Identification of Human Intestinal Microbiota of 92 Men by Data Mining for 5 Characteristics, i.e., Age, BMI, Smoking Habit, Cessation Period of Previous Smokers and Drinking Habit.

The intestinal microbiota compositions of 92 men living in Japan were identified following consumption of identical meals for 3 days. Fecal samples were analyzed by terminal restriction fragment length polymorphism with 4 primer-restriction enzyme systems, and the 120 obtained operational taxonomic units (OTUs) were analyzed by Data mining software focusing on the following 5 characteristics, namely, age, body mass index, present smoking habit, cessation period of previous smokers and drinking habit, according to the answers of the subjects. After performing Data mining analyses with each characteristic, the details of the constructed Decision trees precisely identified the subjects or discriminated them into various suitable groups. Through the pathways to reach the groups, practical roles of the related OTUs and their quantities were clearly recognized. Compared with the other identification methods for OTUs such as bicluster analyses, correlation coefficients and principal component analyses, the clear difference of this Data mining technique was that it set aside most OTUs and emphasized only some closely related ones. For example for a selected characteristic, such as smoking habit, only 7 OTUs out of 120 were able to identify all smokers, and the remaining 113 OTUs were thought of as data noise for smoking. Data mining analyses were affirmed as an effective method of subject discrimination for various physiological constitutions. The species of bacteria that were closely related to heavy smokers, i.e., HaeIII-291, were also discussed.

INTRODUCTION

The human intestinal microbiota (HIM) is closely related to our health and plays a crucial role in nutrient absorption, development of our immune systems and excretion of physiological waste materials. In order to analyze and compare the HIM of each subject, various factors (e.g., diets and drugs) that can directly affect the composition of the HIM must be controlled. Particularly, it is essential to unify the dietary factors, because daily eating habits vary among individuals. However, a fixed diet itself may also affect the composition of the HIM, so we cannot apply a fixed diet to long-term feeding experiments. In the present study, we attempted a new method of screening the HIM data, namely, Data mining analysis (DM), which has been popularly utilized in other technical fields, such as in commercial business management as described by Berry and Linoff [1]. DM has been valuable as a means of sharp identification of a characteristic from a large amount of data; in other words, DM is able to eliminate noise that is not directly related to a selected characteristic. We previously attempted to introduce a new method of numerical analysis for HIM data, and our previous report [9] examined whether the method works or not. The obtained results were only meaningful for smoking habit of subjects and for less data compared with the data in the present report. So, the results reported here were much expanding and deepening.

MATERIALS AND METHODS

In the present study, we designed identical meals that were subsequently fed to 92 healthy male volunteers living in Japan for 3 days. All dietary components, including drinks, were unified as already reported by Jin et al. [2]. In order to standardize the diet, the identical meals designed by Fujicco Co. (Kobe, Japan) were delivered to and consumed by all subjects. The average calorie composition was 1,879 kcal/day. Fecal samples were analyzed by terminal restriction fragment length polymorphism (T-RFLP) using 4 primer-restriction enzyme systems.

All volunteers who participated in this study provided full informed consent. The subjects were 21–59 years old (average: 36.8 years old), and body mass index (BMI) ranged from 17.3 to 30.2 kg/m2 (average: 22.6 kg/m2). After ingestion of identical meals for 3 days, fecal samples were collected in containers and stored at home in the subjects’ freezers until they were brought them to the laboratory. The studies were performed in accordance with the protocol approved by the RIKEN Research Ethics Committee.

Approximately 40–100 mg of fresh fecal samples were suspended in 500 µL of TE buffer (10 mM Tris-HCl [pH 8.0], 1 mM EDTA) and centrifuged at 20,800 ×g for 5 min, following which the supernatant was discarded; this washing step was repeated three times. Subsequently, 600 µL of TE, 600 µL of Tris-saturated phenol, 100 µL of 10% sodium dodecyl sulfate and 0.3 g of glass beads (diameter, 0.15–0.21 mm) were added, and the samples were treated at 7,000 rpm for 20 sec on a MagNA Lyser (Roche, Penzberg, Germany). Following this step, the samples were incubated at 70°C for 10 min; this crushing step was repeated twice. Six-hundred microliters of upper layer was transferred into a 1.5-mL tube, followed by addition of 600 µL of glacial isopropanol and 60 µL of 3 M sodium acetate. After gentle shaking, the solutions were centrifuged at 20,800 ×g for 5 min and then the supernatant was discarded by decantation. The DNA pellet was washed with 70% glacial ethanol and dried on a Centrifuge Evaporator (Eyela, Tokyo, Japan) for 15 min. The pellet was then dissolved in 200 µL of TE. Next, the DNA preparation was purified using a High Pure PCR Template Preparation Kit (Roche, Penzberg, Germany) and quantified with a NanoDrop 2000c (Thermo Scientific, Wilmington, DE, USA). Amplification of the fecal 16S rRNA, restriction enzyme digestion, size-fractionation of the T-RFs, and T-RFLP analysis, were carried out as has been previously described by Sato et al. [3], Matsuki et al. [4] and Nagasaki et al. [5]. PCR was performed on the FAM-labeled 516f (5′-TGCCAGCAGCCGCGGTA-3′; E. coli positions 516-532) or 27f (5′-AGAGTTTGATCCTGGCTCAG-3′; E. coli positions 8-27). The resulting 16S rRNA amplicons were further treated for 1 hr with 2 U of BslI or HaeIII (New England Biolabs, Ipswich, MA, USA) for the PCR products that were amplified with the 516f primer and with 2 U of MspI or AluI (TaKaRa Bio Inc., Shiga, Japan) for the PCR products that were amplified with the 27f primer. The digestion products were fractionated using an automated sequence analyzer (ABI PRISM 3130xl DNA Sequencer, Applied Biosystems, Carlsbad, CA, USA) and the GeneMapper software (Applied Biosystems). The major T-RFs were identified by computer simulation, which was performed using a T-RFLP analysis program (MiCA3 by Shyu et al. [6], http://mica.ibest.uidaho.edu/ [7]), a phylogenetic assignment database for T-RFLP analysis of human colonic microbiota (PAD-HCM, Matsumoto et al. [8]) and MicrobiotaProfiler (InfoCom Corporation, Tokyo, Japan). The obtained data of operational taxonomic units (OTUs) were as follows: 27-BslI (27 refers to the number of OTUs, B106 to B990, B--- represents the base pair number), 33-HaeIII (HA—), 20-27f-MspI (M—) and 40-27f-AluI (A—). The amounts of each OTU represent the fluorescence intensity and quantities of each OTU group. These OTU data are reproducible and can be used for further statistical analyses.

All 120 OTUs (27B+33HA+20M+40A) were combined with the answers of the 92 subjects and analyzed with a DM software (IBM-SPSS Clementine14). We applied the Classification and Regression Tree (C&RT) method of DM algorithm, which is the most typical way to construct a Decision tree (Dt). A Dt is a decision supporting pathway that forms a tree-like graph. C&RT algorithms is processing the optimization of Gini coefficient between a characteristic and OTUs. The Gini coefficient: is famous for quantitative evaluation of the impurity of a group and is defined at a node, , as

, where and are categories of the target field. The C&RT divides subjects into two subsets so that the subjects within each subset are more homogeneous than in the previous subset. It is quite flexible and allows unequal misclassification costs to be considered compared to the other algorithms of DM. A major specialty of C&RT is the use of a single selected OTU for each step of Dt construction.

RESULTS

The details and segments of the 5 characteristics of the subjects, i.e., age, BMI, present smoking habit, cessation period of previous smokers and drinking habit are shown in Table 1.

Table 1.

Characteristics and segments of the 92 male volunteers

Present smoking habit

After performing the analyses, DM provided the Dt shown in Fig. 1, which identified explicitly the various smoking groups, i.e., node of subjects. The root node, which is also referred to as Node-0, and always sits alone at the left end of the figure, is the starting point of tree construction, and the Dt grows toward the right to divide the subjects appropriately according to the cited characteristic, i.e., smoking habit, A or B. The details of the Dt and the pathways to reach the Terminal node, which does not split or grow further, indicate clearly the names and quantities of OTUs. Selected OTUs played roles in dividing the various smoking groups. DM processing is able to construct 5 or more steps for a Dt, but we focused on up to the 3rd step of Dt here, because the later steps had less effects for the characteristic and were easily affected by OTUs of upstream.

Fig. 1.

Decision-tree (Dt) of ‘smoking habit’ with 120 OTUs. OTUs: 27B + 33HA + 20 M + 40A; Arrows: Terminal nodes of ‘B’, smokers; Dotted arrow: gathered node of nonsmokers, ‘A’. Each box is called ‘node’, a group of subjects, of which components were shown. Along the pathway of Dt, name and cutoff value of OTU, which was estimated with C&RT method, and played a role of dividing, were indicated. Upper side of Fig. 1. were less amount of OTUs quantities, and lower side did higher amount comparatively.

Table 2 shows the details of identification for the 16 present smokers. We also had a similar list of nonsmokers (A), but the list was omitted due to space limitations. Regarding the utilization of HA291 twice and the position of Node-5 (N-5) in the lower left part of Fig. 1, the comparative quantities of HA291 of the subjects was attractive in understanding the hidden phenomena of heavy smoking. Figure 2 shows the actual aspects of them.

Table 2.

Comparison of smoking habits with the 120 OTUs

Fig. 2.

Circumstances of the ‘OTU-HA291 quantities’ with 5 smoking categories. N-5: Node-5 shown in Fig. 1 and Table 2.

Age and BMI

The data for age and BMI are continuous numerical values, which differs from category values or nominal partitions, e.g., A or B, so the structures of obtained Dts were similar but rather different. The figures were much more precise and complicated to exhibit compared with Fig. 1. Moreover, continuous numerical values could not be correlated directly with DM estimation as shown in Table 2. However, a merit of using these continuous values was that we could easily divide the subjects at any point within cited area, so the 92 men were separated into 2 parts, both by age and BMI, i.e., younger and elder groups and skinny or obese ones, and compared. The obtained OTUs utilized at the 1st to 3rd steps of the Dt are described later at Major OTUs related to the 5 characteristics.

Cessation period of previous smokers

The subjects contained 76 present nonsmokers, but 19 of them were previous smokers with various cessation periods, i.e., 1 month to 26 years. So we tried to determine whether these periods could be identified by DM. In Table 3, only 2 groups, i.e., 5 AP nodes and 1 AS node, are shown. Furthermore, at the time of this DM processing, omitting HA291, we applied a total of 119 OTUs, because HA291 was already known to have a vigorous relation with present smokers and could not be used to uncover this delicate hidden characteristic. Although this attempt was successful, DM provided 2 misclassified cases of subjects, i.e., at N-19 and N-33 in Table 3.

Table 3.

Comparison of cessation periods of previous smokers with 119 OTUs

Drinking habit

Within the 92 subjects there were 45 men who had the habit of drinking in the evening. The DM results are shown in Table 4. The 45 drinkers were grouped into 5 nodes, and detailed concerning the amounts they drank were also collected and shown in the table. ‘3-1-2’, which means that the subject drank 1 large bottle of beer (633 mL/bottle, 5% alcohol (AlOH); 32 mL of AlOH/bottle) and 2 ‘Go’ of sake (Japanese rice wine, 180 mL/‘Go’, approximately 14% AlOH; 25 mL of AlOH/Go) per drinking day on an average of 3 days/week. There were no misclassified drinkers or nondrinkers. However, the detailed DM processing for Table 4 was a little different from the normal DM processing used for Fig. 1 and Table 2, and the sequences of the 120 OTUs were partly replaced with 20M+40A+27B+33HA. When we applied the other sequences of OTUs, i.e., 27B+33-HA+20M+40A, 3 drinkers had been misclassified into the nondrinkers groups. Therefore, we need to learn more about the order for the sequence of OUT fields and to accumulate more processing experience. It was also important to trace whether there were any effects on the species and the amount of AlOH, and the total AlOH intake per week was calculated and shown in Fig. 3 based on the 5 nodes of drinkers in Table 4.

Table 4.

Comparison of drinking habits with the 120 OTUs

Fig. 3.

Drinking habits of 45 men out of 92 subjects. N-7 – N-21 were cited in Table 4.

Major OTUs related to the 5 characteristics

After the various DM processings, the major OTUs related to the 5 characteristics were listed in Table 5, in which ‘Dt 1st step’ indicates the closest relation between the characteristic and the 120 OTUs (except previous smokers, i.e., 119 OTUs), and ‘Dt 2nd step’ and ‘Dt 3rd step’ indicates the next and less closest relations. Age and BMI were each divided into 2 parts by dividing them in half. To understand the details of Table 5, it is better to compare the pathway smoking with Fig. 1, in which the actual Dt is shown.

Table 5.

Major OTUs related to the 5 characteristics

Tracing of HA291 to accession number

We tried further to trace the accession numbers of HA291, which had close relations with the heavy smokers, i.e., it was used twice in Fig. 1. Using Microbial Community Analysis III of the University of Idaho [7], simple tracing of HA291 produced 1036 bacterial 16S rRNA gene sequences. Then the data for HA291 were compared with those of other 3 restriction enzymes, i.e., BslI, 27f-MspI and 27f-AluI, and closely related OTUs were selected by DM. These OTUs were possibly included the same bacteria. To find out the same accession number, the lists were scanned and crosschecked, and ultimately 28 bacteria were pursued as a pair of OTUs and were shown in Table 6. However, all of them were uncultured species, and 3 of them were recognized as rumen bacteria.

Table 6.

The 28 bacteria concordant with HA291

DISCUSSION

Regarding present smoking habit in Fig. 1, the 7 solid arrows indicate all nodes of the smokers’ (B: yes) group, and the 1 dotted arrow shows 64 subjects, i.e., 84% of the nonsmokers (A: no), gathered at a node, N-15, in the top right part of Fig. 1, where 1 smoker was misclassified. The Dt also provided practical values for the dividing point. The cutoff values were indicated, and similar steps were repeated. A major specialty of DM is the use of a single selected OTU for each step of Dt construction. So only 7 OTUs out of 120 were utilized in Fig. 1, and the other 113 were neglected as a result. Therefore, we were able to accept the fact that there was a large gathering node, i.e., 65 subjects in N-15, as this Dt was focused only on dividing the subjects based on whether they smoked or not.

At the top of Table 2, a misclassified smoker is shown in N-15. Regarding the amount of smoking of this subject, ‘5-2Y’ was the lowest amount among the 16 smokers. So the reason for misclassification was partly apprehensible. On the other hand, when we applied 80 OTUs, except 27f-AluI, very clear identification, with no misclassification, was observed as reported previously [9].

In N-5 located in the lower part of Table 2, all 4 heavy smokers, who smoke 20 cigarettes/day, were gathered together. This indicated actually that the OTUs had very close relations with the amounts of smoking and that their HIMs were sensitively following their individual characteristics. Comparing these results with our previous report [9], the gathering of 4 heavy smokers with HA291 at N-5 was the same. But the OTU in the upper 2nd step of the Dt in Fig. 1 (A87) was different from that in the previous report (B469) [9]. Then A87 and the following OTUs, i.e., A238, B650, B657, HA175 in Fig. 1, ultimately channeled a new pathway to reach the other Terminal nodes. This difference means that construction of the Dt and the ability to make a detailed identification are closely related to the contents of applied OTUs, which can be easily understood using the fundamental algorithms of the C&RT system.

The actual quantities of HA291 are shown in Fig. 2, and the enlarged inset for N-5 indicates the details of the 4 heavy smokers. Focusing on those who had larger quantities of HA291 than the subjects in N-5, we realized that elder subjects, i.e., 50–55 years, who were nonsmokers stayed at the top of the right hand side. Since HA291 was a group of bacteria containing various species, the right upper end of Fig. 2, where the quantities of HA291 were larger than in N-5, contained some species of bacteria that differed from those in N-5.

Concerning the cessation period of previous smokers in Table 3, when focusing on 2 nodes, i.e., N-19 and N-33, there were 2 ‘b’ subjects in N-19 and 2 ‘-’ subjects in N-33. Comparison with the other nodes suggested that the 2 ‘b’ men in N-19 might be previous smokers but with blank answers and that 2 ‘-’ men in N-33 might be hidden smokers, because the other member of N-33 were men with a short cessation period. Additionally, DM successfully identified the AP, AS, BG and BH groups of Table 1. We also realized that there were some remaining signatures of previous smoking in the HIM as shown in Table 5, even after 26 years had passed.

Regarding the subjects’ drinking habit, which are shown in Fig. 3, comparison of the detailed amounts of drinking AlOH with the 5 nodes identified in Table 4 ultimately revealed that there was no differences between the amounts. Some additional factors might affect the classification of the 45 drinkers. Drinking habit was also clearly identified by DM processing with selected OTUs as shown in Table 4.

Concerning major OTUs related to the 5 characteristics in Table 5, when the whole area and divided areas for age and BMI were compared, we were unable to find the same OTUs within them. The major OTUs within each area were mutually independent and existed in one group only. The species of HIM had tight relations with each characteristic, even when divided by age and BMI. In other words, the original 120 OTUs would have many possibilities to classify different characteristics, that is not only these 5 but also some other features such as diseases.

Regarding the tracing of HA291 to accession number in Table 6, we have to recognize that application of the 120 OTUs to the HIM might produce many more possibilities to identify various human characteristics. However, all the bacteria shown in Table 6 were found to be uncultured bacteria. On the other hand, the Dt and OTUs were able to identify the relations between subjects and characteristics as shown in Table 5.

Comparing DM with the other analyzing methods for the HIM, the most unique point is the introduction of DM identification and predictive analyses. Previously, bicluster analyses have been popular [2, 10, 11]. Application of bicluster analyses to our data revealed the following 3 shortcomings. The first was that a typical OTU like HA291 was not found in the Dendrogram. Second, all clusters showed a classified feature but not the visible reason for reaching the each divided cluster. Third, the obtained clusters were tightly attributed to the data; therefore, if a slight addition or subtraction took place, the next cluster became very different from the former one, which depended on the weighting of all data, and which included a large amount of noses.

On the other hand, DM showed clear reasons for the Dt constructions, and the sequential roles of selected OTUs and their simple utilization could be understood quantitatively. Moreover, once the structure of the Dt was constructed, as long as the basic concepts of the data were active, all of the subsequent new data could be run on the same Dt. For example, the Dt shown in Fig. 1 is able to classify new data of the men and predict who smokes and who does not, when we obtain now OUT data.

We checked the usual correlation coefficients (Pearson’s product moment) between OTUs and the smoking habit, and especially about HaeIII, HA291 had the 6th positive amount of the value. And another OTUs, which played main roles for Dt construction in Fig. 1, did not have larger correlation coefficients. Furthermore, we obtained various different score plots with principal component analyses, but the outstanding OTUs like HA291 were not clearly recognized.

The major differences of these comparisons were in how noise included in the data was handled. Namely, DM ignores the noise and evaluates the limited OTUs relating to a focused characteristic, but the former 3 analyses utilize and evaluate all data including a large amount of noise. These methods then reveal the general aspect of the HIM, including all the noise. On the other hand, DM in the present study ultimately neglected most of the OTUs data and emphasized only a few selected OTUs after the C&RT processings. But some of these ignored OTU fields would act in main roles for other characteristics as shown in Table 5. Noise is defined simply based on the relationship with a focusing characteristic.

Concerning the roles of the HIM for health, we have to understand the fact that an OTU contains different species of bacteria and that HA291 had a close relation with heavy smokers as shown in Table 2. So, we have to determine the best combination for a characteristic and restriction enzyme systems for T-RFLP as shown in Table 5. The HIM of the individual subjects exhibited a wide range of differing aspects and possessed more systematic relations with the characteristics and health of the host than ever presumed.

Many studies concerning the HIM have been reported, especially for bacteria flora [12, 13], obesity and BMI [14, 15], and aging and maturation [16, 17]. The species contained within the HIM are related, and their mutual effect mechanisms have been thought to be very complicated. Therefore, detailed studies of each bacterium needs to be performed. Here we are able to reveal new features of the HIM and to understand their separate roles in detail.