Highly sensitive terahertz spectroscopy in microsystem

Abstract

We herein present a microfluidic system dedicated to THz spectroscopy of aqueous solutions. This device is able to reach a state of the art sensitivity of 5 mg mL⁻¹, while requiring only small sample quantities and a low power source. Hydration of BSA, lysozyme and chymotrypsine proteins is studied inside microchannels and hydration numbers are computed, showing that the developed system is well suited for quantitative analysis. Moreover, coupled with advanced chemometrics algorithms, this system can become a tool for fundamental research and for the understanding of biochemical processes. Here, the hydration shell structure is discussed by chemometric analysis of the measured absorption spectra. Two spectral behaviours are observed and can be explained by Molecular Dynamics simulations. This methodology could be considered as a key element of a lab-on-a-chip for biological liquid metrology.

---

Introduction

Terahertz spectroscopy, which deals with electromagnetic waves in the 100 GHz–10 THz frequency band,¹ is now known to be of great interest to the biological and chemical sciences.^2,3 Numerous studies have shown that terahertz waves could detect molecular signatures in gases, distinguish healthy and cancerous tissues⁴ or probe DNA hybridization.⁵ Conformational changes in proteins can also be observed with this real-time, marker-free and non-ionizing technique.^6,7

However, studying liquids with THz waves is difficult due to the significant THz absorption by water, resulting from the excitation of both water dipolar moments and the hydrogen bond network. For instance, sending a 1 THz wave through a 1 mm-wide water sample divides its power by 10⁹. To circumvent this problem it is possible to use more powerful THz sources,⁸ with however the risk of sample heating. Performing reflection spectroscopy is another way to overcome this issue but data extraction, based on the reflected signal phase analysis, is complex.^9,10 Finally, a third way is to reduce the sample volume, which enables us to perform transmission measurements with the usual integrable and low-power sources.^11–14 In this case, accurate volume control of the solutions is needed and this is why microfluidic systems have already been proposed to perform THz spectroscopy.^15–17

Here, we demonstrate that integrating THz waveguides directly into the microfluidic system makes it possible to reach a higher sensitivity than previously developed microfluidic systems dedicated to THz spectroscopy, in which the spectroscopy and fluidic aspects are less intimately integrated. With this configuration, it is possible to get sensitivity in the state of the art conventional setups, with the added benefits brought by microfluidics : (i) to consume small sample quantities, (ii) to accurately control sample injection and measurement in a real-time way and (iii) to be integrated in a multiparametric, multiprobe and multifunctional microfluidic platform.

Previous studies have shown that protein hydration, which is the interaction between a protein and water, plays an important role in protein function.^18,19 Moreover, intermolecular interactions of hydrated proteins with sugars could be involved in the fascinating biopreservation phenomenon.^20,21 Studying hydration with THz spectroscopy has been done previously and has shown that the behaviour of water molecules changes in the vicinity of the protein.^22,23 These water molecules are usually referred to as “bound water” or “hydration water”, whereas water molecules far from the proteins are designated as “bulk water”. The bound water properties depend on protein surface chemistry, as the water behaviour around the hydrophilic and hydrophobic sites is different. Furthermore, backbone and side-chain hydration regions are also expected to be distinguishable.²⁴ From a THz spectroscopy point of view, the absorption of the hydration water decreases, compared to bulk water, due to the rigidity of the hydrogen bonds, which is increased between protein and water.^9,25–29 However, some studies show that the hydration shell absorption around proteins seems to increase due to collective vibrations.^30,31 A clear understanding of these phenomena requires more THz measurements, comparison with other spectroscopy techniques and the use of computational molecular dynamics simulations.

This paper deals with sub-THz measurements on aqueous solutions of lysozyme and Bovine Serum Albumin (BSA), which are commonly used in THz spectroscopy. First, the microsystem used will be presented. Then, measurement sensitivity will be discussed and the measured data will be used to characterize the hydration of the two proteins by computing their hydration number (number of water molecules in the hydration shell). In a third part, chemometrics algorithms will be introduced to improve hydration characterization. Lastly, the hurdles to overcome, should one wish to increase the measurement frequency in the microsystem, will be discussed.

1 THz microfluidic system

To perform liquid-absorption measurements, a two-function microsystem has been developed. On the one hand, a microfluidic circuit drives the samples toward the analysis area. On the other hand, an integrated electromagnetic circuit guides the THz waves through the sample to be analyzed. The whole system is presented in Fig. 1 (left) and each function is detailed in the following sections. This microsystem is based on a glass/silicon/glass technological process. Glass is used to ensure both the transparency of the microsystem and the wave propagation with low losses (tanδ ∼ 5 × 10⁻³ at 300 GHz). To fabricate the microchannels, silicon has been chosen for its rigidity and its low losses in the terahertz range (tanδ ∼ 10⁻⁴ at 400 GHz). In contrast with glass, silicon can be accurately deep-etched, resulting in well defined narrow channels. This is crucial, because power absorption depends exponentially on the microchannel width. The use of silicon will also be useful for future microelectronic component integration to get a fully autonomous lab-on-chip for biological metrology. All the technological parameters and steps used to manufacture the BioMEMS have previously been reported.³² The independence of microfluidic and electromagnetic design has to be emphasized. Indeed, these two circuits can be modified independently without changing the technological process parameters, leading to a strong versatility in function design. This is due to the process where the whole silicon wafer is reported on the glass wafer (with THz waveguides). Then, silicon deep-etching is performed and leads to a high-density microchannel network with thin walls (width = 200 μm), which is an important refinement to limit absorption of the empty microsystem. With this clean-room process, 23 microsystems can be realized on a 3-inch glass wafer.

Fig. 1 THz BioMEMS with electromagnetic and microfluidic circuits.

The THz waveguide

To guide the electromagnetic wave, a single-wire propagation line has been used. This so-called “Planar Goubau Line” (PGL) enables the quasi-TEM propagation of the THz waves in a surface-wave configuration called “Goubau mode”. In this configuration, the electromagnetic fields are strongly confined around the metallic wire, whose width can be decreased down to the micrometer-scale.³³ The confinement of THz waves can reach several micrometers, which corresponds to the energy exponential decay length perpendicular to the Goubau line. This confinement depends on the wire width and on the dielectric permittivity of the propagative medium (air/glass, silicon/glass or liquid/glass).

Here, the Goubau mode is excited by an external Vector Network Analyzer (VNA), dedicated to coplanar waveguide excitations in the 0–110 GHz and 140–220 GHz frequency bands. This is the reason why a coplanar-waveguide/Goubau-line transition has previously been developed.³⁴ The Goubau line and coplanar transitions are represented in Fig. 1 (upper right).

This electromagnetic circuit is made by gold deposition on a transparent glass wafer. The Goubau line width is 5 μm wide and 500 nm thick, and propagation losses are from 2 dB mm⁻¹ to 3 dB mm⁻¹ in the 50–220 GHz frequency band.³³ Propagation on Goubau lines has been achieved for higher frequencies up to 800 GHz³⁵ and measured losses at 0.63 THz were about 0.6 dB mm⁻¹.³⁶ Computations show that propagation on Goubau lines is possible in the 0.8–2 THz frequency band.³⁷

The microfluidic circuit

To drive the sample, the 200 μm-deep microchannels are made with silicon walls sandwiched between two glass wafers.

The microfluidic circuit is composed of a 1 mm-wide main channel and a 200 μm-wide measurement channel, where the THz probing will occur. Fig. 1 (bottom right) shows this two-channel configuration, which allows both a quick injection of the probed samples and a measurable transmitted signal through the 200 μm-wide sample.

The system technological process characterization has shown that pressures up to 37 bars can be applied inside the BioMEMS. Such a microsystem has already been used to probe enzymatic reactions in real-time³⁸ and to characterize ethanol hydration layers.³⁹ Measurements on proteins are presented in the next section. They will show that the BioMEMS is an adapted tool to perform quantitative physical analysis of the hydration phenomenon directly in a chip.

2 Protein hydration studies

Material and methods

Lyophilized powder of BSA (66, 43 kDa), α-chymotrypsine from bovine pancreas (EC 3.4.21.1; 24 kDa) and lysozyme from chicken egg white (EC 3.2.1.17; 14, 6 kDa) purchased from Sigma-Aldrich Chemical Co. have been solubilised in pure deionised Milli-Q water (Millipore, resistivity: 18 MΩ cm) at different concentrations and from independent weightings. All solutions were prepared with a total volume of 5 mL and were clear and without precipitates.

Measurements have been carried out by an Agilent XF8510 VNA in the 0–110 GHz frequency band. The raw obtained parameters are the complex transmission parameter S₂₁ and the complex reflection parameter S₁₁. Transmitted and reflected powers can be deduced from the squared modulus of these parameters. The chronological order of sample injection has been randomly chosen in order to avoid any measurement bias. Between each measurement, water is injected as a reference sample in order to check measurement reproducibility.

Measured S₂₁ parameters are shown in Fig. 2 for BSA solutions. The uncertainty concerning the S₂₁ parameter measurement is δS₂₁ ∼ 0.02 dB. The corresponding S₁₁ parameters are not shown because they are below −10 dB on the whole frequency band and their dependence on BSA concentration is weak compared to S₂₁.

Fig. 2 S ₂₁ transmission parameter as a function of frequency for several BSA concentrations. Measurement uncertainty is δS₂₁ ∼ 0.02 dB. Inset: zoom on the 80–85 GHz frequency band.

Absorption extraction protocol

Since the electromagnetic propagation mode is a quasi-TEM mode, it can be assumed that the transmitted power P_t through the system is

P_t = P_iexp(−K) (1)

where P_i is the incident power and K represents the total losses in the system. The K contributions can be from the sample in the microchannel and from the losses k of the empty microsystem. If it is assumed that these two contributions are independent, it can be written that

P_t = P_iexp(−αL + k) (2)

where α is the sample absorption and L the sample width (and thus the measurement channel width). The incident and transmitted power can be expressed as functions of S-parameters with

The sample absorption is thus given by

This extraction protocol suffers from several assumptions, such as neglecting internal reflections and their dependence on the sample inside the microchannels. However, since sample absorption is high, these reflections are weak, as can be shown by numerical computations. Moreover, k is an unknown offset. This is the reason why the next results will be given in arbitrary units. However, the previous assumptions do not affect the quantitative differential analysis that will be done, and a calibration protocol of the system allows us to get the quantitative absorption values, as will be shown in Section 3.

Water/protein solution absorption

Fig. 3 shows the absorption of BSA and lysozyme solutions as a function of protein concentration for a frequency of 100 GHz. Increasing the protein concentration decreases the solution absorption. This is due to replacement of water molecules by proteins whose absorption is negligible in this frequency range. The absorption of a 5 mg mL⁻¹ protein concentration can be detected and is the sensitivity of the microsystem for protein spectroscopy. This value has to be compared with previous sensitivities obtained in terahertz spectroscopy in microfluidic systems: a 100 mg mL⁻¹ sensitivity was obtained by George et al.¹⁶ and a 50 mg mL⁻¹ sensitivity by Kitagawa et al.²⁷ For studies that do not use microfluidic systems, sensitivities are in the order of magnitude of 5 mg mL⁻¹ when measuring absorption. The most significant literature studies give the following sensitivities with corresponding molar fractions x_m :

• 60 mg mL⁻¹ (x_m = 3.1 × 10⁻³) for sugar/water solutions by Arikawa et al.;⁹

• 30 mg mL⁻¹ (x_m = 11.6 × 10⁻³) for alcohol solutions by Jepsen et al.;^10,40

• 10 mg mL⁻¹ (x_m = 12.6 × 10⁻³) for lysozyme solutions by Vinh et al.;²³

• 2 mg mL⁻¹ (x_m = 0.93 × 10⁻⁶) for Subtilisin Carlsberg protein dissolved in dioxan organic and apolar solvent by Mickan et al.;²⁶

• around 1 mg mL⁻¹ (x_m = 1.9 × 10⁻⁶) for λ^*_6–85 protein in water by Ebbinghaus et al.⁴¹

Fig. 3 0.1 THz absorption of BSA/water and lysozyme/water solutions as a function of protein concentration.

In comparison with these studies, one advantage of microfluidics is that the required sample volume decreases. Indeed, considering the 200 μm-deep and 200 μm-wide microchannel and considering that the field extent around the Goubau line is below 100 μm, the sample volume probed is below V = 200 μm × 200 μm × 200 μm = 8 nL. For BSA, the concentration sensitivity of 5 mg mL⁻¹ (x_m = 1.35 × 10⁻⁶) in the V volume leads to a 0.6 picomole detection threshold. A previously reported result for THz spectroscopy with microfluidics without integrated probes obtained a 10 picomole detection limit.¹⁶

The normalized absorption of each BSA/water and lysozyme/water solution is given as a function of concentration and of frequency in Fig. 4 in the 50–110 GHz frequency band. A very good agreement is found between these curves and the results from Kitagawa et al.,²⁷ thus validating the absorption extraction protocol and enabling us to perform hydration quantitative analysis.

Fig. 4 Normalized absorption of BSA/water and lysozyme/water solutions as a function of frequency and concentration.

Hydration number of proteins

As previously stated, the decrease in solution absorption α when protein concentration c (mg mL⁻¹) is increased is due to water molecule replacement by protein. The absorption decrease is called Δα = α_H2O − α = α_prot, where α_H2O is the water solution absorption. With x as the protein volume fraction, it becomes Δα = xα_H2O. Besides, x can be expressed as a function of the partial specific volume of the protein⁴²V_s: x = cV_s. However, experimentally, the measured absorption decrease is higher than the computed Δα using the previous expressions. This can be explained by assuming that water molecules in the vicinity of the protein (hydration water) no longer absorb the electromagnetic waves. This assumption^13,39 will be discussed experimentally in the next section. In this case, for a solution volume V:

Δα = α_prot + α_hydr = xα_H2O + x_hydrα_H2O (5)

where x_hydr = N_hydrM_H2O/(VN_Aρ_H2O) is the hydration water volume fraction in solution. M_H2O is the water molar weight, N_A is the Avogadro number, ρ_H2O is the water density and N_hydr is the number of hydration water molecules in solution. With N_h the hydration number (number of hydration water molecules in the first hydration shell per protein) and N_prot the number of protein molecules in solution, it becomes N_hydr = N_protN_h. Considering that N_prot/V = cN_A/M_prot, with M_prot being the protein molar weight, it becomes :

Thus, N_h uncertainties can be computed with the following formula :

where δα are absorption uncertainties. δN_h decreases when protein molar weight decreases and when concentration increases. This is why the solution concentrations used for computing N_h will be 50, 75 and 100 mg mL⁻¹. Fig. 5 shows the computed hydration numbers as a function of frequency and solution concentration. Results for BSA and lysozyme are in good agreement with previous studies^{23,28,43–45} and validate the N_h computing method, which can be used for other solvated proteins. Thus, chymotrypsin/water solutions have been analyzed with this extraction process. The corresponding hydration numbers N_h are given in Fig. 5.

Fig. 5 Hydration number N_h for lysozyme, BSA and chymotrypsine as a function of frequency and protein concentration. N^*_h is the corrected hydration number, as defined in Section 3.

It can be observed that N_h seems to depend on frequency. However, N_h uncertainty margins overlap on the entire frequency band and their intersection should give the best estimation for N_h. Besides, the computed N_h decreases when the protein concentration increases. Actually, this is due to hydration shells overlapping: when protein density increases, a hydration water molecule can be shared between several proteins and the N_hydr = N_protN_h equation is no longer respected. Moreover, we will show in the next part that the model used is approximate and only quantifies the number of water molecules in the direct vicinity of the proteins (first hydration layer). However, it has been largely used in literature^25–28 and gives a good approximation of the N_h order of magnitude from THz spectroscopy of proteins.

The hydration of proteins can thus be quantitatively probed in the microsystem. Moreover, due to the accuracy and sensitivity of the developed microsystem, it can also become a tool to increase our understanding of bio-chemical phenomena. Indeed, the next part will show that a combination of microsystem measurements and chemometrics algorithms leads to a better understanding of hydration processes.

3 Ethanol hydration insight

Several previous studies have focused on characterizing alcohol/water mixtures in the THz spectrum.^10,25 Compared to proteins, which have a limited solubility, when working with alcohol one can produce a continuum of samples with concentrations spanning from pure water to pure alcohol. As a result, large datasets can be generated for further analysis. In this part, the sub-THz spectroscopy of alcohol/water mixtures is presented. First, the absorption of the mixtures will be given. Then, the full measurement dataset will be analyzed by chemometrics algorithms in order to extract maximum information from the measured broadband data. From that, an improved model of solvated molecule hydration will be postulated.

Alcohol/water mixture absorption

Ethanol/water, isopropanol/water and propanol/water mixtures have been prepared with alcohol volume ratios x from 0 to 1 in steps of 0.05. Ethanol, propanol and isopropanol are from Carlo Erba Reagents. Pure deionised water is obtained from a Barnstead Easy-Pure RoDi purification system with a resistivity of 18.2 MΩ cm.

By a calibration step, using absorption Debye models for alcohol and water from previous studies,^46–48 the absorptions of the injected mixtures are obtained as shown in Fig. 6. As observed for the protein study, increasing the alcohol content decreases the solution absorption. Indeed, water molecules are progressively replaced by alcohol molecules, whose absorption is weaker. As already observed, pure isopropanol absorption is weaker than propanol absorption, which is weaker than ethanol absorption. It is possible to use the measured absorptions to characterize alcohol hydration number in quite a similar way as presented in the previous section. This study has already been reported for ethanol/water mixtures.³⁹

Fig. 6 Ethanol/water, propanol/water and isopropanol/water mixture absorptions as a function of frequency for several ethanol volume ratios. The arrow indicates the increase of alcohol volume ratio in steps of 0.05.

Here, the study goes further, using multivariate analysis i.e. exploiting simultaneously all frequencies in the spectral domain. Note that in the next section, the raw (non-calibrated) absorption dataset will be used.

Linear decomposition of the dataset

Measured absorption spectra are rather flat and do not show obvious features that can be considered as molecular signatures. However, it is possible to apply advanced algorithms to these spectra to better characterize the measured mixtures. These algorithms, coming from the field of chemometrics, are commonly used for vibrational spectroscopy.^49–51 Their goal here is to decompose the solution spectra as a sum of the pure spectral contributions of each solution component, multiplied by the component volume ratio: where n is the number of components in the solution, x_i the volume ratio of the i^th component and α_i the absorption spectrum of the i^th component. This decomposition of the spectra is only possible with highly sensitive, low-noise and reproducible measurements because the algorithm extracts the maximum information from the spectra details. Thus, the chosen frequency band (10–90 GHz) corresponds to the measurements’ higher signal to noise ratio. Note that, due to some multiple wave reflections in the microsystem, some ripples appear in the absorption spectra. However, since they hardly depend on the probed liquid inside the microchannels, their influence on the chemometrics algorithms is negligible.

The first step in the decomposition is to determine the number n of pure components observable in the spectral dataset. This is made by using a Principal Component Analysis (PCA) algorithm, based on singular value decomposition.⁵² Then, an iterative algorithm (MCR-ALS : Multivariate Curve Resolution – Alternating Least Squares^53–55) is used to decompose the whole dataset to determine x_i for each mixture and α_i for each frequency. Methodology concerning this study has been reported elsewhere.⁵⁶ Here we want to present the study results in the case of ethanol/water mixtures. They show that four species can be observed in the absorption spectrum of an ethanol/water mixture. The spectra of each component, and the corresponding volume fraction in each mixture are given in Fig. 7 (left). Note that the absorption spectra are non-calibrated. Indeed, they correspond to the convolution of the liquid calibrated absorption and the microsystem instrument function. However, the instrument function does not depend on the liquid inside the microchannels and so, the differences observed between the non-calibrated absorption spectra are directly correlated to the liquid calibrated absorption spectra.

Fig. 7 Decomposition of the measurement dataset by MCR-ALS. Concentration of the four species as a function of ethanol volume ratio. Non-calibrated absorption of the four species as a function of the frequency.

Component 1 disappears when the ethanol volume ratio increases. The concentration of component 2 increases with ethanol volume ratio. The highest absorption in the whole frequency band is the component 1 absorption, whereas the lowest one is the component 2 absorption. As a consequence, component 1 is bulk water and component 2 is ethanol. Components 3′ and 3′′ show more complex behaviour. They correspond to the hydration shell of ethanol. 3′ is the first hydration layer, strongly linked to ethanol, and whose absorption is quite the same as ethanol (far below water) and 3′′ is the second hydration layer with an intermediate absorption. To validate this explanation, Molecular Dynamics (MD) simulations have been performed for the ethanol/water mixtures using the DL POLY program.⁵⁷ The all-atom optimized potentials for liquid simulations (OPLS) force-field⁵⁸ was used to model the intra and intermolecular interactions of ethanol molecules. Rigid water molecules were modeled using the TIP4P/2005 force-field.⁵⁹ Results of MD are shown in Fig. 7 (right) and show a very good agreement with the chemometrics results. The whole methodology used for MD simulations has been reported elsewhere.⁶⁰ Investigations on isopropanol/water and propanol/water give the same behaviours.

The conclusion of such a study is that two hydration layers can be probed in the investigated frequency band. The first layer absorption is very weak, whereas the second layer absorption is an intermediate between water and ethanol. Since the absorption of water molecules directly linked to the solvated molecule is negligible compared to bulk water, the computed N_h in the previous section appears to be an approximation of the number of water molecules in the first hydration layer. Extending the MCR-ALS results to solvated proteins (protein volume ratio is x), the number of water molecules in the whole hydration shell (first and second layer) can be estimated by writing the absorption of a diluted protein solution :

α = xα_prot + x_hydr1α_hydr1 + x_hydr2α_hydr2 + (1 − x − x_hydr1 − x_hydr2)α_H2O (9)

where ‘hydr1’ and ‘hydr2’ stand for the first hydration layer and the second hydration layer, respectively, and thus x_hydr = x_hydr1 + x_hydr2. As a consequence, eqn (6) becomes: where N^*_h is the corrected hydration number, corresponding to the whole hydration shell. Indeed, the previously computed hydration number underestimates the total quantity of hydration water molecules. To estimate this bias, the α_hydr2/α_H2O ratio can be approximated by 0.6 ± 0.05, considering Fig. 7. Moreover, x_hydr2/x_hydr can be computed in the case of ethanol from previous studies³⁹ where the water molecule distribution function is presented. In this case, when assuming a spherical hydration shell model with x = 0.02 (low ethanol concentration with no hydration shells overlapping), x_hydr2/x_hydr ∼ 0.8. Note that this value is close to the one given by MD simulations in Fig. 7 for a small ethanol volume ratio. As a consequence, it seems that the corrected hydration number N^*_h is twice the one obtained (N_h) by the model used in the previous section (Fig. 5).

Further work will be dedicated to increasing the investigated frequencies. However, the transmitted signal decreases when frequency increases and will limit the use of the microsystem beyond 220 GHz. This can be explained by the substrate mode phenomenon: beyond a critical frequency (around 180 GHz in the microsystem), the whole electromagnetic wave is no more guided by the Goubau line but progressively by the substrate and thus it is no longer entirely detected in the transmitted signal. This drawback can be overcome by integrating corrugated Goubau lines in the microsystem.⁶¹ Since Vector Network Analyzers are now able to deliver frequencies of up to 1 THz, the described study has to be extended to 1 THz, in order to look at the frequency dependence of hydration probing in the microfluidic system.

Another perspective of this work is to study the influence of pH on protein hydration shells.⁶² Here, protein were dissolved in pure deionised water in order to avoid the presence of supplementary ions, which could interact with THz waves. However, the next work will be to modify the solution pH and to observe the variations in hydration properties, as a function of the iso-electric point of the protein, for instance. Hydration characterization could also lead to protein conformation monitoring by THz waves in the microsystem.

Conclusion

These studies have demonstrated that the developed microsystem is able to reach a sensitivity comparable to the sensitivity obtained with macroscale setups, with the added benefits of using small sample quantities. This sensitivity makes it possible to perform quantitative analysis of the protein hydration. The use of chemometrics algorithms on our measurements improves our understanding of the hydration bio-chemical phenomenon. They enable us to clarify the extraction model chosen for computing the hydration number. Indeed, we believe chemometrics is well suited to exploring and analyze high-throughput measurement data obtained by the “lab-on-a-chip” approach. This combination hints at a new way to develop liquid metrology platforms dedicated to biology, liquid physics or chemistry.

References

1. M. Tonouchi, Nat. Photonics, 2007, 1, 97–105.
2. P. H. Siegel, IEEE Trans. Microwave Theory Tech., 2004, 52, 2438–2447.
3. A. McIntosh, B. Yang, S. Goldup, M. Watkinson and R. Donnan, Chem. Soc. Rev., 2012, 41, 2072–2082.
4. C. Yu, S. Fan and E. Pickwell-MacPherson, Quantitative Imaging in Medicine and Surgery, 2012, 2, 33–45.
5. M. Nagel, F. Richter, P. Haring-Bolivar and H. Kurz, Phys. Med. Biol., 2003, 48, 3625–3636.
6. A. Markelz, S. Whitmire, J. Hillebrecht and R. Birge, Phys. Med. Biol., 2002, 47, 3797–3805.
7. A. G. Markelz, IEEE J. Sel. Top. Quantum Electron., 2008, 14, 180–190.
8. D. M. Leitner, M. Gruebele and M. Havenith, HFSP J., 2008, 2, 314–323.
9. T. Arikawa, M. Nagai and K. Tanaka, Chem. Phys. Lett., 2008, 457, 12–17.
10. U. Moller, D. G. Cooke, K. Tanaka and P. U. Jepsen, J. Opt. Soc. Am. B, 2009, 26, A113–A125.
11. J. Kindt and C. Schmuttenmaer, J. Phys. Chem., 1996, 100, 10373–10379.
12. T. Globus, D. Woolard, T. Crowe, T. Khromova, B. Gelmont and J. Helser, J. Phys. D: Appl. Phys., 2006, 39, 3405–3413.
13. J. Xu, K. Plaxco and S. Allen, Protein Sci., 2006, 15, 1175–1181.
14. C. C. Cooksey, B. J. Greer and E. J. Heilweil, Chem. Phys. Lett., 2009, 467, 424–429.
15. V. Mille, N. Bourzgui, F. Medjdoub and B. Bocquet, European Microwave Conference, 2004, 1, 169–172.
16. P. A. George, W. Hui, F. Rana, B. G. Hawkins, A. E. Smith and B. J. Kirby, Opt. Express, 2008, 16, 1577–1582.
17. V. Matvejev, C. De Tandt, W. Ranson, J. Stiens, R. Vounckx and D. Mangelings, Prog. Electromagn. Res., 2011, 121, 89–101.
18. P. W. Fenimore, H. Frauenfelder, B. H. McMahon and R. D. Young, Proc. Natl. Acad. Sci. U. S. A., 2004, 101, 14408–14413.
19. H. Frauenfelder, G. Chen, J. Berendzen, P. W. Fenimore, H. Jansson, B. H. McMahon, I. R. Stroe, J. Swenson and R. D. Young, Proc. Natl. Acad. Sci. U. S. A., 2009, 106, 5129–5134.
20. N. Jain and I. Roy, Protein Science, 2009, 18, 24–36.
21. A. Lerbret, F. Affouard, P. Bordat, A. Hedoux, Y. Guinet and M. Descamps, J. Non-Cryst. Solids, 2011, 357, 695–699.
22. M. Heyden and M. Havenith, Methods, 2010, 52, 74–83.
23. N. Vinh, S. Allen and K. Plaxco, J. Am. Chem. Soc., 2011, 133, 8942–8947.
24. B. Jana, S. Pal and B. Bagchi, J. Chem. Sci., 2012, 124, 317–325.
25. Y. I. Khurgin, V. A. Kudryashova and V. A. Zavizion, Radiotekhnika I Elektronika, 1996, 41, 737–743.
26. S. P. Mickan, J. Dordick, J. Munch, D. Abbott and X. C. Zhang, Biomedical Applications of Micro- and Nanoengineering, 2002, 4937, 49–61.
27. J. Kitagawa, T. Ohkubo, M. Onuma and Y. Kadoya, Appl. Phys. Lett., 2006, 89, 041114.
28. J. Xu, K. Plaxco and S. Allen, J. Phys. Chem. B, 2006, 110, 24255–24259.
29. S. Pal, J. Peon and A. Zewail, Proc. Natl. Acad. Sci. U. S. A., 2002, 99, 1763–1768.
30. U. Heugen, G. Schwaab, E. Brundermann, M. Heyden, X. Yu, D. M. Leitner and M. Havenith, Proc. Natl. Acad. Sci. U. S. A., 2006, 103, 12301–12306.
31. B. Born and M. Havenith, Journal of Infrared Millimeter and Terahertz Waves, 2009, 30, 1245–1254.
32. S. Laurette, A. Treizebre and B. Bocquet, J. Micromech. Microeng., 2011, 21, 065029.
33. A. Treizebre and B. Bocquet, Int. J. Nanotechnol., 2008, 5, 784–795.
34. A. Treizebre, B. Bocquet, Y. S. Xu and R. G. Bosisio, Microwave Opt. Technol. Lett., 2008, 50, 2998–3001.
35. L. Dazhang, J. Cunningham, M. Byrne, S. Khanna, C. Wood, A. Burnett, S. Ershad, E. Linfield and A. Davies, Applied Physics Letters 95, year.
36. D. Gacemi, J. Mangeney, T. Laurtent, J. Lampin, T. Akalin, K. Blary, A. Degiron, P. Crozat and F. Meng, Opt. Express, 2012, 20, 8466–8471.
37. S. Laurette, A. Treizebre, N. Bourzgui and B. Bocquet, Prog. Electromagn. Res. Lett., 2012, 30, 49–58.
38. A. Abbas, A. Treizebre, P. Supiot, N. E. Bourzgui, D. Guillochon, D. Vercaigne-Marko and B. Bocquet, Biosens. Bioelectron., 2009, 25, 154–160.
39. S. Laurette, A. Treizebre, F. Affouard and B. Bocquet, Appl. Phys. Lett., 2010, 97, 111904.
40. P. U. Jepsen, U. Moller, F. Eichhorn, H. Merbold, J. R. Folkenberg and S. J. Clark, 2007 Conference On Lasers & Electro-optics/quantum Electronics and Laser Science Conference, 2007, 1–5, 1740–1741.
41. S. Ebbinghaus, S. J. Kim, M. Heyden, X. Yu, U. Heugen, M. Gruebele, D. M. Leitner and M. Havenith, Proc. Natl. Acad. Sci. U. S. A., 2007, 104, 20749–20752.
42. K. Gekko and H. Noguchi, J. Phys. Chem., 1979, 83, 2706–2714.
43. M. Suzuki, J. Shigematsu and T. Kodama, J. Phys. Chem., 1996, 100, 7279–7282.
44. T. Yoshida, H. Inagaki, H. Kamya and I. Ueda, J. Phys. Chem. B, 2001, 105, 256–260.
45. K. Yokoyama, T. Kamei, H. Minami and M. Suzuki, J. Phys. Chem. B, 2001, 105, 12622–12627.
46. T. Sato and R. Buchner, J. Chem. Phys., 2003, 118, 4606–4613.
47. T. Sato and R. Buchner, J. Phys. Chem. A, 2004, 108, 5007–5015.
48. J. Barthel and R. Buchner, Pure Appl. Chem., 1991, 63, 1473–1482.
49. N. Gracia, S. Thomas, F. Thibault-Starzyk, O. Lerasle and L. Duponchel, Chemom. Intell. Lab. Syst., 2011, 106, 210–215.
50. C. Ruckebusch, A. D. Juan, L. Duponchel and J. Huvenne, Chemom. Intell. Lab. Syst., 2006, 80, 209–214.
51. C. Ruckebusch, L. Duponchel, J. P. Huvenne and J. Saurina, Anal. Chim. Acta, 2004, 515, 183–190.
52. E. Malinowski, Factor analysis in chemistry, Wiley-Interscience, New York, 2nd edn, 1992.
53. R. Tauler, Chemom. Intell. Lab. Syst., 1995, 30, 133–146.
54. A. de Juan and R. Tauler, Anal. Chim. Acta, 2003, 500, 195–210.
55. S. Navea, A. de Juan and R. Tauler, Anal. Chem., 2003, 75, 5592–5601.
56. L. Duponchel et al., Analytica Chimica Acta, submitted.
57. W. Smith, C. W. Yong and P. M. Rodger, Molecular Simulation, 2002, 28, 385–471.
58. W. L. Jorgensen, D. S. Maxwell and J. Tirado-Rives, J. Am. Chem. Soc., 1996, 118, 11225–11236.
59. J. L. F. Abascal and C. Vega, J. Chem. Phys., 2005, 123, 234505.
60. F. Affouard et al., Journal of Physical Chemistry B., submitted.
61. S. Laurette, A. Treizebre and B. Bocquet, IEEE Trans. Terahertz Sci. Technol., 2012, 2, 340–344.
62. S. Ebbinghaus, S. J. Kim, M. Heyden, X. Yu, M. Gruebele, D. M. Leitner and M. Havenith, J. Am. Chem. Soc., 2008, 130, 2374–2375.

---