker-Hayes algorithm to remove cosmic rays (66). This algorithm, utilizing a z-score threshold of 7 and a window size of 3, effectively identifies and eliminates spikes caused by cosmic rays, thereby refining the spectral data. Subsequent to cosmic ray removal, we implement a baseline removal strategy as proposed by Zhang et al., utilizing the default settings of the Python BaselineRemoval package to eliminate slow-changing spectral trends (67). The processed spectra are then smoothed to reduce noise further and enhance peak resolution. This is accomplished with the Savitzky–Golay filter, which is applied via the “savgol_filter” function from Scipy (68). The filter parameters are set to a window size of 11 (approximately 10 cm⁻¹) and a polynomial order of 3, optimizing the balance between smoothing and maintaining spectral detail. Finally, an additional noise removal step is applied specifically to spectra from organic samples. This step targets the residual noise floor remaining after baseline removal. Inspired by the approach of Hou et al. (69), we estimate the noise floor using the median of a running median, with a window size of 79. This estimated noise floor is then subtracted from the spectra to yield cleaner and more analytically valuable results.

CaPE.

We apply the CaPE technique to the preprocessed spectra of each PAH/PAC to identify potential CP locations (42). The max-pooling technique is then applied to compress the spectra, extracting the maximum intensity at each CP location for each targeted PAHs/Cs. To analyze organic samples, we strategically apply CP locations previously identified from the PAH/PACs of interest rather than directly determining CP locations from these individual samples. This targeted use of CaPE, combined with max-pooling based on predetermined CP locations, optimizes our ability to detect and analyze PAHs/PACs in various organic matrices, leveraging established spectral features for more reliable and efficient chemical identification.

Similarity calculation.

We utilize a comprehensive array of similarity metrics to determine the extent of similarity between the spectrum of a specific PAH/PAC and the spectra obtained from organic samples. These metrics include the dot product, soft intersection over Union (IoU), Pearson’s correlation, cosine similarity, and 1st diff cosine similarity. Each metric offers a unique perspective on spectral alignment and intensity correlation, thereby providing a multidimensional understanding of spectral similarities. Before calculating similarity scores, all spectra are subjected to l_2 normalization. This normalization process scales the spectra to the unit norm, diminishing the influence of varying signal intensities and ensuring that the similarity metrics reflect true spectral characteristics rather than differences in signal magnitude. The similarity between each PAH/PAC spectrum and the spectra from organic samples is calculated individually using the metrics mentioned above. The final similarity score is obtained by averaging individual similarity values across all recordings within a given organic sample.

The Wasserstein Distance.

To quantify the dissimilarity between the distributions of similarity scores for contaminated and uncontaminated samples, we employ the Wasserstein distance. This metric provides a robust measure of the separation between probability distributions, with larger values indicating greater dissimilarity. Unlike simple summary statistics such as means or medians, the Wasserstein distance takes into account the overall shape and spread of the distributions, making it particularly suitable for assessing the discriminative power of our detection methods. Given two sets of samples X = {x_1, x_2,…, x_n} and Y = {y_1, y_2,…, y_m}, we can compute the 1-Wasserstein distance as follows:

Step 1: Obtain Empirical Cumulative Distribution Functions (ECDFs). First, we construct the ECDFs for X and Y:

$${\widehat{F}}_n\left(t\right)=\frac{1}{n}{\sum }_{i=1}^{n}I\left[x_i\le t\right],{\widehat{G}}_m\left(t\right)=\frac{1}{m}{\sum }_{j=1}^{m}I\left[y_j\le t\right],$$

where I[⋅] is the indicator function.

Step 2: Obtain Quantile Functions. Next, we define the empirical quantile functions (inverse of ECDFs):

$${\widehat{F}}_n^{-1}\left(p\right)=\inf \left\{t\in \mathbb{R}:{\widehat{F}}_n\left(t\right)\ge p\right\},{\widehat{G}}_m^{-1}\left(p\right)=\inf \left\{t\in \mathbb{R}:{\widehat{G}}_m\left(t\right)\ge p\right\}.$$

Step 3: Compute the Wasserstein Distance. For one-dimensional data, the 1-Wasserstein distance has a closed-form solution:

$$W_1\left({\widehat{F}}_n,{\widehat{G}}_m\right)={\int }_0^1\left|{\widehat{F}}_n^{-1}\left(p\right)-{\widehat{G}}_m^{-1}\left(p\right)\right|\mathrm{d}p.$$

Step 4: Discrete Approximation. In practice, we compute this integral using a discrete approximation:

$$W_1\left(X,Y\right)\approx \frac{1}{N}{\sum }_{i=1}^N\left|{\widehat{F}}_n^{-1}\left(\frac{i}{N}\right)-{\widehat{G}}_m^{-1}\left(\frac{i}{N}\right)\right|,$$

where N is a large number (e.g., max(n,m) or n + m − 1).

Experiment Setup.

We systematically evaluated the effectiveness of CaPSim in detecting four PAHs/PACs within placental samples from smokers (340 observations) and nonsmokers (180 observations). The evaluation metric used was the area under the receiver operating characteristic curve (AUROC), which is a comprehensive measure of the model’s ability to discriminate between positive and negative instances across various threshold settings. For each PAH/PAC, the similarity values were calculated between the SERS spectrum of a placenta sample and the SERS spectrum of the specific PAH/PAC. The spectra database was then divided into three training subsets: training, validation, and test sets. The training set was used to train a classification model capable of determining the presence of a PAH/PAC in the placenta sample. The validation set was employed to select the best hyperparameters for the similarity calculations and the classification algorithm. The test set was applied to evaluate the model’s performance on unseen data.

Detection Model.

To accurately detect PAHs/PACs in placental samples, we implemented a logistic regression model utilizing similarity scores as the primary feature. For each PAH/C, the hyperparameters yielding the best performance, specifically the highest validation AUROC, were selected. Finally, these optimized parameters were used to evaluate the model’s test performance. The ground truth labels for spectra from smokers were labeled as “exist” or 1, indicating the presence of PAHs/PACs, while spectra from nonsmokers were labeled as “nonexist” 0, indicating their absence.

Hyperparameter Tuning.

The tuning involved two key steps: similarity calculation and logistic regression model configuration. Building on the work proposed by Ju et al. (45), we employed CaPSim, which uses the dot product as the base similarity metric for similarity. This was applied directly to peaks extracted from the full spectra. Additionally, metrics such as cosine similarity, Pearson’s correlation, first-difference cosine similarity, and soft IOU were evaluated. Normal similarity (NormSim) refers to similarity metrics calculated directly from full PAH/PAC and placenta spectra. In contrast, CaPSim calculated the similarity between the extracted peaks of specific PAH/PAC and those of a placenta sample. CaPSim uses the CaPE algorithm to extract the peaks, including the maximum allowable shift (distance threshold) among peaks tested at {12, 18, 24} indices [approximately {10, 15, 20} [cm] ^(−1)]. While the original algorithm 42 set the extraction of 10 characteristic peaks, we expanded the methodology to include a range of extracted peaks. We adjusted the number of characteristic peaks to extract from {5, 10, 15, 20}, allowing us to evaluate the impact of varying peak counts on the performance and accuracy of our analysis. We also extracted peak features from both the PAH/PAC spectrum and the placenta spectrum only at the peak locations of the PAH/PAC. For the logistic regression model, we considered the l_1 regularization strength at different magnitudes {0.0001, 0.01, 1, 100, and 10,000}. A grid search approach was employed to systematically evaluate all combinations of the hyperparameters, ensuring optimal settings for the entire detection procedure. This comprehensive approach ensured the optimal configuration for the whole detection procedure, enhancing the detection accuracy of PAHs/PACs in the placenta samples. The results of the finalizing model included a regularization strength of 0.0001 using CaPSim with a distance threshold of 18 (approximately 15 cm⁻¹). We also included peak locations from the placenta samples in the peak extraction process and extracted 20 characteristic peak locations to maximize the detection sensitivity and specificity.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

All study data are included in the article and/or SI Appendix.