To test the possibility of predicting bariatric surgery outcome, we analyzed 308 fecal sampla
The human body is colonized by a wide variety of micro-organisms, commonly referred to as the human microbiota. The gut microbiota is a complex ecosystem, which provides major functions to the host, such as regulation of metabolism, immune system modulation, and protection against pathogens [1, 2]. The microbiome is strongly associated with weight and sugar consumption, and as such it serves as a proxy for nutrition and life habits and may also influence them. Such life habits may influence the total body mass and BMI in regular conditions, as well as after bariatric surgery.
Obesity and diabetes are world pandemics [3]. Approximately 8-10% of the population develop complications of morbid obesity, (BMI>35), frequently coupled to some form of diabetes. According to the WHO, of the 57 million deaths in 2008 worldwide, 1.3 million were due to metabolic disorders, particularly those associated with obesity [3] Recently, the gut microbiome of obese individuals has been shown to be very different from the microbiome of slim subjects [4]. Nagpal et al. [5] suggested that some bacteria increase gut permeability and insulin resistance leading to obesity and diabetes. Experimental fecal transplants to mice demonstrated that transplantation of microbiome from obese individuals into slim mice turned them obese [6], showing the importance of the gut microbiome in regulating body weight. Opposite studies of turning obese mice into slim mice have not been successful but one study demonstrated that certain bacteria can prevent weight gain in mice [7].
The introduction of bariatric surgery as a method for losing weight is rapidly adopted as the most efficient method for weight loss and for reducing blood sugar levels [8], however it has drawbacks, including a range of possible complications from nutrition deficiencies to occurrence of life-threatening conditions and a big diversity in the success rates of achieving weight loss and maintaining it [9]. A few studies have shown microbiome changes after bariatric surgery [10]. However, the real potential of the microbiome as a tool for not only monitoring the procedure's outcomes but rather predicting them beforehand has not been explored. An attempt to study and test this potential can result in an essential tool which will assist in the decision whether to consult patient to perform such surgery.
es from patients of 2 main groups: obese who underwent bariatric surgery and naturally slim. For the obese patients (BMI > 35) we sampled the microbiome at five time points (Fig.
1A) – one at enrollment (A, 78 samples), three weeks after a low carbohydrate diet and immediately before the operation (B, 70 samples), and 3 time points following the surgery (two weeks– C, 34 samples; three months – D, 27 samples; and six months E, 16 samples). Not all individuals had been sampled at all time points. This was compared to 83 slim control individuals (BMI 19–25) (For all details, see Supp. Methods). We have collected BMI and sugar A1C information for the same patients in late time points up to a year and a half post-surgery to track their weight loss and the remission of diabetes.
The slim population was younger and had more males compared to the population who underwent surgery (36 +/- 12 vs 48+/-12 and 50% males vs 29%). Overall, the patients’ mean BMI was reduced from 43.3+/-6.8 (Mean+/-SD) to 27.8+/-1.5, which represents an average loss of 84.7% overweight (compared to BMI 25), blood sugar levels were reduced from Hemoglobin A1C of 6.5+/-0.4 to 5.8+/-0.75 or from blood sugar levels of 125+/-11 g/dL to 95.4+/-10, Triglyceride levels decreased from 183+/-20 to 102+/-13. All parameters described are significantly (P < 0.001) lower from their starting point and not different from the slim control (Fig. 1B-D).
The gut microbiome of all donors was analyzed using 16S rRNA gene sequences (emphasizing the 16SMetaVx.V2) [14], and OTU tables were produced using QIIME2 [15]. The OTU tables were then merged to the genus level. Taxa appearing in only one sample were removed. All samples were log-normalized to highlight the differences in rare bacteria and z-scored. The normalized tables were projected on PCA components. The aim of the first step was to homogenize the description level and reduce the dimension. Since multiple OTUs are associated with the same bacteria, and some OTUs are associated with different levels of classification, we averaged all OTUs associated with the same species in each donor (Fig. 2A-C). Note that while information is lost in the process, such a process is essential for the following machine learning. If an OTU was only present in part of the samples, it was given a value of 0 in all other samples. The resulting values have a scale-free distribution, which often masks large changes in relative frequencies of rare bacteria. To handle that, we log-transformed all OTU values and added a minimal constant value (0.1) to avoid log of zero values. This allows for a narrower distribution of values (Fig. 2D). Finally, given the very high correlation between the relative abundance of different bacteria (Fig. 2E), we projected the bacteria to principal components, which capture most of the variance in the bacterial diversity (Fig. 2F).
The projection on the first principal vector (PC1) clearly delineated axes separating the obese from the slim individuals (Fig. 3A,B). The clear separation of the PC1 projections agrees with observed major differences in the microbiome of slim and obese individuals. The large BMI difference between groups (BMI > 35 in obese versus BMI of < 25 in slim) translates to a large difference in the microbiome. We next tested whether diet or bariatric surgery push back the population toward the slim profile. The results are surprisingly opposite (Fig. 3B). The distance between the projection on the first PC of the post-diet and post-surgery and the slim profile keeps increasing and reaches a maximum after a year. The major difference between these projections allows for a simple classification even with a linear SVM of slim vs obese and pre vs post-surgery samples (Fig. 3C,D ROC curves). The main contributions to the classifiers are from the first two PCs for the healthy (H) vs obese (O) and the same with the 5th PC for the pre vs post-surgery (Fig. 3E). Note that higher test AUC can be obtained by non-linear classifiers. However, the linear classifier gives a clear picture of the contribution of each PC to the microbiome development.
One can then project back the correlations between the PC and the state/BMI to the original OTUs, and find OTUs that are correlated to BMI (PC 1 and 3), the OTUs that are over and underrepresented in obese individuals compared to healthy and the OTUs that change significantly after surgery compared to before surgery (Fig. 4). Interestingly, Helicobacter is strongly associated with high BMI (Fig. 4A,B,C) as it has been reported previously in several studies to be associated with inflammation, insulin resistance and BMI [16–19]. The opposite behavior occurs with the bacterial family Succinivibrionaceae which are highly associated with healthy individuals and the after-surgery state. Members of this family ferment carbohydrates to acetate and succinate (succinate being an important intermediate for propionate production) [20]. Short-chain fatty acids (SCFAs) such as acetate and propionate have multiple beneficial roles. They are known for their contribution to improved insulin sensitivity and glucose homeostasis [21] and for their protective effect against diet-induced obesity [22]. The inflammatory effects of Helicobacter and the SCFA producing capabilities of the Succinivibrionaceae help explain our observation the first is higher in the obese state and the latter are associated with a healthy lean state.
To test for possible confounding effects, we tested whether the observed changes in the profile may be the result of age or gender, or whether they are related to the total BMI. There are no significant correlations with gender and a very weak correlation of PC3 and 5 with age (Fig. 5E, F).
Since the first three PC represent the main effect of the current BMI (i.e., BMI at time of sampling), we tested whether the other PCs represent other aspects of the population. Specifically, we tested if other PCs can be used to predict a future change in BMI. We performed an L1 (Lasso) regression of the projection on the first PCs of the A point and the change in BMI between point A and points E (six months after surgery), 1Y and 1.5Y. The prediction was tested using a Leave One Out (LOO) methodology, and the Spearman correlations on the test values between the predicted change and the observed change (on the LOO test)), as well as the Area Under Curve of a predictor of whether a patient will have a more/less than average reduction in BMI. The AUCs are significant for all points (Fig. 5B-D). The Spearman correlation is highly significant for point E (where we have more information than other points). The prediction of future BMI change is also determined by a single main PC (the 4th PC Fig. 5A and G), further showing the natural decomposition of the microbiome into elements correlated with the host behavior. When looking at which OTUs are contributing to PC4 (Fig. 5A) we found Holdemania to be strongly correlated with future weight loss. Holdemania was previously found to be overrepresented in lean individuals [23]. Coprococcus which also correlates with greater BMI loss in the future is a known butyrate producer [24]. The beneficial effect of butyrate to weight loss may be due to its role as a substrate in gluconeogenesis [25], and recently, oral butyrate supplementation has even been shown to reduce adiposity and improve insulin sensitivity [26].
Another possible candidate to affect the microbiome is the sugar level. We tested the correlation between the projections on the PCA and the A1C. Indeed, a clear correlation is found between A1C and PC7 (Fig. 6B for correlations and 6A for the projections). Using the projection, diabetic and non-diabetic patients, as defined by A1C (above 6.5 and below 5.7) can be separated with a high AUC (0.75 average over test p < 0.01) (Fig. 6B). This correlation might be used for predicting the disease in healthy subjects both from risk groups and in general. The main OTU negatively affecting PC7 are members of Erysipelotrichaceae which are known to be associated with obesity [27] (Fig. 6A).