SERSomes for metabolic phenotyping and prostate cancer diagnosis

Summary

Molecular phenotypic variations in metabolites offer the promise of rapid profiling of physiological and pathological states for diagnosis, monitoring, and prognosis. Since present methods are expensive, time-consuming, and still not sensitive enough, there is an urgent need for approaches that can interrogate complex biological fluids at a system-wide level. Here, we introduce hyperspectral surface-enhanced Raman spectroscopy (SERS) to profile microliters of biofluidic metabolite extraction in 15 min with a spectral set, SERSome, that can be used to describe the structures and functions of various molecules produced in the biofluid at a specific time via SERS characteristics. The metabolite differences of various biofluids, including cell culture medium and human serum, are successfully profiled, showing a diagnosis accuracy of 80.8% on the internal test set and 73% on the external validation set for prostate cancer, discovering potential biomarkers, and predicting the tissue-level pathological aggressiveness. SERSomes offer a promising methodology for metabolic phenotyping.

Graphical abstract

Highlights

• •SERSomes profile the metabolic phenotypes of various biofluids
• •SERSomes improve diagnostic accuracy with deep learning models
• •SERSomes facilitate metabolite biomarker discovery

Bi et al. develop a hyperspectral surface-enhanced Raman spectroscopy-based technique, SERSome, to profile the metabolic phenotypes of various biofluids, including cell culture medium and human serum, offering a methodology for accurate diagnosis and metabolite biomarker discovery with rapidity, low cost, and high sensitivity for fundamental studies and clinical applications.

Introduction

Blood-based testing has recently emerged as a genuinely viable alternative to tissue biopsies in cancer treatment¹ due to its improved flexibility, reliability, and reduced invasiveness.² Since only few diseases can be diagnosed by a single biomarker or a small handful of biomarkers, any technique that is low cost, rapid (ideally within minutes), and capable of detecting an extensive range of molecules should be highly desirable in clinics. A cheap and fast analysis with a small sample volume is a general goal of every blood test approach for large-scale and longitudinal testing in the clinical setting to enable swift therapeutic decisions and intra-operative adjustments.^3,4

Molecular phenotyping, a multiplexed detection strategy, enables a system-wide readout of the physiological status⁵ and the fundamental feature of cancer development in general.⁶ Here, the molecular phenotype in the metabolite level can be leveraged to reflect the most downstream response to the carcinogenesis.⁷ However, among the types of biological information that have been investigated as potential candidates in blood-based tests, metabolite phenotyping has yet to achieve practical success, though some well-established approaches, such as mass spectrometry (MS) and nuclear magnetic resonance spectroscopy, are capable of generating multivariate metabolic profiles.⁸ These methods often require expensive instruments and extensive pretreatment procedures, impeding their clinical applicability.^9,10,11

Raman spectroscopy, offering label-free and non-destructive molecular fingerprinting,^12,13 has great potential for molecular phenotyping and has been employed to detect various biomolecules.¹⁴ Ramanome, a set of Raman spectra collected from a cell population, has been used to profile cellular responses to particular stresses and investigate biological mechanisms based on spectral signatures.^15,16 However, it is not sensitive enough to examine small metabolites in bioliquids due to their intrinsically low Raman cross-sections. Surface-enhanced Raman spectroscopy (SERS), assisted by plasmonic nanomaterials, can detect molecules down to the single-molecule level and quantify a wide range of molecules, exhibiting high potential in metabolic phenotyping for pan-cancer or systems-wide characterizations.^17,18

One of the major reasons for this limited success in metabolic phenotyping by present SERS approaches is, ultimately, poor reproducibility.¹⁹ The label-free method relies on stochastic adsorption of molecules to the metal plasmonic nanoparticles (NPs) for significant enhancement of the molecular signatures; generally leading to an uncontrollable random distribution of adsorbed molecules among the electromagnetic hotspots within the NPs.²⁰ This hinders the comprehensive and unambiguous identification of the metabolic composition if only single or a couple of spectra have been acquired for each sample. Another non-negligible reason lies in the lack of proper pretreatment of the samples (specifically, large molecules, including proteins and nucleic acids) that were not removed in previous works.²¹ As a result, the metabolites are impeded from approaching the hotspots; thus, the signatures of metabolites cannot be fully presented.

In this work, we developed an ultrasensitive and tractable SERS-based technique to robustly profile molecular phenotypes at the metabolite level in complex biological fluids, such as cell culture medium and human serum. Rapid single-step filtration of the biological fluids is applied in advance to remove macromolecules so that the metabolites have maximal access to the hotspots, achieving a high signal-to-noise ratio of the metabolic signatures. We refer to the spectral set of each sample as a “SERSome,” comprising multiple SERS spectra (e.g., 200) acquired at various locations within the sample, as shown in Figure 1. In this way, the molecular composition is comprehensively fingerprinted, embodying rare events and weak signals that might be overlooked if only handful of spectra are acquired for one sample. In addition, this method requires no more than 10 μL of biofluidic metabolite extraction and completes measurements in 15 min. We demonstrate that this method exhibits remarkable sensitivity, from the nanomolar to the millimolar range, and allows the identification of subtle differences and correlations in SERS features between cultured cells during growth. Additionally, we reveal metabolic changes in the human serum of prostate cancer (PCa) and benign prostatic hyperplasia (BPH) patients using SERSomes, unveiling potential biomarkers. Furthermore, we show that classification based on SERSomes dramatically improves the diagnosis of PCa and the prediction of tissue-level aggressiveness compared to conventional biomarkers, like prostate-specific antigen (PSA) and other clinical measures. Thus, this work introduces SERSomes as a powerful tool for the rapid and sensitive characterization of metabolites in complex biological fluids, with immediate applications in cancer liquid biopsies and broader implications for basic research and clinical practice.

Figure 1.

Concept of SERSome

The SERS spectral set, collected pointwise from the bioliquid sample mixed with the SERS NPs, is referred to as a “SERSome,” effectively reflecting the molecular fingerprint phenotype at the metabolite level. Various analyses, including quantification and correlation assessments, can be conducted using SERSomes of different samples to unveil the underlying biological mechanisms and realize clinical diagnosis.

Results

Workflow of SERSome measurements

The goal of metabolic SERSome is the rapid generation of a comprehensive metabolite profile of a biological fluid with high sensitivity. Briefly, biofluid samples are first filtered to remove large molecules (such as proteins, nucleic acids, and polysaccharides) and then mixed with the metallic NPs. Metabolites adsorbed or adjacent to the NPs exhibit a significant enhancement in their Raman spectra. Multiple measurements of the sample are then obtained by pointwise excitation and detection across different regions in the sample using the confocal Raman system. Due to the high level of complexity of the biofluid composition, the molecules close to NPs differ in type and number in the different probed volumes, producing spectral variation across different acquisitions. Consequently, multiple measurements of the sample are required to obtain a spectral set as a SERSome specific to the sample that can robustly characterize its metabolite composition. For each spectral set, calculation of the Pearson correlation coefficient (PCC) between different averaged spectrum subsets is used to determine the minimum number of spectra needed for adequate robustness.²² This process is similar to the collection of a fingerprint as the security code on a cell phone, during which more complete information about the fingerprint is collected by constantly changing the angle of the fingers. Similarly, the SERSomes can be considered as the metabolite fingerprints of biofluids and leveraged to explore the metabolic variation with time and treatment. Correlation among the metabolic patterns and the heterogeneity among different samples can also been unveiled based on the SERSomes (Figure 1).

Typically, measurements were performed using a quartz capillary (inside diameter, 1 mm) containing a 10-μL mixture of plasmonic NPs and the biofluidic sample in a confocal Raman system equipped with a 638-nm laser (12.67 mW) and a 10× objective lens (Figure 2A). The spectra were collected in the pointwise scanning mode with adequate spacing between adjacent acquisitions. Specifically, 10 μm was chosen as the step size in this work, though larger ones were also feasible. This avoided measuring the same metabolite molecules repeatedly, thus enabling comprehensive metabolic profiling (Figure S1). For metabolite detection, citrate-reduced Ag NPs without additional surface modification were selected because they can provide surfaces capable of adsorbing a wide variety of metabolites.²³ For certain types of samples, the same amount of Ag NPs (concentration, 0.4 nM as synthesized) was applied for reproducible sampling. Additionally, they are easy to synthesize (and possibly mass produced), highly stable in a variety of biofluids, and, most importantly, exhibit excellent Raman enhancement for single-molecule detectability (Figure S2). The Ag NPs showed an extinction peak at 443 nm (Figure 2B), a zeta potential of −29.6 mV, and a hydrodynamic diameter of approximately 89 nm (polydispersity index, 0.411) (Figure 2C). It is worth noting that the hydrodynamic diameter was slightly larger than the size (61 ± 13.8 nm) determined from transmission electron microscopy (TEM) (Figure 2D) owing to the contribution of larger NPs to light scattering.

Figure 2.

SERSome measurement

(A) A mixture of metabolites and SERS NPs is introduced into a quartz capillary and measured using a confocal Raman system operating in one-dimensional scanning mode.

(B–D) Characterizations of Ag NPs used for SERSome measurements: (B) absorbance spectrum, (C) histogram of the hydrodynamic diameters measured via dynamic light scattering, and (D) TEM image.

(E–J) Representative SERSomes derived from various biofluidic samples with diverse molecular compositions: (E) cysteamine + riboflavin, (F) phenylalanine + nicotinic acid, (G) adenine + uric acid, (H) HepG2 cell culture medium, (I) human serum, and (J), human urine. Each SERSome is represented by a heatmap of a spectral set consisting of 100 spectra. For clarity, a mean spectrum (black line, n = 100) is also shown at the bottom.

To provide an intuitive illustration of the capabilities of metabolic SERSomes, we initiated our study by preparing three artificial metabolite mixtures (Figures 2E–2G), such as cysteamine + riboflavin, as well as three types of genuine biofluid samples, including cell culture medium (Figure 2H), human serum (Figure 2I), and human urine (Figure 2J). These samples were combined with Ag NPs, and the measurements were conducted as described above. It can be seen that the SERSomes derived from these diverse samples effectively profiled distinct molecular phenotypes, in accordance with their respective molecular compositions. In addition, our approach unveiled rare fingerprints with the SERSomes, which might have been overlooked if relying solely on average spectra for metabolic profiling (Figure S3). Therefore, SERSomes show immense potential in a wide range of biological and clinical applications.

Quantification via SERSomes with specifically prepared samples

To demonstrate the effectiveness of SERSomes, we prepared three distinct samples, denoted S1, S2, and S3. Each sample contained 20 common metabolites, such as cytosine and adenine, acquired from commercial sources at various concentrations within a range of 0.015–1.5 mM (see Table S1 for precise compositions). This setup aimed to mimic biofluids with relatively simple compositions (e.g., serum-free cell culture). The metabolite concentration range was chosen mainly in consideration of detection sensitivity of the benchmark MS technique. As shown in Figure 3A, 200 Raman spectra were collected within each sample along a designated line, a process completed in approximately 5 min. It is evident that, for each sample, notable differences in peak positions and intensities among the spectra were observed, as expected due to the inherent stochasticity in molecular enhancement across the probed volumes (Figures 3A and 3B). To determine the minimum number of spectra needed for adequate robustness in the final profile, we computed the PCC between two averaged spectra derived from two randomly selected subsets of the entire spectral set. This analysis revealed that PCC curves rapidly converged to values close to 1 when approximately 40 spectra were averaged together (Figure 3C). Consequently, to ensure maximal robustness in final averaged spectra, we analyzed the full complement of 200 spectra for each sample. Next, we monitored the Raman intensities at specific wavenumbers (649, 723, 786, and 1,492 cm⁻¹) corresponding to SERS-specific peaks of four metabolites: guanine (ring breathing mode²⁴), adenine (ring breathing mode²⁵), cytosine (ring breathing mode²⁶), and tyrosine (ring vibration²⁷), respectively (Figures 3A and 3B). Although the intensity at specific bands of these four components varied among the 200 spectra within each sample, the collection of values for each component obtained from the 200 measurements in each sample (e.g., the mean value) demonstrated the expected trends among different samples (Figure 3D). To further validate this observation, the concentrations of each of the four metabolites were determined using MS. As displayed in Figure 3D, the intensities of the specific peaks exhibited a high degree of correlation with the MS results (R² ≥ 0.90), reinforcing the accuracy of our approach (also refer to Figure S4).

Figure 3.

Metabolic SERSomes of specifically prepared biofluidic samples

(A) SERSome heatmaps (top, consisting of 200 spectra) of three artificial samples (S1, S2, and S3), showing variations in metabolite concentrations, including guanine, adenine, cytosine, and tyrosine, and the corresponding averaged spectra (bottom). The composition of S1–S3 is shown in Table S1.

(B) Typical SERS spectra (black) from S3 as indicated in (A). The standard SERS spectra of guanine (red), adenine (orange), cytosine (green), and tyrosine (blue) are presented at the bottom, and their specific SERS bands are marked with color-matched stripes here and in (A).

(C) Pearson correlation coefficient (PCC) versus the number of spectra per spectral set.

(D) The correlation between the mean intensity of the SERS bands associated with the four metabolites (n = 200) and the concentration measured by mass spectrometry (MS), indicating the quantification potential of SERS spectra.

Next, we aimed to ascertain the minimal resolvable concentration achievable with SERSomes. It is well known that SERS can detect molecules at much lower concentrations compared to conventional MS. Indeed, our findings indicate that concentrations as low as nanomolar levels could be resolved by this method, a capability beyond the detection limits of conventional MS (Figure S5; Table S2).²⁸ Thus, overall, these results illustrate that SERSomes are capable of robustly determining the identity and relative abundance of multiple metabolites in biological fluids from concentrations spanning from the nanomolar to the millimolar range.

Cell metabolic variation revealed by SERSomes

Comprehensive characterization of the metabolic changes in cell cultures under various conditions is important for a number of applications and particularly relevant for drug development endeavors.²⁹ Prior to analyzing actual samples, we assessed the effects of different amounts of originally included or xenobiotic metabolite molecules doped into the culture medium (i.e., DMEM without fetal bovine serum). Our results showed that concentration variations of these molecules can be well reflected by the intensities at characteristic peaks obtained from SERSome measurements (Figure S6). We also confirmed stable spectral fingerprints of metabolites under typical working conditions (i.e., pH and ionic strength). Especially for the salt concentration far exceeding the normal conditions (>100 mS/cm), reproducible profiling can still be guaranteed by using our method (Figure S7). To this end, we used SERSomes to examine the metabolic differences in the extracellular milieu surrounding liver cancer HepG2 cells as a function of time, both in the presence or absence of methotrexate (MTX), a common clinical anti-cancer drug that disrupts the synthesis of tetrahydrofolate (Figure 4A).³⁰ After the initial day, the cells were further cultured for 3 days only (i.e., D0–D3) in order to prevent contact inhibition. Then, 400 μL of the culture medium samples with cell incubation were collected daily, and macromolecules potentially interfering with metabolites approaching the hotspots on the NP surface due to factors such as protein corona were removed by filtration using a 3-kDa-cutoff membrane (Figure S8).^31,32 Next, 10 μL of each sample was mixed with Ag NPs at a ratio of 1:9 for optimal enhancement (Figure S9A). Again, for Raman measurement, we obtained 200 spectra and confirmed that this number was sufficient for subsequent SERSome analysis (Figure S10). Heatmaps of the complete spectral sets are shown for the HepG2 group and HepG2 + MTX group from D0 to D3 (Figure 4B), along with several individual spectra from the set of HepG2+MTX on D1 (Figure 4C). Both indicate significant fluctuations in peak positions and intensities among the different spectra in this measurement, emphasizing the necessity of analyzing them only in toto. This observation has also been confirmed by the t-distributed stochastic neighbor embedding (t-SNE), a method for visualizing high-dimensional spectral data in a lower-dimensional space (i.e., 2D). As shown in Figure 4D, increasing the number of spectra in one SERSome from 5 to 200 spectra per set substantially enhances the classification accuracy among different samples. With 200 spectra acquired for each SERSome, overlapped distributions are observed between two individually cultivated plates of cells during the initial 2 days (D0 and D1). However, following MTX treatment, the SERSomic patterns started to segregate into two distinct groups on D2, further separating on the subsequent day (D3) because of the enhanced response to MTX. This emphasizes the significance of collecting multiple spectra in the spectral set for SERSome analysis.

Figure 4.

Differences in cell metabolism in cultured cells

(A) Bright-field images of HepG2 cells during culture (D0, no cells, not provided; D1; D2; and D3). MTX was introduced to the HepG2+MTX group after D1. Scale bar: 100 μm.

(B) SERSome (200 spectra) heatmaps for each sample.

(C) Representative spectra from D1 in the HepG2+MTX group, as indicated in (B).

(D) t-SNE plots generated using varying numbers of spectra per SERSome (i.e., 5, 10, 50, and 200 spectra).

(E) Correlation analysis among the SERS peaks. Line colors and endpoints indicate changes during the culture process, with orange denoting increased intensities, blue indicating decreased intensities, and purple signifying intensity independence with time. The darkness of the line color and the size of the endpoints represent the extent of correlation, with the darker lines and larger endpoints indicating stronger correlation between two bands.

(F) ANOVA between HepG2 and HepG2 groups on each day. Colors represent the maximum p values at each wavenumber from three parallel experiments after Bonferroni correction.

(G) The intensity fold change (FC) of HepG2+MTX compared with HepG2 (${\log }_{2}FC={\log }_2\left(I_{HepG2+MTX}/I_{HepG2}\right)$). Only FCs with a consistent trend in the three parallel experiments are displayed as numbers with minimum absolute value, while others are assigned as zero.

(H) Histogram illustrating the percentage of wavenumbers with corrected p values smaller than 0.05 and nonzero FCs.

(I) Screened SERS bands with corrected p values smaller than 0.05 and nonzero FCs (black, HepG2; red, HepG2+MTX). Error bars, standard deviation, n = 200.

Notably, significant correlations can be found in each group throughout the culture process and among different SERS bands, as demonstrated by the Pearson correlation analysis involving all SERSomes (D0–D3) (Figure 4E). To elucidate the potential metabolic variations, we inferred the detailed metabolites based on the important SERS bands according to the vibrational modes reported in the literature (Table 1). For example, SERS bands at 1,103, 1,108, 1,118, 1,354, and 1,399 cm⁻¹, primarily attributed to phenylalanine,³³ threonine,³⁴ glucose,³⁵ lipids,³⁶ adenine,³⁷ and guanine,³⁶ respectively, exhibit decreasing intensities in both HepG2 and HepG2+MTX because of the consumption of essential amino acids and other chemicals from the culture medium for cell proliferation.^38,39,40 Metabolite conversions can also be evidenced through negatively correlated SERS bands during culture.⁴¹ The decrease in intensity at 1,118 cm⁻¹, accompanied by an increase at 1,399 cm⁻¹, aligns with glucose consumption in the tricarboxylic acid (TCA) cycle, which drives biosynthetic pathways (e.g., fatty acids synthesis) in cancer cells.⁴² This negative correlation was weakened in the MTX-treated group because of reduced TCA cycle activity induced by MTX.^43,44 In the HepG2+MTX group, the SERS bands at 1,266 cm⁻¹ intensify over time, accompanied by several negative correlations (e.g., 1,354 cm⁻¹), indicating MTX-induced adenosine release⁴⁵ into the extracellular space for anti-proliferation effects,^46,47 which was consistent with the result obtained from the cell viability test (Figure S11).

Table 1.

Assignments to some typical SERS bands for HepG2 SERSomes

Wavenumber	Vibration mode	Assignment
1,103 cm−1	symmetric ring breathing	phenylalanine33
1,108 cm−1	C-C-O out-of-phase stretch	threonine34
1,118 cm−1	C-C stretching	glucose35
1,252 cm−1	alkyl = C–H cis stretching	lipid48
1,266 cm−1	H bending and ring stretching	adenosine45
1,278 cm−1	symmetric stretching of PO43–	lipid49
1,317 cm−1	C=C ring breathing	guanine34
1,339 cm−1	C=C ring breathing	adenine50
1,354 cm−1	CH3CH2 wagging	guanine, adenine, tryptophan36,37,51
1,399 cm−1	COO− symmetric stretching	fatty acids36
1,592 cm−1	C-N stretching	nucleotides52,53
1,658 cm−1	C-NH3+ angle bending deformation	glucosamine54

Obvious differences can also be found between groups with or without MTX treatment. Therefore, we further conducted a comparative analysis of SERSomes from the two groups on a daily basis using two statistical methods: one-way analysis of variance (ANOVA) and fold change (FC). Here, data from three parallel experiments were used to minimize randomness. As shown in Figure 4F, in the initial 2 days, few bands exhibit ANOVA p < 0.05, calculated by the maximum p values among the three experiments after Bonferroni correction. However, starting from D2, more bands present a different distribution in intensity between HepG2 and HepG2+MTX, with a substantial increase on D3. Similar trends can be observed in the FC analysis (Figure 4G). Figure 4H summarizes the percentage of significantly varied SERS wavenumbers each day, illustrating drastic metabolic changes caused by MTX treatment. By the combination of ANOVA and FC, we screened five major bands with corrected p < 0.05 and the nonzero FCs on the last 2 days (D2 and D3) (Figure 4I). Here, reduced intensities at 1,244–1,253 and 1,275–1,280 cm⁻¹ may indicate decelerated lipid production^48,49,55 resulting from suppressed proliferation process caused by MTX. Reduced utilization of nucleotides^52,53 and glucosamine⁵⁴ may also be inferred by the less reduced intensity at 1,555–1,627 and 1,651–1,664 cm⁻¹ on D2, respectively. It is worth mentioning that some small peaks, such as 1,275–1,280 cm⁻¹, which are observable in only a subset of spectra, underscore the superiority of SERSomes in capturing rare signal events and enhancing the information obtained from label-free SERS metabolite detection (Figure S12).

Collectively, these findings confirm the accuracy and effectiveness of using SERSomes for molecular phenotyping, particularly in elucidating intriguing metabotype differences arising from MTX presence. This method, characterized by its multiplex nature and high sensitivity, provides a more comprehensive understanding of the metabolic changes induced by this drug.

PCa diagnosis using serum SERSomes

Inspired by the successful applications of SERSomes in cultured cell medium, we conducted experiments to assess their capability to extract metabolic information from human serum for the diagnosis of PCa, which is one of the most common cancers. Currently, though serum PSA is now used as the only biomarker for PCa screening,⁵⁶ it is still challenging to effectively distinguish between PCa and BPH using the current preoperative methods.⁵⁷ Therefore, we proposed an intelligent system with SERSomes for PCa diagnosis and pathological stratification prediction based on serum that is relatively invulnerable to external interfering factors and minimally invasive to obtain. Over approximately 2 years, we collected serum samples from 241 participants (85/83/73, PCa/BPH/control) as the training cohort, 26 participants (7/9/10, PCa/BPH/control) as the validation cohort, and another distinct 68 participants (24/23/21, PCa/BPH/control) as the test cohort (Figure 5A). Here, for the training cohort, power analysis has been performed in advance with a pilot cohort of 21 participants (7/7/7, PCa/BPH/control) and indicate a minimum number of 75 for each group (75/75/75, PCa/BPH/control) required for classification with the prediction power of 0.8 at a false discovery rate of 0.1. This confirmed that the current sample size of the training cohort was sufficiently large (Figure S13).

Figure 5.

Diagnosis and biomarker discovery of PCa using serum samples

(A) The demographic information for the training, validation, and test cohorts. Age distribution is summarized by the maximum, mean, and minimum values.

(B) A schematic of the algorithm for PCa diagnosis based on SERSomes. The workflow involves (i) colloidal background subtraction using BRNet, followed by (ii) generation of auxiliary SERSomes using DANet. Subsequently, (iii) PCDNet is trained for diagnosis, and (iv) model interpretation is implemented to identify salient bands.

(C) Confusion matrices of the diagnosis result using the proposed approach and PSA values in the test group.

(D) A 2D visualization of deep features extracted from PCDNet using t-SNE.

(E) ROC curves depicting the performance of PCa diagnosis under different implementation settings for the test cohort.

(F) Ablation studies in the test cohort on data augmentation and background removal and the relation between accuracy/PCC and the number of spectra per SERSome. PCC is displayed as the mean (solid orange line) with the standard deviation (yellow shade) from all SERSomes.

(G) Confusion matrices illustrating the pathological stratification prediction of GSs based on multimodal data (SERSomes and clinical data), SERSomes only, and clinical data only.

(H) ROC curves corresponding to pathological stratification prediction based on different data modalities.

(I) Saliency spectra for layers 1, 2, and 3 for PCa diagnosis, indicated by the solid line for the mean value and the shade for the standard deviation (PCa, n = 22; BPH, n = 14; control, n = 19).

To ensure a robust measurement of human serum SERSomes, we sequentially collected 200 spectra for each sample using 10-μL metabolite extraction obtained from 100 μL serum by ultra-filtration. To characterize the sensitivity of this method in serum, a spike-in experiment was also performed in advance, showing that a nanomolar level of metabolite concentration difference can be well profiled (Figure S14). In the first step, we employed an end-to-end measurement calibration deep model, Background Removal Network (BRNet) (see details in Table S3) to eliminate the background signals generated by Ag NPs. This calibration is essential to minimize systematic errors introduced during sample collection and measurement in real-world applications.⁵⁸ BRNet was trained by the BR dataset consisting of artificially constructed spectrum pairs with a known background to expose and subtract the background from real measurements (Figure 5Bi). Notably, BRNet outperformed existing methods that determine the background contribution by non-negative linear squares and the intensities of characteristic peaks (Figure S15).^59,60 It is also worth mentioning that this approach can be applied to other SERS research with different diseases and sample types as long as background removal is required to eliminate the systematic error and improve the data quality. Further, to address training data scarcity issues, we introduced a class-by-class trained 1-dimension diffusion model, Data Augmentation Network (DANet), to augment the training set by generating auxiliary SERSomes^61,62,63 (Figure 5Bii).

Considering the SERSomes embodied both spatial and spectral features in a 2D spectrogram format (matrix size = 200 spectra × 724 wavenumbers/spectra), in analogy to a Mel spectrogram,⁶⁴ we employed the Prostate Cancer Diagnosis Network (PCDNet) (see details in Table S4), which utilizes a convolutional neural network to extract spectrogram texture for the multiple classifications of PCa, BPH, and control (Figure 5Biii).

As a result, SERSomes empowered by deep learning exhibited an obvious improvement in discriminating PCa and BPH compared to other screening methods, such as PSA and the prostate imaging-reporting and data system, achieving an overall accuracy of 80.8% (Figures 5C and S16). In the multi-class receiver operating characteristic (ROC) analysis, the areas under the curve are 0.868 for PCa, 0.835 for BPH, and 0.994 for control, indicating that SERSomes from PCa patients are more likely to be correctly predicted as PCa with a probability of 0.868. In addition, the accuracy of distinguishing between PCa and non-cancers (BPH and control) using SERSomes combined with our system (82.35%) has also been significantly improved (19.12%) compared with using PSA (63.23%) (Figure 5C) Specifically, the sensitivity, specificity, and positive and negative predictive values for PCa were 91.67%, 77.27%, 68.75%, and 94.44%, respectively (Figure S17). Moreover, to demonstrate the clinical generalizability, we recruited additional samples from two different institutions as an external validation set, which were measured using the identical protocol followed by BRNet, Gaussian normalization in the spectral dimension, and calibration by the Combining Batch algorithm.⁶⁵ As the result, our method presented consistently superior predictability of PCa (Figure S18). All of the above highlights the potential of our system for more accurate and minimally invasive preoperative screening of PCa, which could mitigate overdiagnosis and overtreatment.⁶⁶ We visualized the distribution of the 6,144-dimension feature vector extracted by the spectral-spatial encoder of PCDNet using t-SNE (Figure 5D), demonstrating that spectrogram features form stable and separatable clusters. Furthermore, the consistency between the deep feature clustering and the PSA value distribution (the area of the dots related to the PSA value) validated the reliability of PCDNet.

To further verify the necessity of each module in the proposed system, in the test cohort, we carried out ablation studies by implementing the above operations without background removal or data augmentation, which can only achieve an overall average accuracy of 58.8% and 69.1%, respectively (Figures 5E and 5F). Specifically, the diagnosis accuracy increased sharply (5.8% ± 2.5%) as the number of spectra per SERSome increased from 1 to 25 and gradually (3.4% ± 1.3%) from 50 to 200, consistent with the result of the PCC (Figure 5F), emphasizing the importance of acquiring an adequate number of spectra in a single complex biofluid sample to comprehensively profile the metabolic phenotype. Moreover, the whole algorithm is robust, as proven by reproducible results from 9 repeated computations with consistent accuracy on reassigned training and test sets (Table S3). Results from the training cohort and the validation cohort can be found in Figure S19.

The Gleason score (GS) serves as a routine protocol for pathological stratification of PCa,⁶⁷ typically obtained only from invasive tissue biopsy (see details in Figure S20). If a minimally invasive technique can help to predict the GS, it would be highly preferred. In this regard, we explored the fusion of spectral features with some clinical measures (T stage, N stage, and PSA value; Figure S21) to predict the GS (Figure 5Biii). As shown in Figures 5G and 5H, the accuracy of GS prediction using SERSomes fused with clinical indicators reached 71.4%, exhibiting a 33.3% improvement over using clinical indicators alone. Remarkably, using SERSomes alone also lead to a 28.6% increase in prediction accuracy. When distinguishing highly malignant PCa (GS = 8, 9) from the moderately malignant cases (GS = 6, 7), the accuracy of using SERSomes fused with clinical indicators was 90.48%, significantly outperforming the use of clinical indicators alone (71.43%). This not only demonstrated that SERSomes contain important information of pathological prediction but also established that our method is transferable and capable of extracting sufficient features. In summary, these findings, though serving as a first try and definitely requiring further improvement regarding accuracy by collecting more training samples, indicated the potential of predicting highly malignant PCa using minimally invasive methods in the future and underscored its clinical significance.

Uninterpretable “black boxes” have long been a fundamental issue for deep learning methods.⁶⁸ In this regard, we implemented an interpretable network to score the importance of spectrogram features and infer potential metabolite biomarkers (Figure 5Biv). We applied Gradient-weighted Class Activation Mapping++ (Grad-CAM++) to retrospectively excavate the most important features contributing to the diagnosis model.⁶⁹ For each SERSome, a saliency map was generated to reveal the significant features at each wavenumber and spectral position (Figure S22). By matching serum SERSomes with the detailed SERS bands indicated by saliency spectra compressed from the saliency maps on the spatial dimension (Figure 5I), we uncovered a potential PCa biomarker (Table 2). Specifically, bands at 480 and 1,197 cm⁻¹ can be attributed to serum uric acid, an antioxidant associated with increased PCa risk due to the production of reactive oxygen species and impairment of the antioxidant defense mechanism.^70,71,72 Hypoxanthine, inferred from peaks at 718, 1,368, and 1,394 cm⁻¹, and uracil, with a peak at 1,190 cm⁻¹, are closely relevant to the perturbed purine and pyrimidine metabolism pathways involved in DNA and RNA synthesis, consistent with previous research.^73,74,75 Steroids, including testosterone, which might be evidenced by the peak at 1,540 cm⁻¹,^76,77 has also been found previously to be closely related to the risk of PCa.⁷⁸ Salient bands at 417 and 1,570 cm⁻¹ indicate the dysfunctional tryptamine metabolism in PCa, inducing a combined modulation of immune escape and DNA repair mechanisms with its downstream metabolites.^79,80 Salient maps enhance the interpretability of the deep learning models, offering dual benefits in diagnosis and biomarker discovery.

Table 2.

Assignments to some typical SERS bands for serum SERSomes

Wavenumber	Vibration mode	Assignment
480 cm−1	C–N–C ring vibrations	uric acid81
1,197 cm−1	C-N stretching	uric acid82
718 cm−1	C–H bending	hypoxanthine83
1,368 cm−1	C=O stretching	hypoxanthine84
1,394 cm−1	C-N stretching	hypoxanthine84
1,190 cm−1	in-plane ring stretching	uracil85
417 cm−1	out-of-plane benzene ring deformation	tryptamine86
1,125 cm−1	C-H and NH3+ bending, in-plane benzene ring deformation	tryptamine86
1,540 cm−1	C=O stretching	steroid76,77
1,570 cm−1	NH2(NH3+) bending	tryptamine86

In short, our proposed system can be potentially integrated into the standard workflow for SERSome-based blood tests. It effectively addresses challenges such as systematic errors, data scarcity, complex spectral-spatial features, and relationship between SERSomes and physiological characteristics in clinical diagnosis and biomarker discovery. It is worth mentioning that, based on Grad-CAM++, we were able to visualize the important features for classification to evidence the potential biomarkers. This addressed the challenge of “black box” mode and lack of interpretation in current deep learning-based methods for analytical chemistry.⁸⁷ In addition, our system has been proven to outperform the conventional algorithms (including the decision tree classifier, K-neighbors classifier, Gaussian naive Bayes classifier, and multi-layer perceptron classifier) based on the averaged spectra in diagnosis accuracy (Figure S23).

Discussion

In this work, we introduced the concept of metabolic SERSomes as a powerful tool for sensitive and reliable metabolic phenotyping in biological fluids. With proper pretreatment (i.e., ultra-filtration) of the biosamples, which has often been neglected by previous studies, we were able to focus on the metabolites without the interference from other macromolecules, such as proteins. The integration of SERSomes with deep learning represented an advancement in molecular phenotyping in (1) how to treat the multiple measurements for one certain sample and (2) how to use the spectral set for metabolic profiling.

For the first issue, the multiple measurements, instead of using the averaged spectra as in previous studies, were treated as different perspectives of one sample in its whole profile (i.e., SERSome) to fully capture the rare signals generated by trace molecules or molecules with low affinity and cross-sections. Here, the number of spectra in the SERSome was calculated in advance for reproducible profiling. This enabled a more robust and comprehensive metabolic phenotyping.

Another issue is realized with the assistance of deep learning models. The algorithm was specially designed and used the SERSome as a 2D image. Therefore, all features in the SERSome can be potentially excavated by the deep learning model, thus benefitting classification and biomarker discovery. This has been proven by comparison with other conventional AI methods based on the mean spectra or only a handful of spectra, showing improved diagnostic accuracy.

For a demonstration, we chose two scenarios: extracellular metabolism by culture medium and PCa by human serum. As a result, metabotypic variations in cell culture media in response to drug treatments and in the serum of individuals undergoing carcinogenesis were successfully detected and quantified with nanomolar sensitivity. SERSomes achieved an impressive accuracy of 80.8% on the internal test set and 73% on the external validation set for PCa diagnosis, outperforming conventional clinical measures. We also prospected the potential of incorporating SERSomes with clinical data for the prediction of tissue-level aggressiveness with minimally invasive methods. For real-world clinical use, the diagnostic accuracy should be further improved by recruiting more samples in the training cohort, developing more advanced deep learning algorithms, and incorporating other screening methods with little or no invasiveness.

Compared with existing methods such as MS, this technique requires less pretreatment and measurement time. In addition, the cost is relatively low, since only ordinary consumables and microliters of NP suspensions are required, and it is likely to be further reduced by using a portable Raman system or even a microspectrometer.⁸⁸ In the future, this method should be further improved regarding sensitivity to detect a lower abundance of molecules and a comprehensive SERS database to facilitate model interpretation on the potential biomarkers. For real-world applications, urgency should be raised for the wide commercialization of Raman devices for improved spectral resolution with lower cost, broader manufacture, and downsizing. In a word, SERSome, offering a transformative approach to metabolite analysis, holds great promise for a wide range of applications in both fundamental research and clinical settings.

Limitations of the study

In terms of the biological mechanisms regarding the metabolic phenotypes, many more experimental validations should be performed in the future, such as investigating PCa-related cell lines to unveil the underlying cell-level pathology and explore the correlation between cell-level metabolic changes and systematic variations. Further efforts should also be made to incorporate our approach with other assays for higher diagnostic accuracy in future projects.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Chemicals, peptides, and recombinant proteins
Citrate trisodium (98%)	Aladdin	S189183; CAS No.: 68-04-2
Silver nitrate (AR, 99.8%)	Aladdin	S116264; CAS No.: 7761-88-8
Potassium chloride (AR, 99.5%)	Aladdin	P112134; CAS No.: 7447-40-7
Methotrexate	BBI Life Sciences	A600612; CAS No.: 59-05-2
Deposited data
Code for serum data analysis	This paper	https://doi.org/10.5281/zenodo.11068334
Experimental models: Cell lines
Human hepatocyte carcinoma cells	ATCC	HB-8065
Software and algorithms
Python (version 3.9.16)	Python Software Foundation	https://www.python.org
PyTorch (version 1.13.1+cu116)	Python Software Foundation	https://www.python.org
Numpy (version 1.24.2)	Python Software Foundation	https://www.python.org
Matplotlib (version 3.7.1)	Python Software Foundation	https://www.python.org
pytorch_grad_cam (version 1.4.6)	Python Software Foundation	https://www.python.org
CUDA 11.4	Nvidia Developer	https://developer.nvidia.com/
cuDNN 8.0.5	Nvidia Developer	https://developer.nvidia.com/

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Jian Ye (yejian78@sjtu.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •All data is available from the lead contact upon request.
• •The codes and data including the model, the training engine, the testing script as well as the model weight and partial test dataset have been deposited at Zenodo and are publicly available as of the date of publication. DOI is listed in the key resources table.
• •Any additional information required to reanalyze the data reported in this work paper is available from the Lead Contact upon request.

Experimental model and study participant details

Cell lines

Human hepatocyte carcinoma cells (HepG2) were obtained from the American Type Culture Collection (ATCC). Cells were maintained in Dulbecco’s Modified Eagle Medium (DMEM) supplemented with 10% fetal bovine serum (FBS) and 1:100 penicillin-streptomycin (10⁴ units/mL) at 37°C and 5% CO₂. 2 × 10⁶ cells were cultured in T25 flasks in a total volume of 10 mL of medium, which was not changed during culture. After the adherence of cells, MTX was added to reach a final concentration of 5 μM at D1. Herein, MTX was directly resolved in DMEM without any other solvents. 400 μL of the culture medium was taken from each culture flask every day. For the analysis of extracellular metabolism, the culture medium was centrifuged at 200 g for 3 min to remove the dead cells and at 300 g for 20 min to get rid of the cell debris, subsequently. Both two groups (HepG2 and HepG2 + MTX) were performed in 3 replicates.

Human subjects

A total of 424 clinical samples were collected for this study (Table S4). Herein, for the training set, the validation set and the internal test set, 231 samples (training: 85 PCa and 83 BPH, validation: 7 PCa and 9 BPH, test: 24 PCa and 23 BPH) were obtained from the urology department in Ren ji Hospital, School of Medicine, Shanghai Jiao Tong University, and 104 samples (training: 73 controls, validation: 10 controls, test: 21 controls) were collected from the Physical Examination Center of Ren ji Hospital, School of Medicine, Shanghai Jiao Tong University. For the external validation set, 37 samples (15 PCa and 22 BPH) were obtained from the urology department in Ningbo Hangzhou Bay Hospital and 52 samples (25 PCa and 27 BPH) were obtained from the urology department in Jiading District Central Hospital Affiliated Shanghai University of Medicine and Health Sciences. All the samples were from distinct participants. The collection and usage of clinical samples have been approved by the ethics committee of Ren ji Hospital, School of Medicine, Shanghai Jiao Tong University (KY2021-030). Consent documents were obtained from all participants included in this study. All participants excluded other tumor history, anti-tumor medication history and metabolic syndrome. Clinical data of each participant including age and PSA value were collected for study analysis. PCa patients also underwent pelvic enhanced MRI and/or whole-body PET-CT examination. The T stage and N stage of PCa patients were determined by urologists according to the American Joint Committee on Cancer (AJCC) 8^th staging system. Each BPH and PCa patient has been confirmed by transrectal prostate biopsy. The corresponding diagnosis was made by pathologists according to the H&E staining of the biopsy tissue. And the Gleason score was therewith determined for those with prostate cancer. The control participants have been excluded from BPH and PCa, and were aged from 50 to 70 years old which was comparable to the BPH and PCa group. To collect the serum for SERSome measurement, 3–5 mL fasting venous blood were collected by experienced nurses in the morning and placed at room temperature for stratification. After the centrifugation of the blood at 3000 rpm for 10 min under ambient condition, 500 μL serum was obtained and stored at −80°C until use. SERS measurements were acquired within 28 days upon serum collection. Before the SERS measurement, all samples were thawed in an ice box with crushed ice at 2°C–4°C. Power analysis was employed to compute the sample size of clinical serum samples based on the SERS spectra.

Method details

Instrumentations

A JEM-2100F transmission electron microscope (JEOL, Tokyo, Japan) was used to obtain the NPs’ morphology. UV-vis spectra were collected on a UV1900 UV-Vis spectrophotometer (Aucybest, Shanghai, China). The hydrodynamic diameter and the zeta potential of the NPs were characterized by a Zetasizer Nano ZSP (Malvern, UK). An NP tracking system (ZetaView, Merkel, German) was used to measure the concentration of the NPs. The mass spectroscopy for the 20-metabolite mixture was implemented with a UPLC-MS system (Acquity UPLC H-class/Xevo TQ-XS).

Synthesis of the SERS NPs

The citrate-reduced silver (Ag) NPs were synthesized according to Lee and Meisel’s method⁸⁹ with minor modifications. Briefly, 9.2 mg AgNO₃ was dissolved in 75 mL ultrapure water and heated to boiling under stirring. Then, 2 mL of citrate sodium solution (10.1 mg/mL) was added dropwise. The mixture was further kept boiling for 1 h without exposure to light and then cooled to room temperature under vigorous stirring. The product was stored in 4°C refrigerator and protected from light exposure before being used.

SERS measurement

For the specifically prepared metabolite mixtures, 20 metabolite molecules were revolved in pure water (Table S5). To collect the metabolic substances from the mixture with other macromolecules, the 3 kDa-cutoff filter (Millipore, Amicon Ultra-4, PLBC Ultracel-3) was used to centrifugate the culture medium and serum at 10000 rpm for 10 min. The metabolite extraction was mixed with the Ag NPs with the ratio of 1:9 for culture medium and 1:1 for serum in the first step. Then the mixtures were ultrasonicated for 10 s to ensure monodispersion and complete mixing and set still for full adsorption of molecules onto the surface of the Ag NPs, e.g., for 1 h used in this work. Before measurement, the mixtures were slightly ultrasonicated to avoid NP precipitation. Finally, 200 SERS spectra were collected for each sample on the confocal Raman system (Horiba, XploRA INV) for scanning along a quartz capillary (inside diameter: 1 mm, outside diameter: 2 mm) containing 10 μL of the NP-metabolite extraction mixture with the incident laser wavelength of 638 nm (power: 12.67 mW), the acquisition time of 5 s per spectrum and the step size of 10 μm between adjacent spectra through a 10× objective lens. Herein, each spectrum consisted of 724 datapoints within the spectral range of 200–2000 cm⁻¹ by using the 600 gr/mm grating in the spectrometer.

Quantification and statistical analysis

Before analysis, all SERS spectra have been preprocessed by baseline removal with the software LabSpec 6 (method: poly; degree: 10; max points: 200). The Pearson correlation coefficient (PCC) is defined as $\text{PCC}=\frac{\text{cov}\left(\mathrm{X},\mathrm{Y}\right)}{\sqrt{\text{var}\left(\mathrm{X}\right)\text{var}\left(\mathrm{Y}\right)}}$, where $\text{cov}\left(\mathrm{X},\mathrm{Y}\right)$ represents the covariance between $\mathrm{X}$ and $\mathrm{Y}$, $\text{var}\left(\mathrm{X}\right)$ represents the variance of $\mathrm{X}$, and both $\mathrm{X}$ and $\mathrm{Y}$ were the average of random sampling from the measured spectra.

The signals of cytosine, adenine, guanine and tyrosine in the synthetic samples were read by the max value within the band of 765–820 cm⁻¹, 695-750 cm⁻¹, 625-680 cm⁻¹, and 1465-1520 cm⁻¹, respectively.

For the cell culture medium, Pearson correlation was applied to evaluate the time-dependency and intra-independency of the SERS intensities. t-Distributed stochastic neighbor embedding (t-SNE) algorithm was used to visualize the distribution variability among different samples and the perplexity was optimized according to the spectrum number used for each sample (i.e., perplexity = 5 for 5 spectra/set, perplexity = 10 for 10 spectra/set, perplexity = 40 for 50 and 200 spectra/set). One-way analysis of variance and fold change were utilized to screen SERS bands with significant changes between two groups.

The SERSome data of human serum $\mathit{S}\in R^{724\times 200}$ consists of 200 spectra scanning from different spatial positions. Each spectrum $\mathit{s}\in R^{724}$ includes 724 spectral components (i.e., wavenumbers/data points collected by the Horiba instrument) ranging from 600 to 1800 ${\text{cm}}^{-1}$. The whole process regarding the analysis of human serum SERSomes includes BRNet, DANet, PCDNet, explainable clinical deep learning and the prediction of Gleason score.

The NP background removal neural network—BRNet: In real applications, there is sometimes a dilemma about the NP backgrounds. Specifically, in order to obtain high signal-to-noise ratio signals of the target molecules, the NPs are designed to aggregated slightly thus to form more intensive hotspots. While at the same time, the NP background will also be enhanced most probability generated by the residue moieties of reagents on the NP surface and these NP background peaks might overlap with target signals and interfere with the downstream data analysis. The removal of the NP background has been a great challenge for a long time since the background signals vary from different spectra. Conventional background removal approaches, for example, estimate the background contribution in the mixed signal referred to one certain background pattern or some of the peaks,^59,60 as the result, prone to overestimate or underestimate the background. Therefore, the BRNet (details in Table S6), which denoted as $f_{BR}$, was proposed to remove the background signal precisely and automatically. First, we constructed a BR dataset consisting of 100 pure background data (the signal of NPs only), 1,000 no (weak) background data (SERS spectra of real biological fluids with low or negligible NP signals) and 5,000,000 “mixed spectra”. Herein, the “mixed spectra” were artificially generated by linearly adding background signals on sample signals. Each data pair in the BR dataset consisted of “mixed spectra” and its corresponding background components, as the input and the output of the model, respectively. To accurately remove the NP background, we modified 1D-ResUnet as BRNet to understand the patterns of the NP background and precisely infer the background signal without any prior knowledge of background spectra. In detail, BRNet consists of three components, i.e., the 1D residual CNN encoder, 1D transpose convolutional up-sample decoder and layer-wise skip connection. The learning criterion was to minimize the Mean Squared Error (MSE) loss between the predicted background and the ground truth (GT) background. As shown in Equation 1, we obtained the background-removed spectrum ($\tilde{\mathit{s}}$) by subtracting the inferred background spectrum ($f_{BR}\left(\mathit{s}\right)$) from the original spectrum ($\mathit{s}$). The background-removed SERSomes were concatenated by the background-removed spectra in the spatial dimension, as show in Equation 2.

$$\tilde{\mathit{s}}=\mathit{s}-f_{BR}\left(\mathit{s}\right)$$ (Equation 1)

$$\tilde{\mathit{S}}=\left[{\tilde{\mathit{s}}}_{\mathbf{1}},{\tilde{\mathit{s}}}_{\mathbf{2}}\dots {\tilde{\mathit{s}}}_{\mathbf{200}}\right]$$ (Equation 2)

The diffusion model data augmentation—DANet: The 1D-diffusion model was extended to synthesize spectra since the conventional augmentation methods for sequential data do not fit well with spectral format. In this study, a 1D-ResUnet with 8 Resnet blocks was implemented as the denoising model (${ϵ}_{\theta }$) which iteratively reduces the noise with $T$ = 2000 times from the 1D-random input ${\mathit{s}}_T\sim N\left(0,\mathit{I}\right)$ to generate auxiliary spectra. The denoising models were trained class by class, i.e., three diffusion models of synthetic PCa, BPH and control. To accelerate the training process, the input spectra were down-sampled to shrink the input scale. The diffusion model has been trained by 7000 steps with Adam optimizer with the learning rate of $4\times {10}^{-6}$. The synthetic spectra were then up-sampled to the original spectral length. The t-step shrinkage parameter was denoted as ${\alpha }_t=\sqrt{1-\frac{0.02t}{T}}$ and the t-step noise variance as ${\beta }_t=\frac{0.02t}{T}$. As shown in Equation 3, the spectrum of ($t-$ 1)-step (${\mathit{s}}^{t-1}$) is inferred from the $t$-step spectrum (${\mathit{s}}^t$) by denoising. The synthetic SERSomes were concatenated by the $T$-step denoised spectra from Gaussian random inputs, as shown in Equation 4.

$${\mathit{s}}^{t-1}=\frac{1}{{\alpha }_t}({\mathit{s}}^t-{\beta }_t{ϵ}_{\theta }\left({\mathit{s}}^t,t\right)+{\beta }_t\mathit{z},\mathit{z}\sim N\left(0,\mathit{I}\right),\mathit{t}\in \left(0,T\right)$$ (Equation 3)

$$\hat{\mathit{S}}=\left[s_1^0,s_2^0\dots s_{200}^0\right]$$ (Equation 4)

The SERSome-based deep learning diagnosis for prostate cancer—PCDNet: The medical information used for PCa diagnosis is consistently and coupled embedded in the spatial-spectral features of SERSome data. Fully understanding the spectral and spatial information and their association can strengthen the SERSome for diagnosis. Actually, this has been widely practiced in acoustics that the time-frequency feature of sound was jointly rephrased in the visual form of Mel spectrograms and processed by the computer vision methods.⁶⁴ The PCDNet (details in Table S7) architecture was adapted from the typical Resnet, which was generally applied as a visual encoder ($f_{PCD}$). It consists of an initial convolution block followed by 4 residual blocks, integrated with flattening operation as the spectral-spatial encoder ($f_{enc}$) and a final fully connected classification layer. The residual blocks contained skip connections between the input and output of each residual block, benefiting stable training and gradient propagation.⁹⁰ Herein, the learning criterion was to minimize the cross-entropy loss between the predicted classes and the ground truth (GT) classes (PCa, BPH and control). The prediction was inferred from the SERSome inputs, as shown in Equation 5. t-SNE was done based on 6144-D features vectors extracted by $f_{enc}$. Then, every SERSome was mapped to a 6144-D latent feature vector for t-SNE analysis and visualization.

$$pred=f_{PCD}\left(\mathit{S}\right)=softmax\left({\mathit{W}}^T{f}_{enc}\left(\mathit{S}\right)\right)$$ (Equation 5)

The models were trained on the training sets and validated after every training epoch. The initial learning rate was 10⁻⁴ and decay every 10 epoch to 98% of previous learning rate. The models were saved at the epoch when validation accuracy was highest and training accuracy larger than 75%. Then the models were evaluated on the test sets.

Mining spectral features—explainable clinical deep learning: Explainable AI (XAI) is an emerging development in deep learning, which enables the knowledge that deep models learn from medical big data to be distillated to clinical experts. Spectral data have good interpretability because of the stable and explicit correlation between characteristic bands. However, traditional statistical methods, though having screened some differentiated bands, are not necessarily to find the features literally conducive to classification. Spectral data combined with deep learning can learn the principles for diagnosis, and interpretable models can visualize the features that the deep models pay attention to. The feature peaks from interpretable deep learning are more reliable. The Gradient Class Activation Map++ (Grad CAM++)⁶⁹ is applied to the well-trained PCDNet to visually show the inference procedure of the deep model via saliency maps. The feature maps ($\mathit{A}$) were encoded by $f_{enc}$, as shown in Equation 6. The saliency maps were the weighted summation of $\mathit{A}$, and the weights (${\mathit{\omega }}_k^c$) were calculated by the first, second and third-order back propagation gradients, as shown in Equations 7 and 8. The saliency maps were then activated by each class ($c$), denoted as $L^c$. We calculated the spatial average of $L^c$ and obtained the saliency spectrum, as shown in Figure 5B.

$$\mathit{A}=f_{enc}\left(\mathit{S}\right)$$ (Equation 6)

$$\begin{array}{c}{\mathit{\omega }}_k^c=\sum _i{a}_i^{k,c}\text{Relu}\left(\frac{\partial pre{d}^c}{\partial {\mathit{A}}_{\mathit{i}}^{\mathit{k}}}\right)\text{,}\\ a_i^{k,c}=\frac{{\partial }^{2}pre{d}^c}{{\left(\partial {\mathit{A}}_{\mathit{i}}^{\mathit{k}}\right)}^{\mathbf{2}}}/\left(2\frac{{\partial }^{2}pre{d}^c}{{\left(\partial {\mathit{A}}_{\mathit{i}}^{\mathit{k}}\right)}^{\mathbf{2}}}+\sum _j{A}_j^k\frac{{\partial }^{3}pre{d}^c}{{\left(\partial {\mathit{A}}_{\mathit{i}}^{\mathit{k}}\right)}^{\mathbf{3}}}\right)\end{array}$$ (Equation 7)

$${\mathit{L}}_i^c=\sum _k{\mathit{\omega }}_k^c{\mathit{A}}_i^k$$ (Equation 8)

Prediction of Gleason score: The clinical information (T-stages, N-stages and PSA) is used to predict the Gleason score via Multi-Layer Perceptron (MLP). The SERSome-based method employed the PCDNet. During the process of multimodal fusion, the spectral features were encoded by transferring $f_{enc}$ from the PCDNet and fused with clinical indicators. The cases of GS = 6, 8 & 9 are insufficient so that operations of copy and noise addition are implemented to balance the training data.

A ROC analysis and AUC were used to evaluate model performance for diagnosis and pathological stratification of PCa. The multiclass classification accuracy is defined as $\frac{\sum _i^N{\mathit{f}}_{\mathbf{1}}\left(\text{prediction}=\text{label}\right)}{N}$, where $f_1(\ast )$ is true-or-false indicator so that $f_1\left(\mathrm{T}\text{rue}\right)=1$ and $f_1\left(\text{False}\right)=0$. For multiclass tasks, the macro averages of ROC and AUC were used as the metrics for each class. The ROC curves were plotted by using the true-positive rate (TPR, namely, sensitivity) versus the false-positive rate (FPR, namely, 1 − specificity) with changing decision thresholds. For a model $f$ and a given ROC curve TPR = f(FPR), where FPR∈[0,1], the AUC is defined as: $\underset{0}{\overset{1}{\int }}f\left(x\right)dx$. Normalized confusion matrices were also applied to quantify the classification results. The above mentioned multivariate and statistical analyses were performed using scikit-learn (version 1.2.2) and scipy (version 1.9.1) packages in Python (version 3.9.16). All SERSomes were processed on Intel(R) Xeon(R) Platinum 8276M CPU @ 2.20GHz and a total of two Nvidia 3090 graphics processing units.

Published: May 21, 2024