Origin identification of Chinese Maca using electronic nose coupled with GC-MS

Maca (Lepidium meyenii Walp.), originated in the high Andes of Peru, is rich in nutrients and phytochemicals. As a new resource food in China, Maca suffers marketing disorders due to the limitation of basic research. Due to the close relationship of Maca quality and origin of place, it’s of scientific, economic and social importance to set up a rapid, reliable and efficient method to identify Maca origin. In the present study, 303 Maca samples were collected from 101 villages of the main producing area in China. Using electronic nose and BP neutral network algorithm, a Maca odor database was set up to trace the origin. GC-MS was then employed to analyze the characteristic components qualitatively and semi-quantitatively. As a result, very significant differences (p < 0.01) were detected in the volatile components of Maca from different areas. This study not only constructs a network model to forecast the Maca origin, but also reveals the relationship between Maca odor fingerprints and origins.

with the genuine. Without the characteristic chemical database of Maca, quality control is hard to be executed. Thus it is of importance to establish a reliable and efficient method to identify Maca origin.
Maca smells special, and the types and compositions of Maca volatiles might be characteristic. In this study, electronic nose was used to collect the odor fingerprints of 303 Maca samples of the main producing area. By using the pattern recognition toolbox, a BP neural network was constructed to identify the Maca origin, and its performance was tested against a few sets. Further GC-MS analysis identified the volatile types and compositions of different Maca samples. The microscopic and macroscopic results altogether revealed the correlation between Maca odor fingerprints and volatiles. The mass identification of Maca samples at low cost will solve the fake and inferior problems of the Maca market, and provides scientific support to standardize the Maca industry.

Results and Discussion
Maca sampling and powder preparation. There are many factors, such as planting conditions, place of origin, drying method, color type, that influence the metabolites of Maca 2,5,6,17,19,20 . In the present study, a total of 303 Maca samples were collected from 101 villages, 19 towns, six cities and two provinces of the main producing area in China (Table S1). These samples had similar planting cycles, i.e. breeding in greenhouse in March to April, transplanting to fields in June, and being harvested in December or the next January and February. Because Maca grew at high altitude, no fertilizer, pesticide or weeding was applied, and therefore the growth conditions are close to the wildness. Moreover, all samples were fresh roots, and were pretreated using the same method. Since color types have little effects on Maca 5,20,21 , the origin of Maca was determined to be the only key factor related to the type and concentration of Maca metabolites.

Selection and optimization of the electronic nose parameters.
To meet the standards of electronic nose for better accuracy and efficiency, the parameters were set as follows: carrier, high purity air; sampling time, 120 s; delay time of data acquisition, 18 min; loading amount, 4 mL; loading rate, 1 mL/s; sampler incubation time, 120 s; sampler incubation temperature, 50 °C; and syringe temperature, 60 °C. Under the optimized conditions, representatives of Maca samples released the strongest signals (Fig. S1). The instrument stability was also assessed against 70 Maca samples. The data collected by ten odor detectors all showed a relative standard deviation of less than 15% (Table S2). It indicates that the odor detection system is accurate, efficient and reliable. Using this system, the odor fingerprint of each Maca sample was produced by the electronic nose equipped with 17 sensors (Table S3). In comparison to previous Maca studies that scanned a large area across different countries or different regions of the same country or focused on one or several specific metabolites 5,11,22,23 , this study collected the odor data of 303 Maca samples of two neighboring Maca-producing provinces and would provide more valuable and systematic information for the study of Maca.

BP neural network of Maca odors.
A two-level BP neural network was constructed to identify the Maca origins (Table 1) based on their odor fingerprints. The level 1 (cities) consisted of an input layer of 17 neurons, a hidden layer of 15 neurons, and an output layer of 6 neurons. The level 2 (towns) was subordinate to the level 1, in which the output layer of Huidong (the only city containing six towns) had 6 neurons. Cross-entropy cost function was used to test the performance of the BP neural network. When the iteration occurred 150 times, the cross-entropy was down to 0.10137, and the signals were transferred back to the input layer. Further optimization www.nature.com/scientificreports www.nature.com/scientificreports/ was conducted using the conjugate gradient descent algorithm. With 156 times of iterations, the gradient was decreased to 0.21163, and the best network was constructed (Fig. 1a). Receiver operating characteristic curve (ROC) was used to assess the performance of the BP neural network. The larger the area under curve (AUC), the better the BP neural network (Fig. 1b). After optimization, the AUC values of the six cities were all higher than 0.8. It indicated that the BP neural network we constructed is reliable for Maca origin prediction.
In the food industry, a trained BP neural network with a correct forecast rate of more than 70% is recognized to be reliable and efficient 24 . In this study, we used confusion matrix to evaluate the predictability of BP neural network. For initial test, 297 Maca samples of six cities (Xichang, Chuxiong, Dali, Kunming, Lijiang and Shangri-La assigned to number 1 to 6) were subjected to the analysis, and the corresponding output values for training, validation, test and overall were 84.1%, 78.4%, 83.6%, and 83.2% respectively (Fig. 2a). Due to the small area of Maca plantation in Kunming and Chuxiong, fewer samples were collected, which caused low recall and www.nature.com/scientificreports www.nature.com/scientificreports/ inaccurate prediction inaccuracy of these areas. By decreasing the sample sizes of other cities 16 , the weights of Chuxiong and Kunming were improved, with the increased recall rates of 71.4% and 75.0%, and the recognition accuracy was improved to 83.6% (Fig. 2b). To predict the majority accurately, Chuxiong and Kunming were excluded, and Maca samples of the other four cities were randomly selected for training (211 samples) and test (72 samples) of the BP neuron network. As a result, 85.3% of training samples (Fig. 3a) and 82.6% of test samples (Fig. 3b) were predicted to be correct. It indicated that the level 1 of BP neural network is reliable to predict the Maca origin. www.nature.com/scientificreports www.nature.com/scientificreports/ The level 2 network (Huidong, the only city containing six towns) was constructed. After 36 times of iteration, the minimum cross-entropy (0.041636) was achieved. Six more iterations had no improvement, and the best network was then constructed. ROC analysis indicated that the level 2 BP neuron network performed well, as each town had an AUC value of close to 1.0. To improve the accuracy and applicability of the network, samples from Wudongde were excluded (with an AUC value of 0). When assigned Baishan, Duge, Lama, Manyingou, Qianxin and Wudongde to number 1 to 6, the accurate rates of training, validation, testing, and overall were 83.8%, 92.9%, 93.0%, and 86.5%, respectively. Using the level 2 BP neuron network of Huidong, the origin of 81.8% samples was accurately predicted.
Using the main Maca-producing area of China as a large sampling site, the Maca odor fingerprints were collected by electronic nose, and a BP neural network of two levels based on Maca odor was constructed to differentiate Maca of different origins (down to towns). By retaining most of the natural ingredients 25 , this method is simple, rapid, and accurate to identify a large number of Maca samples across low geographic span and of different phenotypes, and thus is applicable for quality control in industries and markets.

Identification of the Maca volatiles by GC-MS.
In this study, we used SPME-GC/MS to analyze the Maca volatiles. The performance of three solid microectractors (85 μm PA, 95 μm CAR WR/PDMS, and 75 μm CAR/PDMS) was compared using the Maca sample YXHD-1B. Qualitative and semi-quantitative analysis indicated that these microextractors varied in the detection of sorts and contents of Maca volatiles. Of them, CAR/ PDMS having capability of detecting 80 chemicals (vs. 76 and 72 chemicals) and a total content of 671.2 ng/mg (vs. 101.8 and 128.7 ng/mg) was selected for further GC/MS analysis.
The number of compounds of each Maca sample was then subjected to one-way ANOVA analysis. As shown in Fig. 4, the compound numbers of Maca samples from Shangri-La and Lijiang were significantly higher (p < 0.01) than that of Xichang and Dali. With towns as the units, the Maca samples from Weixi, Xiaozhongdian, Hutiaoxia, Ninglang, and Yulong had more compound numbers detected than those from Huidong and Heqing (p < 0.01). Further analysis indicated a correlation between the compound number and sampling altitude, i.e. more compounds were detected in the high-altitude area of Shangri-La and Lijiang.
To distinguish different Maca origins, the compounds only detected in a specific area and surroundings were defined as characteristic. As shown in Table 2, 4-methyl-tridecane in Heqing, n-hydroxy methyl-2-phenylacetamide in Jiantang, nonanal in Jiantang and Xiaozhongdian, benzoic acid in Huidong and Heqing, 6-hydroxy-4(1H)-pyrimidinone in Hutiaoxia and Weixi, and p-xylene in Ninglang, Yulong, Weixi and Xiaozhongdian were only detected and treated as characteristic compounds.

PLS-DA analysis of the Maca volatiles.
In previous studies, some Maca metabolites have been used to identify Maca origin. For example, the Maca samples from Junin and Ancash were differentiated based on the contents of glucosinolates 18,28,29 . By combining the Maca characteristic fingerprints and PLS-DA analysis, the Maca samples from the northeast of Yunnan, the northwest of Yunnan, and Peru were successfully separated 30 . However, these studies are limited by further identification into lower levels.  (Table S4) Table S3, the sensors varied in the responses to different types and concentrations of chemicals, corresponding to the mechanism of electronic nose 31 . Of them, sensor LY2/LG was sensitive to 2-hexanone,
Multi-regression analysis. In this study, we employed electronic nose and GC-MS to detect the Maca odor and volatiles, respectively, and constructed relationships between characteristic chemicals and the origin of Maca. Correlation analysis indicated that some volatiles and sensors had relationships. Considering the sensitivity of electronic nose to some category of chemicals 32 , the volatiles detected by GC-MS were classified into aromatics, esters, aldehydes, olefins, ketones, ethers, alcohols, heterocyclics and sulfides based on the functional groups 33 . Multi-regression equations were then constructed with the response values of electronic nose as dependent variable and different chemicals as independent variables. As shown in Table 3, all sensors had significant relationships with specific Maca volatiles, such as esters, aldehydes, ethers, alcohols, heterocyclics or sulfides. It suggested that the electronic nose coupled with GC-MS is applicable and accurate to predict the origin of Maca. To meet the market requirements and produce better functional Maca, further studies will be conducted to undermine the biosynthesis pathways of Maca volatiles and to identify the hub metabolites for metabolic engineering of Maca. www.nature.com/scientificreports www.nature.com/scientificreports/

Conclusions
Electronic nose coupled with GC-MS was used to identify the origin of Maca from macroscopic and microscopic levels: electronic nose revealed that the odor fingerprint of each Maca sample was distinctively different and specific, while GC-MS confirmed the roles of several characteristic volatiles. Further correlation and multi-regression analyses indicated the close relationship between odor types and components and the roles of each specific volatile. This study proposed a simple, reliable, and efficient method to identify the Maca origin, and revealed the relationship between Maca odors and volatile chemicals. It is of great importance in both scientific and industrial fields.  Table S1.

Methods
Eight to ten Maca fruits of each sample were washed by tap water, and the fibrous roots and top were removed. After three washes with deionized water, the fruits were sliced to 2-mm-thick pieces and dried at 42 °C in a Yiheng oven (Shanghai, China) to the water content of 6−9%. The Maca slices were then ground into powders, sieved through 60 mesh, and stored at −20 °C before use.
Detection and analysis of Maca odor. Maca powders (1.5 g) of each sample were placed in a 20-mL headspace bottle, and each sample had triplicate. The odor was detected by an electronic nose (FOX 4000, Alpha MOS, Toulouse, France) coupled with an automatic sampler (HS-100) and a high purity air generator (AG 2301). The electronic nose equipped with 17 detectors (Table S3) is sensitive to strong oxidizing gases, toxic gases, combustible gases, flammable gases, aromatic compounds, and organic compounds. The parameters were set to meet standards that the maximum response values are within 0.25-0.85 and the detectors with maximum values more than 0.95 and minimum values less than 0.05 are as few as possible. To verify the optimized parameters, one Maca sample of each color variety was selected for odor analysis. And the instrument stability was tested by randomly selected 30 Maca samples of six cities. By using the optimized electronic nose parameters, odor characteristics of all Maca samples were detected by 17 sensors (Table S3), and used to construct the BP neuron network. The multi-layer feed-forward algorithm (Fig. S2) was used to identify and classify the Maca origin according to the odor fingerprints detected by electronic nose. The BP neural network was constructed by using the pattern recognition toolbox of MATLAB R2015a.
GC-MS analysis of the volatiles. Maca powders (1.0 g) of each sample were placed in a 20-mL headspace bottle, and sealed immediately with a silicon lid. After equilibrium in an heated oscillator at 65 °C for 10 min, the volatiles were extracted at 350 rpm for 30 min and placed at the GC injection port for 2 min for desorption. The mass spectrometer (GCMS-TQ8040, Shimadzu, Tokyo, Japan) was equipped with an autosampler AOC-6000 and a InertCapPure-Wax column (30 m × 0.25 mm × 0.25 µm). Solid phase extraction was conducted with the fibers 85 μm PA, 95 μm CAR WR/PDMS, and 75 μm CAR/PDMS (Supelco). Helium (>99.99%) was used as the carrier gas at a constant flow rate of 1 mL/min. The injection volume was 1 mL in splitless mode. The initial oven temperature was held at 50 °C for 5 min, ramped to 250 °C at a rate of 10 °C/ min, and held at 250 °C for 10 min. The temperatures of injector, pressure, split ratio, and electron impact ion source were set to 250 °C, 83.5 kPa, 5:1, and  Table 3. Multi-regression analysis of the odor fingerprints and different Maca volatiles. X 1 , aromatics; X 2 , esters; X 3 , aldehydes; X 4 , olefins; X 5 , ketones; X 6 , ethers; X 7 , alcohols; X 8 , heterocyclics; and X 9 , sulfides.