Serum samples were provided by The First People’s Hospital of Yunnan Province. All subjects had given informed consent to be included before they participated in the study. The study was conducted in accordance with the Declaration of Helsinki, and the protocol was approved by the Ethics Committee of Yunnan Normal University (Number: ynnuethic2021-13). Serum samples from 92 patients with lung cancer and 155 samples from healthy people were collected. The information of patients with lung cancer and healthy people was listed in Table 1 . 50μl serum samples was placed on a glass slide and dried in a vacuum oven at room temperature (25°C) for 40 minutes. The purpose of vacuum pumping is to accelerate the drying speed. Drying in the vacuum drying oven can ensure that the sample will not be polluted and oxidized, and the organic substance will not be destroyed within 40 minutes. Then the serum was removed from the slide for measuring ATR-FTIR spectrum. Before sampling, the glass slides were soaked with aqua regia for 1 hour, washed with water and then soaked in acetone for 1 hour, and finally washed with ultrapure water and dried for use.

ATR-FTIR spectra were measured in the range of 4000-600cm^-1 by a Frontier spectrometer (Perkin Elmer, UK), coupled with an ATR accessory and a deuterated triglycine sulfate (DTGS) detector. Each spectrum was an accumulation of 32 scans at a resolution of 4cm^-1. The dried serum sample was transferred to the crystal plate, and then pressed with pressure tip to ensure the best contact with the crystal surface. The air background spectrum was recorded before each sample scan and automatically deducted when the sample was tested. After each measurement, the crystal surface was cleaned with ethanol and ultrapure water, and then dried with a dust-free paper. Three IR spectra were collected for each serum sample and the resulting spectra were averaged using OMNIC 8.2 software (Thermo Scientific).

The influence of noise and irrelevant information can be eliminated by proper preprocessing of the original spectra. This increases the accuracy of the analytical model and improves the signal-to-noise ratio (17). Baseline correction (BL) is a necessary processing method in infrared spectroscopy, which is helpful for further qualitative or quantitative analysis (18). Savitzky-Golay (SG) smoothing is adopted to increase the spectral quality by eliminating random noise. Derivative processing can eliminate background interference and spectral overlap, and minimize baseline drift caused by the differences in optical setups (19). Multiplicative scatter correction (MSC) is aimed to effectively eliminate the influence of scattering and improve the spectral information to obtain a relatively ideal spectrum (20). Standard normal variate (SNV) is used to reduce baseline shifting or tilt due to scattering and the change of light distance (21). It subtracts the average intensity from the spectral intensity to achieve offset correction, and then divides the standard deviation to reduce the multiplicative effect (22).

The spectral band area was measured using OriginPro 9.1 software (OriginLab Corporation, Northampton, MA). The obtained results were presented as mean ± SEM (standard mean error). For statistics, independent sample t test was processed using SPSS 19 software (SPSS, Inc., Chicago, IL) and GraphPad Prim 9.0 (GraphPad Software Inc., CA, USA). The statistical significance was signified as less than or equal to p < 0.05*, p < 0.01**, p < 0.001***, and p < 0.0001****.

Principal component regression (PCR) and partial least squares-discriminant analysis (PLS-DA) were performed to analyze the spectral data using Unscrambler X 10.4 software (Camo Software AS, Oslo, Norway). PCR is a regression analysis based on PCA (23). It decomposes the X matrix by PCA, and then takes the transformed new variables as predictive variables for multiple linear regression (MLR) (24). PLS-DA is a linear supervised classification technique combining partial least squares (PLS) regression with linear discriminant analysis (LDA) (25). It can find variables and directions from the multivariate space to distinguish the categories in the calibration set (26).

After preprocessing the spectral data, they were randomly divided into calibration set (69 serum samples from patients with lung cancer and 116 serum samples from healthy people) and validation set (23 serum samples from patients with lung cancer and 39 serum samples from healthy people) according to the ratio of 3:1 for model work. The performance of the regression model was evaluated by calculating the square of the correlation coefficient (R²) and the root mean square error (RMSE) (27). The sensitivity, specificity and accuracy were used to evaluate the judgment ability of the diagnostic model. The corresponding formula is as follows:

Figure 1 shows the IR spectra after baseline correction and SG smoothing (9-point) of serum samples from patients with lung cancer and healthy people. The average IR spectra of them are shown in Figure 2 . It can be seen that the main components of serum are protein, lipid and nucleic acids. The amide I protein (1700-1600cm⁻¹) band mainly originated from the α-helix structure at 1646cm^-1 (28). The amide II protein (1560-1500cm⁻¹) band mainly originated from the N-H functional group at 1542cm^-1 (29). The peak at 1740cm⁻¹ was attributed to the C=O stretching vibration from ester carbonyl in triglycerides (25). The spectral band of 3000-2800cm⁻¹ was mainly correlated with the lipid-related C-H asymmetric stretching vibration of CH₃ at 2959cm^-1 and CH₂ at 2930cm^-1 (30, 31). The spectral band of 1250-1000cm⁻¹ was correlated with the P=O asymmetric stretching vibration at 1243cm^-1 and symmetric stretching vibration at 1079cm^-1 of P O 2 − in nucleic acids (32). It could be observed that the absorbance of average IR spectrum in serum from patients with lung cancer was significantly increased at nucleic acids band compared with healthy people. However, there were no significant differences in amide I, amide II and lipid bands in average IR spectra.

To further analyze the differences between serum of patients with lung cancer and healthy people in these four bands, we showed the statistical analysis results of the spectral band area of serum samples in Figure 3 . It can be observed that the spectral band area of patients with lung cancer was significantly increased in amide I, amide II and nucleic acids bands compared with healthy people (p < 0.0001). Although the increased area in lipid band was not significant compared with the other three bands, there was still a statistical difference between the two groups of serum samples (p < 0.05). According to Beer-Lambert law, the increase of the absorbance of the spectral band indicates the increase of the corresponding functional group concentration (31). Therefore, the concentrations of protein, lipid and nucleic acids molecules in serum of patients with lung cancer were increased compared with those in healthy people. This may be due to the aerobic glycolysis in cancer cells that produces a large number of biosynthetic intermediates such as lipid, protein and nucleotide, which are used for the growth and proliferation of cancer cells (33).

In order to evaluate the classification effect of these four bands on serum of patients with lung cancer and healthy people, PCR and PLS-DA were performed on the spectral data after preprocessing using full cross-validation. A model with a low value of RMSE and a high value of R² closer to 1 is considered to be a good model (24, 34).

Tables 2 , 3 show the calibration and validation results of PCR and PLS-DA. Although good classification accuracy appeared under the PCR model, the values of R² and RMSE were lower than those of PLS-DA. The PLS-DA model for first derivative spectral data in nucleic acids band (1250-1000cm^-1) showed the best calibration model with R c 2 of 0.8949 and RMSEC of 0.3136. The values of R v 2 and RMSEV in its corresponding validation model were 0.8153 and 0.4180, respectively. The results of this model showed 80% sensitivity, 91.89% specificity and 87.70% accuracy.

Figure 4 shows the score plot of Factor-1 and Factor-2 using PLS-DA model for first derivative spectral data in nucleic acids (1250-1000cm⁻¹) band. The first two factors indicate that 65% (X1 36%, X2 27%) of the X variance, explains 58% (Y1 50%, Y2 8%) of the sample classification level. It can be seen that serum samples are distributed into two clusters along the Factor-1. The red cluster is mainly composed of serum samples from patients with lung cancer, and the black cluster is mainly composed of serum samples from healthy people. In this model, 80% of patients with lung cancer were correctly identified, 91.89% of healthy people were correctly separated, and the total accuracy rate was 87.10%.

Figure 5 shows the loading plot of Factor-1 for identifying the peaks with high weights in classifying samples. There are positively weighted peaks around 1176cm^-1, 1130cm^-1, 1085cm^-1 and 1043cm^-1, of which 1176cm^-1 was related to the vibration band of sugar-phosphate, 1130cm^-1 was assigned to the C=O stretching vibration of ribose in RNA (32), 1085cm^-1 was ascribed to the symmetric phosphate vibrations (35), 1043cm^-1 was attributed to the stretching vibration and bending vibration of C-O in carbohydrates (14). Two positively weighted peaks at 1226cm^-1 and 1026cm^-1, of which 1226cm^-1 was due to the asymmetric stretching vibration of P O 2 − in nucleic acids (36), 1026cm^-1 was related to the stretching vibration of C-O and bending vibration of C-H in aromatic amino acids (37). Therefore, the loading plot of Factor-1 showed that P O 2 − in nucleic acids play a key role in distinguishing the serum patients with lung cancer from that of healthy people. This may be due to DNA damage caused by oxidative chemical mutagenic aberrations in serum of patients with lung cancer (38).