1. Introduction

Glucose monitoring is critically important in biomanufacturing. The industry has recently undergone a significant growth thanks to the strong demand for biologic vaccines and drugs. These biologic medications are the core of many cutting-edge technologies to tackle various ineradicable diseases, such as pandemic viral infections [1], cancers [2], and autoimmune disorders [3]. Biological products are manufactured by living organisms, e.g. Chinese hamster ovary cells [4], which are cultured in production bioreactors with controlled nutrient supply [5]. Glucose is the primary nutrient for living cells, and needs to stay at certain concentrations in the bioreactor so that the cells can be grown with maximized productivity, without sacrificing the product quality [6]. More specifically, the glucose concentration is usually maintained at 5–40 mM [6], beyond which the biologic production can suffer from low cell density (too little glucose) or toxic metabolic byproducts (too much glucose) [6,7]. To hold the constant balance between cell productivity and quality, a rapid and effective glucose-concentration monitor is desirable in biomanufacturing.

Raman spectroscopy is a promising tool to achieve the goal. Unlike the conventional methods (like ELISA or mass spectrometry) [8], which entail disruptive sampling and long-hour processes, Raman spectroscopy allows real-time (online) bioprocess monitoring because of its non-invasive nature and high molecular specificity [5]. Measuring Raman signals produced by the vibrating glucose molecules in the cell culture media has been demonstrated as an effective process analytical technology (PAT) in the biopharmaceutical industry [5–8]. However, determining glucose concentration with Raman signals is a challenging work. This is mainly due to the low Raman scattering cross section of glucose [9], making it very difficult to record repeatable signals. The challenge is even greater in the biomanufacturing reactor, where the complex ingredients of cell culture media severely interfere the weak Raman intensity of glucose [5]. These obstacles mandate increased sampling frequency, multivariate modeling, even sophisticated machine learning tactics to improve the prediction accuracy of glucose concentration [5–8]. These extra computing costs compromise the benefits brought by Raman spectroscopy.

In this study, we advance the Raman-based PAT for glucose monitoring in the cell culture media. The new PAT is made by a surface-enhanced Raman spectroscopy (SERS) built with InGaN quantum wells (QWs), whose exceptional carrier confinement provides abundant surface charges for the resonance effects contributing to SERS intensity [10–12]. Because of the additional resonance charges from QWs, SERS signals recorded on the nitride surface exhibit not only enhanced intensity, but also improved stability in space and in time [11], which is strongly desired for practical applications. We demonstrate that the two-dimensional InGaN thin layers can deliver excellent specificity of SERS signals, allowing quick, accurate and cost-effective determination of glucose concentration in the complex cell-culture environment.

2. Materials and methods

The nitride SERS biochips were grown on sapphire substrates by metal-organic chemical vapor deposition (MOCVD, AIXTRON 200/4 RF). Details of the QW structure were given in our previous report [10]. In order to verify the advantages of InGaN QWs, two nitride SERS samples were prepared, i.e. the one with three-repeat QWs (3QW) and the one without QWs (0QW). Except for the QWs, the two samples were prepared with identical growth and fabrication conditions. To induce the SERS effect, the QW surface was covered with an Au nanostructure, which was fabricated by a 45-nm Au deposited using an e-beam evaporator, followed by a rapid annealing at 300 °C for 140 seconds in a N₂ atmosphere. Glucose solutions with controlled concentrations were made by dissolving glucose powder in the cell culture medium, CD FortiCHO (catalog number A1148301), purchased from Thermo Fisher Scientific Inc. Raman spectra were excited by a 532-nm single frequency diode-pumped solid-state laser (LASOS Lasertechnik GmbH, power: 18 mW, exposure time: 1 second). An objective (50x PlanApo, Mitutoyo) with a numerical aperture of 0.55 was utilized to focus the excitation light on the specimens. Before each measurement, 0.5 µL of glucose solution was drop-casted on the SERS substrate and dried naturally. The output light was filtered using an edge filter (Semrock) and then collected by a spectrometer (Shamrock 500i, Andor). Scanning electron microscopy (SEM) images were recorded with field emission HITACHI S-4300 at the acceleration voltage of 10 kV. All of the characterizations were carried out at room temperature.

3. Results and discussion

Figure 1(a) shows the layer structure of the 3QW sample. For the 0QW, the growth was stopped at the 2-µm GaN. The nanoporous Au surface, displayed in Fig. 1(b), was formed by the rapid annealing process to induce the SERS effect [13]. Wafer-scale uniformity is facilely attainable with the MOCVD growth (Fig. S1 in Supplementary Materials). Although the samples used here are ∼1 cm long, for practical applications, the detecting area can be less than 1 × 1 mm² because of the small laser spot size (50 µm). Commercial potential of the nitride SERS biochip can be estimated by the production of GaN-based light-emitting diodes (LEDs, size: ∼1 × 1 mm²), whose market prices are now around $\$$0.32 USD/chip [14]. Compared to that of LED, the production of our SERS device is much simpler, requiring no p-type layer, no photolithography, no etching, no thick metals, no wire bonding and no epoxy packaging. The cost for the nitride biochip is therefore expected to be lower than $\$$0.32 USD/chip.

 

Fig. 1. (a) Layer structure of the SERS biochip with 3-repeat InGaN quantum wells (QWs). (b) SEM image of the Au surface roughened by the rapid annealing at 300 °C for 140 seconds in N₂.

Download Full Size | PDF

The 3QW sample exhibits a peak at 557 nm in the photoluminescence (PL) spectrum under the excitation wavelength of 532 nm [10]. The same excitation source was used for the SERS characterizations in this study. To ensure efficient charge coupling between Au and QWs [10], the QWs were capped with a thin (1.6 nm) GaN layer. For SERS applications, PL of the QWs is often undesired because the luminescence background can severely interfere the weak Raman signals [15]. This is not an issue after the coverage of 45-nm Au, on which the PL peaks were not detectable. In fact, the relatively thick Au layer not only blocks direct optical excitation of the QWs, but also intensifies the Raman signals of glucose (see Fig. S2 in Supplementary Materials).

Figure 2(a) and 2(b) respectively presents the Raman spectrum of glucose (87 mM) in deionized water and the cell culture medium recorded on the 3QW sample. Although the target peak at 1125 cm^-1 (corresponding to the C-O-H bending of glucose) is visible in both solutions, the two figures show distinct spectral features. In Fig. 2(a), since the Raman peaks of H₂O are at the wavenumbers beyond 3000 cm^-1 [16], the water solution produces limited background noise in the spectral range for glucose, allowing multiple peaks to be detected. These peaks are commonly reported in the literatures [17–22]. On the other hand, dissolving glucose in the culture medium leads to a much reduced signal-to-noise ratio, blotting out many feature peaks, as seen in Fig. 2(b). The result is caused by the various constituents in a cell culture medium, including salts, amino acids, vitamins, and many others [5], all of which contribute to the much stronger background intensity (indicated by the 10-fold difference in the vertical scale bars of the figures). Such interfering noises make the quantification of glucose concentration a challenging task during the cell culture process. Although not as clear as that in water, the 1125-cm^-1 peak in Fig. 2(b) is noticeable. The result in Fig. 2(b) is not attainable with regular Raman spectroscopy for biopharmaceutical manufacturing [6,7], which often requires sophisticated spectra processes (e.g. derivatives, smoothing, peak-search algorithms) to reveal the change with glucose concentration.

 

Fig. 2. Raw SERS spectra of the glucose (87 mM) dissolved in: (a) deionized water; (b) cell culture medium, showing the reduced signal-to-noise ratio in the cell culture medium.

Download Full Size | PDF

Figure 3 shows the concentration-dependent (1-87 mM) Raman signals of glucose in the culture medium recorded on the 0QW and 3QW samples. Each spectrum at a concentration was the average of six measurements at different locations, all within a dried droplet (0.5 µl) on the chip. With the strong background removed [11], the 1125-cm^-1 peaks become more noticeable comparing to the one in Fig. 2(b). The effect of QWs is manifested in Fig. 3(c) and 3(d), which present the dependence of peak intensity at 1125 cm^-1 on concentration. With identical intensity scale of the two figures, one can see the glucose on 3QW not only delivers stronger signals, but also higher uniformity (i.e. smaller error bars, which are the standard deviations of the six measurements). More importantly, the peak intensity on 3QW increases linearly from 1 mM to 20 mM and levels off at the higher concentrations. The curve exhibits a typical Langmuir adsorption isotherm [23], which is a dynamic model describing a monolayer of molecules covering an adsorbent surface [24].

 

Fig. 3. SERS spectra of the glucose in cell culture medium (diluted from 87 mM to 1 mM) recorded on the nitride surface with: (a) no quantum wells (0QW); (b) 3-repeat quantum wells (3QW). The peak intensity at 1125 cm^-1 as a function of concentration on 0QW and 3QW are given in (c) and (d), respectively. The classic Langmuir adsorption isotherm is seen on 3QW, indicating that the surface is covered by a monolayer of glucose molecules with a linear concentration-dependence at 1-20 mM.

Download Full Size | PDF

The Langmuir curve in Fig. 3(d) indicates that the monolayered glucose molecules on the Au/QW surface are mostly exposed to the SERS hot spots, which are significantly densified by the QWs. Our recent studies show that the electrons confined in QWs can contribute to SERS via two routes: i) Enhance the charge transfer from the discrete energy levels in QWs to the Au surface, building up the collective electron vibration for Raman scattering [10,11,25]; ii) Promote the localized surface plasmon resonance by the coupling between Au and QWs, in which the electrons on Au and those in QWs are oscillating at the same frequency [10–12]. These two contributive routes from QWs densify the hot spots, significantly expanding the SERS-active region. In other words, the QWs can form a hot “surface” by interconnecting the densified hot spots. Detailed elucidation and experimental evidences on the SERS hot surface realized by QWs were provided in our recent reports [10–12]. Compared with the conventional SERS hot spots [11], the hot surface exhibits enhanced sensitivity and specificity for SERS sensing, as the molecular vibration modes are more likely to be boosted by the SERS resonance that only occurs at a certain frequency [12]. Therefore, the 1125-cm^-1 peak gradually goes up with concentration as more and more glucose molecules occupy the sorption sites, and the climbing intensity (before 20 mM) eventually stops at the maximum as the surface is fully covered by the molecules (i.e. no more exposure to the surface). The linear curve at 1-20 mM overlaps the working range (5-40 mM) of industrial cell cultivations [6]. Preserving the linear region in Fig. 3(d) is crucial in translating Raman spectra to glucose concentration [19]. Extending the linear region beyond 20 mM (e.g. 40 mM) can be attained by reducing the droplet volume (thus fewer molecules), with which the full coverage of glucose monolayer occurs at the concentration higher than 20 mM. Detecting resolution (Δn) on the 3QW chip can be calculated by the equation: Δn = 3Δd/s [26], where Δd is the noise level (6.78 a.u. in our case, determined by the standard deviation of the spectral intensity on bare surface) and s is the slope of the dotted fitting line shown in the figure. With the slope of 69 a.u./mM, the Δn of 3QW is 0.29 mM, smaller than that (0.38 mM) of 0QW. As shown in Fig. 3(c), the poor Δn of 0QW is accompanied by the larger deviation from linearity (smaller value of R²) and the absence of Langmuir isotherm. These inferior sensing performances can be explained by the much fewer hot spots on the QW-free surface [10–12]. Since only few of the glucose molecules are exposed to the sparse hot spots on 0QW, the SERS intensity cannot faithfully reflect the glucose concentration. The results in Fig. 3 demonstrate that the hot surface formed on 3QW makes it possible for SERS sensing in a complex environment. The number of QWs do affect the linear relationship between SERS intensity and analyte concentration, and should be selected based on the trade-off between surface-charge density and QW crystal quality [10]. Although increasing QW number can boost SERS intensity, too many QWs will deteriorate the crystal quality of InGaN (thus SERS intensity) because of the lattice strain built up on GaN [27].

4. Conclusion

The feature and advantage of our QW-based biochip for glucose detection are listed in Table 1, which summarizes the performances achieved by various SERS techniques. It is seen that most of the biochips require binder molecules to increase the Raman intensity of glucose. Even with the binders, none of the previous techniques performed the detection in complex biological media, which is essential for practical applications. Our SERS chip with roughened Au surface and QW-induced hot surface allow the binder-free glucose detection in a noisy culture media, without the disturbing issue of Ag oxidation [28]. The robust, simple and quick nitride SERS biochip is an important step towards the glucose-monitoring in biomanufacturing, as well as the real-time quantitative analysis in other medical applications.

Table 1. SERS detection of glucose reported by research groups, highlighting the features of demonstrated in this study.

View Table