Non-invasive continuous real-time in vivo analysis of microbial hydrogen production shows adaptation to fermentable carbohydrates in mice

Real time in vivo methods are needed to better understand the interplay between diet and the gastrointestinal microbiota. Therefore, a rodent indirect calorimetry system was equipped with hydrogen (H2) and methane (CH4) sensors. H2 production was readily detected in C57BL/6J mice and followed a circadian rhythm. H2 production was increased within 12 hours after first exposure to a lowly-digestible starch diet (LDD) compared to a highly-digestible starch diet (HDD). Marked differences were observed in the faecal microbiota of animals fed the LDD and HDD diets. H2 was identified as a key variable explaining the variation in microbial communities, with specific taxa (including Bacteroides and Parasutterella) correlating with H2 production upon LDD-feeding. CH4 production was undetectable which was in line with absence of CH4 producers in the gut. We conclude that real-time in vivo monitoring of gases provides a non-invasive time-resolved system to explore the interplay between nutrition and gut microbes in a mouse model, and demonstrates potential for translation to other animal models and human studies.

specific nutrients is far less well established. While a variety of cross-sectional methods can be applied to analyse changes in intestinal microbiota in rodents at specific time points, longitudinal measurements in rodent and human studies mainly rely on sampling of the faeces, long after food-microbiota interactions have already taken place throughout the gastrointestinal tract. Continuous measurements of fermentation gas emissions are already in place for ruminants like cattle and sheep [10][11][12] , as they are known to fully rely on microbiota fermentation in rumen and hindgut to digest cellulose, being distinct from monogastric organisms including rodents and humans. Furthermore, recent studies showed strong correlations between dynamics of metabolite production and microbiota composition and activity in dairy cows 13,14 . Measurements of H 2 and CH 4 as indicators of human gut microbial activity in vivo have been used before [15][16][17][18][19] , but these are in fact single-time-point gas measurements that lack the information that continuous analysis can provide. Therefore, our study objective was to apply a simple non-invasive method to monitor the effect of diet on intestinal microbiota in real time using a human-relevant model, which we envisioned as a powerful tool to better understand the direct impact of nutrition on the microbiota and by extension of diet-microbiota interactions on human health.
C57BL/6J mice are the most widely used model in medical and nutritional health research and have shown their validity in dissecting microbe-host interactions and causality testing. However, analysis of fermentation gases in mice and other rodent models is a largely unexplored area. As is the case in humans, single-time-point measurements of H 2 (refs [20][21][22][23] ) and CH 4 (refs [24][25][26] ) have been reported for mice and rats. This is critical, because only continuous measurements allow to faithfully study the time-resolved kinetics of digestion and metabolism of nutrients reaching the gut microbiota.
Indirect calorimetry makes use of the measurement of oxygen (O 2 ) and carbon dioxide (CO 2 ), as well as food and water intake and locomotor activity, to analyse energy metabolism. We have equipped a commercially available indirect calorimetry system with sensors for H 2 and CH 4 , allowing continuous measurements of release of these gases non-invasively in real time. We applied this extended system to explore the adaptation of gut microbiota to highly-and lowly-digestible carbohydrates. To the best of our knowledge, this is the first time that food-microbiota interactions have been studied continuously, non-invasively and in real time in a murine model.

Results
In vitro reflects in vivo diet digestibility. To confirm the difference in digestibility of the two starches incorporated into our experimental diets (Table 1), an in vitro model that mimics food digestion for the oral, gastric and small intestinal phases was used. The lowly-digestible starch diet (LDD) showed a slower and 14% less complete carbohydrate digestion than the highly-digestible starch diet (HDD; Fig. 1). In addition, we quantified food intake and faecal energy content in female and male mice habituated to the experimental diets (Table 2). Daily faecal mass was increased in all mice fed LDD, whereas faecal energy density was increased in LDD females only. LDD mice lost on average twice as much energy in faeces compared to HDD mice. With similar food and energy intake, the diet digestibility was 6% lower in LDD vs HDD fed mice (Table 2). Taken together, both in vitro and in vivo analyses showed a reduced digestibility of the LDD vs HDD. Gross energy density (kJ g −1 ) a 18.9 19.5 Calculated energy density (kJ g −1 ) b 17.9 17.9 Protein (en%) b 20 20 Carbohydrate (en%) b 55 55 Fat (en%) b 25 25  Measuring H 2 production in real time. Reduced digestibility likely also affects colonic fermentation, for which H 2 has been used as a marker in mice 20 . However, measurement of its continuous production in response to the diet has not yet been possible. We therefore adapted and extended an indirect calorimetry system to allow H 2 and CH 4 production to be studied in real time, by introducing the respective sensors in series with the O 2 and CO 2 sensors already present in the system (Fig. 2a). To determine if the small quantities of H 2 originating from microbial carbohydrate fermentation in mice could be detected by our system, we measured gas concentrations in cages with and without chow-fed mice over 24 h. Stable signals for all gases were seen in the absence of mice (Fig. 2b), and the concentrations were clearly decreased for O 2 and increased for CO 2 in mouse-occupied cages, as expected (Fig. 2c). H 2 increased (Fig. 2c), while CH 4 concentrations were not altered by the presence of a chow-fed mouse in the cage. The adapted indirect calorimetry system was therefore suitable for simultaneous respirometry and H 2 production measurements in real time, however under the conditions tested, CH 4 production appeared to be absent based on measured ambient levels well above the lower detection limit of the CH 4 sensor (Fig. 2b).
H 2 production indicates extent of carbohydrate digestibility. Since the contrasting digestibility of the experimental diets was expected to result in sustained differences in H 2 production as a consequence of fermentation in the large intestine, we fed female and male mice, as a proof-of-concept, either the HDD or the LDD for three weeks and measured H 2 , CH 4 , O 2 , and CO 2 levels continuously during several days (Study 1). Calculation of energy expenditure, based on 24 h O 2 consumption and CO 2 production, revealed no differences between dietary groups (females 1.59 ± 0.08 vs 1.63 ± 0.08 kJ h −1 in HDD and LDD respectively, P = 0.2094; males 1.80 ± 0.13 vs 1.76 ± 0.13 kJ h −1 in HDD and LDD respectively, P = 0.5470). However, 24 h mean respiratory exchange ratio (RER) was lower in LDD-vs HDD-fed male mice (0.85 ± 0.03 vs 0.88 ± 0.03 respectively, P = 0.0097), indicating higher fat oxidation and lower carbohydrate oxidation in LDD mice. Overall, these observations agree with indirect calorimetry data reported for mice fed diets containing carbohydrates similar to the carbohydrates used here 27 . Both LDD-fed females (Fig. 3a,b) and males (Fig. 3d,e) constantly produced more H 2 than HDD-fed mice. A distinct pattern of H 2 production became apparent in LDD-fed mice, with H 2 levels being higher in the active dark phase and lower, but still clearly present, in the inactive light phase (Fig. 3a,b,d,e). This was fully consistent Figure 1. In vitro digestibility of starches in experimental diets. Triplicate samples of the lowly-and highlydigestible starch diets (LDD and HDD, respectively) were digested in vitro, and free glucose concentrations were determined at indicated time points. Statistical comparisons were made with two-way ANOVA with Bonferroni's post hoc test; ***P ≤ 0.001. Values are plotted as mean ± s.d. with the circadian food and starch intake (Fig. 3c,f). Importantly, the difference in H 2 production between HDDand LDD-fed mice was explained by the type of starch rather than the amount of starches ingested, as cumulative starch consumption was similar between the groups (Fig. 3c,f). Together, this data provides proof-of-concept for measuring H 2 production in real time as an indicator of carbohydrate digestibility.

Females
H 2 evolution reflects adaptation to dietary carbohydrates. As we could show that H 2 production can be sensitively and continuously measured, we next questioned whether it would be possible to measure adaptation to the diet in vivo in real time. For this, we provided HDD or LDD to mice that had no previous exposure to these diets and we followed H 2 production continuously. We introduced the new diets in one of two conditions; the first condition was as a single meal challenge given to fasted mice, followed by ad libitum access to the diet the next day, as a second fasting-refeeding challenge (Study 2, Fig. 4a). The second condition was by replacing the standard chow diet directly with HDD or LDD ad libitum (Study 3, Fig. 4b). H 2 production was significantly increased in LDD-compared to HDD-fed mice as early as 4 h after fasted mice gained ad libitum access to the experimental diet (Fig. 4a). The direct switch from chow to HDD or LDD without fasting had similar results, with LDD-fed mice producing significantly more H 2 after 53 h of access to the LDD compared to mice receiving HDD (2-way ANOVA, Fig. 4b). In both conditions, i.e. fasted or directly switched to HDD or LDD, cumulative H 2 production became significantly higher already within 12 h upon access to LDD vs HDD (Fig. 4c,d), and H 2 production patterns in LDD-fed mice closely followed the patterns of LDD intake (Additional File 1: Fig. S1). Interestingly, mice that continued on the chow diet after a period of food restriction exhibited a spike in H 2 production (Fig. 4a), while consuming similar amounts of starches compared to the HDD and the LDD groups. H 2 production in HDD-fed mice remained lower compared to mice on LDD or chow, as expected. Importantly, LDD-induced H 2 production increased gradually before reaching its maximal levels (up to 0.89 ml h −1 ), revealing the process of adaptation to the lowly-digestible starch. LDD-vs HDD-fed mice thus showed a differential adaptation, likely in their microbiota, based on increased H 2 production.

Alterations in intestinal microbiota by dietary carbohydrates.
Since the production of H 2 fully depends on intestinal microbial communities and their metabolism, we further investigated the changes in the microbiota induced by the LDD to validate our observations. As an additional parameter of fermentation, we first assessed SCFA levels in intestinal digesta after 3 weeks of exposure to the HDD or the LDD (Study 1). Total caecal SCFA levels were similar between LDD-and HDD-fed mice (35.6 ± 13.9 vs 34.9 ± 11.9 μmol g −1 , respectively), including valeric and isobutyric levels (data not shown), whereas total SCFA in colon were higher in LDD-compared to HDD-fed mice (25.6 ± 9.6 vs 9.6 ± 4.1 μmol g −1 , P = 0.0059). Acetic acid (Fig. 5a) and propionic acid ( Fig. 5b) were the two most abundant SCFA, and both were significantly elevated in LDD-fed mice in colon and caecum contents, respectively. Butyric acid was the most differentially produced SCFA, enriched by 13.8-fold in LDD colon content (Fig. 5c). Finally, isovaleric acid, a product of microbial protein fermentation, was the least abundant of the measured SCFA in all groups and was significantly lower in caecum of LDD-vs HDD-fed mice (Fig. 5d).
We next compared the overall changes in faecal microbiota communities induced by HDD or LDD after exposure to the diets for 3 weeks (Study 1) and 4.5 d (Study 3). Principal coordinates analysis (PCoA) using the UniFrac unweighted distance matrix revealed a clear separation between the two dietary groups (Fig. 6a). These observations were supported by Adonis analysis, using either weighted or unweighted UniFrac distances, as diet explained a large part of the variation in microbiota composition (20% and 29%, respectively, P < 0.001, Table 3). H 2 volume was the second most important variable, followed by duration of intervention and age, and body weight, with minor but significant effects (Table 3). Of note, duration of intervention and age of mice are dependent variables due to study design. In order to control for the effects of duration of dietary exposure, we also analysed Studies 1 and 3 separately. After 3 weeks of intervention, diet and H 2 production were the only significant variables, with H 2 explaining up to 34% of the variation ( Fig. 6b and Table 3). However, only diet contributed significantly to the variation after 4.5 d of intervention in adult mice ( Fig. 6c and Table 3). Additionally, α-diversity appeared to decrease with duration of intervention irrespective of the dietary intervention, with no consistent effects of the diet (Additional File 2: Fig. S2). This is in line with the differences in age of these mice, namely the young mice showing lower α-diversity than the older mice.
We then aimed to identify which microbial taxa were significantly associated with the observed differences in β-diversity. The microbiota of mice fed LDD vs HDD for 3 weeks was enriched in Bacteroides, Parasutterella, Roseburia, and Alloprevotella, along with two other families ( Fig. 7a and Additional File 3: Fig. S3a). In comparison, Lactobacillus, Rikenella, Odoribacter, Enterorhabdus, and Desulfovibrio among others appeared enriched in HDD-vs LDD-fed mice ( Fig. 7a and Additional File Fig. S3b). Similar differences were seen after 4.5 d of exposure ( Fig. 7b and Additional file 3: Fig. S3c,d). While fewer taxa were affected by the short-term dietary intervention, changes in genus level were consistent for both groups (Additional file 4: Fig. S4). Moreover, H 2 production was the only (environmental) variable that was significantly correlated with specific bacteria taxa after three weeks of intervention, with five genera correlating positively with H 2 production and eight genera showing a negative correlation ( Fig. 8 and Additional file 5: Fig. S5). Eleven of these 13 genera were also significantly influenced by diet ( Fig. 7a and Additional file 3: Fig. 3a,b). Finally, Archaea (some of which are CH 4 producers) could not be detected in any of the samples despite the use of primers targeting both bacterial and archaeal 16S rRNA genes equally well. This agrees with the absence of CH 4 detection in these mice and under these nutritional challenges.

Discussion
The goal of this study was to measure real-time interactions between diet, gut microbes, and the host. We implemented H 2 and CH 4 detection in an indirect calorimetry system to track fermentation continuously in mice. H 2 production revealed a time frame for microbiota adaptation to starch of low digestibility, which corresponded with shifts in microbial community composition induced by diet. Thus, measuring H 2 production allowed us to non-invasively study effects of the diet on the intestinal microbiota in real-time.
The difference in starch digestibility as part of the experimental diets was confirmed both by in vitro and in vivo measurements, but did not significantly alter total intake of digestible energy (gross energy minus faecal energy losses) between dietary groups within a sex. The lower digestibility of the starch in the LDD thus suggests that partially undigested starch reached the large intestine which was subsequently partially fermented by the intestinal microbiota providing energy substrates, e.g. SCFA, to the host. Energy of undigested starch can be lost after fermentation in the form of products not utilizable by the host, such as H 2 . However, previous studies considered energy loss in the form of H 2 and CH 4 negligible, representing less than 0.2% of total energy expenditure in humans consuming non-starch polysacharides 28 . Studies in rats fed various types of resistant starch also indicated that energy loss through fermentation is minimal, although the actual H 2 output was not measured directly 29 . Here, our data show that H 2 is produced constantly on a lowly-digestible starch diet. Although the volume of H 2 produced by the mice in our study may be little in terms of energy loss, it is plausible that carbohydrates that give a higher level of fermentation could further increase the H 2 output, which might represent a significant factor to take into account over a lifetime.
H 2 production was detected in mice under all conditions tested, with the amounts produced clearly being influenced by the form of carbohydrate consumed. Mice fed moderately fermentable carbohydrates have been shown to produce H 2 (ref. 20 ). Even in conditions where little fermentation is expected, such as feeding corn starch-based chow 30 or pure sucrose 31 , H 2 production has been seen in rats. In line with our data on three diets with a different carbohydrate profile, this illustrates that H 2 production can directly reflect subtle changes in carbohydrate fermentation. Interestingly, H 2 production was clearly associated with the food intake pattern. This is in contrast with data reported in humans, where H 2 and CH 4 peaked at rather unpredictable times after food intake despite the proper control of the meal schedule 28 . This might be due to e.g. differences in dietary meal composition, time resolution of sampling, intestinal transit time, or other differences in intestinal physiology between humans and mice. More recently, using gut capsule technology, a similar H 2 pattern as in our mice was also observed in a human pilot trial based on dietary fibre differences 32 .
We demonstrated that real-time monitoring of H 2 production can be used to investigate transient effects of diet in time and explore the process of adaptation rather than the end stage only. So far, studies mainly investigated, by measuring H 2 and other fermentation parameters at selected time points, longer timeframes ranging from 1 day to several weeks 25,[33][34][35] . In our study, significant differences in H 2 production appeared within 12 h upon access to LDD. This timeframe was clearly influenced by fasting and whether the diet was provided ad libitum or in a restricted amount. We speculate these early increases in H 2 output to reflect immediate effects of diet on microbial metabolism preceding changes in community structure. Another observation was that mice fed chow produced H 2 , although at low levels. Real-time monitoring newly revealed that a period of food restriction followed by refeeding led to a marked and acute increase in H 2 production once chow became available again. A likely explanation is excessive eating after food deprivation, causing a larger amount of not fully digested chyme to enter the large intestine and thus increasing substrate availability to the microbiota. In addition, a 24 h fasting period alone has been shown to produce shifts in microbial diversity 36 and microbiota configuration 37 . Such changes could in turn alter fermentation stoichiometry and microbial function in response to the diet and ultimately lead to a higher H 2 output. Our analysis indicates (short-term) effects of fasting and refeeding on microbial activity, which should be careful taken into account in nutritional studies focussing on changes in microbiota composition and function.
As could be expected, the driver of the experimental differences, the dietary starch digestibility, was the most important factor explaining the variation in microbiota, showing colinearity with our measured in vivo H 2 production. Although current knowledge of the dynamics of H 2 within the gastrointestinal tract is limited, it is well documented that H 2 is exclusively produced during fermentation by hydrogenogens 6 . Among the major hydrogenogenic bacteria are Bacteroidetes and clostridial members of Firmicutes 38 . In line with this, we observed that LDD, a source of carbohydrates for caecum and colon, stimulated the fermentative Bacteroidetes bacteria, more specifically the genus Bacteroides. This is consistent with the dose-dependent increase in caecal Bacteroidetes density in response to amylose 39 and similar findings for amylose on a high-fat background 40 . Here we extend these findings and show, for the first time in vivo, a positive correlation between Bacteroides and H 2 production in mice.
Interestingly, after the short-term exposure to LDD in adult mice, Bacteroidetes were not significantly increased compared to the HDD group. This might be associated with the shorter duration of the treatment and possibly with more firmly established microbial communities in adulthood. However, most genera induced by diet in adult mice after 4.5 days correspond to those induced in mice after weaning, which were exposed for 3 weeks.
Another consistent shift in microbial community composition was the promotion of Deltaproteobacteria, particularly Desulfovibrio and Bilophila, in HDD-fed mice. Deltaproteobacteria are the major representatives of colonic sulphate reducing bacteria (SRB) 41 including Desulfovibrio. SRB along with methanogens and acetogenic bacteria are the only gut microbes able to use H 2 as an electron donor to produce H 2 S and acetate. Although not a SRB itself, taurine-respiring Bilophila species can also produce H 2 S. Additionally, there is evidence of CH 4 production in rats 26 and mice, with the presence of methanogens in humanized microbiota mouse models 42 and by high fat dietary feeding 43 . The fact that we neither detected CH 4 nor Archaea suggests that H 2 was preferentially used to produce H 2 S in mice fed readily digestible starch. H 2 S is a potentially toxic product of bacterial metabolism [44][45][46] and it has been implicated in human health and disease 19 and, more recently, thermogenesis 47 . Moreover, H 2 S has been reported to inhibit the production of SCFA and specifically to impair butyrate oxidation, depriving colonic cells from their main energy source 45,48 . In line, we report a dramatic difference in colonic butyrate in HDD-fed mice. Apart from Deltaproteobacteria, we observed increased abundances of Odoribacter, a known H 2 S producer 49 and Rikenella, a desulphatase-secreting bacterium 50 , under HDD-feeding. Members of the genus Rikenella are able to cleave sulphate from mucin glycans, making them available for microbial degradation 51 and potentially acting as a  donor of sulphate to H 2 S producers. Based on these facts we speculate that the lack of fermentable carbohydrates favoured the presence of hydrogenotrophs associated with the production of H 2 S, which could have led to the decreased H 2 output and colonic SCFA levels that was observed in mice fed highly-digestible starch. The major taxon increased in HDD-fed mice in our study belonged to the genus Lactobacillus. In contrast, diets supplemented with resistant starch tended to enrich the Lactobacillus population in mouse caecum, but much less at high doses of resistant starch 39 . Incidentally, hydrogenase genes, which encode enzymes for the reversible oxidation of H 2 , were recently shown to be completely absent in Bacilli and bifidobacteria 38 . Considering the lack of a correlation between H 2 production and Lactobacillus in our study, new questions emerge about the ability of Lactobacillus to thrive in H 2 -poor environments.
The increase in isovaleric acid, a product of branched-chain amino acid catabolism 52 , in HDD-fed mice, suggests a shift of microbiota towards protein fermentation. Bacteria from the genera Enterorhabdus 53 and Parvibacter 54 , both significantly induced by HDD-feeding, have the ability to ferment amino acids. Additionally Olsenella, only present in two samples in the HDD group, is documented to grow on tyrosine and produce p-cresol 55 , supporting our hypothesis for a shift to protein fermentation. This might have important implications for the host, since products of protein fermentation such as phenols, ammonia, certain amines, and H 2 S, are considered to play important roles in the initiation or progression of bowel diseases, inflammation, DNA damage, and cancer 56 .
Altogether, our results emphasize H 2 as a key factor within the intestinal microenvironment and the usefulness of knowing its production dynamics to understand the interplay between host, diet, and the intestinal microbiota. At the same time, we are aware that our approach to study such interactions may have conceivable limitations. It has been argued that changes in gas evolution (and other indirect markers of fermentation) cannot accurately indicate changes in fermentation 57 , and even "real-time", carefully controlled measurements have failed to show quantitative changes in H 2 and CH 4 production proportionally linked to consumption of fermentable carbohydrates 15,28 . We completely agree with these authors that the measured outcomes, H 2 and CH 4 , not only reflect the type of carbohydrate consumed, but are the end result of a very complex fermentation stoichiometry that depends on the host's capacity to digest and absorb nutrients, the dominance and metabolic activity of microbial species, and their interactions. However, the conclusion that fermentation gases are extremely limited parameters to study carbohydrate fermentation is largely based on human data, where eating pattern, environment, genetic variation, and the gut microbe interact and ultimately determine an individual's response to the diet 58,59 . When these and other factors can be better controlled, as it is the case with animal models, the analysis of carbohydrate fermentation through H 2 and CH 4 quantification has much to offer. The fact that in vitro models to measure H 2 and CH 4 evolution are still developing and proposed as a tool to unravel the mechanisms behind the association between microbiota and host health 60 is encouraging.
Overall, the applications of gas analysis within an indirect calorimetry system go beyond the arena of carbohydrate quality and nutritional studies, and may be used as a diagnosis tool in clinical practice 19,61,62 . It opens up new avenues not only in preclinical research in rodents, but also has potential in human-diet-microbiota interaction studies if such sensor technology is incorporated into indirect calorimetry chambers or ventilated hood systems.

Conclusions
Using our customized indirect calorimetry system we were able to continuously quantify H 2 production in mice as a reflection of the starch digestibility of the diet. H 2 monitoring also allowed us to catch the earliest stages in the adaptation to carbohydrates of different digestibility, revealing a nuanced process with high inter-individual variation. Importantly, in vivo H 2 production was significantly correlated with specific microbial taxa, including Bacteroides and Parasutterella. The implemented H 2 and CH 4 sensor-technology described here opens yet unmet avenues to study the effects of nutrition on microbiota in real time, not only in rodents, but potentially also in humans. Figure 8. Specific bacterial genera correlate only with in vivo H 2 production. Spearman's rank correlation coefficients of faecal microbiota, H 2 production, food and starch intake, body weight, and fat mass of mice exposed to HDD or LDD for 3 weeks after weaning (n = 12 per diet, Study 1). Non-red and non-blue cells all have a Spearman's correlation value of 0 with FDR P value > 0.13. Nomenclature of microbial genus level taxa is based on highest achievable taxonomic resolution at phylum, class, order, family or genus level.

Methods
Coupling of hydrogen (H 2 ) and methane (CH 4 ) sensors into indirect calorimetry system. A PhenoMaster indirect calorimetry system (TSE Systems, Bad Homburg, Germany) was extended by coupling a Sensepoint XCD H 2 gas analyser (Honeywell Analytics, Hegnau, Switzerland) and a CH 4 gas analyser (ABB Automation GmbH, Frankfurt am Main, Germany) in a closed circuit in series in front of a Siemens High-Speed Sensor Unit containing the O 2 and CO 2 sensors. This order was chosen to prevent dilution of the sample with reference air, which is required by the Siemens unit. The H 2 sensor has a stability of <±2% full scale deflection (fsd)/yr representing <2 ppm/yr as it was adjusted to a measuring range from 0 to 100 ppm. The CH 4 sensor has a zero drift of ≤1% of span per week and a measuring range from 0 to 500 ppm. A two point calibration of both H 2 and CH 4 analysers was performed within 24 h before each animal experiment. The calibration procedure was carried out using three gas mixtures (Linde Gas Benelux BV, Dieren, The Netherlands): zero (20.947% O 2 and N 2 ), span H 2 (98.8 ppm H 2 and synthetic air), and span CH 4 (0.521% CO 2 , 450 ppm CH 4 , and N 2 ). The zero calibration mixture was flushed through the system for 10 min and ADC signals were assigned H 2 and CH 4 values of 0 ppm. Thereafter, each of the span gas mixtures was run for 10 min and ADC signals assigned 98.8 ppm H 2 and 450 ppm CH 4 , accordingly. For O 2 and CO 2 calibration, the routine indicated in the TSE manufacturer's manual was followed, using an additional gas mixture (0.999% CO 2 and N 2 ) for the span calibration point. Animals were measured as previously described 63  were housed in Makrolon II cages enriched with wood chips and wood shavings, with free access to drinking water, at 23 °C ± 1 °C and a 12:12 h light:dark cycle. Standard rodent chow (RMH-B, AB Diets, Woerden, The Netherlands) was provided exclusively and continuously since weaning, unless specified. Three different studies were conducted to investigate diet-host-microbiota interactions upon provision of diets containing starches with differences in digestibility (the experimenter was not blinded to the diets that the animals were given).

Study 1 (long-term exposure, post-weaning).
Mice were mated and the offspring reassigned to a foster dam 1 or 2 days after birth to obtain standardized litters. Males and females were stratified by body weight at post-natal day (PN) 21, housed individually and randomly assigned to be fed a highly-or a lowly-digestible starch diet (HDD and LDD, respectively; see below). The randomisation was achieved by generating a column of random numbers in a spreadsheet and sorting each diet and animal number according to the column of random numbers from smallest to largest. From PN36-42, a subgroup of mice was measured in the indirect calorimetry system with ad libitum access to the experimental diets (males: n = 12 per diet, females HDD n = 12, LDD n = 11). Fresh faecal pellets were sampled on PN39 (n = 6 per diet and sex) and stored at −80 °C for intestinal microbiota analysis. Another subgroup of female mice was culled on PN42 for collection of caecum (n = 6) and colon contents (n = 5 HDD, n = 7 LDD), and the faeces produced during the last week before sacrifice were collected for gross energy measurements (see In vivo diet digestibility). Before sacrifice, food was removed 1 h after the start of the light phase and animals decapitated 2-6 h after removal of food. Caecum and colon contents were immediately frozen in liquid nitrogen, and stored at −80 °C until analysis.
Study 2 (short-term exposure with fasting, adult). Eight-month-old female mice were individually housed in indirect calorimetry cages. After a 2-day adaptation period, mice were allowed a restricted amount (1.1 g) of chow 1 h before the onset of the dark phase to induce a fasting state in early morning, as published 64 . At the end of the light phase at 18.00 h, mice were re-fed with a restricted amount (1.1 g) of chow, or first-time HDD or LDD (the refeeding diet was assigned randomly; n = 4 per dietary group). Shortly before the following dark phase mice received access to the same diet they were allocated the day before, but now ad libitum. Indirect calorimetry measurements continued for an additional 5.5 d.
Study 3 (short-term exposure without fasting, adult). Ten-month-old female mice were individually housed in indirect calorimetry cages. After a 2-day adaptation period, mice were provided clean bedding and given ad libitum first-time access to either HDD or LDD (random assignment, n = 6) shortly before the dark phase and for the remaining experimental period. Measurements continued for an additional 4.5 d. Faecal pellets produced after the introduction of the new diets were collected from the bedding at the end of the experiment and stored at −80 °C. Experimental diets. Both the HDD and the LDD satisfy the nutrient requirements for rodent growth and lactation (AIN-93G) 65 , with appropriate levels of mono-and poly-unsaturated fatty acids. The macronutrient composition was 20.1 energy percentage (en%) protein, 54.9 en% carbohydrates, and 25 en% fat (Table 1), with starch being the sole source of carbohydrates. The starch fraction (Cargill BV, Sas van Gent, The Netherlands) of the HDD was composed of 100% amylopectin (which is highly digestible), while that of the LDD was a mixture of 60% amylose (which resists complete digestion) and 40% amylopectin. The diets were pelleted by Research Diet Services BV, Wijk bij Duurstede, The Netherlands.
In vivo diet digestibility. Total faeces produced from PN 35-42 (Study 1) were recovered from the bedding of a subgroup of randomly selected animals (n = 4 per sex and diet). Food intake was recorded over the same period. Gross energy in faeces and food was determined in blinded samples using a C7000 bomb calorimeter (IKA, Staufen, Germany) and diet digestibility was calculated as published 66 . In vitro carbohydrate digestibility. The in vitro digestibility of starches in the experimental diets was determined in blinded samples in triplicate, as published 67 . Briefly, 5 intact pellets of each diet were cryoground to homogeneous particle size and weighed separately into 3 tubes (70 mg). Each sample was digested in a 15-ml tube by adding cocktail solutions (modified from Versantvoort et al. 68 ) and digestive enzymes at 37 °C in three sequential steps to represent the oral (5 min), gastric (2 h), and duodenal (6 h) phases. Two blanks containing only enzymes and solutions were included. Samples were taken at several time points during the gastric and duodenal phases and centrifuged. Clean supernatants were recovered and free glucose content was determined by the glucose oxidase peroxidase method 69 . Starch digestion was expressed as the percentage of total glucose released based on the amount of starches in the diets.
Quantification of SCFA in intestinal digesta by gas chromatography (GC). Short-chain fatty acids in caecum-and colon-contents were determined as previously reported 70 , with some modifications. Samples (about 25 mg) were weighed, thawed, homogenized in 100 μl of ultrapure water, and centrifuged for 3 min at 21,382 g. To 50 μl of supernatant, 100 μl of 2-ethylbutyric acid solution (0.45 mg ml −1 ) were added as internal standard. An external standard curve was prepared containing 50 μl of a mixture of acetic, propionic, butyric, valeric, isobutyric, and isovaleric acid at concentrations ranging from 0.002 mg ml −1 to 0.8 mg ml −1 , to which 100 μl of internal standard were added. Blanks containing only water or water and internal standard were included for quality control. HCl and oxalic acid were added to all samples, blanks, and standards in order to protonate the SCFA. Gas chromatography was performed on a FOCUS GC apparatus coupled to a flame ionization detector (Interscience, Breda, The Netherlands). Samples were injected (1 μl) into an CP-FFAP CB column (25 m × 0.53 mm × 1.00 μm; Agilent Technologies, Santa Clara, CA, USA). Helium served as carrier gas at a pressure of 40 kPa. The initial oven temperature was 100 °C with 0.5 min hold, ramped to 180 °C at 8 °C min −1 with 1 min hold, and finally ramped to 200 °C at 20 °C min −1 with 5 min hold. Peak identities and areas were analysed with Xcalibur software (version 2.2; Thermo Scientific, Waltham, MA, USA). Concentrations of SCFA were normalised to the internal standard and expressed relative to original sample weight.

Microbiota analysis.
Microbial DNA was isolated from faecal pellets using the Maxwell ® 16 Instrument (Promega, Leiden, The Netherlands). Faecal pellets were added to a bead-beating tube with 350 μl Stool Transport and Recovery (STAR) buffer, 0.25 g of sterilized zirconia beads (0.1 mm), and three glass beads (2.5 mm). Faecal pellets were homogenized by bead-beating three times (60 s × 5.5 ms) and incubation for 15 min at 95 °C at 100 rpm. Samples were then centrifuged for 5 min at 4 °C and 14,000 g and supernatants transferred to sterile tubes. Pellets were re-processed using 200 μl STAR buffer and both supernatants were pooled. DNA purification was performed with a customized kit (AS1220; Promega) using 250 μl of the final supernatant pool. DNA was eluted in 50 μl of DNAse-RNAse-free water and its concentration measured using a DS-11 FX+ Spectrophotometer/Fluorometer (DeNovix Inc., Wilmington, USA). The V4 region of 16S ribosomal RNA (rRNA) gene was amplified in duplicate PCR reactions for each sample in a total reaction volume of 50 μl. The master mix contained 1 μl of a unique barcoded primer, 515F-n and 806R-n (10 μM each per reaction), 1 μl dNTPs mixture, 0.5 μl Phusion Green Hot Start II High-Fidelity DNA Polymerase (2 U/μl; Thermo Scientific, Landsmeer, The Netherlands), 10 μl 5× Phusion Green HF Buffer, and 36.5 μl DNAse-RNAse-free water. The amplification program included 30 s of initial denaturation step at 98°C, followed by 25 cycles of denaturation at 98 o C for 10 s, annealing at 50 °C for 10 s, elongation at 72 °C for 10 s, and a final extension step at 72 °C for 7 min. The PCR product was visualised in 1% agarose gel (~290 bp) and purified with CleanPCR kit (CleanNA, Alphen aan den Rijn, The Netherlands). The concentration of the purified PCR product was measured with Qubit dsDNA BR Assay Kit (Invitrogen, California, USA) and 200 ng of microbial DNA from each sample were pooled for the creation of the final amplicon library which was sequenced (150 bp, paired-end) on the Illumina HiSeq. 2000 platform (GATC Biotech, Constance, Germany).
Microbiota data processing and analysis. Data filtering and taxonomy assignment were performed using the NG-Tax pipeline 71 . Briefly, an OTU table was created for each sample with the most abundant sequences. Low abundance OTUs were discarded, using a minimum relative abundance threshold of 0.1%. Two distinct in-house assembled mock communities were included in the library and were compared with their theoretical composition for quality control (Additional file 6: Fig. 6). Calculations for αand β-diversity analyses were performed using the publicly available Microbiome R package (version 1.2.1) 72 . Adonis permutational multivariate analyses of variance (PERMANOVA) using either the weighted or unweighted Unifrac distances were performed with the Vegan package (version 2.5-2) and were used to determine the amount of variation explained by the grouping variables. Linear Discriminant Analysis (LDA) Effect Size (LEfSe) was applied to determine the differences between the microbial communities of HDD-and LDD-fed mice using a publicly available pipeline (http://huttenhower.sph.harvard.edu/galaxy/) 73 ; the threshold for the logarithmic LDA score was set to 2.0. P values for Kruskal-Wallis and Wilcoxon tests for the LEfSe analysis were set to 0.05. For non-parametric Student's t-tests, reads were transformed to their relative abundances and tests were carried out with 999 permutations using QIIME (version 1; http://qiime.org/index.html) 74 . Statistical significance was determined using the Benjamini-Hochberg false discovery rate (FDR) adjustment. Data analysis. Statistical analysis was performed using GraphPad Prism 5.04 (GraphPad, San Diego, CA, USA), unless stated otherwise. All data was tested for normality using the D' Agostino and Pearson omnibus test and its distribution was normalized by log transformation when applicable. Comparisons between two groups were Scientific REPORTS | (2018) 8:15351 | DOI:10.1038/s41598-018-33619-0 made using unpaired two-tailed Student's t-tests (for data with normal distribution) or two-tailed Mann-Whitney U tests (VH 2 during light phase between HDD and LDD). Comparisons between more than two groups were made by one-way analysis of variance (ANOVA) with post hoc Bonferroni's test for multiple comparisons. Time course data (H 2 evolution) was analysed by repeated measures two-way ANOVA with Bonferroni's post hoc test. When sample sizes being compared were similar and relatively large (n > 5), similarity of variances was not taken into account. All data is reported as mean ± s.d. Statistical significance was set at 5%, with levels indicated as *P ≤ 0.05, **P ≤ 0.01, ***P ≤ 0.001, and ****P < 0.0001. Sample size was not determined statistically as the effect size was unknown, but it was based on our previous results on the use of indirect calorimetry to assess metabolic flexibility 64,75 . Ethics approval and consent to participate. All animal experiments were approved by the Animal Experiments Committee (DEC 2014085.h) and performed in accordance to EU directive 2010/63/EU.

Data Availability
The 16S rRNA gene sequencing dataset supporting the conclusions of this article is available in the European Nucleotide Archive (ENA) database with accession code PRJEB23475 at http://www.ebi.ac.uk/ena/data/view/ prjeb23475. The authors declare that all other data supporting the findings of this study are available within the paper and its additional files 1-6, or from the corresponding authors upon reasonable request.