Effects of Fibroblast Growth Factor 21 on Lactate Uptake and Usage in Mice with Diabetes-Associated Cognitive Decline

Fibroblast growth factor 21 (FGF21) is an endocrine hormone that exerts beneficial effects on glucose and lipid metabolic homeostasis. However, the impact of FGF21 on type 1 diabetes–associated cognitive decline (DACD) and its mechanisms of action remain unclear. In this study, we aimed to evaluate the effects of FGF21 on lactate uptake and usage in a mouse model of streptozotocin-induced DACD. Six-week-old male C57BL/6 mice were divided into the control, diabetic, and FGF21 (which received 2 mg/kg recombinant human FGF21) groups. At the end of the treatment period, learning and memory performance, nuclear magnetic resonance–based metabonomics, and expressions of various hippocampal protein were analyzed to determine the efficacy of FGF21. The results showed that compared to the control mice, the diabetic mice had reduced long-term memory performance after the hyperglycemic insult; decreased hippocampal levels of lactate dehydrogenase-B (LDH-B) activity, bioenergy metabolites, and monocarboxylate transporter 2 (MCT2); and increased lactate levels. Impaired phosphoinositide 3-kinase (PI3K) signaling was also observed in the diabetic mice. However, FGF21 treatment improved LDH-B activity, β-nicotinamide adenine dinucleotide, and ATP levels, and increased MCT2 expression and PI3K signaling pathway, which in turn improved the learning and memory defects. These findings demonstrated that the effects of FGF21 on DACD were associated with its ability to improve LDH-B-mediated lactate usage and MCT2-dependent lactate uptake. Further, these beneficial effects of FGF21 in the hippocampus were mediated by the PI3K signaling pathways.


Introduction
Type 1 diabetes mellitus (T1D) is an autoimmune disease characterized by hyperglycemia and insulin de ciency due to pancreatic β-cell destruction. Notably, T1D is associated with cognitive impairment [1].
Diabetes-associated cognitive decline (DACD) is characterized by a decline in cognition, including learning and memory defects, as well as an impairment of language, understanding, and implementation, which seriously affect a patient's quality of life [2][3][4]. The pathogenesis of DACD is known to include oxidative stress, advanced glycation end products, and mitochondrial dysfunction [5,6]. Some metabolic disturbances (i.e., hyperglycemia), polyols, lipids, and glycation end products have also been implicated in cognitive impairment [7][8][9][10][11][12]. Considering the dramatic increase in the global number of patients with diabetes, research on potential therapeutic targets has gained great signi cance.
Fibroblast growth factor 21 (FGF21) is an endocrine hormone that functions as a regulator in insulin sensitivity as well as glucose and lipid metabolism [13]. Previous studies showed that the administration of FGF21 contributed to metabolic improvements, such as weight loss, insulin sensitivity, and blood glucose and insulin levels [14][15][16]. Except for the peripheral organs, FGF21 has been detected in different brain regions, including the cortex, hippocampus, striatum, and hypothalamus, and is involved in different neurophysiological functions [13]. Geller  Metabonomics is a systemic biological approach used to screen for characteristic metabolites and their perturbed pathways in complex diseases by utilizing information related to pathological factors or by monitoring the effects of drugs and their mechanisms of action [22][23][24][25][26]. Nuclear magnetic resonance (NMR) analysis has high reproducibility and quantitative advantages, because it requires little sample pretreatment [27]. We have previously examined disturbances in the glutamate-glutamine cycle in the hippocampal regions of db/db mice by using NMR-based metabonomics [28]. Further, we integrated molecular biology approaches to identify the glycolytic pathway and lactate levels, which were closely related to the pathogenesis of DACD [29][30][31][32]. By applying the same strategies, we also analyzed the basic FGF-remodeled metabolic phenotypes in the db/db mice, which involved reductions in oxidative stress and apoptosis mediated by reactive oxygen species signaling [33].
These previous studies clearly demonstrated that a combination of metabonomics and molecular biology approaches can identify precursor-product relationships that drive perturbations, thereby providing us a better understanding of therapeutic targets as well as highly reliable results. In the present study, we established a streptozotocin (STZ)-induced type 1 DACD mouse model. Thereafter, we integrated the aforementioned technologies to explore whether exogenous FGF21 administration could have a protective effect on learning and memory after a hyperglycemic insult and to determine its relationship with PI3K-dependent lactate uptake and lactate dehydrogenase-B (LDH-B)-dependent lactate usage.

Subjects
Six-week-old male C57BL/6 mice (weight, 18−22 g) were purchased from Shanghai Slack Laboratory Animal Co., Ltd., and raised in speci c-pathogen-free conditions at the Laboratory of the Experimental Animal Center of Wenzhou Medical University, with regulated humidity and temperature and a 12/12 h light-dark cycle. During the entire experiment, mice were fed with standard rat chow and tap water. All the experiments followed the National Institutes of Health Guide for the Care and Use of Laboratory Animals. The Institutional Animal Care and Use Committee of Wenzhou Medical University also approved the experiments (document number: wydw2012-0083). We also took steps to minimize the suffering and numbers of mice used.
In the experiment, mice were randomly divided into three groups: control, diabetic, and FGF21-treated diabetic (FGF21) groups. First, T1D was induced using streptozotocin (STZ, i.p., 50 mg/kg), which was freshly prepared in 0.1 M citrate buffer (pH 4.2-4.5), for ve consecutive days. The control mice were administered an identical volume of the vehicle. Random blood glucose level in a blood sample drawn from the tail was measured 72 h later using a strip-operated blood glucose monitoring system (One Touch Ultra, Lifescan). Mice with blood glucose levels higher than 11.10 mM were de ned as diabetic mice [34]. After inducing hyperglycemia for 9 weeks, the diabetic mice were subdivided into two groups that received either the vehicle or FGF21 (2 mg/kg, i.p.) daily for four consecutive weeks. The control and untreated diabetic groups received sterile saline. The recombinant human FGF21 used in the study was expressed and puri ed according to previously described procedures [35].

Morris water maze test
The Morris water maze (MWM) test was performed as previously described [36,37]. Brie y, after the drug or vehicle administration, the mice in each group were selected to perform a learning task in the MWM.
The maze consisted of a white pool (diameter, 200 cm; depth, 50 cm) lled with a mixture of water and a non-toxic white dye (temperature, 26 ± 2°C). The behavior of the mice was traced using a camera connected to a XR-XM101 analysis system (Xinruan Information Technology Co. Ltd, Shanghai). The pool had four quadrants, and each orientation was designated as a starting position (North, South, East, or West). The pool also had a camou aged escape platform submerged 2 cm below the water surface.
The platform was placed in the middle of the second quadrant. The mice were trained for 4 days, and on each training day, the mice were placed in the four quadrants for 20 min intervals. The trial was started by placing a mouse at one of the starting points, with its back facing the platform. The trial was terminated when the mouse reached and stood on the platform, but the mouse was left on the platform for an additional 10 s. However, if a mouse could not nd the platform within 60 s, it was guided and placed on the platform for 20 s. During the spatial navigation test, all the mice were trained four times each day. On the 5th day, the probe trial was carried out, wherein the mice were allowed to swim for 60 s, with the platform removed. The swimming distance, latency to nd the platform, the number of crossings over the platform, and the time in the target quadrant were measured. After training, each mouse was removed from the pool, dried, and returned to its cage.

Hippocampal samples preparation
Twenty-four hours after the end of the MWM test, the mice were euthanized and their hippocampal tissues were dissected immediately, snap-frozen in liquid nitrogen, and stored at -80°C. The preparation of the brain samples and acquisition of 1 H-NMR spectra were performed as previously described [38,39].
Brie y, the frozen hippocampal tissues were weighed carefully and placed in centrifuge tubes. Methanol (4 mL/g) and water were added into the tube immediately, homogenized at 4°C, and mixed by vortexing. Chloroform (2 mL/g) and distilled water were then added into the tube and mixed again. The tube was kept on ice for 15 min, and then centrifuged for 15 min at 4°C. The obtained aqueous extracts were then dissolved in 600 µL of 99.5% D 2 O for 1 H-NMR spectroscopy.
Multivariate pattern recognition analysis of 1 H-NMR spectra The detailed acquisition protocol of 1 H-NMR spectra has been previously described [40, 41]. All NMR spectra were phased and corrected, and then data-reduced to 1,100 integrated regions corresponding to the region of δ 10-0 by using the Topspin 2.1 software. The δ 4.6-5.0 region was removed to eliminate artifacts related to residual water resonance. The spectral segments were normalized to the total of the spectral intensity. The normalized integral values were subjected to multivariate pattern recognition analysis using the SIMCA-P + V12.0 software package (Umetrics, Umea, Sweden). Supervised partial least squares-discriminant analysis (PLS-DA) was performed for class discrimination and biomarker identi cation [42]. Metabolites were assigned using Chenomx Pro ler, a module of Chenomx NMR Suite version 7.7 evolution edition (Chenomx, Inc., Edmonton, AB, Canada) and the Human Metabolome Database (version 3.6).
Data were visualized using a principal component score plot to provide the most e cient representation of the data, wherein each point represented an individual spectrum. The score plots were used to visualize the separation between the groups. The loading plots, which were colored according to the absolute value of the correlation coe cient, could identify how the metabolites contributed to the separation of the groups. The scores and loading plots complemented each other. A leave-one-out crossvalidation and permutation test (200 times) on the rst component was also applied to estimate the robustness and credibility of the PLS-DA model. If the plots showed that the Q 2 regression line had a negative intercept and that the R 2 values on the left were thus lower than those of the original point, the PLS-DA model was considered robust and credible. Meanwhile, three additional parameters were calculated: R 2 X and R 2 Y, which explained the variance in the matrix of spectrum data and class membership, respectively, and Q 2 , which indicated the predictive capability of the model. These LDH activity assessment LDH activity was determined using a protocol detailed in our previous study.
[32] Brie y, to determine the LDH lactate to pyruvate (L→P) activity, we used an LDH kit (Jiancheng Bioengineering Institute, Nanjing) according to the manufacturer's instructions. LDH L→P activity was detected at 25°C by measuring sample absorbance at 430 nm wavelength. For determining LDH P→L activity, we used an assay designed on the basis of a procedure documented in the literature [44]. Protein lysates were diluted at a concentration of 0.05 mg/mL in 500 mM potassium phosphate buffer (PPB) and were added to the LDH assay reagent containing 100 mM pyruvate and β-nicotinamide adenine dinucleotide (NADH) in 500 mM PPB (5 mg/20 mL; pH 7.5). Changes in NADH absorbance were measured at 25°C, at 1-min intervals for 10 min, at 340 nm wavelength. LDH total activity was calculated in IU/mg protein and expressed as the P:L ratio (LDH P→L/LDH L→P).

ATP levels measurement
The mouse hippocampal tissue/cell samples were homogenized and lysed, and the ATP content was detected using the Enhanced ATP Assay Kit (Beyotime Institute of Biotechnology, Shanghai) according to the manufacturer's instructions.

NAD + and NADH contents determination
The mouse hippocampal tissues samples (10-20 mg) were homogenized, and the NAD + /NADH Assay Kit with WST-8 (Beyotime Institute of Biotechnology, Shanghai) was used to detect the NAD + and NADH contents.

Reverse transcription-polymerase chain reaction analysis
The mouse brain tissue samples (10-20 mg) were homogenized in TRIzol reagent (Invitrogen, Carlsbad, CA) to extract total RNA. The reverse transcription (RT) of RNA into cDNA was performed by using the PrimeScript TM Real-Time Reagent Kit (Takara, RR037A). The polymerase chain reaction (PCR) analysis was performed using CFX96 Touch™ Real-Time PCR Detection System (Bio-Rad, Hercules, CA). The primer sequences are provided in supplemental Table S1.
C6 cells were cultured with normal glucose, high glucose (HG) for 48 h. Then astrocyte-conditioned medium (ACM) were collected. After ltration, SH-SY5Y cells were respectively treated with ACM and FGF21 for further 48 h. Refer to Figure 6C for the CM testing process [45,46].

Western blot analysis
We used buffer containing 1% protease and phosphatase inhibitors (Beyotime Biotechnology Institute) to extract the total protein from brain tissues and cell. Protein samples were loaded and electrophoresed, and then transferred to polyvinylidene uoride membranes. After electrophoresis, the membranes were incubated in 1×TBST containing 5% skimmed milk for 2 h. Thereafter, the membranes were placed in the primary antibody and incubated overnight at 4°C. The primary antibodies used were as follows: MCT1 . After washing with 1× TBST three times, the membranes were incubated with the secondary antibodies (goat anti-mouse IgG (H+L), 1:5000, SA00001-1, Proteintech; goat anti-rabbit IgG (H+L), 1:5000, SA00001-2, Proteintech) for 1 h at room temperature. We used the ChemiDoc XRS + Imaging System (Bio-Rad) to detect the bands and Image-Pro Plus 6.0 software to analyze the gray values of the bands.

Histological analysis
The brain tissues were isolated and xed in 4% paraformaldehyde for at least 24 h. Thereafter, they were embedded in para n and sectioned using a slicing microtome (Leica, Germany).
For the MCT2 immunohistochemical staining experiments, para n sections were placed in an incubator at 60°C for 1 h. The dewaxed and dehydrated para n sections were then incubated in 3% H 2 O 2 for 10 min and incubated in boiling citrate buffer for 5-10 min. Thereafter, the sections were blocked with 5% bovine serum albumin for 1 h. The para n sections were then incubated with the MCT2 antibody (1:50; SC-166925, Santa Cruz Biotechnology) overnight at 4°C. After washing the sections three times in 1× phosphate-buffered saline (PBS), they were incubated with the secondary antibody (goat anti-mouse IgG, 1:200; SA00001-1, Proteintech) at 37°C for 1 h and stained using a DAB kit (ZLI-9017, ZSGB-BIO). Images of the stained sections were captured using a Nikon ECLIPSE Ti microscope.
For the LDH-B immuno uorescence experiments, most of the protocols were similar to those used in the aforementioned immunohistochemical analysis. The only difference was that the sections were incubated with the uorescent secondary antibody after washing three times in PBS.

Cell immuno uorescence
The cells were spread in a 24-hole plate containing the cover glass at a density of 2 × 10 4 /hole, after washing the sections three times in PBS, they were incubated with the 4% Polyformaldehyde for 45 min. Then 0.5% Trition-X-100 punched for 15 min, the plates were blocked with 5% bovine serum albumin. The plates were then incubated with the MCT2 antibody (1:200; SC-166925, Santa Cruz Biotechnology) overnight at 4°C. After washing the plates three times in PBS, they were incubated with the uorescent secondary antibody for 1 h.

Statistics
Data are presented as mean ± SEM. All data were analyzed using one-way analyses of variance (ANOVA) followed by Scheffe's post-hoc test. Values of p < 0.05 were considered statistically signi cant. SPSS for Windows, Version 13.0 (SPSS Inc., Chicago, IL) was used for all statistical tests. The gures were generated using Prism 5.0 software (GraphPad Software, Inc., San Diego, CA).

Results
Effects of FGF21 on blood glucose, body weight, learning and memory performance in diabetic mice In the present study, we established an STZ-induced T1D mouse model and treated these mice with 28 consecutive days of FGF21 (2 mg/Kg) after STZ administration for four weeks. We found FGF21 could reduce high blood glucose in T1D mice but fail to restored weight loss caused by diabetes (Fig. 1A-B). Oral glucose tolerance tests or insulin tolerance tests also showed improving glucose tolerance in the mice treated with FGF21 ( Fig. 1C-F). To investigate whether FGF21 administration can alleviate long-term learning and memory defects in diabetic mice, we used the MWM test and RT-PCR analysis to examine the related indicators. Figure 2 shows the swimming behavior indicators and immediate early gene (IEG) levels in the differently treated mice. As expected, the diabetic mice showed a longer escape latency, as well as a shorter platform crossing time and target quadrant time, than did the controls, suggesting impairments in their long-term spatial memory formation due to the hyperglycemic insult. However, the FGF21-treated mice displayed signi cant improvements in escape latency compared to the untreated diabetic mice (Fig. 2A). We removed the platform after the last training session, the mice in the three groups showed no differences in swimming distance ( Fig. 2B-C), suggesting similarity in their activity in the probe trial. Nevertheless, FGF21 treatment could signi cantly improve the 24-h memory performance of diabetic mice, as shown by their platform crossing times and target quadrant time ( Fig. 2D-E).
In addition, the molecular indicators underlying the process of long-term (24-hour) memory include the expressions of IEG products, such as early growth response 1 (Egr-1), proto-oncogene (c-Fos), activityregulated cytoskeletal-associated protein (Arc or Arg3.1), as well as several presynaptic (SYP) and postsynaptic proteins (PSD-93 and PSD-95). We found reduced expression levels of these proteins in the hippocampal extracts of diabetic mice. Nevertheless, FGF21 administration could restore both their mRNA (Fig. 2F) and protein levels (Fig. S1). Thus, both the MWM test and molecular indicators demonstrated that FGF21 administration could signi cantly alleviate the hippocampal-dependent spatial learning and long-term memory defects in diabetic mice.
First, we performed a PLS-DA to examine the metabolic patterns in the three groups. The score plot showed a clear separation among the three groups (Fig. 3D), indicating their different metabolic pro les.
We then performed permutation tests with 200 iterations to assess the possibility of model over tting.
The obtained parameters were R 2 X (cum) = 0.525, R 2 Y (cum) = 0.650, and Q 2 (cum) = 0.579, and the pvalue of cross-validated ANOVA was less than 0.001, which indicated that the PLS-DA model was highly reliable (Fig. 3E, Table S2).

Characteristic metabolite alterations in the DACD mice
To identify the characteristic metabolites in the hippocampal extracts that were responsible for the differences among the groups, we applied a PLS-DA for pairwise comparison. The scores and corresponding loading plots with correlation coe cients (|r| > 1.0 is labeled) are shown in Figure 4. The PLS-DA score plots demonstrated signi cant differences between the diabetic and control groups (R 2 X (cum) = 0.504, R 2 Y (cum) = 0.863, Q 2 (cum) = 0.691, p < 0.001; Figure 3A), the FGF21 and control groups (R 2 X (cum) = 0.570, R 2 Y (cum) = 0.870, and Q 2 (cum) = 0.771, p < 0.001; Figure 3B), and the diabetic and FGF21 groups (R 2 X (cum) = 0.458, R 2 Y (cum) = 0.716, and Q 2 (cum) = 0.437, p < 0.01; Fig. 4C). The quality parameters of each PLS-DA plot are shown in Table S2.
Simultaneously, we calculated the relative integrals of the metabolites in the three groups and compared their levels (Table S3). Compared to the control group, the diabetic group showed upregulated levels of lactate, creatine, taurine, and myo-inositol. In contrast, 15 metabolites were downregulated in the diabetic group, including acetate, NAA, GABA, glutamate, succinate, glutamine, glutathione, choline, phosphocholine, ADP, AMP, glycerophosphocholine, tyrosine, and phenylalanine. However, compared to diabetic mice, the FGF21-treated mice showed decreased levels of NAA, GABA, aspartate, creatine, taurine, and phenylalanine, but increased levels of lactate, acetate, succinate, choline, phosphocholine, glycerophosphocholine, and myo-inositol. Interestingly, the levels of some metabolites (i.e., lactate, myoinositol, NAA, GABA, and phenylalanine) in the FGF21 mice did not recover to the control levels, thereby suggesting that the pharmacological activity of FGF21 is complex and diverse.
Furthermore, to con rm characteristic alterations in the metabolic pro les after drug treatment, some metabolites with |r| > 1.0 in the loading plots and p < 0.05 in metabolic levels were ranked according to their variable importance in projection (VIP) scores ( Fig. S2-3), in which VIP > 1 was set as the threshold value. Lactate had the highest VIP value in the plot, suggesting that the glycolytic pathway closely correlated with the effect of FGF21 treatment in DACD mice.

FGF21 promotes lactate usage via LDH-B-dependent glycolysis activity
In glycolysis, LDH-A favors the reduction of pyruvate to lactate in the presence of NADH (P→L). LDH-B is responsible for the opposite reaction, i.e., the oxidation of lactate to pyruvate in the presence of NAD + (L→P), followed by the TCA cycle and ATP production in neurons (Fig. 5A) [47]. In the enzymatic activity assay, we observed a greater increase in P→L conversion and a greater decrease in L→P conversion in the diabetic mice than in the controls, which resulted in a total increase in P→L conversion and lactate accumulation. However, FGF21 treatment could restore these alterations, as re ected by the enhancement in the L→P pathway and lactate usage (Fig. 5B-C).
LDH has ve isoenzymes, namely LDH 1-5, and each isoform is composed of M and H subunits, which are the gene products of LDH-A and LDH-B, respectively [32]. As expected, the hippocampal PCR data showed that LDH-A mRNA levels were upregulated by about 100%, while LDH-B mRNA levels were downregulated by over 50% in the diabetic mice at 15 weeks. However, FGF21 treatment could restore all the alterations in the mRNA levels (Fig. 5D). Moreover, the changes in the levels of ATP and NAD + /NADH, which are glycolytic L→P direction products, con rmed the altered LDH activity (Fig. 5E-F). Furthermore, the results of immuno uorescence staining of the hippocampal regions con rmed that the decreased expression of LDH-B in diabetic mice was reversed by FGF21 treatment (CA3 region; Fig. 5G). Western blot analysis also showed the same LDH-A and LDH-B expression trends (Fig. 5H), which suggested that the changes in LDH protein are consistent with those in its activity and mRNA levels. Taken together, our data indicated that the higher lactate levels in the diabetic hippocampi were the result of elevated lactate production and reduced lactate usage. Nevertheless, FGF21 treatment could increase lactate usage in neurons by promoting lactate to pyruvate conversion, followed by ATP and NADH production. FGF21 promotes lactate uptake via the PI3K-dependent MCT2 pathway Lactate is transported through the cell membrane by different proton-linked monocarboxylate transporters (MCTs) [48]. In the CNS, MCT4 is expressed exclusively on astrocytes, while MCT1 is mainly detected in endothelial cells of blood vessels and glial cells, which can supply lactate to the surrounding neurons [48,49]. MCT2 is found in the axons, dendrites, and cell bodies of neurons [50]. We performed western blot and histological analyses to compare the protein levels of MCT1, MCT2, and MCT4 in the hippocampi of the mice. The results showed signi cant suppression of MCT2 protein levels in the diabetic mice, and these levels were restored after FGF21 treatment, indicating the restoration of lactate uptake into the neurons after drug administration (Fig. 6A-B). To determine if the enhancement in MCT2 protein synthesis occurs due to translational or transcriptional regulation, we also examined their mRNA levels via RT-PCR. However, as shown in Supplementary Fig. 4, FGF21 exerts no effects on MCT2 mRNA levels.
The activation of PI3K is involved in the effects of brain-derived neurotrophic factor (BDNF) and insulin growth factor-1 (IGF-1) on MCT2 expression [51,52]. To determine the mechanism through which FGF21 enhanced MCT2 expression, we examined the signaling pathways by monitoring their phosphorylation levels in the hippocampal extracts. Thereafter, we determined the phosphorylation levels of Akt at Ser473, mTOR at Ser2448, S6 ribosomal protein at Ser235/236, using western blot analysis. Our results revealed greater suppression of phosphorylation protein expressions in the hippocampus of diabetic mice than in the controls, but this suppressed expression was restored after FGF21 administration (Fig. 6A).
In addition, we established an in vitro model in which cell condition culture with high glucose (HG) was applied, that was devoid of exogenous insulin, mimicking diabetic environment (Fig. 6C). The expression of MCT2 protein in SH-SY5Y cells was remarkably reduced under 75 mM glucose astrocyte-conditioned culture (ACM, Fig. 7A-B). In addition, 75 mM HG-ACM treatment can signi cantly reduce the synaptic proteins, i.e. PSD-95, SYP, and EGR-1. c-Fos, while FGF21 (100 ng/ml) can restore these expressions (Fig. 7C-D), which were consistent with the in vivo results. Furthermore, FGF21 treatment also reversed LDH-B expressions and ATP levels in SH-SY5Y cell (Fig. 7E-G). Using an MCT2 inhibitor α-Cyano-4hydroxycinnamic acid (α-CHC, 2.5 mM), we also found that restoration of LDH-B and ATP by FGF21 in HG situation were potently abrogated (Fig. 7E-G).
To validate the role of PI3K/AKT/mTOR in FGF21-mediated alleviation of bioenergy dysfunction induced by HG, we used PI3K inhibitor LY294002 (10 µM) and mTOR inhibitor Rapamycin (20 ng/mL) in vitro. Consistently, we found that activations of PI3K, m-TOR and MCT2 expressions by FGF21 in SH-SY5Y cells were markedly compromised (Fig. 6D-G). The results of immuno uorescence were consistent with the western blotting data (Fig. 6H-I). This in vitro model further strengthened the notion that MCT2 plays an important role in FGF21-mediated improvement in cognition function. Taken together, these results suggested that the enhancement of PI3K-dependent MCT2 protein might be the vital molecular mechanism underlying the effects of FGF21 in DACD mice (Fig. 8).

Discussion
FGF21, which is a hormone that belongs to the FGF family, exerts several bene cial effects on diabetes and associated complications [53]. Furthermore, exogenously administered FGF21 can cross the blood brain barrier via simple diffusion, making it a potential therapeutic medicine for CNS diseases [54]. In the present study, we found that FGF21 could rescue blood glucose levels, and improve learning and memory impairments in mice with STZ-induced T1D. These results are consistent with the previous studies [53]. Taken together, the present ndings demonstrated that FGF21 administration to DACD mice 1) restored long-term learning and memory defects; 2) improved hippocampal bioenergetic metabolite levels; 3) promoted LDH-B-dependent lactate usage in neurons; 4) increased PI3K-dependent MCT2 levels (Fig. 8).

Bioenergetic pathway alterations in DACD mice
Following STZ injection for 9 weeks, we observed reduced MWM test performance and decreased synaptic plasticity and IEG molecule levels in the hippocampal regions of diabetic mice than in those of the controls. Both results indicated the presence of hippocampal-dependent learning and long-term (24 hours after last training) memory defects in the diabetic mice, i.e., the occurrence of DACD. Therefore, the metabolic changes in the hippocampus might be directly related to the pathogenesis of DACD. The NMRbased metabonomic analysis of hippocampal extracts also revealed that the levels of energy metabolites (ADP, AMP, and creatine phosphate) showed a downward trend, suggesting a decline in energy production in the diabetic mice. As phosphate compounds, both adenosine phosphate and creatine phosphate are high-energy reservoirs, which are crucial for sustaining brain functions, i.e., the maintenance of ion gradients and uptake of neurotransmitters [55]. The human brain is known to consume 20% of the energy in a resting state, while only accounting for 2% of the body weight [56]. Therefore, the decline in energy production in the hippocampal regions may lead to the pathogenesis of DACD.
Glucose metabolism is a catabolic process that produces energy. Glucose is rst converted into pyruvate, which is then transformed into lactate by anaerobic glycolysis, or is metabolized in the TCA cycle to produce additional reducing equivalents that drive the electron transport chain and ATP production. Importantly, the disruption of the glucose pathways can result in damages to neuronal function [57,58]. AD is a common pathology of cognitive dysfunction among the elderly and is characterized by the accumulation of cerebral amyloid-β peptide and tau protein phosphorylation [59]. Multiple studies have suggested that energy metabolism impairment precedes neurodegeneration and cognitive decline in patients with AD [60].
Similarly, the decreased succinate levels in hippocampal regions indicated inhibited aerobic glucose metabolism in the diabetic mice. This accumulated lactate may result from reduced lactate usage in neurons or increased lactate production in the diabetic state, which were consistent with the ndings of our previous studies [28, 31, 61]. Using [2-13 C]acetate and [1-13 C]glucose as tracer substrates, we previously identi ed inhibited pyruvate recycling in the neurons of diabetic animals [7,30]. Through in vitro primary cell culture, we further con rmed the decreased usage of lactate in neurons but not in astrocytes [62,63]. All these data suggested that the reduced usage of lactate in neurons was a crucial factor contributing to neuronal dysfunction and bioenergetic de cits in the diabetic mice.
The astrocyte-neuron lactate shuttle (ANLS) hypothesis proposes that in astrocytes, glucose is predominantly metabolized to lactate, which is also a primary fuel source for neuronal and synaptic activities [64-67]. Lactate is an effective neuroprotective metabolite, and exogenous lactate administration can rescue neuronal activity during glucose deprivation [46]. Currently, lactate is recognized as a source of energy as well as a signaling molecule or receptor agonist, modulating neuronal excitability, plasticity, and cognition [68, 69]. Suzuki et al. identi ed the involvement of glycogen breakdown and lactate release from astrocytes in the long-term memory process [70]. In addition, the effects of lactate are mediated by NADH/NAD + -dependent NMDA receptor activity, which involves synaptic plasticity-related genes such as Egr-1, Arc, and c-Fos [71,72]. Therefore, the effects of lactate on learning and memory are associated with direct or indirect redox or energy-dependent mechanisms, which may be attributed to NADH and ATP production in the glycolytic pathway that is also an underlying target of the drug. FGF21 alleviates DACD through promoting neuronal lactate usage FGF21 functions as an energy metabolism regulator and performs a pleiotropic function in glucose and energy metabolism [14][15][16]. Growing evidence has revealed the potential role of FGF21 in CNS diseases, in which it promotes mitochondrial antioxidant enzyme action and biogenesis, thereby enhancing neuronastrocyte interactions [73]. We also identi ed that FGF21 treatment could improve learning and memory impairments in diabetic mice using the MWM test and synaptic plasticity data, which were consistent with those of previous studies [74][75][76]. Furthermore, we determined that FGF21 modulated hippocampal lactate and succinate levels, indicating that its therapeutic effects involved glycolysis remodeling.
Alterations in the glycolysis rate result in changes in the levels of pyruvate, lactate, and NADH. The greater the LDH-A/LDH-B activity ratio was enzymatically controlled, the more lactate was converted to pyruvate. Our data revealed that FGF21 alleviated learning and memory defects, with corresponding changes in the LDH-A/LDH-B activity ratios. LDH has different homo-or hetero-tetrameric isoforms transcribed by LDH-A and LDH-B, and the ANLS hypothesis suggests they are predominantly present in astrocytes and neurons, respectively [77]. In this study, FGF21 directly contributed to the L→P conversion in neurons, accompanied by elevated ATP and NADH production.
Our results are similar to those recently reported by Sun et al.[46] They observed that FGF21 can rescue the ANLS system by increasing the expressions of LDH-A, LDH-B, MCT1, and MCT2, which are all reduced in an AD mouse model. Additionally, in an aging mouse model, scientists found decreased LDH-B expression in the hippocampus, which may be a major factor leading to a decline in cognition [78]. In the study, FGF21 ameliorated DACD by promoting the L→P conversion in neurons, i.e., the usage of lactate under the activity of LDH-B. Interestingly, we also found that FGF21 could stimulate the activity of LDH-A and increase the lactate content in the hippocampus. This nding was consistent with that of a previous oncological study, which found that FGF21 could directly phosphorylate LDH-A on the tyrosine residue [79,80]. FGF21 promotes neuronal lactate uptake via PI3K/Akt/mTOR-dependent MCT2 expressions MCT2 is predominantly expressed by the neuronal postsynaptic membrane. It allows the uptake of lactate, pyruvate, and ketone bodies across the plasma membrane as energy substrates [50]. Decreased hippocampal expressions of MCT2 have been reported in previous studies on AD and diabetic animal models [81, 82]. The MCT2-related depletion of the lactate pool limits the activities of LDH-B, causes glycolytic shifts (L→P conversion), affects NADH and ATP production, thereby leading to learning and memory defects. Recently, Netzahualcoyotzi and Pellerin con rmed that neuronal MCT2 is involved in 24hour and 7-day long-term memory formation, but not in 90-minute short-term memory formation [83]. BDNF, insulin, and IGF-1, which are implicated in the long-term memory process and modulate synaptic plasticity, can promote the protein expression of MCT2 [51,52]. Similarly, our study also showed that FGF21 increased MCT2 expression, but not the expressions of the glial transporters MCT1 and MCT4.
Thus, one of the mechanisms through which FGF21 alleviates long-term memory defects might be the promotion of MCT2-dependent lactate uptake, in parallel with ATP and NADH production.
As an endocrine regulator, FGF21 has many bene cial effects on cell metabolism by stimulating glucose and lipoprotein uptake in liver cells [84,85]. When FGF21 binds with its receptor, it contributes to the phosphorylation of downstream signaling molecules, such as MAPK, PI3K, signal transducer and activator of transcription, and phosphoinositide phospholipase C-γ [13,86]. In our study, we identi ed the role of the concomitant activation of the PI3K pathway in regulating the protein expression of MCT2 in the hippocampal regions, and both these molecules have been classically implicated in translation regulation similar to BDNF, insulin, and IGF-1 [51,52].
As a downstream protein in the PI3K pathways, MCT2 can interact with glutamatergic AMPA receptor (GluR) subtype GluR2, which is involved in glutamatergic transmission and long-term potentiation [87].
Our observation that FGF21 could upregulate MCT2 in protein levels was consistent with the nding of a previous study that energy metabolism was coupled to synaptic plasticity after BDNF treatment [88]. We hypothesize that to meet high energy demands, the translational regulation of MCT2 are increased to facilitate energy supply and neuronal function under the condition of enhanced synaptic transmission or pathological insult, i.e., in diabetes.

Conclusion
The present study demonstrated that FGF21 treatment in mice with STZ-induced DACD could restore long-term learning and memory defects by promoting LDH-B-dependent conversion of lactate to pyruvate, followed by ATP and NADH production. The study further suggested that the regulation of PI3Kdependent MCT2 expression was a potential target mediating the bene cial effects of FGF21 on DACD. Considering the critical roles of MCT2 and LDH-B on lactate uptake and usage, respectively, our results support the possibility of a link between neuronal bioenergetics and synaptic plasticity. Furthermore, our ndings suggest that the combination of metabonomics and molecular biology is helpful for studying potential molecular mechanisms and drug targets involved in various pathologies. Several limitations should be considered: First, although our data showed that FGF21 treatment for 4 weeks effectively improved learning and long-term memory defects in the mice with DACD, we do not know whether a 4- week-treatment period is optimal for producing the best therapeutic effects. Second, among the four types of FGF receptors, FGF21 can interact with at least FGFR 1-3 [13,86]. However, our understanding of which FGFR mediates FGF21 signal transduction is incomplete. Lastly, the protein β-klotho is required for FGF21 to exert its effects in vivo [89]. Therefore, further studies are warranted to determine the protein involved in the FGF21-mediated stimulation of hippocampal neurons.

Declarations
Ethics approval All the experiments followed the National Institutes of Health Guide for the Care and Use of Laboratory  tolerance test and AUC. n = 9. * p < 0.05, ** p < 0.01, *** p < 0.001 versus CON group; # p < 0.05, ## p < 0.01, ### p < 0.001 versus DM group.    Table S2. Metabolite assignments are the same as in Figure 3.