Abstract

   Ketogenic diets (KDs)—very-low-carbohydrate and very-high-fat
   diets—have gained popularity as therapeutic against obesity and type 2
   diabetes. However, their long-term effects on metabolic health remain
   understudied. Here, we show that, in male and female mice, a KD
   protects against weight gain and induces weight loss but over time
   leads to the development of hyperlipidemia, hepatic steatosis, and
   severe glucose intolerance. Unlike mice on conventional high-fat diet,
   KD-fed mice remain insulin sensitive and display low-insulin levels.
   Hyperglycemic clamp and ex vivo glucose-stimulated insulin secretion
   assays revealed systemic and cell-intrinsic impairments in insulin
   secretion. Transcriptomic profiling of islets from KD-fed mice
   indicated endoplasmic reticulum (ER)/Golgi stress and disrupted
   ER-Golgi protein trafficking, which were confirmed by electron
   microscopy showing a dilated Golgi network consistent with defective
   insulin granule trafficking and secretion. Together, these results
   suggest that long-term KD leads to multiple aberrations of metabolic
   parameters that caution their systematic use as a health-promoting
   dietary intervention.
     __________________________________________________________________

   Although a keto diet limits weight gain and evokes weight loss in obese
   mice, it impairs metabolic health in males and females.

INTRODUCTION

   A ketogenic diet (KD) is a very-high-fat, low-carbohydrate diet that
   has been used to manage refractory epilepsy for the past hundred years
   ([43]1, [44]2). The exact mechanisms by which a KD confers seizure
   control are unknown, but there are a few hypotheses including lowering
   and stabilizing blood glucose (BG), inducing favorable changes to brain
   metabolism, neurotransmitter levels and synaptic activity, and
   anticonvulsant effects of ketone bodies themselves ([45]3–[46]5). While
   a KD does not have a defined macronutrient content, as a treatment for
   epilepsy, it is generally first prescribed at a 4:1 ratio of 4 g of fat
   for every gram of carbohydrate or protein translating to ~90% of
   calories coming from fat ([47]2, [48]6). This high proportion of energy
   intake from fat causes the body to use fat rather than glucose as the
   primary fuel source and to enter a state of increased ketone body
   production, known as ketosis ([49]1, [50]7).

   Ketone bodies are produced primarily in the liver when glucose and
   insulin levels are low such as during prolonged fasting or when
   consuming a ketogenic diet. In the 1960s, it was proposed that this
   state of low glucose and insulin, associated with the consumption of a
   KD, could treat obesity and related metabolic health conditions by
   favoring fat usage over fat storage ([51]7). In addition, for similar
   insulin- and glucose-lowering reasons and potential benefits of ketone
   bodies themselves, a KD has been proposed to treat cancer and
   Alzheimer’s disease, increase longevity, and improve metabolic health
   ([52]8, [53]9). However, a breadth of evidence to support these claims
   is lacking. A recent study found that a KD increases cellular markers
   of aging in mice ([54]10), and in 2019, Journal of the American Medical
   Association (JAMA) released a publication titled “The Ketogenic Diet
   for Obesity and Diabetes—Enthusiasm Outpaces Evidence” ([55]11). While
   a KD successfully control seizures in some people, its benefits will
   not necessarily translate to treating other diseases. In particular,
   pre-existing conditions—such as obesity, diabetes and
   hyperlipidemia—might interact with a KD in ways that have not been
   fully characterized in the predominantly young and limited group of
   epilepsy patients that have typically been studied ([56]6, [57]12).
   Therefore, harmful interactions between a KD and metabolic health
   conditions may negate the generally accepted belief that a KD is a
   clinically safe dietary intervention. From a nutritional and
   cardiometabolic standpoint, it is conceivable that there could be
   long-term ramifications to consuming a diet primarily composed of fat
   with very few carbohydrates that have been largely underexplored
   ([58]13–[59]17).

   Previous studies in mice fed a KD have reported conflicting findings on
   metabolic health. For example, compared to mice on a chow diet, most
   studies observed a lower body weight (BW) on KD ([60]18–[61]22), but
   others saw weight gain ([62]9, [63]23, [64]24). In contrast to the
   weight benefits found in many KD studies, high levels of blood
   cholesterol and nonesterified fatty acids (NEFAs) have been observed in
   multiple strains of mice on KD ([65]18, [66]20, [67]22, [68]25,
   [69]26). Furthermore, the effects of a KD on glucose homeostasis have
   led to discordant and partial results. While some studies reported
   improvements in static BG without a glucose challenge ([70]21, [71]25,
   [72]27) and improved glucose tolerance not only in wild type but also
   obese and diabetic models (ob, db, and stz) ([73]19, [74]28), others
   reported insulin resistance ([75]29), glucose intolerance ([76]24,
   [77]30, [78]31), and loss of β cells ([79]30, [80]31). A recent study
   reported diminished glucose-stimulated insulin secretion (GSIS) in 10
   of the 13 inbred strains of mice ([81]32). The inconsistencies in the
   macronutrient compositions of different KDs, together with differences
   in total caloric intake—since many studies use a calorie-restricted
   KD—and duration of the feeding protocol could potentially explain
   incongruencies between studies. Overall, possible cardiovascular risks
   associated with heightened lipid levels and inconclusive results on
   glucose homeostasis warrant further investigations into the long-term
   effect of KD consumption to define the conditions in which it can be
   helpful or harmful.

   Another important factor when considering the effects of KD resides in
   the study design, namely, whether a KD was used to prevent versus treat
   metabolic health conditions. In humans, KD is a very popular weight
   loss (WL) intervention, an effect that has been replicated in mice
   ([82]18, [83]33, [84]34). However, WL per se may not be sufficient to
   normalize metabolic complications. For example, serum dyslipidemia and
   liver steatosis seem to persist despite reductions in weight under KD
   ([85]18, [86]33). Moreover, the effects of KD-induced WL on glucose
   regulation are also unclear: Some studies noted improvements in glucose
   tolerance testing ([87]18, [88]33, [89]34), while some did not
   ([90]35). Noticeably, hypoglycemia, a known complication of treating
   type 2 diabetes, was observed in mice on KD ([91]28, [92]33, [93]34),
   which raises concerns about the safety of using KD in a population
   taking diabetes medication ([94]36). Last, studies of KD-induced WL
   describe a reduction in high-fat diet (HFD)–linked hyperinsulinemia,
   yet we noted that in most, insulin levels dropped below that of
   chow-fed animals, which could be detrimental in the context of
   regulating fat storage and gluconeogenesis. Our search for preclinical
   studies characterizing KD-mediated WL only identified five studies thus
   far ([95]18, [96]28, [97]33–[98]35). With the mainstream popularity of
   KD as a therapeutic for obesity and diabetes, it is critical to further
   establish the broad metabolic effects of WL under KD, its effects on
   pre-existing metabolic conditions, and possible sex differences in
   these responses using a well-defined diet and a systematic metabolic
   characterization.

   This study investigated the effects of long-term ad libitum feeding of
   a ketosis-causing KD on overall metabolic health in male and female
   C57BL/6J mice over the course of almost 1 year. The effects of a KD on
   BW and on multiple parameters of metabolic health were compared to
   those of a standard low-fat diet (LFD), a low-fat moderate protein
   (LFMP) diet with protein content matched to the KD, and the 60% HFD
   classically used in diet-induced obesity studies. The inclusion of both
   sexes allowed a direct analysis of sex differences in metabolic
   response to the different diets. Last, we also tested the metabolic
   effects of KD-induced weigh loss in obese males and females.

RESULTS

A 90% fat KD prevents weight gain as compared to an HFD in both male and
female mice

   Adult C57Bl/6J male and female mice were placed on one of four
   long-term diet regimens—a 10% fat LFD, a standard 60% HFD, an 89.9% fat
   KD, or another type of LFD with 10% fat and 10% protein to match the
   protein content of the ketogenic diet (LFMP) ([99]Fig. 1, A and B).
   Male and female mice were maintained on the different feeding
   interventions for 36 and 44 weeks, respectively. To validate that this
   KD was ketogenic, blood levels of the most prevalent blood ketone body
   β-hydroxybutyrate (BHB) were measured, and the respiratory exchange
   ratio (RER) was also investigated. BHB levels were significantly
   elevated in the LFD, LFMP, and HFD groups in the fasted state as
   compared to the fed state ([100]Fig. 1, C and D). However, in KD fed
   mice, the levels of BHB were elevated regardless of feeding:fasting
   status and were also significantly higher than any other group
   regardless of feeding status. We also monitored the RER in indirect
   calorimetry chambers to interrogate substrate usage in males on LFD,
   HFD, and KD. RER records shows that mice on LFD alternate between more
   carbohydrate burning during the feeding phase to more fat burning
   during the fasting state (fig. S1A). In contrast, HFD RER oscillations
   are severely dampened with an average value of 0.78. Mice on KD display
   a flat RER with even lower RER values of 0.73 suggestive of constant
   ketosis regardless of time-of-day or feeding status (fig. S1A).
   Together, these data suggest that the KD used in this study produces a
   constant state of ketosis.

Fig. 1. Mice on KD are protected from weight gain as compared to mice on HFD.

   [101]Fig. 1.
   [102]Open in a new tab

   (A) Diets and color scheme used throughout the paper. (B) Macronutrient
   composition of diets used. (C) Male fed and fasted beta-hydroxybutyrate
   levels taken at 4 weeks [overnight (OVN) fast] and 6 weeks [ad libitum
   (AL) fed]; n = 7 to 10 per group. (D) Female fed and fasted
   beta-hydroxybutyrate at 15 weeks on diets (n = 4 to 5 per group). (E)
   Male BW pooled from three independent cohorts (n = 30 to 50 per group).
   (F and G) Male body composition (n = 5 to 6 per group): (F) fat mass
   and (G) lean mass. (H) Average 24-hour food intake from 36 weeks of
   data collection from three independent cohorts (n = 8 to 14 cages per
   group). (I) Female BW pooled from two independent cohorts (n = 10 to 20
   per group). (J and K) Female body composition (n = 5 to 10 per group):
   (J) fat mass and (K) lean mass. (L) Average 24-hour food intake from 26
   weeks of data collection from two independent cohorts over (n = 2 to 4
   cages per group). Statistics: One-way ANOVA was used for (H) and (L),
   and mixed effects ANOVA was used for all other comparisons, post Hoc
   testing used Tukey’s HSD test. *P < 0.05, **P < 0.01, ***P < 0.001,
   ****P < 0.0001; # only used for within group comparisons, #P < 0.05,
   ##P < 0.01, and ###P < 0.0001. Data are represented as means ± SEM.

   Although the KD is 89.9% fat, male and female C57Bl/6J mice on KD were
   prevented from weight gain as compared to mice on a 60% HFD ([103]Fig.
   1, E and I). Both males and females on 60% HFD rapidly gained weight
   with BW divergence between the HFD and the LFDs observed as early as 1.
   BW gain on KD was slower and diverging from LFDs only after over 10
   weeks in males ([104]Fig. 1E) and 11 weeks in females ([105]Fig. 1I).
   BW remained significantly lower in KD mice than HFD mice at all times
   ([106]Fig. 1, E and I). At the end of 36 weeks, male mice on HFD
   weighed 54 g versus 43 g on KD ([107]Fig. 1E). In female mice, after 44
   weeks, mice on HFD weighed 60 g versus 35 g for mice on KD ([108]Fig.
   1I).

   In males, weight gain on KD was mainly due to changes in fat mass as no
   changes in lean mass were observed across 26 weeks ([109]Fig. 1, F and
   G). This is different to mice on HFD, in which weight gain was also
   predominantly due to increased fat mass, but changes in lean mass were
   also observed with mice on HFD gaining 19 g of fat mass and 8 g of lean
   mass after 26 weeks ([110]Fig. 1F). Further, in HFD-fed males, fat mass
   accrual was maximal after 12 weeks and plateaued between weeks 12 and
   26, but lean mass increased throughout the whole study ([111]Fig. 1, F
   and G). In both males on HFD and KD, fat mass was higher than in the
   LFD group from week 4 at any time onward. However, lean mass was only
   greater in the HFD, and lean mass was not different between LFD and KD
   groups. After 26 weeks on the diets, although the mice on HFD had more
   fat and lean mass, the body fat percentage was not different between
   HFD- and KD-fed males (42.7 ± 1.6% fat for HFD 35.7 ± 5.1% fat for KD)
   (fig. S1B). In females, weight gain on a KD was due to increases in
   both lean and fat mass with mice gaining roughly 3.5 g of lean mass and
   13 g of fat mass ([112]Fig. 1, J and K). Similar to the males, females
   on HFD gained most of their weight through increase in fat mass but had
   a higher proportion of weight gain coming from fat with a ~29.5-g
   increase in fat mass and 7-g increase in lean mass ([113]Fig. 1, J and
   K). After 27 weeks, females on HFD had 63.7 ± 1.2% body fat versus
   46.7 ± 3.7% fat on KD (fig. S1).

   The differences in weight between groups can at least partially be
   explained by differences in food intake ([114]Fig. 1, H and L). Average
   food intake in male and female mice was highest in mice on HFD followed
   by mice on KD. This difference between HFD and KD was significant in
   male mice with mice on HFD consuming 13.4 kcal versus 11.1 kcals on KD
   and 8.4 kcals and 10.3 kcals for LFD and LFMP, respectively ([115]Fig.
   1H). Furthermore, throughout 36 weeks of food intake monitoring, the
   mice on HFD consumed more than the other groups, although this
   difference was more pronounced early on, it remained for the duration
   of the study (fig. S1D). Throughout the study, mice on KD also consumed
   an intermediate amount of calories that was higher than the LFD groups
   and lower than the HFD group (fig. S1D). In contrast to the males,
   female mice consumed less food overall with mice on HFD consuming 10.2
   kcals versus 8.9 kcals on KD and 7.8 kcals and 7.6 kcals for LFD and
   LFMP, respectively. However, in female mice, the only significant
   difference in food intake was between the HFD and LFMP with HFD versus
   LFD having a marginally significant difference. This lack of
   significance between HFD and KD and HFD and LFD is likely because of
   the low cage number since food intake was measured as average cage
   intake each week ([116]Fig. 1L). There were no differences in food
   intake between the LFD groups neither in males nor in females
   ([117]Fig. 1, H and L). During the first week that the females were on
   the intervention diets, the HFD group consumed significantly more than
   the other groups, and while this difference declined, the HFD group
   still consumed more than other groups throughout the study (fig. S1D).

KD-fed mice have severe plasma hyperlipidemia and males have liver steatosis
and dysfunction

   Because ~90% of the calories in the KD are from fat, we assessed lipid
   homeostasis in KD-fed mice. In both males and females, plasma
   triglycerides (TGs) were significantly higher in KD-fed mice than in
   any other group including HFD-fed mice ([118]Fig. 2, A and B). In
   general, TG levels were lower in females than in males, yet the TG
   increase on KD was similar in both sexes wherein TG levels were 1.7
   times higher in KD than HFD fed mice (336 versus 193 in males and 152
   versus 92 in females). Similar to the changes in TG, plasma NEFA levels
   were significantly higher in KD-fed mice versus other diet groups—2.75
   and 1.8 times higher in KD versus HFD in males and females,
   respectively ([119]Fig. 2, C and D). There were not sex differences in
   NEFA levels ([120]Fig. 2, C and D). Plasma cholesterol (Cho) responses
   to the diets were different than that of TG and NEFA, with similar
   elevations in Cho levels in both HFD- and KD-fed mice as compared to
   mice on either of the LFDs (fig. S2, A and B). Together, we found, that
   in contrast to HFD, KD significantly raises plasma TG and NEFA levels
   in both male and female mice, suggesting an overabundance of fat is
   still present in the body, although mice on KD are leaner than mice on
   HFD.

Fig. 2. A long-term KD causes dyslipidemia regardless of sex, and males have
steatosis and liver dysfunction.

   [121]Fig. 2.
   [122]Open in a new tab

   (A and B) Plasma TG and (C and D) NEFAs in [(A) and (C)] males (n = 5
   to 10 per group) and [(B) and (D)] females (n = 9 to 10 per group)
   after 32 and 15 weeks on diets, respectively. Blood samples were
   collected in the fed state. (E) Representative images from males (top)
   and females (bottom) on diet for 28 and 33 weeks, respectively. Scale
   bars, 20 μm. (F and G) Liver scoring by a pathologist blinded to
   conditions (male, n = 5 per group; female, n = 3 to 6 per group). (F)
   Steatosis as estimated percent visual field containing fat. (G)
   Inflammation score from 0 to 3 based on average number of inflammatory
   foci per visual field with five visual fields scored as described in
   ([123]90). (H) Liver TG in males from two independent cohorts on diets
   for 28 and 38 weeks (n = 10 to 11 per group) and females (n = 8 to 10
   per group) on diets for 33 weeks. (I and J) Plasma alanine
   aminotransferase (ALT) levels in males [(I), n = 5 to 10 per group] and
   females [(J), n = 9 to 10 per group] after 32 and 15 weeks on diets
   resp. (K to M) Liver bulk RNA sequencing (males, 38 weeks on diets,
   n = 4 to 5 per group. (K) UpSet plot comparing all DE genes. [(L) and
   (M)] Heatmap of (L) the top 200 most significant DE genes and (M)
   selected genes associated with steatosis, fibrosis, and metabolism
   between LFD and KD. Color intensity represents the z score based on
   log[2] expression values. Statistics: [(A) to (F) and (H) to (J)] Data
   are represented as means ± SEM. One-way ANOVA with Tukey’s post hoc
   testing with P values as: *P < 0.05, **P < 0.01, ***P < 0.001, and
   ****P < 0.0001. (G) Kruskal-Wallis test with Dunn’s multiple comparison
   tests for post hoc analyses. [(K) to (M)] RNA sequencing statistics
   were determined using DESeq2 analysis with a Wald test and
   Benjamini-Hochberg correction. Significant genes had an adjusted P
   value of <0.05.

   The hyperlipidemic profile of KD-fed mice prompted us to assess liver
   health and function in male and female mice. Hematoxylin and
   eosin–stained (H&E) sections of livers collected after 6 months of
   dietary interventions were examined by a board-certified liver
   pathologist blinded to the feeding regimen. In males, steatosis, as
   estimated by the percentage of area made up of lipid droplets, was
   higher in mice on HFD and KD as compared to mice on LFD and was not
   different between HFD- and KD-fed mice ([124]Fig. 2F). In females, only
   the livers from HFD mice had significant steatosis with no differences
   between LFMP- and KD-fed females ([125]Fig. 2F). The visual
   inflammation scoring did not reveal marked or significant differences
   between any groups other than a possible trend to higher inflammation
   in both HFD- and KD-fed mice in both sexes ([126]Fig. 2G). These
   results were confirmed using biochemical quantification of hepatic TG
   content ([127]Fig. 2H). Male mice on long-term 6 to 9 months KD and HFD
   had significantly elevated liver TG content as compared to mice on LFD
   (three to four times higher), whereas female mice on KD did not have
   elevated liver TGs ([128]Fig. 2H). Together, these results suggest that
   steatosis is similar between males and females on long-term HFD but
   only appears in males under KD, suggesting sex differences in the liver
   response to KD. Last, we measured plasma alanine transaminase (ALT) as
   a readout of liver function. In males, after 32 weeks of their
   respective diets, HFD- and KD-fed mice had roughly 7 and 4.5 times
   higher ALT compared to LFDs, respectively ([129]Fig. 2I). In addition,
   ALT was significantly lower in KD versus HFD males. In females, after
   48 weeks on their respective diets, mice on HFD and KD also had
   elevated ALT. ALT was not different between mice on LFD and KD, but the
   HFD group had ALT levels that were 3.8 and 2.8 times the levels of mice
   on LFMP or KD, respectively ([130]Fig. 2J). Overall, these findings
   indicate liver dysfunction and steatosis under KD feeding in males but
   not females, suggesting that there are sex differences in the liver
   response to KD.

   To evaluate how HFD versus KD affected hepatic lipid handling at the
   molecular level, we compared the liver transcriptome of KD-, HFD-, and
   LFD-fed males. Gene expression profiles were very similar between mice
   on HFD and KD, with only 171 differentially expressed (DE) genes
   (adjusted P < 0.05) between mice on KD and HFD compared to 2351 for HFD
   versus LFD and 2371 for KD versus LFD ([131]Fig. 2K). The gene
   expression profiles of the top 200 DE genes between KD and LFD were
   virtually identical for HFD and KD ([132]Fig. 2L), and many of these
   genes are associated with steatosis, fibrosis, and inflammation in
   agreement with our histology and biochemical findings ([133]Fig. 2M).
   Thus, at the gene expression level, while both HFD and KD livers are
   different from mice on LFD, the livers from mice on HFD and KD are
   almost indistinguishable. Because of the constant state of ketosis
   observed in KD mice, we specifically interrogated the expression of
   genes involved in ketogenesis and glucose metabolism. Genes linked to
   ketone body metabolism—such as 3-hydroxy-3-methylglutaryl-CoA synthase
   2 (Hmgcs2), which controls the rate limiting state in ketogenesis, and
   β-hydroxybutyrate dehydrogenase 1 (Bdh1)—were as expected increased
   under KD, yet, interestingly, similar expression patterns were also
   observed in HFD ([134]Fig. 2M). In addition, genes involved in de novo
   lipogenesis such as Chrebp and Dgat2 were down-regulated in both HFD
   and KD livers as compared to LFD ([135]Fig. 2M). Expectedly, genes
   related to glycolysis and glucose metabolism were lower in KD as
   compared to either HFD or LFD ([136]Fig. 2M). Overall, liver
   histological, biochemical, and molecular analysis revealed that,
   although mice on a KD are leaner than mice on an HFD, the males still
   suffer steatosis and inflammation that likely impair their liver
   function.

Mice on KD have severe glucose intolerance and impaired insulin secretion

   We next examined how KD affects glucose homeostasis compared to a
   traditional 60% HFD and LFDs. In the first 4 weeks of the intervention,
   mice on KD had lower fasting BG levels compared to HFD and lower
   postprandial glucose compared to both HFD and LFD ([137]Fig. 3A).
   However, when tested after a longer time, between 27 and 32 weeks for
   males and 15 weeks for females, fasting glucose was similarly elevated
   in HFD- and KD-fed mice compared to LFDs ([138]Fig. 3, B and C). Fed BG
   remained significantly lower in KD-fed versus HFD-fed males and females
   and LFD males likely due to the very low amount of carbohydrates in the
   diet ([139]Fig. 3B). These results suggest that while a KD may be
   beneficial for regulating (hyper)glycemia in the short term, long term
   KD feeding is associated with fasting hyperglycemia, similar to that
   observed under HFD, and fed hypoglycemia. We next assessed the ability
   of the mice to restore glycemia after an intraperitoneal bolus of
   glucose. Intraperitoneal glucose tolerance test (GTT) revealed glucose
   intolerance similar to HFD-fed mice after 11 weeks in males and 31
   weeks in females ([140]Fig. 3, D and G). However, much more severe
   glucose intolerance was observed in males after 33 weeks on the diet:
   KD males reached a maximum glucose of 539 mg/dl versus 427 mg/dl in
   HFD, and BG only decreased by 95 mg/dl versus 168 mg/dl in HFD after 2
   hours despite KD mice receiving less glucose based on their lower BW
   ([141]Fig. 3E). In females, we also observed a deterioration of glucose
   response over time on KD with glucose intolerance being more severe
   after 31 weeks than 15 weeks ([142]Fig. 3, F and G).

Fig. 3. A long-term KD causes severe glucose intolerance and suppresses
insulin secretion.

   [143]Fig. 3.
   [144]Open in a new tab

   (A to C) Fasted and fed BG. Fasted glucose was measured at 7 a.m.
   (onset of dark cycle) following a 14-hour fast, and fed glucose was
   measured at 11 a.m. during the dark cycle. (A) Males after 4 weeks
   (n = 8 to 10 per group), (B) 27 and 32 weeks on diet from two
   independent cohorts (n = 5 to 10 per group). (C) Females after 15 weeks
   on diets (n = 4 to 9 per group). (D to G) Glucose tolerance tests
   (GTTs) with a bolus of glucose (1.5 mg/g) given after an overnight fast
   (over the light phase) in: (D) males after 11 weeks on diets (n = 8 to
   10 per group), (E) males after 33 weeks on diets (n = 5 to 6 per
   group), (F) females after 15 weeks on diets (n = 9 to 10 per group),
   (G) females after 31 weeks on diets (n = 8 to 10 per group). (H to M)
   Insulin tolerance tests with insulin (0.75 U/kg). Bar graphs [(I), (K),
   and (M)] show the slopes in the first 15 min (I) or 30 min [(K) and
   (M)]. [(H) and (I)] Males after 11 weeks on diets (n = 5 per group).
   [(J) and (K)] Males on diets for 31 weeks (n = 5 to 10 per group). [(L)
   and (M)] Females on diets for 29 weeks (n = 5 to 10 per group). (N to
   Q) Plasma insulin levels at baseline and 30 min after a bolus of
   glucose. Data in (N), (P), and (Q) are from the blood collected during
   the GTT presented in (D), (F), and (G), respectively. Data shown in (O)
   are from a separate experiment after 30 weeks on diets (n = 8 to 10 per
   group). Statistics: One-way or two-way ANOVA with Tukey’s HSD for post
   hoc testing was used. Paired t tests were used to compare insulin
   levels between 0 and 30 min within a group. The number of symbols
   represents P values as: *P < 0.05, **P < 0.01, ***P < 0.001, and
   ****P < 0.0001. Data are represented as means ± SEM.

   To assess how quickly glucose dysregulation develops on a KD, we
   performed longitudinal GTT assessments in males. We found that glucose
   intolerance was present as early as 1 week on KD (fig. S3D), and this
   glucose intolerance became more severe over time (fig. S3, D to F).
   Severe glucose intolerance was not limited to intraperitoneal
   injections as it was also observed upon oral administration of glucose
   (oGTTs) and upon oral mixed meal tolerance tests in mice on KD (fig.
   S3, A to C). Together, these findings show that male and female mice on
   long-term KD develop glucose intolerance that worsens over time.

   To test whether glucose intolerance in KD-fed mice was due to insulin
   resistance, we performed insulin tolerance tests (ITTs). After 11 weeks
   on the diet interventions in males, fasted BG was similar between HFD
   and KD groups and significantly higher than LFD groups ([145]Fig. 3H).
   Upon insulin injection, BG dropped markedly within the 15 min with a
   delta or slope that was not different between KD and the LFDs groups
   (−118.2 mg/dl in LFD, −97.7 mg/dl in LFMP, and −114.0 mg/dl in KD) but
   significantly less in HFD (−63.2 mg/dl in HFD) ([146]Fig. 3, H and I).
   A similar response was observed in females at 29 weeks ([147]Fig. 3, L
   and M). When an ITT was performed in males after at 32 weeks on their
   respective diets, the response to insulin was maintained in KD and LFD
   mice while the HFD-fed males hardly responded ([148]Fig. 3, J and K).
   Together, these results confirmed the well-characterized insulin
   resistance of males fed HFD. On the contrary, the significant drop in
   BG upon insulin injection showed that KD-fed male and female mice
   remained insulin sensitive and do not develop insulin resistant as
   HFD-fed mice do.

   To determine whether glucose intolerance in KD-fed mice was instead due
   to a defect in insulin secretion, we measured plasma insulin levels in
   response to a glucose bolus using an in vivo GSIS assay. We found that,
   in males and females on LFDs and KD, baseline-fasted insulin levels
   were low and not different, in contrast to the HFD groups that had high
   insulin both prior and 30′ after the glucose injection, as would be
   expected with insulin resistance ([149]Fig. 3, N to Q). LFD- and
   LFMP-fed male and female mice had a significant increase in insulin
   upon glucose injection at all longitudinal time points ([150]Fig. 3, N
   to Q, and fig. S3, G to I; 0′ versus 30′ within diet comparisons).
   Despite being numerically modest, such an increase was sufficient to
   restore normoglycemia in about 1 houqr (see GTTs [151]Fig. 3, D, F, and
   G, which are from the same experiments as the respective panels in
   [152]Fig. 3, N, P, and Q).

   Longitudinal assessment of GSIS over time spent on KD further revealed
   that the insulin secretion deficit in response to glucose occurred
   after 4 to 8 weeks of KD (fig. S3, G to I) and from at least 15 weeks
   onward in females ([153]Fig. 3, P and Q), suggesting that a KD impairs
   islets function over prolonged feeding. The lack of glucose-evoked
   changes in insulin levels are consistent with the lack of BG regulation
   seen in the GTTs and corresponds to the further impairment of glucose
   tolerance over time.

   Together, these results suggest that altered glucose homeostatic
   response in HFD-fed versus KD-fed mice is not caused by the same
   mechanism. While impaired glucose regulation in HFD-fed mice is likely
   linked to insulin resistance in the context of high insulin levels,
   impaired insulin secretion likely underlies glucose intolerance in
   KD-fed mice.

KD-fed mice have normal islets size and insulin content

   The severe glucose intolerance and lack of glucose-evoked insulin
   secretion in the context a normal ITT in KD mice prompted us to examine
   the pancreas. We hypothesized that reduced islets number and/or insulin
   content may be involved in the lack of GSIS.

   Because insulin secretion was impaired in mice on KD, we assessed
   differences in pancreatic islet size or number by histologic analysis.
   Since insulin secreting β cells make up the majority of cells in murine
   islets, we used insulin-/proinsulin-stained area to demarcate islets by
   immunochemistry. The average islet area was much larger in HFD- than
   KD- or LFD-fed mice, but there was no difference in size in KD compared
   to LFD islets ([154]Fig. 4, A and D), while the number of islets per
   section was similar in all three groups ([155]Fig. 4C). When examining
   islets size distribution by group, we found that mice on HFD had
   significantly fewer small islets and at least twice as many larger
   islets compared to the other groups ([156]Fig. 4B and fig. S4B). We
   also found that the percent insulin positive area was not different
   between LFD and KD but was higher in the HFD group (fig. S4A).
   Mirroring the findings in males, there were no differences in average
   size, size distribution, or number of islets between females fed a KD
   or a LFMP ([157]Fig. 4, E to G). Islet size and number were also
   comparable between males and females. To complement these histological
   findings, we quantified pancreas mass and insulin content. Pancreas
   mass was significantly increased in both HFD-fed males and females
   ([158]Fig. 4, H and J) but was similar between mice on LFD and KD
   ([159]Fig. 4, H and J). In agreement with our histological findings,
   HFD pancreas had more insulin per milligram tissue in both male and
   female mice ([160]Fig. 4, I and K), while male mice in general had more
   insulin than female mice. The differences in pancreas size and insulin
   content also translated to the HFD having more total insulin per
   pancreas as compared to the KD and LFD mice (fig. S4, C and D).
   Overall, these results confirm the established effect of HFD feeding
   leading to islet hyperplasia and show that, unexpectedly, KD feeding
   was not associated with overt loss in islet mass or insulin content
   that could underly insulin secretion deficiency. Thus, because mice on
   KD can both produce and respond to insulin, we proceeded to test
   whether a defect in secretion could explain their glucose intolerance.

Fig. 4. A HFD but not a KD increases islet size, pancreas weight, and insulin
content in the pancreas.

   [161]Fig. 4.
   [162]Open in a new tab

   (A to G) Analysis of chromogenic insulin/proinsulin staining in the
   pancreas with hematoxylin counterstain. Qupath was used to quantify
   islet area based on detection of the brown insulin stain. [(A) to (C)]
   Males, 8 to 12 slices were taken throughout the pancreas from n = 4 to
   6 animals per groups for a total 5157 islets analyzed on LFD, 6718 on
   HFD, and 6389 on KD. [(E) to (G)] Females, six to seven slices were
   taken throughout the pancreas (n = 5 per group), for a total of 3749
   LFMP and 3428 KD islets analyzed. [(A) and (E)] Average islet area.
   [(B) and (F)] Frequency distribution of islet sizes as number of islets
   per slice in each bin. Bins correspond to islet diameter 100
   μm^2 = diameter of 11.3 μm or the size one β cell (and the minimum size
   needed to identify a nucleus and clear borders in the analysis); 170
   μm^2 = diameter of 14.7 μm; 490.87 μm^2 = diameter of 25 μm; 1963
   μm^2 = diameter of 50 μm; 4417.86 μm^2 = diameter of 75 μm; 7853.981
   μm^2 = diameter of 100 μm; and 17,671 μm^2 = diameter of 150 μm. [(C)
   and (G)] Average number of islets per slice. (D) Representative images
   from male islets. Scale bar, 100 μm. (H) Male pancreas weight after 6
   to 9 months on respective diets (n = 17 to 18 per group pooled from
   three independent cohorts) and (I) insulin content after 6 months on
   the diets (n = 5 to 6 per group). (J) Female pancreas weight (n = 4 to
   8 per group) and (K) insulin content after 6 months on respective diets
   (n = 3 to 4 per group. Statistics: One-way or two-way ANOVA with
   Tukey’s HSD for post hoc testing was used. The number of symbols
   represents P values as: *P < 0.05, **P < 0.01, ***P < 0.001, and
   ****P < 0.0001. Data are represented as means ± SEM.

Mice on KD lose rapid insulin secretion in vivo and isolated islets have
impaired insulin secretion

   To delve further into the underlying cause of glucose intolerance in
   KD-fed mice, we performed an in vivo hyperglycemic clamp experiment in
   awake, freely moving mice. Throughout the 120-min clamp protocol, 50%
   glucose was infused through a jugular vein catheter to obtain 30 min of
   clamped BG at around 250 mg/dl for 30 min. To that end, BG was measured
   at regular intervals throughout, and the infusion rate titrated
   accordingly. Upon initial infusion, BG rapidly increased in all groups
   and peaked at 30 min ([163]Fig. 5A). The glucose infusion rate (GIR)
   was then progressively lowered to bring the BG to the target 250 mg/dl,
   and the GIR was later increased after about 60 min to maintain the
   clamped target; a BG of 250 mg/dl was obtained in LFD- and HFD-fed mice
   between 90 and 120 min of the test ([164]Fig. 5C). However, in KD-fed
   males, BG remained significantly elevated for 90 min despite the GIR
   being reduced to the minimum infusion rate possible from 50 min onward
   ([165]Fig. 5C). BG in KD mice only started to decrease toward the end
   of the experiment ([166]Fig. 5A). From 90 to 120 min, BG was
   effectively clamped at an average of ~230 mg/dl in HFD- and LFD-fed
   mice, but BG in KD-fed mice was significantly higher at 314 mg/dl and
   could not decrease to the target level during the course of the
   experiment ([167]Fig. 5B) even in the context of significantly lower
   GIR ([168]Fig. 5D). In parallel, we measured plasma insulin levels
   throughout the study ([169]Fig. 5E). Upon glucose infusion, insulin
   levels significantly and rapidly rose in LFD and HFD (87 and 68%
   increase respectively), reaching max level at 15 min ([170]Fig. 5, E
   and F). Strikingly, insulin levels in KD-fed mice initially decreased
   from 0 to 10 min; a slow rise in insulin levels was not apparent until
   30 min post–glucose infusion (GI) start, and insulin levels peaked at
   90 min that coincide with the time at which BG begins to decline
   ([171]Fig. 5A). Insulin levels overall were higher in HFD than LFD
   during the clamp ([172]Fig. 5E), which is indicative of insulin
   resistance. Because mice on KD have similar insulin levels to the
   control mice both at baseline and at the peak, this suggests that their
   glucose intolerance is caused by impaired insulin secretion rather than
   insulin resistance.

Fig. 5. Mice on KD lose rapid insulin secretion.

   [173]Fig. 5.
   [174]Open in a new tab

   (A to F) Measures taken during hyperglycemic clamp after 25 weeks on
   the respective diets. n = 4 to 6 male mice per group. Time courses of
   BG (A), the GIR (C), and plasma insulin levels (E) obtained via tail
   vein sampling. Average BG (B) and average GIR (D) during the 90 to
   120 min when BG was clamped. (F) Insulin levels before and 15 min into
   the glucose infusion. (G to I) Ex vivo GSIS (n = 3 per group). (G)
   Insulin level over time during the different phase of the perifusion
   assay, wherein insulin level is normalized to 100 insulin equivalents
   (IEQ; 1 IEQ represents an islet with a diameter of 150 μm). (H) Total
   insulin secretion during the GSIS as calculated by the area under the
   curve (AUC) from (G). (I) Peak insulin secretion during high glucose
   and KCl stimulation [insulin levels from (G) at 18 and 69 min,
   respectively]. Statistics: One-way ANOVA (B, D, and H) or two-way ANOVA
   [(A), (C), (E), (G), and (I)] with post hoc testing using Tukey’s HSD
   except in (H), which used Dunnett’s multiple comparison test to compare
   HFD and KD with the LFD group only. The number of symbols above data
   denote the following P values: *P < 0.05, **P < 0.01, ***P < 0.001, and
   ****P < 0.0001. Data are represented as means ± SEM.

   To assess cell-autonomous differences in insulin secretion, islets from
   males fed LFD, HFD, and KD for 60 weeks were isolated and tested by
   dynamic ex vivo GSIS using a perifusion system (Vanderbilt islet and
   pancreas analysis core). Baseline insulin secretion at 5.6 mM glucose
   (roughly equal to 100 mg/dl) was similar across all three groups.
   However, upon stimulation with 16.7 mM glucose (roughly 300 mg/dl),
   both KD and HFD islets secreted less insulin than LFD islets ([175]Fig.
   5G). Specifically, the initial peak in insulin secretion at 18 min was
   lower in both HFD and KD islets versus LFD islets (LFD: 4.7 ng/100
   Insulin EQuivalent (IEG)/min versus HFD: 0.99 ng/100 IEQ/min KD: 1.9
   ng/100 IEQ/min), and insulin levels remained lower for KD and HFD
   throughout the high-glucose stimulation ([176]Fig. 5G area marked by
   gray background). To determine whether this deficiency in insulin
   secretion is a result of improper glucose sensing versus dysfunctional
   insulin secretion, in the final step of the GSIS, islets were exposed
   to 20 mM potassium chloride (KCl), which depolarizes β cell in the
   absence of high glucose, thus allowing us to assess maximal insulin
   secretory capacity independent of glucose sensing. Under KCl, the HFD
   and KD islets secreted significantly less insulin than the LFD islets
   ([177]Fig. 5, G and I). However, to determine whether this reduced
   insulin secretion under KCl is a result of glucose sensing versus
   aberrant insulin secretion, we must compare peak insulin secretion
   under glucose versus peak insulin secretion under KCl. For HFD islets,
   peak insulin secretion under KCl was higher than under 16.7 mM glucose
   (1.5 versus 0.99 ng/100 IEQ/min) ([178]Fig. 5, G and I), suggesting a
   deficiency in glucose sensing in addition to aberrant insulin
   secretion. Conversely, for the KD islets, the peak insulin secretion
   under KCl and high glucose were similar (1.82 and 1.9 ng/100 IEQ/min),
   suggesting a deficiency in insulin secretion machinery but no problem
   with glucose sensing. Together, the results of dynamic GSIS testing
   suggest that insulin secretion is compromised in both HFD and KD
   compared to LFD yet possibly through different mechanisms. These
   results show that impaired insulin secretion in KD is linked to a
   secretory defect but not a glucose sensing defect as was the case in
   the HFD islets.

Transcriptomic and EM analysis of KD islets reveals ER-Golgi stress and
aberrant Golgi function

   Our findings thus far have established that the pancreases from KD-fed
   mice have impaired insulin secretion as compared to those from LFD/
   LFMP and HFD fed mice. To gain insight into the molecular causes behind
   the secretory dysfunction in KD islets versus HFD islets, we analyzed
   the bulk transcriptome of isolated islets harvested from male mice fed
   LFD, HFD, or KD for 36 weeks. We found the most gene expression changes
   when comparing islets from mice on KD versus those from mice on LFD
   with a total of 1666 DE genes. Of those 1666 genes, 1083 were uniquely
   changed between KD and LFD, 397 distinguish KD versus both HFD and LFD,
   175 changed in both KD and HFD compared to LFD, and a very small 11
   genes were different between all dietary condition ([179]Fig. 6A). In
   contrast, there were only 242 DE genes in HFD versus LFD ([180]Fig.
   6A), indicating that the KD altered the islet transcriptional profile
   much more than an HFD when compared to LFD. Last, comparing KD to HFD,
   there was a total of 1127 DE genes. Among those, 697 were changed only
   under KD versus HFD, whereas 397 changed in both KD compared to HFD and
   LFD, and a small 22 DE genes changed between KD versus HFD and HFD
   versus LFD) ([181]Fig. 6A). Overall, transcriptomics analysis revealed
   408 (397 + 11) DE genes that distinguish the islets from KD-fed mice
   from islets of mice fed either HFD or LFD—that is expression of these
   408 genes were significantly different in KD versus both LFD and HFD
   conditions ([182]Fig. 6A).

Fig. 6. Islet transcriptomic analysis reveals that KD causes ER Golgi stress
and impaired protein transport.

   [183]Fig. 6.
   [184]Open in a new tab

   Bulk RNA sequencing from islets isolated from male mice after 36 weeks
   on interventions. (A, C, and F) Heatmaps showing z scores of
   log[2]-normalized expression. (A) Heatmap of all DE genes. The Venn
   diagram inset shows number of genes per comparison, and the colors
   correspond to the colored labels on the heatmap. (B) The 397 genes that
   were DE in both KD versus LFD and KD versus HFD were run through
   Metascape, which identified ontology categories that were enriched in
   one group versus the other as displayed in the heatmap; the purple
   indicates increased expression in the corresponding group, and the
   color intensity represents −log[10](P value) of the enrichment as
   calculated by Metascape. (C) Using the Gene Ontology database, we
   identified genes involved in ER and Golgi stress genes and plotted the
   genes that were significantly different in KD versus LFD. Heatmap of
   selected genes associated with ER-Golgi stress. (D) Significantly
   up-regulated pathways identified using Reactome Pathway analysis of DE
   genes between KD and LFD. False discovery rate < 0.05 for all pathways.
   (E) Pathway of insulin granule generation and processing. Colors in (E)
   correspond to colors in (D) and (F). (F) Heatmap of selected genes
   involved in protein processing and vesicular transport. (G) Electron
   micrograph images of pancreatic β cells showing Golgi dilation in KD.
   (H) Mechanism linking transcriptomic findings with reduced insulin
   secretion. Statistics: DE genes were identified using DESeq2 analysis
   with a Wald test and Benjamini-Hochberg correction. Significant genes
   had an adjusted P value of < 0.05.

   To identify potential mechanisms behind the KD-specific islet secretory
   dysfunction, we performed pathway enrichment analysis in Metascape
   using this list of roughly 400 genes that differentiate KD islets from
   HFD and LFD islets. Pathways linked to translation and peptide
   metabolism, ER stress, and ER to Golgi transport were specifically
   enriched in KD versus both LFD and HFD islets ([185]Fig. 6B). In
   addition, the most significantly up-regulated gene in KD as compared to
   both HFD and LFD was Creb3l3 (fig. S5, A and B), a member of the CREB3
   transcription factor family that is activated by ER and Golgi stress
   ([186]37). Because these analyses pointed at a potential ER and Golgi
   stress signature, we therefore performed a targeted analysis of genes
   related to “Golgi apparatus” [Gene Ontology (GO):0005794] and “response
   to endoplasmic reticulum stress” (GO:0034976) ([187]Fig. 6C). Genes
   (158 of 1280; 12.3%) linked to Golgi and almost 20% (41 of 224) of
   GO-annotated ER stress genes showed significantly altered expression in
   KD islets versus LFD or HFD islets ([188]Fig. 6C). This targeted
   analysis further evidenced severe ER-Golgi stress in the KD islets.
   Zooming out, we used the bigger list of KD versus LFD 1666 DE genes and
   Reactome to pathway analysis. Reactome identified significantly
   up-regulated pathways in KD linked to (i) translation (orange, 10
   pathways), (ii) metabolism (blue, 4 pathways), (iii) Golgi-ER transport
   and coat protein complex II (COPII)-mediated vesicle transport (green,
   pathways), (iv) protein localization (dark green, 2 pathways), and (v)
   amino acid metabolism (yellow, 3 pathways). The similar enrichment of
   genes related to vesicle transport, ER/Golgi stress, and translation
   was also observed when KD and HFD islets were compared (fig. S5C). A
   color-coded schematic ([189]Fig. 6E) shows that these altered pathways
   can influence insulin secretory capacity throughout the process of
   insulin synthesis and granules secretion. Targeted
   investigation/further probing the expression of genes involved into
   genes that are specifically involved in vesicle transport between the
   ER and Golgi and to the plasma membrane (as identified by Reactome and
   literature search) shows that many of these genes are significantly
   different between KD and both HFD and LFD ([190]Fig. 6F). Together, the
   gene expression analysis suggests a unique dysfunction of the ER-Golgi
   transit in KD islets, which may underlie the KD-specific deficit in
   insulin secretion as both other groups secreted insulin in response to
   glucose during the clamp studies. Because up-regulation in protein
   trafficking mainly occurs pre-Golgi, we hypothesized that an impairment
   in Golgi processing causes an overall disruption in the production of
   mature insulin granules and, thus, looked specifically at genes
   involved in Golgi organization and transport. Rab8b, which was
   down-regulated in KD islets as compared to LFD and HFD islets, plays a
   role in transporting proteins from the Golgi to the plasma membrane,
   suggesting reduced protein export from the Golgi ([191]38). In further
   support of impaired Golgi functioning is our observed down-regulation
   of Rab30 and Rab39, which are necessary for Golgi apparatus
   organization ([192]38, [193]39). Last, we compared the 175 genes that
   were DE in KD versus LFD and HFD versus LFD, to identify a “high-fat”
   transcriptomic signature. We found that HFD and KD had reduced
   expression of genes related to immune function (fig. S5D), suggesting
   that the “high-fat” signature is immune related and does not alter
   vesicle transport or ER/Golgi stress.

   To establish the relevance of these transcriptional findings, we
   obtained electron microscopy images of islets from LFD and KD islets.
   Islets, which represent less than 1% of the surface area of the
   pancreas, were identified at low magnification through a flower-like
   structures centered around a capillary (fig. S6A). Β cells were
   identifiable by the presence of insulin granules, and some of them with
   characteristics insulin crystals (fig. S6A). Insulin granules were
   present in β cells from both KD and LFD islets. We interrogated marked
   visual differences in the pool of readily releasable insulin granules
   at the basal membrane that could explain the lack of rapid release of
   insulin in KD-fed mice. However, with these granules only making up
   about 1% of all insulin granules and not having any known defining
   features ([194]40, [195]41), we could not determine whether there were
   differences between LFD and KD (fig. S6A). However, when zooming on the
   ER-Golgi, the Golgi apparatus in KD β cells looked vastly different
   than in LFD, appearing dilated and vesiculating ([196]Fig. 6G and fig.
   S6B). This dilation, swelling, and fragmentation of the Golgi would be
   consistent with deficient insulin secretion observed under KD feeding.

   Together, we propose that a KD causes Golgi disorganization that
   impairs protein transport leading to compensatory up-regulation of
   trafficking in the ER and ERGIC and a down-regulation of proteins
   transported out of the Golgi in addition to significant ER-Golgi stress
   with both the stress and impaired protein trafficking suppressing
   insulin secretion and production ([197]Fig. 6H).

While a KD induces WL, insulin secretion becomes severely impaired.

   A KD is currently being used to treat obesity, diabetes, and metabolic
   syndrome despite few preclinical studies establishing its efficacy and
   effects on overall metabolic health. Our observation that over time, KD
   feeding impairs insulin secretion prompted us to evaluate insulin
   regulation in obese mice fed a 60% HFD subjected to KD-induced WL. Male
   and female mice rendered obese by 14 or 15 weeks of 60% HFD feeding,
   respectively, were then switched to a LFD (LFD or LFMP) or a KD.
   Noticeably, after 15 weeks of HFD, males had reached their maximal BW,
   whereas females weight continued to increase throughout the study (37.0
   g at week 15 and 55.5 g at week 32; [198]Fig. 7D). In males, while a KD
   induced immediate WL ([199]Fig. 7A), the mice on KD lost much less
   weight than the mice switched to LFMP and even began to gain some
   weight back after week 20. After 9 weeks of WL, the mice on LFMP had
   lost almost four times as much weight as the mice on KD (LFMP lost 20.9
   g and KD lost 5.4 g) ([200]Fig. 7A). Females switched to the LFD and KD
   rapidly and markedly lost weight for the first 3 weeks after starting
   on the new diets and then gained a small amount of weight back,
   although their weight still remained far below that of mice on HFD in
   ([201]Fig. 7D). Similar to the males, the females on LFD lost much more
   weight than those on KD, with females on KD ending up with a higher
   weight than before starting the KD (LFD lost 9.7 g, and KD gained 0.6
   g). Overall, in both males and females, KD induced WL, but the LFDs
   were a more effective WL intervention than the KD. In both males and
   females, the early WL was accompanied by an initial drastic drop in
   food intake when switching from HFD to either KD or LFD/ LFMP (fig. S7,
   E and I). While food intake increased in mice on KD and LFD/LFMP, the
   intake remained lower than that of mice that remained on HFD showing
   that reduced food intake can at least partially explain reductions in
   weight (fig. S7, E, F, I, and J).

Fig. 7. A KD for WL is less effective than a LFD and causes impaired insulin
secretion.

   [202]Fig. 7.
   [203]Open in a new tab

   (A and D) Body weight during weigh gain (WG), WL, and WRG intervention
   in males (A) and females (D). (A) 14 weeks of WG on 60% HFD (n = 30),
   followed by 9 weeks of WL on KD (red) or LFMP (gray) (n = 7 to 10 per
   group), and WRG on HFD (KD→HFD, purple and LFMP→HFD, green) in males.
   (B) GTT after 7 weeks of WL [week 21 dotted line on the BW graph in
   (A)] with glucose (1.5 mg/g). (C) In vivo GSIS: Plasma insulin during
   GTT in (B) and pre-WL. (D) Fifteen weeks of WG on 60% HFD (n = 28),
   followed by 16 weeks of WL on KD (red) or LFD (black) (n = 5 to 10 per
   group). At week 31 (16 weeks of WL), females were switched back to HFD
   (KD→HFD, purple and LFD→HFD, green) leading to WRG. (E) GTT after 14
   weeks of WL [week 29 dotted line on the BW graph in (D)] with glucose
   (1.5 mg/g). (F) In vivo GSIS: Plasma insulin during GTT in (E) and
   pre-WL. (G and I) GTT with glucose (1.5 mg/g) and (H and J) plasma
   insulin during GTT (in vivo GSIS) in males after 4 weeks of WRG [dotted
   line at week 27 in (A); (G) and (H)] and in females after 10 weeks of
   WRG [dotted line at week 41 in (D); (I) and (J)]. (K and L) Following a
   year-long dietary regimen on either LFMP or KD, KD mice were switched
   to LFMP for 4 weeks. (K) BW before and after switching the diets. (L)
   GTT [glucose (1.5 mg/g)] before and 4 weeks after switching to LFMP.
   Statistics: One-way or two-way ANOVA with post hoc Tukey’s HSD. Number
   of symbols represents significance level: *P < 0.05, **P < 0.01,
   ***P < 0.001, and ****P < 0.0001. Data are represented as means ± SEM.
   ns, not significant.

   Because a KD caused high lipid levels in our prevention study, we also
   measured Cho and TG levels to determine whether KD-induced WL caused
   favorable changes in plasma lipids. We found that, while WL under LFD
   and LFMP was associated with lower Cho and TG levels, mice that lost
   weight under KD remained hyperlipidemic (fig. S7, G, H, K, and L).
   These findings suggest that while a KD can induce WL, it generally
   cannot reverse metabolic dysfunction.

   Once BW had begun to stabilize—7 and 14 weeks into the WL interventions
   in males and females respectively—in vivo GTT and GSIS were performed.
   In both males and females, although the groups switched to KD had lost
   weight, the mice were as glucose intolerant as mice on HFD who had not
   lost any weight ([204]Fig. 7, B and E). The measurement of blood
   insulin levels revealed that mice switched to LFMP and LFD no longer
   had hyperinsulinemia and had normal GSIS and glucose tolerance response
   ([205]Fig. 7, B, C, E, and F). In contrast, while the males and females
   on KD now also had low insulin levels, insulin secretion in response to
   glucose was gone ([206]Fig. 7, C and F). These results show that
   regardless of sex and regardless of BW at initiation of the WL
   intervention, KD induces WL but impaired insulin secretion occurs and
   prevents the normalization of glucose homeostasis.

   We next wondered whether a KD could prevent or limit weight regain
   (WRG) following a WL intervention. To test this, a subset of mice that
   had gone through WL on KD or LFD/ LFMP were switched back to HFD
   ([207]Fig. 7, A and D; WRG). Both male and female mice regained some or
   all of the lost weight after switching back to HFD, showing that a KD
   cannot prevent or cure diet induced obesity ([208]Fig. 7, A and D). We
   also assessed whether a KD could cure HFD induced insulin resistance
   and performed in vivo GTT and GSIS at either 4 weeks after switching
   back to an HFD in males ([209]Fig. 7, G and H) or 10 weeks after in
   females ([210]Fig. 7, I and J). In males and females, following WRG,
   there was no protection against glucose intolerance or hyperinsulinemia
   ([211]Fig. 7, G to J). In males and females, any reductions in insulin
   levels that accompanied WL with a KD or LFD were reversed following
   resumption of an HFD ([212]Fig. 7, C, F, H, and J). Thus, while a KD
   causes WL, it does not cure the ensuing diabetes nor does it prevent
   WRG.

   Because we observed severe islet dysfunction on a KD, we tested whether
   the effects of a KD on glucose tolerance were reversible. To our
   surprise, we found that, in male mice, 4 weeks after switching from a
   long-term KD to LFMP, the glucose intolerance phenotype disappeared
   ([213]Fig. 7L), suggesting that at least some of the detrimental
   metabolic effects of KD are reversible. In summary, these findings show
   that while a KD causes WL following diet induced obesity, it cannot
   cure or permanently ablate glucose intolerance or impaired insulin
   secretion. Although the effects of KD are reversible, these experiments
   show that a LFD more favorably affects glucose tolerance, insulin
   secretion, and WL.

   Last, to begin to understand whether the glucose and weight effects
   were a result of the specific macronutrient composition of our KD or
   could be generalized to other low-carbohydrate diets, we also included
   a group of mice on a high-fat high-protein diet (HFHP) for which the
   composition of fat matched that of the 60% HFD and the composition of
   carbohydrate matched that of our KD at 0.1% carbohydrates (fig. S7, M
   to P). The mice on the HFHP had similar BW, glucose tolerance, and
   insulin secretion to the mice on KD (fig. S7, M to O). While the HFHP
   group was phenotypically similar to the KD group, this group had
   significantly lower levels of KBs (KD: 2.0 mM versus HFHP: 0.4 mM),
   suggesting that the impaired insulin secretion we observed on KD would
   likely be observed in other low-carbohydrate HFDs as well.

DISCUSSION

   The ketogenic diet has grown in popularity over the past several
   decades as a tool for improving weight and metabolic health. While a KD
   benefits in treating epilepsy are concrete, its effects on metabolic
   health have been largely understudied. In particular, changes in
   glucose metabolism upon KD are not fully understood. As most people who
   go on a KD will likely consume glucose eventually due to the difficulty
   to strictly adhere to a KD long term ([214]3), this is of critical
   importance for patients using a KD to treat obesity and related
   metabolic conditions such as type 2 diabetes. In addition, the very
   high-fat content of the diet can challenge whole body regulation of
   lipid homeostasis and may increase the risk of cardiovascular events
   ([215]15).

   Our study is one of few long-term KD interventions (almost 1 year) in
   which mice had ad libitum access to KD without calorie restriction. We
   incorporated two low fat diets, a standard LFD with 20% kcal from
   proteins and a LFMP with a lower 10% kcals from proteins that matches
   KD protein content to control for the potential confounding effects of
   reduced proteins. Throughout our investigations, we found that the LFD
   and LFMP produced virtually identical phenotypes, suggesting that the
   effects of KD are independent of the more moderate protein content and
   consistent with studies showing that 10% protein is enough to support
   normal physiology in B6 mice ([216]42, [217]43). We also included a
   comparison group on 60% HFD, the gold standard model of diet-induced
   obesity for which metabolic disturbances are well described and a group
   on 60% HFD with very low 0.1% carbohydrate. Our study was powered to
   detect significant differences, and our results were replicated in two
   to three independent cohorts. We performed extensive characterization
   of whole-body metabolic parameters, including state of the art,
   hyperglycemic clamp, dynamic GSIS, and electron microscopy (EM)
   experiments. Last, we included both sexes and carefully profiled the
   diet:sex interaction known to affect metabolic outcomes ([218]44,
   [219]45). For the most part, the phenotypic effects of KD followed the
   same patterns in male and female mice, yet interesting differences were
   observed as discussed below. Our results set the ground for future
   investigations into the molecular determinants of KD-mediated insulin
   secretion defects.

   Prior research has led to conflicting findings on weight regulation on
   KDs. In our study, both male and female B6 mice fed a KD ad libitum
   gained less weight than mice on an obesogenic 60% HFD but more than
   mice on a LFD. Several rodent studies reported similar prevention of
   excessive weight gain on a KD as compared to an HFD ([220]9, [221]18,
   [222]20, [223]27). Weight gain between KD and chow fed rodents has led
   to more discordant results, with several studies describing lower BW
   under KD ([224]19, [225]22, [226]29, [227]46). Two lifelong studies of
   mice on KD, noted the necessity of restricting caloric intake ([228]23)
   or cycling the KD with a LFD ([229]9) to prevent obesity ([230]9,
   [231]23), and another 8-week study in middle-aged females also reported
   using pair feeding to prevent weight obesity ([232]47). The long-term
   duration of our study and differences in diet compositions across
   studies, for all diets, can all also affect the phenotypic outcomes
   ([233]7, [234]48–[235]50). Future studies could directly investigate
   these parameters that were not explored here.

   In our study, reduced caloric intake of KD and LFD/LFMP groups compared
   to HFD can at least partially explain the lower weight, although
   concomitant changes in energy expenditure were not assessed
   systematically. During our WL intervention, we observed a marked
   decrease in food intake during the first couple of weeks that matches
   the initial drop in weight. However, weight stabilized after mice
   increased their ad libitum caloric intake to levels that were slightly
   below mice maintained on HFD, suggesting that food intake was the main
   driver of weight changes. KD was less effective than chow diets at
   inducing WL as mice on LFD/ LFMP lost more than double the weight that
   mice on KD lost. Moreover, KD did not cure obesity but merely cause a
   transient WL that was reversed upon resumption of an HFD.

   Despite being lighter than mice fed a 60% HFD, males and females on KD
   had significant increases in plasma TGs and NEFA. KD-fed males also
   show hepatic steatosis and increased plasmatic ALT activity, suggestive
   of liver dysfunction. Those findings are in agreement with most other
   studies of KD ([236]18, [237]20, [238]22). With the increasing
   prevalence of metabolic dysfunction–associated steatohepatitis and
   especially in people with obesity, the potential of KD to cause or
   worsen pre-existing hepatic dysfunction is of utmost concern.
   Unexpectedly, females on KD did not display signs of hepatic lipid
   deposition nor liver dysfunction, and these conditions were only
   observed under 60% HFD. Future studies will explore these sexual
   dimorphisms with the potential to uncover novel biology of hepatic
   buffering of lipid spill over.

   The effect of KD on glucose homeostasis in mice is controversial. Many
   reports initially pointed at the benefits of KD on glucose homeostasis
   ([239]18, [240]25, [241]33), yet a number of recent publications report
   similar negative effects on glucose tolerance to what we observed
   ([242]20, [243]24, [244]29–[245]32). This glucose intolerance was
   observed using multiple KDs with varying fat compositions such as with:
   (i) an 89% kcal fat KD with fat from roughly equal amount of lard and
   cocoa butter [1% kcal carb, 10% kcal protein; ([246]30)], (ii) an 89.5%
   kcal fat KD with fat from Primex [0.1% kcal carbohydrate, 10.4% kcal
   protein; Research Diets, D12369B; ([247]24)], (iii) an 93.1% kcal fat
   KD with equal amounts of saturated, monounsaturated, and
   polyunsaturated fats [1% kcal carb, 5.9% kcals protein; ([248]31)],
   (iv) an 91% kcals fat KD from mostly lard (0% kcal carbohydrate, 9%
   kcal protein; ([249]32)], and (v) an 95.1% kcal fat KD with fat from
   lard and butter fat [0.4% kcal carbohydrate and 4.5% kcal protein;
   Bio-Serv (F3666) ([250]20, [251]29)].

   Another shared feature of these KD diets is their very low carbohydrate
   content (less than 1% kcal). To test the effect of the limited 0.1%
   sugar content on glucose tolerance, we used a 60% HFD, with identical
   fat composition to that of the 60% HFD but with a very low 0.1% kcal
   carbohydrate content (HFHP; 60% kcal fat, 0.1% kcal carbohydrate, and
   39.9% kcal protein). This diet uncovered metabolic phenotypes that were
   largely similar to our KD, with weight gain, glucose intolerance, and
   impaired insulin secretion, in the absence of elevated circulating KBs.
   These results suggest that the combined high-fat very-low-carbohydrate
   nature of KDs, rather than the ketogenic effects per se, likely play a
   large role in inducing glucose intolerance and islet dysfunction.
   Hence, it is likely that glucose homeostasis would be dysregulated
   under many KDs regardless of the specific macronutrient content.

   In our study, in vivo GSIS revealed a lack of insulin secretion in
   KD-fed mice in contrast to LFD and HFD groups. Reviewing the literature
   highlighted that regardless of the positive or negative effects of KD
   on GTT, insulin levels in KD mice were low across the board,
   irrespective of diet composition, age, or study duration ([252]18,
   [253]19, [254]46, [255]51). While this is often pointed to as a
   positive effect (i.e., prevention or reversal of hyperinsulinemia), we
   posit that low-insulin levels on a KD are a hallmark of β cell
   dysfunction and can lead to other detrimental metabolic effects.
   Accordingly, two prior studies described reduced β cell mass as a cause
   of low insulin levels ([256]30, [257]31). In our study, however,
   extensive characterization of islets mass did not point at differences
   in islets mass as the underlying determinants of KD-linked insulin
   secretion defect.

   Hyperglycemic clamp further uncovered inefficient rapid insulin
   secretion in KD-fed mice, a phenomenon that has not been described to
   the best of our knowledge. In mice on KD, insulin levels did not rise
   significantly until 90 min into the 120-min clamp study while mice on
   HFD and KD had an immediate rise in plasma insulin that peaked at
   15 min. The exocytosis of insulin containing secretory granules docked
   at the plasma membrane ([258]40, [259]52) is incriminated in the early
   insulin response to glucose, suggesting that this population of insulin
   granules may be specifically affected by a KD. To assess cell
   autonomous defects in insulin secretion, β cells were isolated and
   tested in an ex vivo dynamic GSIS assay. Both HFD and KD islets had
   reduced insulin secretion under high glucose compared to LFD islets.
   Upon KCl-induced depolarization, HFD islets secreted significantly more
   insulin suggestive of impaired glucose sensing. In contrast, islets
   from KD-fed mice released similarly low levels of insulin whether
   stimulated by glucose or KCl, indicating defective insulin secretory
   machinery.

   Integrating in vivo and ex vivo insulin results further highlighted the
   differences in islets physiology between HFD- and KD-fed mice. Despite
   the impaired glucose sensing and insulin secretion of isolated islets,
   HFD-fed mice were hyperinsulinemic and able to secrete insulin during
   the hyperglycemic clamp. This ability to secrete high levels of insulin
   in vivo suggests a compensation for cell autonomous defects, which may
   be linked to islet hyperplasia since mice on HFD have double the number
   of the largest islets and higher pancreatic insulin content compared to
   LFD and KD mice. This effect was not assessed in the ex vivo GSIS
   because islets size was controlled for. It is also very possible that
   the extracellular milieu influences insulin secretory response in vivo.
   In particular, islets from KD-fed mice are exposed to a milieu richer
   in lipids (both NEFA and TG) than HFD islets, and the effects of this
   milieu are likely abated in our ex vivo GSIS studies. A number of in
   vitro studies have established a causal link between long-term fatty
   acid treatment and reduced GSIS ([260]53–[261]55); however, a washout
   period has been shown to repair GSIS ([262]55). Hence, removing KD
   islets from their lipotoxic milieu might have improved their insulin
   secretion in the ex vivo GSIS.

   Transcriptomics analysis highlighted alterations in translation,
   protein transport, Golgi function, and ER/Golgi stress specific to KD
   islets. Pathways and genes associated with ER transport, COPI and II
   transport, and secretion from the Golgi were all elevated suggesting
   increased protein transport pre-Golgi. The transcriptional repression
   of Rab30 and Rab39, which are important for Golgi organization
   ([263]38, [264]39, [265]56), suggested Golgi disorganization, and
   repression of Rab8b impaired Golgi vesicle secretion. From these data,
   we suggest a model in which Golgi-related defects lead to a traffic jam
   of proteins in the Golgi and a compensatory upregulation of pre-Golgi
   transport. ER/Golgi stress markers were also up-regulated under KD. For
   instance, Creb3l3, an ER/Golgi stress marker ([266]57), was the most
   up-regulated gene in KD islets. CREB3 (of the same family as Creb3l3)
   is up-regulated by palmitate in human β cells where it attenuates
   palmitate toxicity and is hypothesized to act as a mediator of the
   Golgi stress response ([267]58). Overall, we speculate that high lipids
   on KD lead to the development of a lipotoxicity stress, which has
   previously been shown to cause ER stress and disrupt ER to Golgi
   protein transport in islets ([268]59, [269]60). Electron microscopy
   images further confirmed aberrations in the Golgi that appeared dilated
   and vesiculating in mice on a KD. A recent study characterized the
   disrupted protein trafficking observed in islets of diabetic leptin
   receptor KO db/db mice and uncovered distended and vesiculated Golgi
   very similar to what we see in our KD mice ([270]61). Misfolded
   proinsulin released from the ER is put forth as cause of Golgi dilation
   and altered morphology ([271]61). Incidentally, misfolded proinsulin
   has been suggested to occur in early stages of type 2 diabetes (T2D)
   ([272]62), so it is possible that increases in misfolded proinsulin
   cause Golgi swelling and further impair insulin trafficking and
   secretion.

   Overall, we propose that extremely high-lipid levels under KD lead to
   ER/Golgi stress in islets, thereby disrupting protein trafficking and
   causing a traffic jam of vesicles in the Golgi. With impaired protein
   transport and ER stress, the pool of readily releasable insulin
   available may be insufficient for the islets to rapidly secrete insulin
   in response to glucose, thus leading to the observed glucose
   intolerance. It has been postulated that, early on, lipotoxicity
   induces an expansion of β cell that precedes β cell failure ([273]63).
   We hypothesize that the extra high fat of a KD as compared to an HFD
   causes mice on KD to bypasses the islet proliferative phase as observed
   in HFD-fed mice and go straight to early β cell failure. Future
   experiments will be conducted to test this hypothesis.

   Many studies have shown that a KD can lower insulin levels in humans
   ([274]64–[275]68) and mice ([276]19, [277]46), and while this is
   generally taken as a sign of improved glycemic control and remission of
   diabetes, our results led us to question whether low-insulin levels are
   safe and whether they actually indicate improved glycemic control.
   Insulin is crucial not only for glucose regulation but also for lipid
   homeostasis through stimulation of plasma lipid clearance and
   inhibition of lipolysis ([278]69–[279]71). Insulin resistance is
   accompanied by elevated blood lipids and ectopic lipid deposition
   ([280]72, [281]73), as observed in HFD-fed mice. These metabolic
   alterations were also observed in KD-fed mice, suggesting that their
   very low levels of insulin might be insufficient to suppress lipolysis
   and promote the uptake and benign storage of the excess dietary fat in
   adipose tissues. Thus, while the lack of insulin on a KD may lead to a
   lower BW and less fat storage, it may worsens metabolic health by
   destabilizing lipid homeostasis.

   While our findings show that a KD both prevents and causes WL, a KD
   does not lead to permanent reductions in BW and, thus, should not be
   treated as a “cure” for obesity or diabetes. Moreover, we observed
   worsening glucose intolerance and impaired insulin secretion the longer
   the animals had been on KD, so the vision of KD as a treatment for
   metabolic disease should be questioned even if initial improvements in
   health are observed. Although we found that glucose intolerance caused
   by a KD is reversible, it is possible that other effects may persist.
   Moreover, in mice with diet-induced obesity, a traditional
   high-carbohydrate LFD caused greater WL than a KD while improving
   glucose intolerance.

   KD is used to manage refractory epilepsy in children. It is rarely a
   lifelong intervention as confirmed by a metanalysis of 45 studies
   reporting that, of children put on a KD, only 45% remain on the diet
   after 1 year and 29% after 2 years ([282]74). Moreover, an
   international consensus group statement recommends discontinuing KD by
   2 years with some experts recommending discontinuation by 6 months
   ([283]2). Because KD is predominantly used in children and discontinued
   after a short time, long-term studies of metabolic effects are
   generally lacking. Nevertheless, high-plasma lipids, pancreatitis, and
   cardiac complications have been reported in patients on KD ([284]2,
   [285]13, [286]15, [287]75–[288]78). Studies around the world have also
   reported higher than average prevalence of diabetes in people with
   epilepsy ([289]79–[290]81). The reasons for this co-occurrence are
   unclear ([291]82, [292]83), but one hypothesis is that epilepsy
   treatments increase the likelihood of diabetes ([293]84).

   In summary, while a KD can prevent and treat obesity, it causes
   hyperlipidemia, hepatic steatosis, and glucose intolerance. Unlike mice
   on HFD, the mice on KD do not have detectable insulin resistance or
   hyperinsulinemia. Instead, they have impairments in insulin secretion
   due to a blockade in protein trafficking resulting from dilation of the
   Golgi apparatus, which also causes ER/Golgi stress.

   Our study shows that while C57BL/6J male and female mice on a KD are
   protected from weight gain as compared to mice on a traditional 60%
   HFD, the mice experience severe glucose intolerance, high plasma
   lipids, and impaired insulin secretion with males also developing
   hepatic steatosis. Further studies in other strains of mice, other
   animal models, and in humans are necessary to determine whether
   KD-linked metabolic derangements are universal. Because the KD is not a
   standardized diet, it has been suggested that saturated versus
   unsaturated fats and different macronutrient compositions can lead to
   different metabolic outcomes. While we were able to replicate our
   results in a group of mice on another type of low-carbohydrate, HFD,
   future studies to elucidate whether the type of fat and macronutrient
   composition of the KD influences the metabolic outcomes will be
   important for making KD safer for all including those who require a KD
   to treat their epilepsy. Despite these limitations, our findings have
   relevant translational ramifications. They suggest that a KD used as a
   long-term dietary intervention may have harmful effects on metabolic
   health—especially in terms of β cell function, plasma lipid levels, and
   liver health—and caution against the systematic use of a KD as a health
   promoting dietary intervention.

MATERIALS AND METHODS

Mice

   All animal experiments were carried out in accordance with the
   guidelines and approved by the Institutional Animal Care and Use
   Committee (IACUC) of the University of Utah (protocol 1764 to A.C.).
   C57Bl6/J male and female mice from The Jackson Laboratory (catalog no.
   000664) were used in all studies. After receiving mice from Jax, they
   were placed on chow and allowed to acclimate for at least 3 weeks. When
   mice were 14 to 15 weeks old, they were placed on their respective
   diets. Mice had 24-hour ad libitum access to food and water and were on
   a 12-hour light/12-hour dark cycle. Lights on a 7:00 a.m. Th prevention
   studies ([294]Figs. 1 to [295]6) used169 males from three independent
   cohorts and 60 females from two independent cohorts. The WL studies
   ([296]Fig. 7) used 30 female and 30 male mice. The male WL study was
   repeated in two additional cohorts.

Diets

   All diets used in the studies were ordered from Research Diets: LFD
   [10% kcal from fat, 70% kcal from carbohydrates, 20% kcal from protein
   (D12450K)]; LFMP [10% kcal from fat, 80% kcal from carbohydrates, and
   10% kcal from protein (D21110806)]; HFD [60% kcal from fat, 20% kcal
   from carbohydrates, and 20% kcal from protein (D12492)]; KD [89.9% kcal
   from fat, 0.1% kcal from carbohydrates, and 10% kcal from protein
   (D16062902)]; and HFHP [60% kcal from fat, 0.1% kcal from
   carbohydrates, and 39.9% kcal from protein (D05022501)].

Body composition

   Body composition was measured using a Bruker Minispec small animal NMR
   (Bruker, catalog no. LF50). Animals were awake and freely moving.
   Measurements were taken during the dark cycles, and animals were in the
   ad libitum fed state.

Blood glucose and blood ketone beta-hydroxybutyrate

   Blood glucose and blood ketone beta-hydroxybutyrate were collected
   using handheld meters. Venous blood was obtained via a small nick in
   the tail. Blood glucose meters (Amazon, Nova Max Plus) had a range of
   20 to 600 mg/dl, so any readings that were high were given a numeric
   value of 600. Ketone body meters (Amazon, KetoneBM) measured only the
   presence of beta-hydroxybutyrate.

GTT and ITT

   Before studies, mice were fasted over night for roughly 15 hours. Blood
   glucose levels were measured at baseline and then at regular intervals
   during the study as noted. Mice were given intraperitoneal injections
   of glucose (1.5 mg/g; Sigma-Aldrich, D-(+)-glucose, catalog no. G7021)
   or insulin (0.75 IU/ kg; Lily Medical, Humulin R U-100) in
   phosphate-buffered saline (PBS). Blood glucose and blood were collected
   as noted in the experiment. For in vivo GSIS, blood was collected using
   a cheek bleed before and during the GTT.

Plasma metabolites

   All blood metabolite measures were performed using plasma collected
   from the submandibular vein. Fed blood was collected in the ad libitum
   fed state during the middle of the dark cycle (12 p.m.), and fasted
   blood, which was taken to measures baseline insulin during in vivo
   GSIS, was collected at 7 a.m. (onset of the dark cycle) following a
   14-hour overnight fast. Blood was collected in EDTA tubes
   (Sarsdtet,16.444.100) and placed immediately on ice. Blood was then
   spun at 1000g for 10 min. Cholesterol, TGs, and nonesterified fatty
   acids were measured using colorimetric kits (Sekisui Diagnostics, TG:
   Cat#236–60, Chol: Cat#234–60, NEFA: Cat#s:999–34691, 991–34891,
   995–34791, 993–35191, 997–76491). Insulin was measured using an insulin
   ELISA (Crystal Chem, Cat#90080).

Liver TGs

   Liver was collected at sacrifices and immediately snap frozen. Tissues
   were homogenized using a mortar and pestle kept cold with dry ice and
   liquid nitrogen. Lipids were extracted using a modified Folch methods.
   Liver (50 mg) was extracted in 2:1 chloroform:methanol mixture
   (chloroform: Sigma-Aldrich, C2432-1L). Methanol mixture was vortexed
   and left at 60°C for 30 to 60 min. The aqueous layer was transferred to
   a new tube, and 125 μl of water was added for every 500 μl of
   chloroform/methanol mixture. The mixture was vortexed and then spun at
   500g for 10 min to cause phase separation. The lower organic layer was
   moved to a new and liquid was evaporated. Chloroform was added back to
   resuspend the lipids (20:1 ratio original tissue weight to chloroform).
   Chloroform mixture (100 μl) was combined with 200 μl of 1% Triton-100
   and re-evaporated at 45°C. The leftover lipid triton mix was
   resuspended in water. Triglycerides were measured using the same TG kit
   as for the plasma (Sekisui Diagnostics, catalog no. 236-60).

Hyperglycemic clamp

Surgery

   Mice were anesthetized using isoflurane (Vet One NDC,13985-528-60; Kent
   Scientific Somnosuite, 22-01), given anesthetics, and skin was shaved
   and prepped with betadine and isopropyl alcohol. A small incision was
   made starting from halfway between the left armpit and midline and
   extended up roughly 50 cm to the next. The jugular vein was isolated
   and the catheter (Instech, C20PU-MJV2013) was introduced into 1- to
   2-mm incision in the jugular. The catheter was slid roughly 10 mm into
   the vein, and ligatures (Fine Science Tools, catalog no. 18030-60) were
   applied above and below catheter insertion to hold it in place. The
   catheter was then threaded around the side of the mice and pulled
   through a small incision sitting atop the scapula. Incisions were
   sutured closed (Covidien, SN1955), The catheter was held in place using
   a backpack harness (Instech, catalog no. CIH62) and filled with lock
   solution. Mice were allowed to recover for 4 days.

Clamp procedure

   On the day of the study catheters were connected to a swivel system
   (Instech, catalog no. KVAH62T), allowing mice to be freely moving
   during the clamp. Blood glucose was measured beginning 90 min before
   the clamp to ensure mice had a steady baseline glucose. Mice were
   infused with 50% glucose in saline starting at a rate of 500 μl/hour
   for 2 min. After the first 2 minutes, the infusion rate was changed to
   an amount in microliters that was 3× the BW (is a 50 g of mouse
   received an infusion of 150 μl/hour). Then, the infusion amounts were
   changed on the basis of BG levels that were taken every 3 to 5 min with
   a target glucose level of 250 mg/dl. Tail bleeds were also performed at
   baseline 0, 5, 10, 15, 20, 25, 30, 60, 90, and 120 so that insulin
   levels could be measured.

Islet harvest for RNA sequencing

   Mouse was euthanized, and the pancreas was perfused using an oxygenated
   buffer of Hanks’ balanced salt solution (HBSS) (Thermo Fisher
   Scientific, catalog no. 14025092) with 25 mM Hepes (Gibco, catalog no.
   15630-080), 8 mM glucose, and 0.2% bovine serum albumin (BSA). Liberase
   (0.1 mg/ml; Sigma-Aldrich, 0540102001) was added tto the buffer for
   pancreas digestions. The pancreas was digested at 37°C for 30 min.
   After 30 min, the samples were flooded with fresh buffer and allowed to
   settle for 10 min, so that the islets would sink to the bottom and the
   debris would float to the top; part of the supernatant was removed,
   fresh buffer was added to replace the volume removed, and samples were
   agitated to mix. This process was repeated until the liquid was very
   clear (four to six times); the islets with the clear liquid were then
   handpicked under a microscope. Islets were then pelleted, washed with
   PBS, and snap frozen.

RNA sequencing and analysis

   RNA was extracted using TRIzol (Invitrogen, catalog no. 15596018)
   chloroform and followed by a QIAGEN RNEasy kit with deoxyribonuclease
   treatment (Islets: micro kit 74004; liver: Mini Kit 74106). The
   University of Utah High Throughput Genomics core performed RNA
   sequencing with the following methods. Total RNA samples (5 to 500 ng)
   were hybridized with NEBNext rRNA Depletion Kit v2 (human, mouse, and
   rat) (E7400) to substantially diminish ribosomal RNA from the samples.
   Stranded RNA sequencing libraries were prepared as described using the
   NEBNext Ultra II Directional RNA Library Prep Kit for Illumina
   (E7760L). Purified libraries were qualified on an Agilent Technologies
   4150 TapeStation using a D1000 ScreenTape assay (catalog no. 5067-5582
   and 5067-5583). The molarity of adapter-modified molecules was defined
   by quantitative PCR using the Kapa Biosystems Kapa Library Quant Kit
   (catalog no. KK4824). Individual libraries were normalized and pooled
   in preparation for Illumina sequence analysis. The following adapter
   reads were used: Adapter Read1 (AGATCGGAAGAGCACACGTCTGAACTCCAGTCA) and
   Adapter Read 2 (AGATCGGAAGAGCGTCGTGTAGGGAAAGAGTGT). Sequencing
   libraries were chemically denatured in preparation for sequencing.
   Following transfer of the denatured samples to an Illumina NovaSeq X
   instrument, a 150 × 150 cycle paired-end sequence run was performed
   using a NovaSeq X Series 10B Reagent Kit (20085594). Data were aligned
   to mm39, and deseq2 analysis was run as described by Love et al.
   ([297]85) by the University of Utah Bioninformatics Core. Pathway
   analysis was performed using Reactome ([298]86) camera analysis on DE
   genes. Metascape ([299]87) was also run on DE genes to identify gene
   ontology groups that were enriched. UpSet ([300]88) plots were used for
   data visualization.

   Fastq files are available online: Liver RNA sequencing [gene expression
   omnibus (GEO): [301]GSE248297
   ([302]www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE248297)] and Islet
   RNA sequencing (GEO: [303]GSE268020;
   [304]www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE268020). The data
   will be ready and available to the public before publishing.

Insulin staining and islet sizing

   Pancreas was collected after 29 weeks on the respective diets from both
   male and female mice. In males, 8 to 12 slices were taken throughout
   the entire pancreas from n = 4 to 6 animals per groups for a total 5157
   islets analyzed on LFD, 6718 on HFD, and 6389 on KD. For females, six
   and seven slices were taken throughout the entire pancreas (n = 5 per
   group), for a total of 3749 LFMP, and 3428 KD islets were analyzed.

   Sections for histology were collected at sacrifice and then placed in
   10% formaldehyde (Ricca, catalog no. 3190-5) for 24 hours, followed by
   70% ethanol (Decon Laboratories, catalog no. 2701). Sections were then
   paraffin embedded and sectioned to 5-μm thickness by the University of
   Utah Research Histology Core. Sections were deparrafinized and
   rehydrated in sequential 3-min soaks in xylene (2×) (Sigma-Aldrich,
   catalog no. 296325), 100% ethanol (Decon Laboratories, catalog no.
   2701), 95% ethanol, 70% ethanol, 50% ethanol, and water. Antigen
   retrieval was performed using 10 mM sodium citrate (Thermo Fisher
   Scientific, BP327-500) with 0.05% Tween 20 (Calbiochem, 655204) pH 6 in
   kept around 95°C for 30 min. A 30- to 60-min blocking step with 2% BSA
   (Sigma-Aldrich, catalog no. A7906) in tris-buffered saline was
   performed, followed by 30% incubation with 3% hydrogen peroxide to
   quench endogenous peroxidase activity. After washing 3× for 5 min with
   PBS + 0.05% Tween 20 the horseradish peroxidase–conjugated insulin
   antibody was used (Abcam, catalog no. 28063 mouse anti-insulin and
   pro-insulin- AB dilution, 1:200) in the dark for 60 min. Antibody was
   washed off as described above and followed DAB application (Thermo
   Fisher Scientific, catalog no. 34002) and a hematoxylin counterstain
   (Vector Labs, catalog no. H-3502). Slides were then dehydrated and
   coverslipped. Slides were imaged on an Axioscan 7 slide scanner
   (Zeiss), and images were processed in Qupath ([305]89). To quantify
   insulin positive area, thresholding for DAB staining was performed, and
   manual corrections were made as needed to remove artifacts or add
   islets that were not counted. A minimum islet size was set to exclude
   islets with an area of less than 100 μm^2; this corresponds to a
   diameter of roughly 11.28 μm, which is slightly larger than a singular
   β cell, and detections smaller than 100 μm^2 often did not contain a
   visible nucleus. We calculated area and perimeter for each unique
   islet. A separate analysis in Qupath for total pancreas area was
   created. We created a positive area detection threshold for hematoxylin
   and measured the total area; regions of adipose tissue surrounding the
   pancreas were removed. Area and perimeter data for islets and the
   pancreas sections were extracted from Qupath. The binned areas used in
   the frequency distribution correspond to islet diameter as follows: 100
   μm^2 = diameter of 11.3 μm or the size one β cell (and the minimum size
   needed to identify a nucleus and clear borders in the analysis); 170
   μm^2 = diameter of 14.7 μm; 490.87 μm^2 = diameter of 25 μm; 1963
   μm^2 = diameter of 50 μm; 4417.86 μm^2 = diameter of 75 μm; 7853.981
   μm^2 = diameter of 100 μm; and 17,671 μm^2 = diameter of 150 μm.

Liver histology and scoring

   Sections for histology were collected at sacrifice and then placed in
   10% formaldehyde (Ricca, catalog no. 3190-5) for 24 hours, followed by
   70% ethanol (Decon Laboratories, catalog no. 2701). Sections were then
   paraffin embedded, sectioned to 5-μm thickness, and stained with H&E by
   the University of Utah Research Histology Core. Slides contained one
   section from each diet group and were scored by a board-certified
   pathologist (K.J.E.) who was blinded to the diet conditions. Steatosis
   was determined by estimating percent area containing lipid droplets,
   and liver inflammation was scored between 0 and 3 based on the scale
   developed by Liang et al. ([306]90). Images were acquired using
   cellSens software and an Olympus DP73 camera attached to an Olympus
   BX53 microscope with UPlanFL objectives.

Electron microscopy

   Mice were anesthetized by isoflurane gas. Mice were anesthetized using
   isoflurane (Vet One, NDC,13985-528-60; Kent Scientific Somnosuite,
   22-01) and fixed by cardiac perfusion with 2.5% glutaraldehyde and 1%
   paraformaldehyde, 6 mM CaCl[2], and 2.4% sucrose in 0.1 M cacodylate
   buffer (pH 7.4; Electron Microscopy Sciences, catalog no. 16537-07).
   Immediately after perfusion, the pancreas was dissected out and soaked
   in fresh fixative, where the organ was sliced into small blocks of
   about 10 mm^3. Tissue blocks were transferred into fresh fixative where
   they were continuously fixed for at least 2 hours at room temperature
   before being stored at 4°C. Sample processing and conventional electron
   microscopy were conducted by the University of Utah Electron Microscopy
   Core Facility using a microwave-assisted fixation and processing
   protocol (Pelco Biowave, 36500). Sequential steps were as follows: 2%
   osmium post-fixation, 4% uranyl acetate en-bloc staining, serial
   dehydration in ethanol/acetone, and embedding in Embed 812 (Electron
   Microscopy Sciences, catalog no. 14120). Specimens in the cured
   epoxy-resin block was sectioned using Lecia ultramicrotome UC6. The
   islet of Langerhans was identified under a light microscope in semithin
   sections stained with toluidine, then ultrathin sections of 100-nm
   thickness were cut and deposited onto carbon-formvar–coated slot grids.
   Sections on grids were further contrast-stained with uranyl acetate and
   lead citrate and examined at 120 kV with JEOL JEM 1400 equipped with a
   GatanOrius digital camera.

Whole pancreas insulin quantification

   At sacrifice, pancreas was carefully dissected and weighed. A small
   piece (10 to 20 mg) was taken and weighed separately. This small piece
   was added to a tube containing 5 ml of acid ethanol extraction buffer
   (1.5 ml of 1 N HCl with 98.5 ml of 70% ethanol). The pancreas was
   homogenized by vortexing with metal screws for 2 hours on two
   consecutive days or until mixture was homogenized. The slurry was then
   centrifuged at 2000g for 3 min to pellet any debris. A 500 ml of
   supernatant was mixed with 500 ml of 1 M tris. The tris extract mixture
   was then diluted 1:100 using Crystal Chem Islet sample diluent and then
   run on the Crystal Chem ELISA using the high range assay.

Ex vivo GSIS

   The ex vivo GSIS was performed by the Vanderbilt islet and pancreas
   core using established protocol that are found online: islet harvest
   ([307]91) and islet perifusion ([308]92).

Islet harvest

   Mice were anesthetized with ketamine/xylazine solution (1 μl/mg BW),
   and after the loss of consciousness, they were placed on a surgical
   stage and skin was prepped with 70% ethanol. The skin and peritoneum
   were opened, and the bile duct was located and closed with a suture at
   the duodenum. Using a bent 30-gauge needle collagenase solution
   [collagenase P (0.6 mg/ml) in HBSS] (collagenase: MilliporeSigma,
   catalog no. 11249002001; HBSS: Thermo Fisher Scientific, catalog no.
   14025) was injected into the bile duct entering the duct near the liver
   to inflate the pancreas. The pancreas was removed and placed in a
   conical tube containing 5 ml of HBSS/collagenase solution, and the
   tubes were placed in a 37°C water bath shaker for 8 min with shaking by
   hand every 4 min and until the tissue had dissolved into a slurry at
   the end of the second shaking. Tubes were placed on ice and topped up
   to 15 ml with HBSS with 10% FBS (Sigma-Aldrich, catalog no. F0926) and
   then centrifuged at 200 rcf for 00:01:00 at 25°C. Liquid was decanted,
   and pellets were washed by adding 10 ml of HBSS with 10% FBS and then
   vortexed to mix before centrifuging again. This process was repeated
   3×. After the final wash, supernatant was discarded, and 10 ml of HBSS
   with 10% FBS was added and mixed. The slurry was transferred to a 50-ml
   conical tube. The original tube was washed with another 5 ml of HBSS
   with 10% FBS and added to the 50-ml tube for a total of 15 ml of
   digest. The digest was vortexed, and 15 ml of Histopaque gradient was
   underlayed using a manual bulb pipette. Tubes were centrifuged at 500
   rcf for 00:10:00 at 25°C. New 50-ml tubes containing 15 ml of HBSS with
   1% FBS were prepared, and islet layer was transferred to the new tubes.
   Tubes were centrifuged tubes at 500 rcf for 00:10:00 at 25°C, the
   liquid was decanted, and pellets were washed 3× with 25 ml of HBSS with
   1% FBS vortexing each time after adding HBSS to mix before
   centrifugation. After the final decanting, 10 ml of HBSS with 10% FBS
   was added to the tube and mixed by pipetting. The mixture was
   transferred to a 10-cm untreated culture dish. Islets were handpicked
   under an inverted microscope at 4×. Islets were then sized and
   handpicked to a 6-cm dish.

Perifusion

   Base perifusion media: 1 liter of deionized water plus the following:
   3.2 g of NaHCO[3] (Sigma-Aldrich, catalog no. S6014-500G), 0.58 g of
   l-glutamine (Sigma-Aldrich, catalog no. G8540-100G), 0.11 g of sodium
   pyruvate (Sigma-Aldrich, catalog no. P2256-25G), 1.11 g of Hepes
   (Sigma-Aldrich, catalog no. H7523), 8.28 g of Dulbecco’s modified
   Eagle’s medium (Sigma-Aldrich, catalog no. 90113), 1 g of RIA-grade BSA
   (Sigma-Aldrich, catalog no. A7888), and 70 mg of l-ascorbic acid
   (Sigma-Aldrich, catalog no. A5960).

   Glucose medium (5.6 mM) was made by combining 550 ml of base media with
   0.5549 g of glucose (Sigma-Aldrich, catalog no. D16). High-glucose
   medium (16.7 mM) was made by adding 0.7522 g of glucose (Sigma-Aldrich,
   catalog no. D16) to 250 ml of base perifusion medium. KCl medium was
   mice by by adding 0.149 g of potassium chloride (Sigma-Aldrich, catalog
   no. BP366-500) to 100 ml of 5.6 mM glucose media. Perifusion setup was
   made up of a circulating water bath, a fraction collector, a
   peristaltic pumps capable of pumping 1 ml/min, and glass columns with
   two fixed-end pieces (Diba Omnifit LC columns (Cole Parmer, catalog no.
   006CC-10-05-FF) as described ([309]92).

   Islets were transferred to a clear 1.5-ml microcentrifuge tube and then
   loaded into perifusion system. Bubbles were removed, before starting
   the system, and fraction collection was set to every 3 min. After
   collection of 10 preliminary fractions using 5.6 mM glucose baseline
   media to wash the islets, the experiment started. Baseline medium was
   used from 0 to 12 min, 16.7 mM glucose medium was used from 12–42
   minutes, and then intake was switched back to baseline media from 42 to
   63 min. Next, 5.6 mM glucose with 20 mM KCl was used from 63 to 72 min,
   and then the medium was switched back to the baseline media for the
   rest of the experiment. Islets were recovered to a 6-cm dish to
   calculate IEQ after the perifusion.

Quantification and statistical analysis

   Graphpad Prism was used for all statistical analyses. A P value of
   ≤0.05 was used as the cutoff for significance. All results and tests
   used can be found in the relevant figure and figure legends. Graphs
   show all individual data points with the bar showing the mean and the
   error bars showing SEM. One-way or two-way analysis of variance (ANOVA)
   were used for all group comparisons with post hoc Tukey’s HSD test
   aside from sequencing, which used Deseq2 analysis and a
   Benjamini-Hochberg correction for the adjusted P value. Sample sizes
   were determined with a power calculation, and animals were randomly
   assigned to groups that were matched for BW and spread.

Acknowledgments