Volatolomics analysis of exhaled breath and gastric-endoluminal gas for distinguishing early upper gastrointestinal cancer from benign

This study was approved by the Ethics Committee on Biomedical Research, West China Hospital of Sichuan University (No. 858, 2021). All subjects were from West China Hospital of Sichuan University and signed informed consent forms. Patients with biopsy-confirmed esophageal cancer, gastric cancer, and benign disease (including gastritis, esophagitis, and stromal tumor) were included in this study. Inclusion criteria were as follows: (1) older than 18 years old, (2) no history of malignancy in the last five years and no anticancer therapy, (3) no ongoing infections, and (4) no signs of liver or renal malfunction. Demographic information was provided in supplementary table 1. All the gas samples were collected from 1 October 2021, to 1 March 2022. All subjects were required to hold a 24 h nucleic acid negative certificate at sampling. Researchers wore masks and medical gloves throughout the sampling and analysis process. After sample analysis, all samples were disinfected with ultraviolet lamps. The study design flow can be seen in supplementary figure 2.

The collection of gas samples was performed by well-trained researchers following the same protocol. All subjects were required to fast for at least 8 h before sample collection and refrain from smoking, drinking, or eating spicy foods the day before sampling. For possible pollution monitoring, environmental air was collected in parallel (supplementary table 2).

Exhaled breath was collected before the UGI endoscopy. Exhaled breath collection was carried out by a Bio-VOC breath sampler (Markes Int. U.K) and a disposable mouthpiece. Subjects were required to exhale slightly into the Bio-VOC sampler until they felt resistance in their lungs. The dead-space air was discharged through a three-way valve, and the end-tidal gas was retained and transferred into a Tedlar bag (500 ml, Dalian Delin Gas Packing Co., Ltd, China). Each subject was sampled five times to collect about 450 ml of breath (figure 1).

Zoom In Zoom Out Reset image size

Gastric-endoluminal gas was collected during the UGI endoscopy. A disposable medical pressure extension tube was inserted into the gastric endoluminal through the biopsy channel of a standard endoscope. The outer end of the tube was directly connected to the Bio-VOC sampler, and the inner end of the tube extended into the gastric cavity. During collection, the gas in the gastric cavity was directly extracted using the Bio-VOC sampler and then transferred into a Tedlar bag (500 ml, Dalian Delin Gas Packing Co., Ltd, China). Each subject was also sampled five times, and the total gas volume was about 450 ml (figure 1). The background analysis of the tube was performed by using a Tedlar bag filled with pure synthetic air as the gastric cavity. Then, using the same sampling process, the gas in the Tedlar bag was extracted by Bio-VOC sampler and transferred into a new Tedlar bag for analysis. After testing, it was found that there was no background interference in the tube. All samples were sent to the laboratory instantly after sampling and analyzed within the same day.

Two kinds of gas samples, exhaled breath gas and gastric-endoluminal gas, were collected from each subject and analyzed by two methods. Firstly, gas samples were directly analyzed by UVP-TOFMS (Aliben science and technology Co., Ltd, Chengdu, China). Gas samples were introduced into the ionization chamber through a PEEK capillary (i. d. 2.10 mm, length 0.50 m). The temperature of the capillary and the ion source was 70 °C. The pressure of the ion source was about 500 Pa, which allowed the sample to be automatically inhaled and then ionized in the ion source. The TOF signal acquisition and accumulation were completed by a 1 GS s⁻¹ analog-to-digital converter, and the acquisition frequency was 20 kHz. Mass spectra were accumulated for 50 s for each sample analysis.

Then further analysis was performed on a Thermo Scientific TRACE 1300 GC system coupled with a TSQ 8000 triple quadrupole system. For analysis of trace VOCs in exhaled breath and gastric-endoluminal gas, a solid-phase microextraction (SPME) device with a divinylbenzene/carboxen/polydimethylsiloxane fiber was used to enrich at 37 °C for 30 min before analysis. Before each batch of sample analysis, the SPME fiber was desorbed at the injection port at 270 °C for 5 min to remove the residual compounds on the fiber. A medium polarity capillary column VF-624 (60 m × 0.25 mm × 1.4 μm) was used to separate VOCs. The GC carrier gas was high-purity helium (99.999%), and the constant flow rate was 1 ml min⁻¹ with a split ratio of 5:1. The oven temperature programs were set as follows: initial 40 °C, held for 5 min, then ramped at 10 °C min⁻¹–160 °C, next ramped at 5 °C min⁻¹–200 °C, and last ramped at 10 °C min⁻¹–220 °C. The MS worked in full scan mode with a mass range of m/z 30–350 and a scan rate of 0.5 scan s⁻¹. The temperature of the ion source and MS transfer line was 250 °C, and the ionization energy of the ion source was 70 eV.

The original data obtained by UVP-TOFMS analysis went through peak identification, integration, and alignment to obtain data matrices containing sample ID and peak (m/z) area. Peak identification was accomplished by our previously developed peak detection algorithm based on wavelet transform and image segmentation [27], and the peak integration was completed with the Monte Carlo method. Peaks (m/z) with missing values greater than 40% in all samples were removed, and peaks with m/z > 200 and m/z < 30 were also removed because of low intensity. The raw data of GC-MS analysis was imported into TraceFinder^TM (5.0, Thermo Fisher, USA) for deconvolution and peak alignment. The data matrix containing retention time, sample ID, and corresponding peak areas was obtained after processing. Pollutants were excluded from the gas sample analysis (supplementary table 2). Compounds were tentatively identified by searching in NIST 17, and the match probability was required to be greater than 80%. Then, the identity was confirmed by analyzing the standard if available.

After obtaining the data matrix for classification model construction, the KNNImputer algorithm was used to fill in the missing values. And data were pre-processed by the StandardScaler algorithm to get a normalized form. Then, RF, XGB, SVM, and LDA were employed as classifiers for model construction. The recursive feature elimination algorithm was used to select the optimal feature subset for models. The data were randomly divided into a train set and a test set in the ratio of 7:3. Train set was utilized to construct models, and test set was used to evaluate the model. Furthermore, Mann-Whitney U test was employed to compare the VOCs between cancer group and benign group. VOCs with p < 0.05 were considered statistically significant. Finally, biomarkers for UGI cancer diagnosis were screened. Biomarkers were compounds that had been selected as important features by three or four models when modeling and had significant differences (p < 0.05) between UGI cancer and benign group.

The data of exhaled breath analysis based on GC-MS was preprocessed to obtain a data matrix with 193 samples and 163 features. The parameters of the classification models based on exhaled breath analysis by GC-MS for UGI cancer diagnosis were provided in supplementary table 3. The ROC performance of four ML models for test dataset predicting is shown in figure 2(a). As shown in figure 2(a), the AUC values of LDA, SVM, RF, and XGB models were 0.895 (95%CI: 0.731, 0.897), 0.959 (95% CI: 0.859, 0.952), 0.943 (95% CI: 0.853, 0.950) and 0.797 (95% CI: 0.680, 0.827), respectively. Among the four models, the SVM model had the best performance in the test set with AUC value of 0.959, sensitivity of 82.6%, and specificity of 91.7%. Eleven biomarkers were screened based on the four models, and their trends were shown in the VOC biomarkers plot of figure 2(a). Among them, dimethyl sulfide, n-propanol, methyl cyclopentane, cyclohexane, trimethylbenzene, and 2,2-dimethyl heptane were significantly increased in the exhaled breath of UGI cancer patients, and chlorobenzene, a-methyl styrene, 3-ethyl styrene, benzothiazole, and 1-methylnaphthalene were decreased. Overall, exhaled breath analysis based on GC-MS was able to differentiate between patients with UGI cancer and benign diseases.

Zoom In Zoom Out Reset image size

The data of gastric-endoluminal gas analysis based on GC-MS was preprocessed to obtain a data matrix with 190 samples and 190 features. The parameters of the classification models based on gastric-endoluminal gas analysis by GC-MS for UGI cancer diagnosis were also provided in supplementary table 3. The classification results of four ML models are shown in figure 2(b). It could be seen from the ROC curve plot in figure 2(b) that the AUC values of four models for test dataset predicting was more than 0.80, indicating that all models based on gastric-endoluminal gas analysis by GC-MS could distinguish UGI cancer patients from benign disease patients. The LDA model was superior to the other three models, with AUC value of 0.935 (95% CI: 0.808, 0.955), sensitivity of 95.7%, and specificity of 82.6%. As illustrated in the VOC biomarkers plot of figure 2(b), thirteen compounds were selected for UGI cancer diagnosis. The intensities of benzene, methylcyclohexane, amyl butyrate, 2-butoxyethyl acetate, trimethylbenzene, 2,2-dimethyl heptane, 3,3-dimethyl octane, and limonene were significantly increased in the gastric-endoluminal gas of UGI cancer patients, and the intensities of 1-butanol, chlorobenzene, 6-methyl-5-hepten-2-one, undecane, and 1-methyl naphthalene were decreased.

Four identical VOC biomarkers, trimethylbenzene, chlorobenzene, 1-methylnaphthalene, and 2,2-dimethyl heptane, were screened by comparing the results of exhaled breath and gastric-endoluminal gas analysis. As shown in figure 3, trimethylbenzene and 2,2-dimethyl heptane increased in UGI cancer patients both in exhaled breath and gastric-endoluminal gas, while chlorobenzene and 1-methylnaphthalene decreased in UGI patients. Notably, the intensities of all VOCs were higher in the gastric-endoluminal gas than in the exhaled breath. VOCs produced by diseased tissues enter the bloodstream by diffusion, then circulate with the bloodstream and are exchanged at the alveoli or airways before being expelled by exhalation. The concentration of VOCs in the metabolic gas of diseased gastric-endoluminal tissue would be greater than in the exhaled breath obtained by multiple diffusions.

Zoom In Zoom Out Reset image size

In this study, UVP-TOFMS was used for rapid analysis of exhaled breath and gastric-endoluminal gas samples. The original spectrum obtained by UVP-TOFMS analysis of gas samples is shown in supplementary figure 3, and it was found that a large number of compounds were detected in both exhaled breath and gastric-endoluminal gas. A data matrix with 85 samples and 166 features was obtained by preprocessing the data of exhaled breath analysis based on UVP-TOFMS. The UGI cancer diagnosis model was also constructed by four ML algorithms, and parameters of the models were provided in supplementary table 4. As illustrated in the ROC curve plot of figure 4(a), the linear model based on LDA and SVM (linear kernel) had better performance in the test set, with AUC > 0.90. Overall, the SVM model had the best performance with the AUC value of 0.994 (95% CI: 0.799, 1.000), sensitivity of 92.3%, and specificity of 100.0%. Five common features were selected based on four models, which were m/z 88, m/z 106, m/z 142, m/z 169, and m/z 195 (figure 4(a)). Combining the qualitative result of GC-MS, VOCs in the exhaled breath of healthy subjects summarized by Costello et al [19], and the standards analysis in UVP-TOFMS, m/z 88, 106, 142, 169, and 195 were tentatively characterized as allyl methyl sulfide, xylene, nonanal, dodecane, and isobornyl acetate, respectively (supplementary table 5). These VOCs had the same trend in UVP-TOFMS and GC-MS analysis (supplementary figure 4). The mass spectra of these standards analyzed by UVP-TOFMS are shown in supplementary figure 5.

Zoom In Zoom Out Reset image size

A data matrix with 87 samples and 160 features was obtained by preprocessing the data of gastric-endoluminal gas analysis based on UVP-TOFMS. The data matrix was used for classification model construction, and the model parameters can be found in supplementary table 4. As illustrated in the ROC curve plot of figure 4(b), the linear models had better performances, with AUC > 0.90. This result was similar to the exhaled breath analysis based on UVP-TOFMS for UGI cancer diagnosis. LDA and SVM (linear kernel) are linear classification algorithms. The ultimate goal of classification is to find an optimal decision boundary. Generally, the classification results would be better when the sample size is small [28]. Both RF and XGB are ensemble tree models, which are prone to overfitting when the sample size is small [29]. Due to the small size of samples enrolled in this study, the performances of these models are as expected. Three biomarkers, m/z 78, m/z 154, and m/z 156, were screened (figure 4(b)). They were tentatively characterized as benzene, 2-ethenyl naphthalene, and undecane, respectively. Benzene and undecane had the same trend in gastric-endoluminal gas analysis based on GC-MS and UVP-TOFMS (supplementary figure 6). The mass spectra of these standards analyzed by UVP-TOFMS are shown in supplementary figure 5.