Gut microbiota mediates intermittent-fasting alleviation of diabetes-induced cognitive impairment

Cognitive decline is one of the complications of type 2 diabetes (T2D). Intermittent fasting (IF) is a promising dietary intervention for alleviating T2D symptoms, but its protective effect on diabetes-driven cognitive dysfunction remains elusive. Here, we find that a 28-day IF regimen for diabetic mice improves behavioral impairment via a microbiota-metabolites-brain axis: IF enhances mitochondrial biogenesis and energy metabolism gene expression in hippocampus, re-structures the gut microbiota, and improves microbial metabolites that are related to cognitive function. Moreover, strong connections are observed between IF affected genes, microbiota and metabolites, as assessed by integrative modelling. Removing gut microbiota with antibiotics partly abolishes the neuroprotective effects of IF. Administration of 3-indolepropionic acid, serotonin, short chain fatty acids or tauroursodeoxycholic acid shows a similar effect to IF in terms of improving cognitive function. Together, our study purports the microbiota-metabolites-brain axis as a mechanism that can enable therapeutic strategies against metabolism-implicated cognitive pathophysiologies.


16S rRNA Microbiome sequencing
The raw sequencing reads were merged and trimmed, following by removing chimera and constructing zero-radius Operational Taxonomic Units (zOTUs) with UNOISE implemented in Vsearch (v2.6.0) 9,10,11 . UNOISE is denoising algorithm to infer accurate biological template sequences from noisy illumina reads, which had comparable or better accuracy and much faster than DADA2 12 . Raw reads were merged with fastq_mergepairs (Vsearch 2 ) using defined parameters of fastq_minovlen = 16 and fastq_maxdiffs = 5. Merged reads were filtered with fastq_filter (Vsearch 2 ) using defined parameters of fastq_truncqual = 4 and fastq_minleng = 400 and primers were chopped from both ends. In order to generate zOTUs, remaining high quality reads were dereplicated, clustering and denoised using derep_fullength, cluster_unoise and uchime3_denovo (Vsearch 2 ) sequentially. Reads were mapped back to the zOTU sequence. The Greengenes (13.8) 13 . 16s rRNA gene database was used for taxonomy annotation of each zOTU using assign_taxonomy.py implemented in Qiime (v1.9.1) 14 .
All the samples were rarefied to 28257 counts (lowest sample depth) to calculated the observed OTU index using alpha_diversity.py in Qiime (v1.9.1) 14 . Using normalized_table.py in Qiime (v1. 9.1) 14 . The raw OTU table was normalized with cumulative sum scaling (CSS) 15 to calculate unweighted Unifrac distance. Permutational multivariate analysis of variance (PERMANOVA), a non-parametric multivariate statistical test, was adopted to detect differences among intervention groups using Adonis function in Vegan 16 .
Constrained analysis of principal coordinate (CAP) in R package Vegan 16 was conducted to identify the influence of mice gene-type and IF on microbiota after setting time as a condition effect. The CSS normalized zOTU table was used to calculate relative abundance and summarized in different levels using Taxonomic-Binning. R in Rhea 17 Specific taxa comparisons among groups was analyzed by analysis of composition of microbiomes (ANCOM) 18 . In brief, ANCOM algorithm,accounts for the underlying dependence structure of microbiota data, makes no distributional assumptions and scales well to compare samples involving thousands of taxa, which has been widely used in recent microbiota researches 18 . Using Correlation.R in Rhea 17 , the Pearson correlation analysis was conducted between centered log-ratio transformed relative abundance of genera and body weight, blood glucose, food intake, water intake, lipopolysaccharide (LPS), leptin, gamma-Aminobutyric acid (GABA), 5-hydroxytryptamine (5-HT), insulin and short-chain-fatty-acids (SCFAs  19 following the official guide. PICRUSt helps to predict metagenome functional content from marker gene, including 16S rRNA surveys and full genomes. Briefly, the zOTU representative sequence were re-mapped with usearch_global in Usearch 20 to the reference OTU sequence in the Greengene (13.5) database as PICRUst utilized the same Greengenes 13.5 assigned OTUs to conduct the prediction. Then, the realigned zOTU table followed the standard PICRUst processing procedure using normalize_by_copy_number.py and predict_metagenomes.py.
Predicted gene was annotated with KEGG 21 at different levels using categorize_by_function.py in PICRUst, and the significantly abundant pathways (at least appearing in 3 samples) were identified by edgeR 22 with FDR-p < 0.1.

Plasma metabolomics
The raw liquid chromatography-mass spectrometry (LC-MS) metabolomics data was processes using commercial software package Progenesis QI 2.0 (Nonlinear Dynamics; Newcastle upon Tyne, UK). Progenesis QI has emerged as a standard software for processing LC-MS metabolomics data and has been widely applied for data deconvolution, peak-picking, alignment, and identifications of metabolites. Samples were analyzed in one batch with a randomized injection order. The stability and functionality of the system were monitored throughout all the instrumental analyses using quality controls, i.e. the pooling of all samples acquired at the beginning of analytical sequence and after every 10 injections.
Data files of the information dependent acquisition scan mode were incorporated in the software for identification purposes to have MS/MS spectra of the most abundant detected metabolites. For the MS/MS detection, all precursors were fragmented using 20-40 eV, and the scan time was 0.2 seconds. During the acquisition, the signal was acquired every 3 seconds to calibrate the mass accuracy. Metabolite features that were detected in ≤ 50% QCs or 80% of biological samples were excluded. Missing values were imputed using knearest neighbor. Metabolite identification was carried out accurate mass (ppm<5) and product ion spectrum (MS/MS ppm<10) matching against different online databases including METLIN the Human Metabolome Database (HMDB, V4.0), NIST and Lipidblast.
The list of microbial metablites (i.e. metabolites whose levels were modified by gut microbiota, n=26) was determined according to Rowan et al., 2017 23 and detailed annotation information is provided in Supplementary Spreadsheet 9.

Integrated multi-omics analysis
Integrated multi-omics data analysis was performed on a priori selected parsimonious set of 36 genes, 17 ANCOM-derived OTUs that differed significantly between db/db and db/db-IF treatment and 26 pre-defined plasma microbial metablites. A detailed data processing workflow and R script are provided as an R markdown file.
First, multivariate predictive modelling on each omics dataset was conducted using partial least square-discriminant analysis incorporated into a repeated double cross-validation framework (rdCV-PLSDA) 24 . The rdCV separates cross validation into an outer "testing" loop and an inner "tuning" (or validation) loop to effectively reduce bias from overfitting models to experimental data, which have shown better results than other cross-validation approaches 25 26 . To gain a robust and reliable estimate of model performance, 200 repetitions of the outer cross validation loop was performed. Data was log-transformed and auto-scaled prior to the rdCV-PLSDA. We further applied permutation analysis (n=1000) to evaluate whether the constructed models outperformed than random classifications.
Second, a multivariate dimension reduction method, DIABLO (Data Integration Analysis for Biomarker discovery using a Latent component method for Omics), was employed for multiple omics integration 27 . DIABLO is a novel R programing based approach that is available in R package 'mixOmics', which is designd for objectively integrating multiple 'omics datasets measured on the same biological samples. This algorithm is based on a variant of the multivariate methodology Generalised Canonical Correlation Analysis. Since each omics dataset has shown good predictive performance, as assessed by rdCV-PLSDA, we applied a full design matrix to seek for linear combinations of variables from each OMICs dataset that are maximally correlated (Supplementary Figure 6A). Subsequently, a tuning procedure (tune.block.splsda function) was applied to determine the optimal number of key predictors in each dataset for a minimum misclassification rate. Model performance was evaluated by 10-fold cross validation. The optimal number of component for each omics dataset was determined by rdCV-PLSDA. DIABLO model was then generated using block.splsda. A global overview of the correlation structure at the component level was represented with the plotDiablo function. A clustered image map that represents the multiomics molecular signature expression for each sample was created using cimDiablo function. The loading weights of each selected variables on each component was represented with plotLoadings function.

R markdown file for multi-omics analysis
This is an R Markdown document that presents the detailed procedure for multi-omics analysis.