Quantitative determination of microbial materials activity based on infrared extinction properties

Wanying Ding; Wanying Ding; Youlin Gu; Youlin Gu; Youlin Gu; Yihua Hu; Yihua Hu; Yihua Hu; Hao Cao; Hao Cao; Guolong Chen; Guolong Chen; Haihao He; Haihao He

doi:10.1364/OE.496673

1. Introduction

With increasing awareness of environmental protection and food safety, microbial materials with spectroscopic sterilization and active food fermentation are attracting increasing attention. Biocontrol fungi such as Candida albicans and Candida aeruginosa are widely used in forestry pest control [1–3]. Aspergillus niger and other fungi are widely found in food, plant products and soil, and are significant industrial fermentation strains that can be used to produce enzyme preparations, organic acids, saccharified feed, etc [4,5]. Biomaterials can also be used as extinction smoke screen materials to weaken infrared detection devices such as infrared detectors [6]. P.S Tuminello et al. calculated the transmittance based on instrumental measurement of the energy attenuation of light passing through Bacillus subtilis solution and obtained the attenuation properties of Bacillus subtilis in the 0.2-2.5 µm band using the Lambert-Beer law [7]. Hu et al. [8,9] investigated the extinction mechanism of bioaerosols using smoke box experiments and explored the potential of bioaerosols as new extinction materials. Microorganisms are vital to human life and nature. Therefore, the determination of activity during the production, storage, and application of microorganisms is crucial.

Several studies have been conducted on the activity of microbial materials. The plate counting method is widely used as a standard method; however; it has certain limitations; that is, it is time- and effort-consuming and has low specificity [10]. With the development of molecular biology, polymerase chain reaction (PCR) and quantitative polymerase chain reaction (qPCR) have become common techniques for microbial detection, significantly improving the efficiency of microbial activity detection [11,12]. Specific primers were designed to extract microbial cell genomic DNA for the quantitative detection of viable bacteria [13].

However, these methods can only be used to assess the quality of a batch of material using destructive representative sample. Spectroscopic analysis technique is a non-invasive and non-destructive batch testing method for samples, which is commonly used to detect the content of compounds such as specific proteins and nucleic acids in samples. For example, a chemometric method using terahertz spectroscopy has been used to quantify the protein content in soybeans [14]. The quantitative analysis of imidacloprid in rice flour samples was successfully achieved using transmission-mode terahertz time-domain spectroscopy [15]. Fourier-transform infrared spectroscopy (FTIR)-based spectroscopy combined with machine learning was used to assess the degree of spoilage on the surface of chicken breasts [16], and most of the above studies were aimed at detecting the content of certain substances such as proteins, sugars, and other compounds in the samples.

Previously, our group investigated the extinction properties of living and dead bioparticles and found that active samples have stronger extinction properties, thus enabling qualitative analysis of dead and viable biomaterials [17]. However, in practical applications, qualitative analysis alone is not sufficient; therefore, this study is a follow-up of a previous study on the quantitative analysis of microbial material activity.

In summary, the studies mentioned above demonstrate that infrared spectroscopy-based analytical methods can provide sufficient information for researchers to develop predictive models for various product properties. However, studies applying infrared spectroscopy to determine the activity of microbial materials, particularly for quantitative determination, remain limited. In this study, the least angle regression (LAR) variable screening method was selected for wavelength selection, and a quantitative determination model was established by combining several regression methods. Finally, a rapid, accurate, nondestructive, and noninvasive quantitative assay of microbial activity was achieved.

2. Materials and method

The main approach of the article analysis is explained in the section. This study is summarized in the following parts: biomaterial preparation, IR spectroscopy acquisition, data processing, and quantitative determination. The detailed experimental design and modeling process of this study are presented in Fig. 1.

Fig. 1. Flowchart of experimental design and modeling process.

Download Full Size | PDF

2.1 Biomaterials preparation

Three microbial materials, AH0302, BM0302 and AP0302 were obtained from the Key Laboratory of Ion Beam Bioengineering at the Chinese Academy of Sciences. They were prepared by bacterial species activation, shake flask culture, large-scale tank fermentation, centrifugation, pure water cleaning, vacuum freeze-drying, and crushing using an ultra-fine Chinese medicine crusher. They were then divided into two parts and stored in a drying dish as a backup. One portion was inactivated, whereas the other remained active without treatment. In this study, a single spore exhibited only two states: dead and viable. Thus, for large batches of microbial material, the activity refers to the ratio of the number of viable spores to the total number of spores.

The material was deactivated by placing it in an autoclave (ZEALWAY GIDP) at 103.4 kPa vapor pressure and 121° high temperature for 30 min. It was then removed and placed in a desiccator and dried at 80° until the sample mass remained at a constant weight. At this point, inactivation was completed, and a completely dead material was obtained.

Proportional mixing was used to acquire samples with different activities, and 11 different viable samples were configured in the range of 0-100% in a gradient of 10%. The materials were weighed using an electronic balance and stirred using an LC-DMS-H magnetic mixer to ensure uniform mixing of the dead and viable materials. Pressed tablets (0.1 g samples) with different activities were prepared using an infrared press with a pressure of 4 tons and maintained for 15 s. Consequently, pressed tablets with a diameter of 10 mm and a thickness of 2 mm (±0.1 mm) were obtained.

2.2 IR spectroscopy acquisition

Microscopic infrared spectra of the sample materials were measured using a BRUKER TENSOR 37 infrared spectrometer equipped with a HYPERION 2000 FT-IR infrared microscope. The reflectance spectra of the pressed sheets were collected in the 4000-600 cm^-1 band. The resolution was 2 cm^-1 and an average value of 128 scans was selected. A gold-plated reflector surface was used as the backing, the angle of incidence was 18°. Five sampling points were selected for reflection spectroscopy for each sample press sheet, and the average of three measurements was selected for each point tested. Compared with Fourier spectroscopy, the spot size of the microscopic infrared spectrometer was 100 µm × 100 µm, the surface of the pressed sheet was smoother and flatter in a very small spot-size range, and the measured reflectance was closer to the ideal reflectance spectrum.

2.3 Kramers-Kronig relation

Based on the measured infrared spectrum, the complex refractive indices (CRI) can be calculated from the Kramers-Kronig relationship [18,19]. For a vertically incident electromagnetic wave, the reflected phase shift of the material is calculated as [20]

(1)$$\Theta (\lambda )= \frac{\lambda }{\pi }P\int_0^\infty {\frac{{\ln R({{\lambda^{\prime}}} )}}{{({{\lambda^{^{\prime}2}} - {\lambda^2}} )}}d{\lambda ^{\prime}}}. $$

where λ is the wavelength of the reflected wave, R(λ) is the vertical incident reflectance of the material, and P is the Cauchy principal value function. In practical measurements, the light was not strictly perpendicular to the microbial material sample at an 18° oblique incidence. However, analysis shows that an angle of incidence of 18° has minimal effect, and the difference between it and the specular reflectance obtained by ideal perpendicular incidence is in the range of 10⁻⁴-10⁻⁵ [17,21].

Eq. (1) shows that the full-band reflectance of the spectrum needs to be measured. Moreover, in the actual experimental measurement, as the working band of the test instrument does not cover the full (0∼∞), then the reflectance outside the test band needs to be obtained by extrapolation of the empirical formula or constant extrapolation [22,23].

(2)$$\Theta (\lambda )= \frac{\lambda }{\pi }P\left( {\int_0^{{\lambda_\textrm{a}}} {\frac{{\ln R({{\lambda^{\prime}}} )}}{{({{\lambda^{^{\prime}2}} - {\lambda^2}} )}}d{\lambda^{\prime}}} + \int_{{\lambda_\textrm{a}}}^{{\lambda_\textrm{b}}} {\frac{{\ln R({{\lambda^{\prime}}} )}}{{({{\lambda^{^{\prime}2}} - {\lambda^2}} )}}d{\lambda^{\prime}}} + \int_{{\lambda_\textrm{b}}}^\infty {\frac{{\ln R({{\lambda^{\prime}}} )}}{{({{\lambda^{^{\prime}2}} - {\lambda^2}} )}}d{\lambda^{\prime}}} } \right)$$

Assuming that the complex refractive index of the material is m(λ)=n(λ)+ik(λ), the real part n(λ) and imaginary part k(λ) of the complex refractive index can be calculated using Eqs. (3) and (4) [20].

(3)$$n(\lambda )= \frac{{1 - R(\lambda )}}{{1 + R(\lambda )+ 2\sqrt {R(\lambda )} \cos \Theta (\lambda )}}$$

(4)$$k(\lambda )= \frac{{ - 2\sqrt {R(\lambda )} \sin \Theta (\lambda )}}{{1 + R(\lambda )+ 2\sqrt {R(\lambda )} \cos \Theta (\lambda )}}$$

2.4 Discrete dipole approximation method

Discrete dipole approximation (DDA) [24–27] is a method for solving electromagnetic waves scattered by small irregular particles. An array of a large number of dipoles is used to imitate a continuous object. The properties of the absorbed and scattered electromagnetic waves of the target were then obtained by solving the polarization of these dipoles under irradiation by the incident electromagnetic waves. DDSCAT is a tool for performing calculations based on the DDA method. The CRI were introduced in the previous section by reflectivity, and its extinction efficiency factor parameter is obtained by complex refractive index calculation.

Assuming the particle volume is V, then the equivalent radius is expressed as ${R_{eff}} = \sqrt[3]{{3V/4\pi }}$.

The extinction coefficient (absorption coefficient and scattering coefficient) were obtained using the vectors defined in DDSCAT. The extinction cross section, absorption cross section and scattering cross section of a material can be calculated by the following equations [24,25]:

(5)$${C_{ext}} = \frac{{4\pi k}}{{{{|{\overrightarrow {{E_0}} } |}^2}}}\sum\limits_{i = 1}^M {{\mathop{\rm Im}\nolimits} (E_{inc,i}^ \ast{\cdot} {P_i})}$$

(6)$${C_{abs}} = \frac{{4\pi k}}{{{{|{\overrightarrow {{E_0}} } |}^2}}}\sum\limits_{i = 1}^M {\left\{ {{\mathop{\rm Im}\nolimits} [{{P_i} \cdot {{(\alpha_i^{ - 1})}^ \ast } \cdot {P_i}^ \ast } ]- \frac{2}{3}{k^3}{{|{{P_i}} |}^2}} \right\}}$$

(7)$${C_{sca}} = {C_{ext}} - {C_{abs}} = \frac{{{k^4}}}{{{{|{{E_{inc}}} |}^2}}}\int {d\Omega \left[ {\sum\limits_{i = 1}^M {[{{P_i} - n(n \cdot {P_i})} ]\exp ( - ikn \cdot {r_i})} } \right]}$$

Where E₀ is the incident electric field, $\lambda$ is the wavelength of the incident electromagnetic wave, $k\textrm{ = }{{2\pi } / \lambda }$, is the cubic angle micrometric element, $n$ is the direction vector of the electromagnetic wave scattering, M is the number of dipoles, E is the dipole moment of the i-th dipole, α_i is the polarizability of the i-th dipole, and ${E_{inc,i}}$ is the intensity of the incident electric field.

Extinction efficiency factor (Q_ext) is defined as the ratio of the extinction cross section of the particle to the geometric cross section of the particle. The same is true for absorption and scattering.Then the efficiency factor is expressed as [24]:

(8)$${Q_{ext}} = \frac{{{C_{ext}}}}{{\pi R_{eff}^2}}$$

(9)$${Q_{abs}} = \frac{{{C_{abs}}}}{{\pi R_{eff}^2}}$$

(10)$${Q_{sca}} = \frac{{{C_{sca}}}}{{\pi R_{eff}^2}}$$

2.5 Quantitative activity determination model

2.5.1 Characteristic wavelength screening

The use of the entire spectral combination, that is, the spectral intensities of all wave numbers directly as input data, sometimes leads to poor correlation performance. To obtain optimal spectral variables and better models, variable selection must be considered and redundant information must be removed. LAR is an iterative algorithm for linear regression problems with fast feature selection and regression coefficient calculation, and is widely used to solve linear regression and Lasso regression problems. The core idea of the LAR algorithm is that the regression target vector is decomposed into linear combinations of several sets of feature vectors, such that the residual vector, which is linearly independent of all features, is minimized [28,29].

LAR selects useful variables by setting the coefficients of irrelevant variables to 0. The model is as follows:

(11)$$\begin{array}{l} \begin{array}{*{20}{c}} {minS\left( {\hat{\beta }} \right) = y - {{\hat{\mu }}^2} = \mathop \sum \limits_{i = 1}^n {{\left( {{y_i} - {{\hat{\mu }}_i}} \right)}^2} = \mathop \sum \limits_{i = 1}^n {{\left( {{y_i} - \mathop \sum \limits_{j = 1}^p {x_{ij}}{\beta _j}} \right)}^2}} \end{array}\\ s.t.\sum\limits_{j = 1}^p {\left| {{\beta _j}} \right| \le t} \end{array}$$

where (x_i₁, x_i₂, …, x_ip) is the wavelength of sample i, y_i is the response of the sample, and β_j is the coefficient of the j-th wavelength of sample i with t ≥ 0 constraint value. At t→0, the LAR algorithm filters the wavelength that best characterizes the properties of sample i by setting the coefficient β_j of the wavelength that has a smaller effect on sample i to 0.

2.5.2 Regression model

Partial least squares (PLS) is a method for multivariate statistical data analysis [30,31]. This method obtains the mutually orthogonal eigenvectors of the independent and dependent variables by projecting the high-dimensional data space of the independent and dependent variables into the corresponding low-dimensional space and then establishes a univariate linear regression relationship between the eigenvectors of the independent and dependent variables. It overcomes the covariance problem, and emphasizes the explanatory and predictive role of the independent variables on the dependent variable in the selection of eigenvectors. Thus, the effect of useless noise on the regression is eliminated, and the model contains the minimum number of variables.

Support vector regression (SVR) is a machine-learning algorithm based on the statistical learning theory maximizing generalization capabilities using the principle of structural risk minimization. SVR introduces loss functions to solve regression problems based on support vector machine (SVM) classification. It is suitable for addressing both linear and nonlinear problems with small sample size and multiple variables. The algorithm offers the advantages of strong learning ability, strong anti-noise ability, strong generalization ability, and fast operation speed.

Extreme learning machine (ELM) is a single hidden layer feedforward neural network learning algorithm [32]. It significantly improves the generalization ability of the network model and increases the training speed. The output weight coefficient variables of the network model can be obtained in a single step. The ELM network model is divided into three layers for training and learning: input, hidden, and output layers [33].

2.6 Software and configuration

DDSCAT 7.3 was used for the extinction efficiency factor calculation. All operations generated by spectral processing, CRI calculation, model building, cross-validation and plotting were performed using Python 3.8.6 environment. It was performed with a PC (CPU 2.60 GHz, RAM 32.0 GB).

3. Result and discussion

3.1 Raw spectrum

The acquired raw spectrum of the AH0302, BM0302 and AP0302 were shown in Fig. 2, which is contained 55 spectral curves, respectively.

Fig. 2. Raw spectrum of three materials samples. (a) AH0302. (b) BM0302. (c) AP0302.

Download Full Size | PDF

These three types of materials have almost the same peak and valley regions, which are related to their properties of the materials themselves. Differentiation of the activity spectra was the largest for AH0302 among the three, followed by BM0302, and the smallest for AP0302. Therefore, it can be inferred that the number of selected wavelengths in the subsequent modeling should be related to it. These materials comprise numerous cells whose main constituents are water, proteins, lipids, and nucleic acids. Water molecules showed significant uptake at 3000-3750 cm^-1 and 1600-1700cm^-1.

Because the tablet surface is not a perfectly flat specular surface, the selection of different test points may lead to differences in the spectra. In infrared spectroscopy, the absorption band below 1200 cm^-1 is known as the fingerprint region and corresponds to the low-frequency vibrational characteristic peak of the sample. This is caused by a combination of multiple motions such as the twisting and oscillation of multiple chemical bonds, which is key for the identification of simple small-molecule compounds. However, this region does not provide much information because the samples are mixtures with complex composition of characteristic peaks. Thus, the difference shows a more confusing spectrum in 1000-600 cm^-1, and the high wavelength band would be ignored in the subsequent calculations.

For clearer observation, the dead (0%) and viability (100%) spectra of the three materials were selected for the conversion analysis of their absorptions, as shown in Fig. 3. The structural changes before and after the inactivation of AH0302 were significant. For example, the regular displacement of the characteristic absorption peak of membrane lipids at 2927 cm^-1 could reflect the stacking characteristics of cell membrane lipid hypomethyl chains and the ordered or disordered nature of lipid bilayers. The relative intensity of AH0302 samples decreased significantly and shifted to the right after inactivation, indicating that the inactivation process disrupted the structure of membrane lipids.

Fig. 3. Comparison of IR spectrum of three materials before and after inactivation

Download Full Size | PDF

In addition, the characteristic absorption peaks of the sample at 1760cm^-1 and the amide I band were significantly shifted. The absorption peaks primarily originated from C = O structures such as carbonyl carboxyl groups, suggesting that the sample structure was oxidized by the inactivation process. The change in the amide band may have been caused by the instantaneous change in the dipole ion due to the hydration of the protein, indicating that the protein was denatured during the treatment. This affected the hydrogen bonding structure within the protein molecule, thus changed its secondary structure composition.

The sample ehibited a significant increase in the relative intensity of the absorption peak at 1121 cm^-1, which is mainly attributed to the antisymmetric stretching and symmetric stretching vibrations of the nucleic acid phosphodiester backbone. The apparent change in the peak shape indicated a high degree of denaturation of the nucleic acid structure by inactivation or changes in the membrane lipid structure lead to full exposure of the internal phospholipid molecules.

These changes indicate that AH0302 is more sensitive to the inactivation process, while the other two materials are not. It is because the functional group of AH0302 changes more significantly during the inactivation process. This was also observed in case of the more pronounced changes in the reflectance spectrum of AH0302 in Fig. 2.

3.2 CRI calculation

Infrared spectra contain irrelevant information and noise, so data smoothing pre-processing is required before calculating the optical constants. Then it is used in the Kramers-Kronig transform for subsequent calculations. Figure 4 shows the CRI results for different active biomaterials based on the Kramers-Kronig relationship.

Fig. 4. CRI values with different activities of the three materials. (a) The n value of AH0302. (b) The n value of BM0302. (c) The n value of AP0302. (d) The k value of AH0302. (e) The k value of BM0302. (f) The k value of AP0302.

Download Full Size | PDF

The real part n of the CRI indicates the dispersion of the medium to the electromagnetic waves, which is related to the wavelength and properties of the absorbing medium. The n varies in the range of 1.16-1.34 for microbial materials. The imaginary part k indicates the absorption of electromagnetic waves by the medium, which is determined by the attenuation rate of the wave propagation in the absorbing medium. The k value of the microbial materials varies as 0.15-0.34.

The CRI trends of the three materials are consistent; however, the value of AH0302 was significantly lower than those of the other two. The curve of AH0302 has the greatest difference and that of AP0302 has the least difference, which is also consistent with what was inferred in Section 3.1. The difference in their reflectance spectra was directly reflected in the value of the CRI. It may be inferred that the above conclusion is valid for the extinction characteristic results.

3.3 Extinction properties calculation

The corresponding structural model was established, and the other parameters were set according to the SEM of the bioparticles, as shown in Fig. 5.

Fig. 5. Structure models of three kinds of microbial agglomerated particles. (a) SEM of AH0302. (b) Particles aggregation model of AH0302. (c) SEM of BM0302. (d) Particles aggregation model of BM0302. (c) SEM of AP0302. (d) Particles aggregation model of AP0302.

Download Full Size | PDF

The wavelengths were chosen from 2.5-8.5 µm, and the CRI were taken from the above calculation. The detailed input parameters for the calculations are listed in Table 1; where the R_p refers to the particle size distribution in the model, and R_eff refers to the equivalent radius of the aggregated particles as input in the DDA calculation.

Table 1. DDA input parameters

View Table | View all tables in this article

The extinction efficiency factors of different materials were obtained using the DDA method. The extinction properties include the absorption and scattering properties, which represent the ability of the material to attenuate electromagnetic waves. The results are shown in Fig. 6.

Fig. 6. Extinction efficiency factor with different activities of the three materials. (a)Q_ext of AH0302. (b) Q_ext of BM0302. (c) Q_ext of AP0302. (d)∼(l) Details of Q_ext at some wavelengths for the three materials.

Download Full Size | PDF

Figures 6(a)-(c) show the overall extinction curves of the three materials. In the 2.5-8.5 µm band, the extinction properties of AH0302 were significantly lower than those of the other two materials. This could be owing to variations in the substances contained in the cells of the different microbial materials themselves. AP0302 exhibited a smaller distinction, implying that it is more difficult to distinguish it visually. However, the turning point in the Q_ext curves of all three materials occurred in the 3-3.5 µm band, which is because of the turning of the imaginary part k value in the CRI here.

Figure 7 shows the derivatives of the extinction efficiency factors, where the vertical coordinate is represented by D1 and the black dashed line represents the vicinity of the turning point. On the left side of the dashed line, biomaterials with lower activity exhibited slower changes in extinction characteristics in this wavelength band. On the right side, participants with stronger activity underwent faster changes. It is speculated that the extinction properties may be reversed owing to the presence of absorption or scattering extremes in different proteins, nucleic acids, and other organelles in the cells of different activity biomaterials in the 3-3.5 µm band.

Fig. 7. Derivative results of extinction efficiency factor. (a) The derivatives of AH0302 Q_ext. (b) The derivatives of BM0302 Q_ext. (c) The derivatives of AP0302 Q_ext.

Download Full Size | PDF

Figures 6(d)∼(l) showed the details of Q_ext curves at important wavelength points for the three materials after wavelength selection (details are shown in Fig. 7, which represents wavelength screening; certain characteristic wavelength points are marked with black arrows in Figs. 6 (a)-(c)). The same color represents the same material, the color from dark to light represents the activity from 0 to 100%, and the colored arrows indicate the direction of gradual increase in activity. As evident, the extinction coefficient of biomaterials in the 2.5-3 µm band decreased with the increase of activity; however, in the 3-8.5 µm band the extinction coefficient was most enhanced with the increase of activity. Therefore, biomaterials, such as smoke screen materials, should maintain high activity in the infrared window band to achieve a good effect by obscuring infrared detection equipment and systems.

However, because of the complexity of the extinction properties at different wavelengths, although certain of the above results show certain trends, practical evidence to directly determine the activity is lacking. Therefore, it is necessary to test and validate a data-set by combining machine-learning methods.

3.4 Quantitative activity determination results

A 3:1:1 ratio was used to randomly divide their respective reference values into a training set (33 elements), a calibration set (11 elements), and a prediction set (11 elements). The LAR algorithm was then applied to select the characteristic wavelengths from all pre-processed data to build the prediction models discussed below.

The first step was to select the characteristic wavelengths. Because of the large number of data points in the spectral collection and a large amount of data, not all wavelength points can provide valid information. Therefore, screening redundant data is extremely important. The characteristic wavelength points were selected using the LAR algorithm, and 10, 14, and 43 characteristic points were selected for the three materials. The modeling process for the wavelength selection is as follows:

Assuming that the spectral data of sample i are X = { x_i₁, x_i₂, …, x_il}, the specific process of the LAR screening of spectral wavelengths is as follows.

(1) Solving the regression variable matrix β_j
The spectral data X of the samples were input into the LAR model and the regression coefficients were solved following the minimization AIC principle to construct the regression variable matrix. In this matrix, 1 indicates that the wavelength is highly correlated with sample i and 0 indicates a low correlation.
(2) Screening variables
The LAR model uses a matrix of regression variables to solve for the characteristic wavelengths of the sample spectra.

(12)$$\widehat X = X\beta$$

where $\widehat X$ is the wavelength filtered by the LAR model, X represents the raw spectral data of the sample, and $\beta = [{{\beta_1},{\beta_2}, \cdots ,{\beta_l}} ]$ is a regression-variable matrix.

The distribution of the characteristic wavelength points is shown in Fig. 8. The number of wavelength points verifies the previous inference that for a curve group with considerable variability, a smaller number of characteristic wavelengths may be selected to achieve the goal. However, for a material such as AP0302, 43 characteristic wavelengths were selected to build a better build the regression model. It is speculated that this is strongly related to the difference in absorption or scattering intensity owing to the internal properties and external structure of the material; however, the exact meaning of these distributions requires futher in-depth study.

Fig. 8. Characteristic wavelength screening. (a) AH0302. (b) BM0302. (c) AP0302.

Download Full Size | PDF

The performance of the developed model was evaluated using three performance metrics: the coefficient of determination (R²) [34], the Root Mean Square Error (RMSE) [35], and the Mean Absolute Error (MAE) [36]. R² shows the link between the actual results and the predictions of the quality parameters. The efficiency of the model can be determined using the RMSE and MAE. In general, a good model should produce high R² values but low RMSE and MAE values.

The data-set was subsequently trained and predicted using several regression models. The PLS, SVR, and ELM models were used in this study. The sensitivity analysis of the main hyperparameters for each model is presented below.

The effects of different hyperparameters on the performance of the prediction models for the three materials are compared in Figs. 9,10, and 11, respectively. Figures 9 (a)∼(c), Figs. 10 (a)∼(c), and Figs. 11 (a)∼(c) demonstrate the effect of the number of principal components of the PLS model on the results, where 3,5,7 principal components were selected for each of the three materials according to the model evaluation criteria described above. It can be observed that AH0302 and BM0302 are less sensitive to changes in the principal components than AP0302.

Fig. 9. Performance comparison with different model hyperparameters for AH0302. (a)∼(c) PLS: Components. (d)∼(f) SVR: C. (g) SVR: Kernels. (h)∼(j) ELM: Hidden neurons. (k)∼(m) ELM: Activation function.

Download Full Size | PDF

Fig. 10. Performance comparison with different model hyperparameters for BM0302. (a)∼(c) PLS: Components. (d)∼(f) SVR: C. (g) SVR: Kernels. (h)∼(j) ELM: Hidden neurons. (k)∼(m) ELM: Activation function.

Download Full Size | PDF

Fig. 11. Performance comparison with different model hyperparameters for AP0302. (a)∼(c) PLS: Components. (d)∼(f) SVR: C. (g) SVR: Kernels. (h)∼(j) ELM: Hidden neurons. (k)∼(m) ELM: Activation function.

Download Full Size | PDF

Figures 9 (d)∼(g), Figs. 10 (d)∼(g), and Figs. 11 (d)∼(g) show the analysis of the two hyperparameters C and Kernels of SVR, respectively, and it is clearly seen that the choice of hyperparameters for the SVR model has significant impact on the results. The optimal values for the parameter C are 50, 10, and 50, while for the choice of Kernels, only the “linear” shows the desired results.

Figures 9 (h)∼(m), Figs. 10 (h)∼(m), and Figs. 11 (h)∼(m) represent the selection process of hyperparameters for the ELM model, with the number of neurons in hidden layer being 10,15,10 for the three materials, and “sigma” was selected for the activation function. As can be seen in subgraphs (h)∼(j), the variation in the number of neurons in hidden layer is very sensitive to the sample prediction results and the three materials show different trends, which is different from the function selection in subgraphs (k)∼(m).

Through the sensitivity analysis of hyperparameters, the main optimal hyperparameters for the different models of the three groups of samples were obtained as shown in Table. 2.

Table 2. Hyperparameters of different models for three materials

View Table | View all tables in this article

The results of the model evaluation parameters are presented in Table 3. Among them, the LAR-PLS model showed the best prediction results for determining the activity of all three materials, with the largest R² value and smallest RMSE and MAE values.

Table 3. Results of model evaluation parameters

View Table | View all tables in this article

The results of these models are shown in Fig. 12 and Fig. 13, where the horizontal axis represents the actual activity values of the biomaterials and the vertical axis represents the predicted activity values of the biomaterials. The points of different colors in the graph represent different biomaterials. Each point on the black dashed line indicates a completely accurate prediction, whereas a point farther away from the dashed line indicates a poor prediction. The dashed lines in different colors represent the corresponding fitting curves.

Fig. 12. Calibration set results of model predictions for three materials. (a)∼(c) AH0302. (d)∼(f) BM0302. (g)∼(i) AP0302.

Download Full Size | PDF

Fig. 13. Prediction set results of model predictions for three materials. (a)∼(c) AH0302. (d)∼(f) BM0302. (g)∼(i) AP0302

Download Full Size | PDF

In addition to the LAR-PLS model, which predicted the best results, the figure also shows that the other two models also differed. It showed better results for the ELM model and worse results for SVR for all three materials. The distribution of points shows that the ELM and SVR models predict approximately the same range. However, the prediction points of the SVR model occasionally have large deviations, resulting in poorer overall prediction results. The results in Section 3.3 show that the effect of 10% activity on the extinction performance is small, thus, the prediction results of this model are acceptable.

The errors in the results obtained after changing the division of the training set for several trials are shown in Fig. 14. These results are within the error range and represent the prediction results of the model. In addition, the error results show the stability of the LAR-PLS model, which is the best and most stable of the three models in terms of prediction results.

Fig. 14. The values of R², RMSE and MAE errors. (a)∼(c) AH0302. (d)∼(f) BM0302. (g)∼(i) AP0302.

Download Full Size | PDF

Based on the results mentioned above, combined with those of previous studies [17,37], the reasons for the variability in extinction characteristics due to the activity of microbial materials are explained as follows.

(1) In the process of inactivation and drying of microbial materials, the internal water content of microbial cells gradually decreases, which leads to the absorption of dead cells in the infrared band is greatly reduced. This can be inferred from the phenomenon that the value of extinction efficiency factor of the material decreases after inactivation.
(2) All three samples showed a significant decrease in relative polysaccharides content after inactivation, indicating that the inactivation process affected the peptidoglycan layer of the cell walls.
(3) The amide I bands of all three samples were significantly shifted to the long wavelength after inactivation, indicating that the secondary structure of the sample protein was disrupted. However, the specific degree of disruption and the changing pattern were different among the samples, which was closely related to the protein type and stability of the phosphoplasmic layer structure. After the inactivation of microbial materials, the internal chemical structure changes significantly; for example, the structure of proteins, DNA nucleic acids changes from an order, curly and compact structure to a disorder, loose, and sparse structure or the position of genes such as hydrophobic proteins is shifted from the interior of the cell to the surface of the molecule, improving the absorption performance of the corresponding band.

4. Conclusion

In this study, the activity of biomaterials was quantified by calculating their extinction properties using infrared spectroscopy and modeling its correlation with the activity of biomaterials. Certain results in this study are consistent with the previous results [17], such as the values of infrared spectra and CRI in the 3∼5 µm band, and the intensity of extinction properties and activity remain consistent in the same waveband. This was a follow-up of a previous study [17].

The extinction property is an important optical parameter that can be measured experimentally. Biomaterials used as smoke screen materials exhibit broadband extinction capability [8,9]. Whether based on spectral theoretical derivation or direct measurement of the extinction performance of the material when diffusing in air, the corresponding models can be developed and the material activity can be predicted according to the method in this study. Later, the transmittance, concentration, and shading area of different active materials in the process of diffusion and settling in air through smokescreen box experiments or field tests will be measured and the corresponding relationship models will be established. The innovation of this study lies in the proposal of a measurable and calibrated parameter for the quantitative determination of the activity of microbial materials, rather than being limited to the analysis of their raw spectra.

The study suggested that a rapid and accurate quantitative determination of microbial activity can be achieved based on extinction properties. It is also valuable to study the properties of microbial materials for electromagnetic attenuation. In future work, we will focus on exploring the relationship between absorption, scattering, and polarization properties and microbial material activity, and further develop the field of quantitative detection of microbial activity using optical properties.

Funding

Advanced Laser Technology Laboratory Foundation of Anhui Province of China (20191003); National Natural Science Foundation of China (62075241).

Disclosures

The authors declare no conflicts of interest.

Data availability

Data availability. The data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.

Reference

1. L. K. Agboyi, G. K. Ketoh, O. K. Douro Kpindou, T. Martin, I. A. Glitho, and M. Tamò, “Improving the efficiency of Beauveria bassiana applications for sustainable management of Plutella xylostella (Lepidoptera: Plutellidae) in West Africa,” Biol. Control 144, 104233 (2020). [CrossRef]

2. U. A. Awan, S. Xia, L. Meng, M. F. Raza, Z. Zhang, and H. Zhang, “Isolation, characterization, culturing, and formulation of a new Beauveria bassiana fungus against Diaphorina citri,” Biol. Control 158, 104586 (2021). [CrossRef]

3. Y. Cai, Y. Nie, Y. Gao, and B. Huang, “Natural populations of the entomopathogenic fungus Beauveria bassiana in Chinese forest ecosystems are diverse and reveal equal frequencies of mating types within phylogenetic species,” Fungal Ecol. 56, 101139 (2022). [CrossRef]

4. Y. Zou, X. Li, X. Xin, H. Xu, L. Mo, Y. Yu, and G. Zhao, “Comparative transcriptomics to reveal the mechanism of enhanced catalytic activities of Aspergillus niger whole-cells cultured with different inducers in hydrolysis of citrus flavonoids,” Food Res. Int. 156, 111344 (2022). [CrossRef]

5. J. Jiang, M. Zhang, T. An, Z. Zu, P. Song, M. Chen, P. Yue, and X. Gao, “Preparation of instant dark tea by liquid-state fermentation using sequential inoculation with Eurotium cristatum and Aspergillus niger: Processes optimization, physiochemical characteristics and antioxidant activity,” LWT 162, 113379 (2022). [CrossRef]

6. Y. Gu, W. Lu, J. Fang, C. Zheng, X. Chen, X. Wang, and Y. Hu, “Research progress on artificially prepared infrared extinction materials and their extinction properties (Invited),” Infrared Laser Eng. 49, 20220313 (2020).

7. C. Gittins, L. G. Piper, W. T. Rawlins, W. Marinelli, J. O. Jensen, and A. N. Akinyemi, “Passive and active standoff infrared detection of bio-aerosols,” Field Anal. Chem. Technol. 3(4-5), 274–282 (1999). [CrossRef]

8. Y. Hu, X. Zhao, Y. Gu, X. Chen, X. Wang, P. Wang, Z. Zheng, and X. Dong, “Significant broadband extinction abilities of bioaerosols,” Sci. China Mater. 62(7), 1033–1045 (2019). [CrossRef]

9. X. Zhao, Y. Hu, Y. Gu, X. Chen, X. Wang, P. Wang, and X. Dong, “Analysis of optical properties of bio-smoke materials in the 0.25–14 µm band,” Chin. Phys. B 28(3), 034201 (2019). [CrossRef]

10. G. Goudarzi, M. Ghafarzadeh, P. Shakib, and K. Anbari, “Culture and Real-Time PCR Based Maternal Screening and Antibiotic Susceptibility for Group B Streptococcus: An Iranian Experience,” Global J. Health Sci. 7(6), 233–239 (2015). [CrossRef]

11. J. Kim and S. Oh, “A colorimetric lateral flow assay based on multiplex PCR for the rapid detection of viable Escherichia coli O157:H7 and Salmonella Typhimurium without enrichment,” LWT 152, 112242 (2021). [CrossRef]

12. K. Stingl, J. Heise, M. Thieck, et al., “Challenging the “gold standard” of colony-forming units - Validation of a multiplex real-time PCR for quantification of viable Campylobacter spp. in meat rinses,” Int. J. Food Microbiol. 359, 109417 (2021). [CrossRef]

13. J. Guo, W. Wang, H. Zhao, Y. Luo, M. Wan, and Y. Li, “A new PMA-qPCR method for rapid and accurate detection of viable bacteria and spores of marine-derived Bacillus velezensis B-9987,” J. Microbiol. Methods 199, 106537 (2022). [CrossRef]

14. X. Wei, S. Li, S. Zhu, W. Zheng, S. Zhou, W. Wu, and Z. Xie, “Quantitative analysis of soybean protein content by terahertz spectroscopy and chemometrics,” Chemom. Intell. Lab. Syst. 208, 104199 (2021). [CrossRef]

15. Z. Chen, Z. Zhang, R. Zhu, Y. Xiang, Y. Yang, and P. B. Harrington, “Application of terahertz time-domain spectroscopy combined with chemometrics to quantitative analysis of imidacloprid in rice samples,” J. Quant. Spectrosc. Radiat. Transfer 167, 1–9 (2015). [CrossRef]

16. E. D. Spyrelli, O. Ozcan, F. Mohareb, E. Z. Panagou, and G. E. Nychas, “Spoilage assessment of chicken breast fillets by means of fourier transform infrared spectroscopy and multispectral image analysis,” Curr. Res. Food Sci. 4, 121–131 (2021). [CrossRef]

17. Y. Gu, Y. Hu, X. Zhao, X. Chen, P. Wang, and Z. Zheng, “Discrimination of viable and dead microbial materials with Fourier transform infrared spectroscopy in 3-5 micrometers,” Opt. Express 26(12), 15842–15850 (2018). [CrossRef]

18. H. C. Booij and G. P. J. M. Thoone, “Generalization of Kramers-Kronig transforms and some approximations of relations between viscoelastic quantities,” Rheol. Acta 21(1), 15–24 (1982). [CrossRef]

19. P. Grosse and V. Offermann, “Analysis of reflectance data using the Kramers-Kronig Relations,” Appl. Phys. A 52(2), 138–144 (1991). [CrossRef]

20. C. F. Bohren and D. R. Huffman, “Classical Theories of Optical Constants,” in Absorption and Scattering of Light by Small Particles (1998), pp. 226–267.

21. L. Le, H. Yihua, G. Y. lin, Z. H. A. O. Yi zheng, Y. Lei, and H. B. kun, “Infrared Extinction Performance of Biological Materials,” Spectrosc. Spectral Anal. 37, 3430–3434 (2017).

22. Y. Gu, Y. Hu, X. Zhao, and X. Chen, “Determination of infrared complex refractive index of microbial materials,” J. Quant. Spectrosc. Radiat. Transfer 217, 305–314 (2018). [CrossRef]

23. M. Segal-Rosenheimer and R. Linker, “Impact of the non-measured infrared spectral range of the imaginary refractive index on the derivation of the real refractive index using the Kramers-Kronig transform,” J. Quant. Spectrosc. Radiat. Transfer 110(13), 1147–1161 (2009). [CrossRef]

24. B. T. Draine and P. J. Flatau, “Discrete-dipole approximation for scattering calculations,” J. Opt. Soc. Am. A 11(4), 1491–1499 (1994). [CrossRef]

25. B. T. Draine and P. J. Flatau, “Discrete-dipole approximation for periodic targets: theory and tests,” J. Opt. Soc. Am. A 25(11), 2693–2703 (2008). [CrossRef]

26. P. Flatau, “Fast solvers for one dimensional light scattering in the discrete dipole approximation,” Opt. Express 12(14), 3149–3155 (2004). [CrossRef]

27. P. Flatau and B. Draine, “Fast near field calculations in the discrete dipole approximation for regular rectilinear grids,” Opt. Express 20(2), 1247–1252 (2012). [CrossRef]

28. B. Efron, T. Hastie, I. Johnstone, and R. Tibshirani, “Least angle regression,” Ann. Statist. 32(2), (2004).

29. H. Wang, P. Chen, J. Dai, D. Liu, J. Li, Y. Xu, and X. Chu, “Recent advances of chemometric calibration methods in modern spectroscopy: Algorithms, strategy, and related issues,” TrAC, Trends Anal. Chem. 153, 116648 (2022). [CrossRef]

30. Y. Miyashita, T. Itozawa, H. Katsumi, and S.-I. Sasaki, “Comments on the NIPALS algorithm,” J. Chemometrics 4(1), 97–100 (1990). [CrossRef]

31. S. Wold, M. Sjöström, and L. Eriksson, “PLS-regression: a basic tool of chemometrics,” Chemom. Intell. Lab. Syst. 58(2), 109–130 (2001). [CrossRef]

32. G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, “Extreme learning machine: Theory and applications,” Neurocomputing 70(1-3), 489–501 (2006). [CrossRef]

33. H. Jiang and W. Zhu, “Determination of Pear Internal Quality Attributes by Fourier Transform Near Infrared (FT-NIR) Spectroscopy and Multivariate Analysis,” Food Anal. Methods 6(2), 569–577 (2013). [CrossRef]

34. A. G. Asuero, A. Sayago, and A. G. González, “The Correlation Coefficient: An Overview,” Crit. Rev. Anal. Chem. 36(1), 41–59 (2006). [CrossRef]

35. A. S. Sant’Ana, B. D. Franco, and D. W. Schaffner, “Modeling the growth rate and lag time of different strains of Salmonella enterica and Listeria monocytogenes in ready-to-eat lettuce,” Food Microbiol. 30(1), 267–273 (2012). [CrossRef]

36. J. Liu, X. Zhou, and S. Zhang, “Seismic behaviour of square CFT beam–columns under biaxial bending moment,” J. Constr. Steel Res. 64(12), 1473–1482 (2008). [CrossRef]

37. Y. Gu, C. Wang, L. Yang, Z. Ou, Y. Hu, L. Li, Y. Zhao, W. Chen, and P. Wang, “Infrared extinction before and after aspergillus niger spores inactivation,” Infrared Laser Eng. 44, 36 (2015).

	Index
Materials	Wavelength (µm)	R_p (µm)	R_eff (µm)
AH0302	2.5∼8.5	2.5∼3.5	6.24
BM0302	2.5∼8.5	1.5∼2.0	6.45
AP0302	2.5∼8.5	2.0∼3.0	6.67

		Calibration			Prediction
Material	Model	R_C²	RMSEC	MAEC	R_P²	RMSEP	MAEP
AH0302	LAR-PLS	0.9518	0.0601	0.0512	0.9699	0.0457	0.0437
	LAR-SVR	0.9461	0.0635	0.0546	0.9496	0.0614	0.0523
	LAR-ELM	0.9482	0.0622	0.0557	0.9539	0.0587	0.0551
BM0302	LAR-PLS	0.9651	0.0531	0.0389	0.9744	0.0455	0.0374
	LAR-SVR	0.7667	0.1415	0.0874	0.7849	0.1359	0.0874
	LAR-ELM	0.9589	0.0576	0.0436	0.9389	0.0702	0.0661
AP0302	LAR-PLS	0.9368	0.0714	0.0582	0.9621	0.0553	0.0438
	LAR-SVR	0.8230	0.1233	0.0756	0.8231	0.1233	0.0756
	LAR-ELM	0.9569	0.0589	0.0450	0.9428	0.0679	0.0554

	Index
Materials	Wavelength (µm)	R_p (µm)	R_eff (µm)
AH0302	2.5∼8.5	2.5∼3.5	6.24
BM0302	2.5∼8.5	1.5∼2.0	6.45
AP0302	2.5∼8.5	2.0∼3.0	6.67

		Calibration			Prediction
Material	Model	R_C²	RMSEC	MAEC	R_P²	RMSEP	MAEP
AH0302	LAR-PLS	0.9518	0.0601	0.0512	0.9699	0.0457	0.0437
	LAR-SVR	0.9461	0.0635	0.0546	0.9496	0.0614	0.0523
	LAR-ELM	0.9482	0.0622	0.0557	0.9539	0.0587	0.0551
BM0302	LAR-PLS	0.9651	0.0531	0.0389	0.9744	0.0455	0.0374
	LAR-SVR	0.7667	0.1415	0.0874	0.7849	0.1359	0.0874
	LAR-ELM	0.9589	0.0576	0.0436	0.9389	0.0702	0.0661
AP0302	LAR-PLS	0.9368	0.0714	0.0582	0.9621	0.0553	0.0438
	LAR-SVR	0.8230	0.1233	0.0756	0.8231	0.1233	0.0756
	LAR-ELM	0.9569	0.0589	0.0450	0.9428	0.0679	0.0554

Quantitative determination of microbial materials activity based on infrared extinction properties

Abstract

1. Introduction

2. Materials and method

2.1 Biomaterials preparation

2.2 IR spectroscopy acquisition

2.3 Kramers-Kronig relation

2.4 Discrete dipole approximation method

2.5 Quantitative activity determination model

2.5.1 Characteristic wavelength screening

2.5.2 Regression model

2.6 Software and configuration

3. Result and discussion

3.1 Raw spectrum

3.2 CRI calculation

3.3 Extinction properties calculation

3.4 Quantitative activity determination results

4. Conclusion

Funding

Disclosures

Data availability

Reference

Data availability

Cited By

Figures (14)

Tables (3)

Equations (12)

Optics Express

	PLS	SVR		ELM
Material	Components	C	Kernels	Hidden Neurons	Activation Functions
AH0302	3	50	linear	10	sigm
BM0302	5	10	linear	15	sigm
AP0302	7	50	linear	10	sigm