# HIGH LEVEL SUMMARY
---
## Summary
This study investigates the role of N-linked protein glycosylation in the pathophysiology of Alzheimer's disease (AD) by integrating spatially resolved metabolomics, lipidomics, and glycomics in post-mortem human brain specimens and transgenic mouse models (5xFAD and PS19) with in vivo stable-isotope ($^{13}\text{C}_6$-glucose) pulse-chase tracing. The authors demonstrate that widespread brain N-glycan hyperglycosylation is conserved across human AD and mouse models and is driven by increased de novo glycan biosynthesis in the endoplasmic reticulum and Golgi rather than impaired turnover, occurring alongside reduced O-GlcNAcylation and hyaluronan levels. Furthermore, genetic knockdown of phosphoglucomutase 3 (*Pgm3*) or pharmacological inhibition of oligosaccharyltransferase (using NGI-1) improves social recognition memory in AD mice independently of amyloid plaque, tau, or astroglial pathology, whereas oral glucosamine supplementation exacerbates hyperglycosylation and memory deficits in AD mice and correlates with accelerated mild cognitive impairment (MCI)-to-dementia progression and reduced survival in a retrospective electronic health record (EHR) analysis.

## Strengths
* The study addresses a critical gap in neurodegenerative metabolism by coupling a $^{13}\text{C}_6$-glucose liquid diet pulse-chase paradigm with ion-mobility MALDI mass spectrometry imaging (MSI) and LC–MS precursor tracing, providing direct in vivo evidence that brain N-glycan accumulation in AD stems from accelerated de novo biosynthesis rather than passive lysosomal storage failure.
* By profiling N-glycans alongside O-GlcNAcylation and hyaluronan, the authors uncover a compartment-specific metabolic redirection of UDP-GlcNAc toward ER/Golgi N-glycosylation at the expense of cytosolic O-GlcNAc and extracellular glycosaminoglycans, reconciling the long-standing paradox of reduced O-GlcNAcylation in AD brains.
* The authors introduce an optimized three-step xylene delipidation protocol for post-mortem human brain tissue that overcomes severe myelin lipid ion suppression, improving spatial N-glycan MALDI–MSI signal recovery by 10- to 20-fold across Braak stages.
* Through bidirectional genetic (*shPGM3*), pharmacological (NGI-1), and dietary (glucosamine) interventions across both amyloid (5xFAD) and tauopathy (PS19) mouse models alongside wild-type controls, the study establishes that hyperglycosylation selectively drives cognitive deficits in the diseased neurometabolic state independently of canonical amyloid or tau burden.
* The translational integration of preclinical mouse models with real-world clinical EHR cohorts highlights a potentially important disease-stage-specific vulnerability to over-the-counter glucosamine supplementation in patients with established MCI or dementia.

## Weaknesses
* In the spatial multiomics violin plots (e.g., Figs. 1c,d, 3c, 5d–g, and 6c,d), statistical comparisons rely on two-tailed Student's *t*-tests across thousands of pooled MALDI–MSI pixels ($n > 2,000$) from only $n = 3\text{–}7$ biological specimens per group; treating spatially autocorrelated pixels as independent replicates constitutes pseudoreplication that artificially inflates degrees of freedom, and the authors might consider re-evaluating these comparisons using biological specimen-level summary statistics or hierarchical linear mixed-effects models.
* The retrospective EHR survival and MCI-to-ADRD transition analyses require at least 1 year of documented post-diagnosis glucosamine use while measuring follow-up from initial diagnosis without time-dependent covariates or landmark analysis (introducing potential immortal time bias), and propensity score matching adjusts only for age and basic demographics without controlling for osteoarthritis indication, healthcare utilization, or regex negation contexts, which could help reconcile the reported mortality increase with neutral or protective associations observed in prospective cohorts like the UK Biobank (Reference Paper: Association of regular glucosamine use with incident dementia: evidence from a longitudinal cohort and Mendelian randomization study, Zheng et al., 2023).
* The intact glycoproteomics workflow omits glycopeptide-specific affinity enrichment (such as HILIC or SAX-ERLIC) and pools $n = 4$ brains into a single composite sample per group ($N = 1$ biological pool) without parent-protein normalization, recovering only 77 human and 137 mouse unique glycopeptides compared to over 10,000 N-glycoforms in established human AD reference datasets and limiting statistical power to support broad claims of global neuronal membrane hyperglycosylation (Reference Paper: Human brain glycoform coregulation network and glycan modification alterations in Alzheimer's disease, Zhang et al., 2024).
* In the $^{13}\text{C}_6$-glucose pulse-chase experiments (Fig. 4d,e and Extended Data Fig. 4), turnover kinetics are inferred from unnormalized post-washout abundance endpoints despite a twofold difference in pre-washout labelled pool sizes between WT and 5xFAD mice, and the figures contain mismatched or duplicated x-axis isotopologue $m/z$ labels alongside identical repeated *P* values across distinct bars that would benefit from a careful audit.
* The manuscript contains several biochemical and methodological inconsistencies—including describing PGM3 as converting glucosamine-6-phosphate to N-acetylglucosamine-6-phosphate (a reaction catalysed by GNPNAT1, whereas PGM3 isomerizes GlcNAc-6-P to GlcNAc-1-P), conflicting descriptions of stereotaxic shPGM3 delivery (lateral ventricles in the text/Methods versus hippocampal injections in Extended Data Figs. 6a and 7a), and discrepancies in NGI-1 dosing (2 nmol vs. 5 nmol), social memory trial counts (4 trials vs. 5 days), and missing direct validation of PGM3 knockdown efficiency.

## Potential Issues And Suggestions
### [Abstract & Introduction: Metabolic Dysfunction and Glycosylation in AD] (Pages: 1-2)
#### Potential Mistakes and Improvements

1. **Alignment of Clinical EHR Cohort Definitions and Stratified Outcomes Across the Abstract, Introduction, and Results (Pages 1410–1411):**
   There is a potential concern regarding how the electronic health record (EHR) patient cohorts and clinical outcomes are summarized in the Abstract and Introduction compared to the data and definitions presented in the Results (pages 1416–1418, Fig. 6f–j) and Methods (page 1423):
   - *Cohort Definitions:* The Abstract describes the analysis as being conducted in *"patients with AD with varying disease severity."* However, the clinical datasets actually comprise two distinct diagnostic groups: mild cognitive impairment (MCI, $n = 41,884$, explicitly defined on page 1416 as *"a prodromal phase rather than a definitive dementia diagnosis"*) and a broader composite cohort of Alzheimer's disease and related dementias (ADRD, $n = 24,481$, which encompasses vascular dementia, Lewy body dementia, and frontotemporal dementia alongside AD; Fig. 6f).
   - *Outcome Stratification:* Both the Abstract (*"associated with accelerated AD progression and worsened survival"*) and the Introduction (*"associated with increased mortality and a higher likelihood of progression from MCI to AD"*) group these outcomes together without specifying their cohort-specific nature. In the Results, worsened survival was observed specifically in the established ADRD cohort ($P < 0.001$, Fig. 6h), whereas glucosamine use had no statistically significant effect on survival in the MCI cohort ($P = 0.252$, Fig. 6i). Furthermore, disease progression was measured specifically as the diagnostic conversion rate from MCI to AD/ADRD (Fig. 6j), rather than progression across severity stages within already-diagnosed AD patients.
   - Revising the phrasing in the Abstract and Introduction to accurately distinguish between the prodromal MCI cohort (used to evaluate conversion to dementia) and the composite AD/ADRD dementia cohort (where reduced survival was observed) would prevent readers from misinterpreting the clinical stratification and scope of the epidemiological findings.

2. **Scope of Neuropathological vs. Behavioural Claims for Hyperglycosylation as a "Driver of AD Pathology" (Pages 1410–1411):**
   In the Abstract (page 1410) and Introduction (page 1411), hyperglycosylation is framed broadly as a *"pathological driver of AD"* and a *"critical driver of AD pathology."* The Introduction (page 1410) motivates this premise by discussing how aberrant glycosylation intersects with canonical neuropathological hallmarks, including microglial inflammatory responses, tau aggregation and propagation, and amyloid-$\beta$ plaque composition and clearance. A potential concern is that, in the subsequent in vivo interventions (Results, page 1416; Extended Data Figs. 8 and 9; Supplementary Fig. 14; Discussion, page 1420), modulating N-glycosylation over the 2-week treatment window (via *Pgm3* knockdown, NGI-1 administration, or oral glucosamine supplementation) significantly altered performance in a single behavioural assay (social recognition memory) *without* altering core neuropathological hallmarks—specifically $\beta$-amyloid plaque burden (in 5xFAD mice), phospho-tau pathology (in PS19 mice), or GFAP+ reactive astrocytosis. Clarifying in the Abstract and Introduction that these interventions specifically modulate cognitive and behavioural deficits independently of canonical amyloid, tau, or astroglial pathology would better align the introductory claims with the experimental evidence presented later in the manuscript.

3. **Precision in Describing Genetic Knockdown Versus Pharmacological Inhibition (Pages 1410–1411):**
   In both the Abstract (page 1410, *"Genetic knockdown of glycan biosynthetic enzymes improves cognitive outcomes in AD mice"*) and the Introduction (page 1411, *"Genetic knockdown of key glycosylation enzymes ameliorated cognitive deficits in AD mice"*), the plural phrasing implies that multiple glycosylation enzymes were genetically knocked down. In the experimental sections (Results, page 1416, Fig. 5; Extended Data Fig. 6; Discussion, page 1420), only a single enzyme—phosphoglucomutase 3 (PGM3)—was targeted via genetic knockdown (lentiviral shPGM3), whereas the second enzymatic target—the oligosaccharyltransferase (OST) complex—was inhibited pharmacologically using the small-molecule inhibitor NGI-1. Updating the Abstract and Introduction to state that *genetic knockdown of PGM3 and pharmacological inhibition of OST* (or *genetic and pharmacological inhibition of glycan biosynthetic enzymes*) improved cognitive outcomes is a potential improvement that would ensure accurate alignment with the experimental methodology.

#### Minor Corrections and Typos

* **Redundant Text Across Adjacent Paragraphs (Page 1410, Right Column, Paragraphs 1–2):** The closing sentences of the first paragraph (*"To this end, perturbations in the entire metabolome have not been comprehensively explored in AD. For example, complex carbohydrate metabolism, such as glycan biosynthesis and processing, remains elusive in the context of AD."*) and the opening sentences of the second paragraph (*"The dynamic metabolome as a hallmark and driver of AD pathophysiology has not been defined... In particular, complex carbohydrate metabolism including glycan biosynthesis and processing in AD remains poorly understood."*) repeat nearly identical points and could be consolidated to improve flow.

### [Spatial Multiomics, Stable Isotope Tracing & Glycoproteomics in Human and Mouse AD] (Pages: 2-6, 8, 17-22)
#### Potential Mistakes and Improvements

1. **Statistical Unit of Analysis and Pseudoreplication in Spatial Multiomics Comparisons (Pages 2–6, 8–10, 17–19, 21–25; Figs. 1c–d, 2d, 3c, 5d–g, 6c–d; Extended Data Figs. 1c, 2a,c, 3a, 5c, 6b–c, 7b,f):**
   - **Potential Concern:** Across the spatial glycomics, metabolomics, lipidomics, and hyaluronan MALDI–MSI violin plots comparing human AD versus normal control brains (Fig. 1c,d), transgenic AD mouse models (5xFAD and PS19) versus wild-type (WT) controls (Fig. 3c; Extended Data Figs. 2a,c, 3a, and 5c), and in vivo interventions (Fig. 5d–g, Fig. 6c–d, and Extended Data Fig. 6b–c), statistical significance ($P < 0.0001$) is assessed using two-tailed Student's $t$-tests where individual tissue pixels ($n > 2,000$ pixels per group across $n = 3\text{–}7$ biological replicates per group, as explicitly noted in the legends of Fig. 1c,d on page 3 and Extended Data Fig. 2a on page 18) serve as the unit of statistical independence. Because thousands of spatially adjacent pixels acquired from the same tissue section are highly autocorrelated technical subsamples that share subject-level biological and analytical variance, treating individual pixels as independent replicates in a standard two-sample $t$-test constitutes pseudoreplication. This artificially inflates the degrees of freedom and drives $P$ values below $0.0001$ regardless of inter-individual biological variability; furthermore, single-pixel MALDI–MSI intensities are markedly right-skewed (as visible in the violin plots in Fig. 1c,d and Fig. 2d), whereas the Methods (page 1423) state that *"Data distribution was assumed to be normal, but this was not formally tested."* Additionally, in the multi-group human Braak-stage progression analysis (page 4, Fig. 2b–e; page 17, Extended Data Fig. 1a–c), the number of independent human donors ($n$) within each Braak stage group (Braak 0, 1–2, 3–4, and 5–6) is not reported, and the violin plots in Fig. 2d and Extended Data Fig. 1c omit formal statistical test results. Similarly, the biological sample size ($n$ animals per group) is omitted from the legends of Fig. 5d–g, Fig. 6c–d, Extended Data Fig. 5c (pages 21–22), and Extended Data Fig. 6b–c, and the specific $m/z$ values of the two representative complex N-glycans are omitted in Fig. 5d–g, Fig. 6c–d, Extended Data Fig. 6b–c, and Extended Data Fig. 7b,f.
   - **Suggested Improvement:** Re-evaluate group comparisons using the biological specimen ($n = 3\text{–}7$ brains per group, or the exact number of independent donors per Braak stage) as the experimental unit of independence—either by comparing specimen-level regional summary statistics (such as mean or median intensity per anatomical region per brain) via two-tailed $t$-tests or ANOVAs, or by fitting a hierarchical linear mixed-effects model that includes biological specimen identity as a random intercept to account for within-section spatial correlation. Furthermore, explicitly report the number of independent biological specimens per group and the specific glycan $m/z$ identities across Fig. 2, Fig. 5d–g, Fig. 6c–d, Extended Data Fig. 1, Extended Data Fig. 5c, Extended Data Fig. 6b–c, and Extended Data Fig. 7b,f alongside specimen-level statistical comparisons.

2. **Kinetic Interpretation, Isotopologue Axis Labelling, and Statistical Reporting in $^{13}\text{C}_6$-Glucose Pulse-Chase Tracing (Pages 5–6, 8, 20; Fig. 4d,e,g; Extended Data Fig. 4a–c):**
   - **Potential Concern (Kinetic Interpretation of the 24-Hour Chase Phase Under Differing Pool Sizes):** On page 5, the authors conclude from the lack of statistically significant differences between WT and 5xFAD isotopologue abundances after a 24-h unlabelled washout (chase) period (Fig. 4e and Extended Data Fig. 4a–c) that *"increased glycan biosynthesis, rather than impaired glycan degradation or salvage, is the predominant driver of hyperglycosylation in AD pathology."* However, comparing raw or unnormalized isotopologue abundances at a single post-washout endpoint without accounting for the pre-washout (end-of-pulse) baseline or total pool size complicates turnover interpretation:
     1. At the end of the 24-h pulse phase ($t = 0$ of the chase; Fig. 4d), 5xFAD brains exhibit roughly twofold higher labelled isotopologue abundance (~40–65 abundance units) than WT brains (~15–30 units). After the 24-h washout (Fig. 4e), 5xFAD isotopologue abundances decline to ~35–45 units, whereas WT isotopologue abundances *increase* from ~15–30 units to ~35–50 units (a pattern also visible across Extended Data Fig. 4a–c), likely reflecting continued label incorporation from slow-turnover upstream precursor pools (such as brain glycogen or hexosamine intermediates) during the chase. Converging to the same absolute abundance after 24 h of washout implies that 5xFAD brains experienced a substantially larger net reduction in $^{13}\text{C}$-labelled glycans during the chase window than WT brains, rather than identical clearance kinetics.
     2. The y-axes in Fig. 4d,e and Extended Data Fig. 4a–c are labelled simply `Abundance` without specifying whether values represent Total Ion Current (TIC)-normalized intensity, raw ion counts, or fractional isotopologue enrichment ($\text{M}+i / (\text{M}_0 + \sum \text{M}_i)$), or whether natural $^{13}\text{C}$ abundance correction was applied. Because the steady-state unlabelled ($\text{M}_0$) glycan pool is markedly expanded in 5xFAD brains (Fig. 3c), static abundance measurements at a single washout time point cannot isolate the fractional degradation rate constant ($k_{\text{deg}}$).
   - **Definitive Flaw (Mismatched and Duplicated Isotopologue $m/z$ Axis Labels):** Inspection of the pulse-chase bar graphs in Fig. 4d,e (page 8) and Extended Data Fig. 4a,b (page 20) reveals multiple x-axis label discrepancies:
     - In **Fig. 4d** (`Enrichment of 1,257`), the 14 x-axis isotopologue $m/z$ labels run from `1,262` to `1,266` and then jump directly to `1,271–1,279` (omitting `1,267–1,270`), whereas in **Fig. 4e** (`24-h washout of 1,257`), the 16 x-axis labels start at `1,264` and run consecutively from `1,264` to `1,279` (omitting `1,262–1,263` while including `1,267–1,270`), preventing direct one-to-one comparison of individual isotopologues between pulse and chase.
     - In **Extended Data Fig. 4a** (`Enrichment of 1688` and washout of `1688`), the x-axis contains a transposed digit (`1965` between `1694` and `1696`) and repeats the mass labels `1699, 1700, 1701` twice (`1696, 1697, 1698, 1699, 1700, 1701, 1699, 1700, 1701, 1702, 1703, 1704, 1705`).
     - In **Extended Data Fig. 4b** (`Enrichment of 1282` and `24hr washout of 1282`), the x-axis isotopologue $m/z$ labels are out of numerical order (`1288, 1289, 1299, 1291, 1295, 1296, 1297, 1298, 1292, 1293, 1294, 1299`, with `1299` appearing twice).
   - **Potential Concern (Inconsistent Statistical Test Descriptions and Repeated $P$ Values):** For the same pulse-chase experimental design ($n = 3$ animals per group), the Fig. 4d,e legend (page 6) states that statistical significance was determined by two-tailed $t$-tests, whereas the Extended Data Fig. 4a–c legend (page 20) states that per-isotopologue $P$ values were determined by two-way repeated-measures ANOVA. Moreover, across both figures, numerous distinct isotopologue bars display identical $P$ values to four or six decimal places (in Fig. 4d, `0.105564` appears on four bars at `1,262`, `1,272`, `1,273`, and `1,274`, `0.012455` appears on two bars at `1,266` and `1,271`, and `0.007846` appears on two bars at `1,277` and `1,278`; in Fig. 4e, nearly all $P$ values are close to `1.0` such as `>0.9999`, `0.9968`, `0.9953`, and `0.9963`; in Extended Data Fig. 4a, `0.0040` appears on nine pulse bars and `0.9987` on all 13 washout bars; in Extended Data Fig. 4b, `0.0189` appears on 11 pulse bars; in Extended Data Fig. 4c, `0.0005` appears on eight pulse bars), which is characteristic of a multiple-comparison post-hoc adjustment (such as Holm–Šídák's or Šídák's test following a two-way ANOVA) rather than independent two-tailed $t$-tests. Conversely, in Fig. 4g (page 8), error bars ($\text{mean} \pm \text{s.e.m.}$) are missing from several isotopologue bars and a single global $P$ value (`0.0353` for UDP-Hexose and `0.0211` for UDP-HexNAc) is displayed above a bracket spanning all 15–17 isotopologues while the legend cites two-tailed $t$-tests.
   - **Suggested Improvement:** Audit the underlying source tables for Fig. 4d,e and Extended Data Fig. 4a–c to correct the x-axis $m/z$ labels and align the tracked isotopologues between pulse and chase panels; explicitly define the `Abundance` metric on the y-axes; quantify fractional turnover within each genotype relative to its end-of-pulse baseline and total glycan pool size; and harmonize the statistical test descriptions (clarifying whether the identical per-bar $P$ values reflect false discovery rate-adjusted $q$-values or step-up multiple-testing procedures and whether the single bracketed $P$ values in Fig. 4g represent the omnibus genotype main effect from a two-way ANOVA).

3. **Lack of Biological Replication and Protein-Level Normalization in Pooled Membrane Glycoproteomics (Page 5):**
   - **Potential Concern:** In the "Glycoproteomic analysis" section (page 5), tissues from $n = 4$ brains per group were pooled into a single composite sample per condition prior to high-salt/carbonate membrane fractionation and nano-LC–MS/MS (*"we pooled n = 4 samples per group to ensure sufficient input material"*), and the text concludes that *"Relative quantification revealed a significant increase in glycopeptide abundance in AD samples compared with their respective WT controls (Supplementary Figs. 7b and 8b)."* Pooling all four biological specimens into a single composite sample per group reduces the effective biological sample size to $N = 1$ pool per condition, precluding any estimation of inter-individual biological variance. Consequently, statistical comparisons across the 137 mouse or 77 human glycopeptides identified within a single pooled run measure variance across different peptide species within the pool rather than biological reproducibility across independent animals or patients. Furthermore, because the workflow relied on bulk membrane ultracentrifugation without glycopeptide-specific affinity or hydrophilic interaction enrichment (such as HILIC or lectin chromatography) and without normalizing glycopeptide intensities to unmodified peptides from the corresponding parent proteins, increased glycopeptide signal in the AD pool cannot fully decouple increased site-specific glycosylation stoichiometry from elevated expression or membrane retention of the parent glycoproteins.
   - **Suggested Improvement:** Explicitly clarify in the text on page 5 (and when contextualizing glycoproteomic coverage in the Discussion on page 1420) that the glycoproteomics analysis represents a comparison across detected glycopeptide species within single pooled samples ($n = 1$ pool of 4 brains per group) rather than across independent biological replicates, and discuss the absence of parent-protein normalization and glycopeptide-specific enrichment (relative to deep intact glycoproteomics studies that resolve thousands of site-specific glycoforms) as methodological limitations.

4. **Reconciling Steady-State Hexosamine Precursor Depletion with Claims of Elevated UDP-GlcNAc Availability and Channelling in Competing Pathways (Pages 2, 3, 5, 8, 11, and 21–22; Fig. 1d, Fig. 4g, Extended Data Fig. 5):**
   - **Potential Concern:** In human AD frontal cortex grey matter, steady-state spatial metabolomics demonstrates a statistically significant *reduction* in the upstream hexosamine biosynthetic pathway intermediates $N$-acetylglucosamine and glucosamine-6-phosphate (page 2; Fig. 1d on page 3). Conversely, on page 5 (and in the Discussion on page 1420), the text refers to *"the observed increase in UDP-GlcNAc"*, states that reduced hyaluronan and O-GlcNAcylation in 5xFAD and PS19 brains occur *"despite elevated UDP-GlcNAc availability"*, and interprets this as *"selective substrate channelling"* and *"compartment-specific redirection of hexosamine flux"* from a shared UDP-GlcNAc pool. However, Fig. 4g (page 8) measures $^{13}\text{C}$-labelled `UDP-HexNAc` ion abundance (`M + 1` to `M + 17`) following a 24-h $^{13}\text{C}_6$-glucose pulse in 5xFAD mice—which reflects de novo isotopic flux into the combined isobaric UDP-GlcNAc/UDP-GalNAc pool rather than total steady-state (unlabelled $\text{M}_0$ plus labelled) UDP-GlcNAc pool size, and does not include PS19 or human AD steady-state measurements. Moreover, while $^{13}\text{C}_6$-glucose pulse-chase tracing was performed for N-linked glycans and total UDP-HexNAc, isotopic incorporation into O-GlcNAc or hyaluronan was not measured, nor were the expression or enzymatic activities of O-GlcNAc transferase (OGT), O-GlcNAcase (OGA), hyaluronan synthases (HAS1–3), or hyaluronidases evaluated; thus, reduced steady-state O-GlcNAc and hyaluronan could alternatively arise from increased cleavage/degradation or reduced synthase activity rather than direct metabolic competition for cytosolic UDP-GlcNAc. Additionally, the caption of Extended Data Fig. 5c (page 22) notes that the hyaluronan MALDI–MSI violin plots display *"absolute intensities"* (though the top y-axis is labelled `Relative abundance`), whereas the rest of the MALDI–MSI datasets are normalized to total ion current (TIC).
   - **Suggested Improvement:** Clarify in the Results and Discussion whether total steady-state UDP-HexNAc abundance (including unlabelled $\text{M}_0$) is elevated in human AD, 5xFAD, and PS19 brains or whether the depletion of upstream GlcN-6-P and GlcNAc in human AD (Fig. 1d) reflects accelerated downstream consumption into UDP-HexNAc and N-glycans; note that flux into O-GlcNAc and hyaluronan was inferred from steady-state abundance rather than directly traced; and confirm that the hyaluronan reductions in Extended Data Fig. 5c hold after standard per-pixel TIC normalization.

---

#### Minor Corrections and Typos

1. **Page 4 (Main text and Fig. 2d,e):** The high-mannose N-glycan species ($\text{Man}_6\text{GlcNAc}_2$, depicted by the 6-circle/2-square cartoon in Fig. 2e) is labelled `1,409 m/z` in the text (*"Violin plots of glycan structures at 1,688 m/z and 1,409 m/z"*) and in Fig. 2d,e, whereas the exact same sodiated high-mannose glycan ($\text{Hex}_6\text{HexNAc}_2\text{Na}^+$) is reported at `1,419 m/z` throughout the rest of the manuscript (page 2, Fig. 1c on page 3, and Extended Data Fig. 2c on page 18). Additionally, in Fig. 2e, the `1,409 m/z` (or `1,419 m/z`) label is missing next to the lower row of spatial glycomics brain images.
2. **Pages 4 and 17 (Text vs. Extended Data Fig. 1c):** On page 4, the text states that violin plots of glycans at `1,079 m/z` and `1,905 m/z` indicate *"an early stage rise in white matter glycosylation without sustained accumulation at later stages (Extended Data Fig. 1b,c)."* However, in Extended Data Fig. 1c (page 17), the two columns are labelled `Grey Matter` (left) and `White Matter` (right), and for `1,905 m/z`, the transient rise at Braak 1–2 appears in the `Grey Matter` plot while the `White Matter` plot shows a progressive decline from Braak 0 to Braak 5–6. Please verify whether the column headers or plotted datasets in Extended Data Fig. 1c were inadvertently swapped.
3. **Page 5 (Article Page 1414, "Glycoproteomic analysis", paragraph 2):** The text states, *"Notably, 55% and 75% of glycopeptides were shared between human and mouse samples (Supplementary Figs. 7c and 8c), respectively."* Based on the subsequent sentence (*"between normal and AD samples"*) and the separate species-specific workflows in Supplementary Figs. 7 and 8, this should state that 55% and 75% of glycopeptides were shared between control (WT/normal) and AD samples within the mouse and human cohorts, respectively.
4. **Pages 6, 7, 8, 9, and 21–22 (Hexosamine Pathway Description and Schematics in Fig. 5a and Extended Data Fig. 5a):** On page 6/7 (article page 1416) and in Fig. 5a (page 9) and Extended Data Fig. 5a (pages 21–22), phosphoglucomutase 3 (PGM3) is described as converting *"glucosamine-6-phosphate to N-acetylglucosamine-6-phosphate,"* and the diagrams depict `Glucosamine-6-phosphate -> N-acetyl-glucosamine -> UDP-N-acetyl-glucosamine` (or `N-Acetyl-Glucosamine` branching directly into `N-Glycans`, `Hyaluronan acid`, and `Protein O-GlcNacylation` in Extended Data Fig. 5a, where `Glucose-6-phosphate` is also misspelled as `Glucos-6-phosphate`). Canonically, glucosamine-6-phosphate *N*-acetyltransferase 1 (GNPNAT1) converts GlcN-6-P to GlcNAc-6-P, PGM3 catalyses the reversible isomerization of GlcNAc-6-P to GlcNAc-1-P, and UAP1 generates UDP-GlcNAc, which serves as the direct branching nucleotide-sugar donor substrate.
5. **Page 8 (Article Page 1417, Fig. 4c,f,g):**
   - In **Fig. 4c**, insert a space after the colon in `1,266:M + 9` (`1,266: M + 9`), and note that **Fig. 4f** uses lowercase `m2`, `m5`, `m6` whereas **Fig. 4c,g** use uppercase `M + n`.
   - In **Fig. 4g**, on the y-axes of both the `UDP-HexNAc` and `UDP-Hexose` graphs, `"Ion abundance"` is misspelled as `"lon abundance"` (with a lowercase `"l"`).
6. **Page 18 (Extended Data Fig. 2c):** The third glycan is labelled `2253 m/z` above the brain images in the figure panel, but is written as `2523 m/z` in the figure legend; likewise, the fourth glycan ($\text{Man}_5\text{GlcNAc}_2$) is labelled `1237 m/z` in the figure panel and legend, which appears to be a typographical error for `1,257 m/z` (as shown in Fig. 3c and Fig. 4b–e).
7. **Page 19 (Extended Data Fig. 3a):** The x-axis labels on all four violin plots read `WT` and `5xFAD`, whereas the MALDI images, figure title, and legend describe comparisons between `WT` and `PS19` mice.
8. **Page 20 (Extended Data Fig. 4a,c):** In **Extended Data Fig. 4a**, the top-right washout bar graph is titled `Enrichment of 1688` instead of `24hr washout of 1688`. In **Extended Data Fig. 4c**, the unlabelled spatial map is labelled `1442: Unlabeled` and the washout bar graph is titled `24hr washout of 1442` instead of `1444` (both the pulse graph and legend refer to `1444 m/z`).
9. **Pages 21–22 (Extended Data Fig. 5c,d):** In **Extended Data Fig. 5c**, the top violin plot (`WT vs 5xFAD`) is missing its x-axis group labels (`WT` and `5xFAD`), and the bottom violin plot (`WT vs PS19`) has a truncated y-axis label. In **Extended Data Fig. 5d**, the y-axis label for the bottom bar chart reads `Relatvie Intensity` instead of `Relative Intensity`, and the legend states `(n = 9 animals per group)` whereas the plotted dot/bar graphs display 10 individual points in the top `WT` group (vs. 9 in `5xFAD`) and 8 individual points in the bottom `PS19` group (vs. 9 in `WT`).

### [Causal Genetic, Pharmacological & Dietary Interventions and Real-World EHR Analysis] (Pages: 5-11, 23-27)
#### Potential Mistakes and Improvements

1. **Potential Immortal Time Bias and Exposure Ascertainment Discrepancies in the Retrospective EHR Survival and Transition Analyses (Pages 7–10, 14; Fig. 6f–j):**
   - **Observation:** On page 7 (article page 1416), the main text states: *"Using physician notes and prescription records, we identified patients who had documented glucosamine usage for at least 1 year following a dementia diagnosis,"* whereas the Methods section (page 1423) states only that clinical notes were searched via a regex-based keyword matching pipeline (*"Notes containing relevant glucosamine terms were flagged as exposures"*). Furthermore, in the survival and cumulative transition analyses (Fig. 6h–j; contextualized by the Methods on article page 1423, which define survival time for all patients as *"the interval from first diagnosis to death"*), follow-up time starts at initial diagnosis (`Time (days) = 0`).
   - **Potential Concern:** Requiring patients in the glucosamine exposure group to accrue at least 1 year of documented usage *after* diagnosis while tracking survival and transition from initial diagnosis ($t = 0$) without an equivalent 1-year survival requirement for unexposed controls introduces a critical potential concern regarding **immortal time bias** (or time-window conditioning). Under a 1-year post-diagnosis documentation requirement, patients classified as glucosamine users must survive and remain under clinical observation long enough after diagnosis to meet that criterion, whereas patients who die or are censored early after diagnosis are systematically classified as non-users. Similarly, if regex keyword extraction from clinical notes flags any note mentioning glucosamine across the multi-year follow-up without time-varying exposure modelling, patients with longer survival or follow-up have more opportunities to have an over-the-counter supplement documented. Conversely, in Fig. 6h, the survival curve for the glucosamine group drops below the control group even within the first 365 days, indicating a discrepancy between the 1-year post-diagnosis exposure criterion stated on page 7 and how the index date ($t = 0$) was defined in the model. In addition, the Methods state that *"Kaplan–Meier analyses estimated survival probabilities over a 15-year window"* using EHR data extracted *"from 2012 to 2024"* (a maximum span of 12 years), whereas the Results (page 1418) and Fig. 6h–j captions describe a *"10-year survival analysis"* (plotted up to ~3,000 days). Clarifying the exact index date relative to glucosamine initiation, reconciling the follow-up window (10-year vs. 12-year vs. 15-year) and how structured prescription records were integrated, and re-evaluating exposure using a time-dependent Cox proportional hazards model or a 1-year post-diagnosis landmark analysis are essential to ensure the validity of the mortality and MCI-to-ADRD progression estimates.

2. **Potential Confounding by Indication, Surveillance Bias, and NLP Exposure Misclassification in EHR Propensity Score Matching (Pages 7–10, 14; Fig. 6f–j):**
   - **Observation:** On page 9 (article page 1418) and in the Methods (page 1423), the 10-year survival and MCI-to-ADRD transition analyses are described as *"adjusting for age, sex and demographics"* via 1:1 nearest-neighbour propensity score matching balancing covariates across *"two sets: (1) age alone and (2) demographic variables"* (though the specific demographic variables in Set 2 are not listed, nor is it stated which matched set produced the primary curves in Fig. 6h–j versus Supplementary Fig. 17). Furthermore, glucosamine exposure was classified by extracting glucosamine-related keywords from unstructured physician notes (`Doctor's note`, Fig. 6f) via regular-expression matching.
   - **Potential Concern:**
     - *Confounding by indication and healthcare utilization:* Because glucosamine is an over-the-counter supplement primarily taken for osteoarthritis and chronic joint pain, patients with documented usage may systematically differ from non-users in physical mobility, chronic pain, body mass index, concomitant NSAID/analgesic use, and baseline cardiometabolic comorbidities—all independent risk factors for accelerated cognitive decline and all-cause mortality. Moreover, patients with greater comorbidity burden or faster clinical progression generally have higher healthcare utilization (more frequent clinic visits and physician notes), which simultaneously increases the probability of having an over-the-counter supplement recorded in unstructured notes and of having a diagnostic conversion from MCI to ADRD detected (`surveillance/ascertainment bias`).
     - *Regex keyword matching and reproducibility:* Simple regex keyword matching without negation or context detection can misclassify clinical notes stating that a patient *stopped*, *denied*, or was *advised against* taking glucosamine as positive exposures, and neither the specific regular-expression patterns/inclusion-exclusion rules, the ICD-9-CM and ICD-10-CM code lists used to define the five cohorts, nor a Code Availability statement is provided. Controlling for baseline comorbidity burden (e.g., Charlson or Elixhauser Comorbidity Index, osteoarthritis diagnosis), baseline visit/note frequency, and negation contexts in the NLP pipeline—along with reporting validation metrics such as positive predictive value from manual chart review and specifying the exact regex terms, ICD code lists, and demographic matching variables—would substantially strengthen the epidemiological conclusions and reproducibility.

3. **Inconsistencies in EHR Statistical Reporting, Cohort Sizes, and Y-Axis Scaling (Pages 7–10; Fig. 6h–j):**
   - **Observation:**
     1. *P-value and hazard ratio discrepancy:* On page 9 (article page 1418), the main text reports that *"glucosamine usage was associated with a 25% increase in mortality risk among ADRD patients (P = 0.0023; Fig. 6h)"*, whereas the plot inside **Fig. 6h** (page 10) directly displays `P < 0.001`. Furthermore, while a relative risk metric (*"25% increase in mortality risk"*) is quoted in the text, no hazard ratio (HR) or 95% confidence interval is reported alongside the Kaplan–Meier curves, nor is it specified whether Fig. 6h–j plots the full unmatched cohort (`24,481` ADRD and `41,884` MCI patients; pages 7–8) or the 1:1 propensity-score-matched subset (`1,896` ADRD and `2,750` MCI glucosamine users plus matched controls).
     2. *Y-axis scaling and endpoint terminology in Fig. 6j:* On page 9, the text states that *"while most individuals with MCI remain stable or even recover, approximately 5% progress to ADRD"* and reports *"a 25% increase in the proportion of patients with MCI transitioning to ADRD in the glucosamine user group (Fig. 6j)."* However, in **Fig. 6j**, the y-axis is labelled `Transition from MCI to AD (%)` with two sets of y-axis tick labels (`0–100` on the far left and `0–0.6` on the plot axis). It is unclear whether the inner y-axis values (`0` to `0.6`) represent proportions (`0%–60%`) or percentages (`0.0%–0.6%`), why the cumulative incidence curves appear compressed or jump vertically near `Time = 0`, and whether the plotted endpoint is specifically `AD` (as written on the y-axis and caption) or all-cause `ADRD` (as described in the main text).
   - **Potential Concern:** Reconciling the $P = 0.0023$ vs. $P < 0.001$ discrepancy in Fig. 6h, reporting the exact matched sample sizes (with numbers at risk) and hazard ratios (with 95% CIs), clarifying whether Fig. 6j tracks conversion to AD or all-cause ADRD, and properly scaling the single y-axis of Fig. 6j are necessary for transparent interpretation.

4. **Anatomical Delivery Site, Pharmacological Dosing, and Mouse Strain Discrepancies in Preclinical Interventions (Pages 7, 9–10, 12, 23–25, 27; Figs. 5–6, Extended Data Figs. 6a, 7a, 9e):**
   - **Intracerebroventricular vs. Hippocampal injection site:** On page 7 (article page 1416), the main text states that 8-month-old female 5xFAD and PS19 mice received `shScrambled` or `shPGM3` lentivirus *"delivered via stereotaxic injection into the ventricles for whole-brain distribution"*, and states that age-matched WT mice received *"the same intracerebroventricular lentiviral shPGM3 intervention in age-matched WT mice (Extended Data Fig. 7a)"* (matching the bilateral lateral ventricle coordinates in Methods on article page 1421: AP, −0.2 mm; ML, ±1.0 mm; DV, −2.2 mm). However, the schematics and captions for **Extended Data Fig. 6a** (page 23) and **Extended Data Fig. 7a** (pages 24–25) explicitly state that PS19 and WT mice received **hippocampal** injections (*"8-month-old PS19 mice receive hippocampal injections of shScr or shPGM3"* and *"8-month-old WT mice injected with hippocampal shScr or shPGM3 lentivirus"*). Clarifying whether PS19 and WT mice underwent intracerebroventricular or intrahippocampal parenchymal delivery is essential for interpreting the whole-brain sagittal MALDI–MSI glycan maps in Extended Data Figs. 6b–c and 7b.
   - **NGI-1 dose, laterality, infusion parameters, and vehicle formulation:** On page 7 (article page 1416), the main text states: *"A single intracerebroventricular administration of NGI-1 (2 nmol in 2 µl) was delivered into the lateral ventricle of 8-month-old 5xFAD mice."* This conflicts with the Methods (article page 1421), which state that *"2 µl of the small-molecule inhibitor NGI-1 (HY-117383, MedChemExpress; 5 nmol dissolved into PBS) was injected bilaterally into the lateral ventricles of 5xFAD mice at a speed of 50 nl min−1."* Beyond reconciling the `2 nmol` (unilateral) vs. `5 nmol` (bilateral) dose and the `50 nl min−1` infusion rate (which requires 40 min per 2 µl injection, or 80 min total for bilateral delivery—10-fold slower than the `0.5 µl min−1` rate used for lentiviral injections in the preceding sentence), NGI-1 is a hydrophobic small molecule with poor aqueous solubility in neat PBS at 1–2.5 mM without an organic co-solvent (such as DMSO or PEG/Tween), and the vehicle control is inconsistently referred to as `"Veh"` (Fig. 5f–g), `"Con"` (Fig. 5j), `"Saline"` (Extended Data Fig. 9e–f), and `"PBS"` (Methods).
   - **Glucosamine control vehicle and missing PS19 mouse strain details:** In **Fig. 6b** (page 10), the experimental schematic labels the oral gavage control as `Glucosamine/PBS`, whereas **Fig. 6c–e**, **Extended Data Fig. 7e–g**, and **Extended Data Fig. 9c–d** label the control group as water (`H2O`), and the gavage vehicle and administration volume are omitted in the Methods (page 1421). Furthermore, under `Human and mouse tissue sources and preparation` (page 1421), only 5xFAD transgenic mice and WT littermate controls (JAX stock #008730) are described; although PS19 mice are used extensively across the study, their formal strain nomenclature, genetic background, supplier/stock number (e.g., JAX stock #008169), and sex are omitted.

5. **Social Recognition Memory Normalization, Trial Count Discrepancies, and Repeated-Measures ANOVA Reporting (Pages 7, 9–10, 13, 23–25; Figs. 5h–j, 6e, Extended Data Figs. 6d–e, 7c–d,g):**
   - **Normalization to Trial 1 and ANOVA assumptions:** In all social memory graphs (Figs. 5i–j, 6e; Extended Data Figs. 6e, 7d,g), interaction time is plotted as a ratio relative to Trial 1 (`Time interacting relative to Trial 1`), fixing every animal's value at Trial 1 to `1.0` with zero variance. Including a zero-variance baseline trial in a two-way repeated-measures ANOVA violates homoscedasticity and sphericity assumptions, and ratio normalization can obscure whether an intervention altered baseline social investigation or locomotor activity in Trial 1 (for instance, in Fig. 5i, `5xFAD-shScr` rises to ~`1.35–1.45` relative to Trial 1 across Trials 2–4, which can occur if baseline Trial 1 exploration was low). In addition, the figure legends do not clarify whether the *P*-values printed above individual trials (`0.0455` and `0.0265` in Fig. 5i; `P = 0.034` and `P = 0.029` in Fig. 5j; Trials 2–4 in Extended Data Fig. 6e) versus the single curve *P*-value (`P = 0.029` in Fig. 6e) represent post-hoc pairwise comparisons (and which multiple-testing correction was applied) or omnibus group/interaction effects.
   - **Protocol discrepancy (4 trials vs. 5 days), session duration, and cohort sizes:** The main text and all behavioural figures depict a **4-trial** social recognition memory paradigm (`Trial 1` to `Trial 4`), whereas the Methods section (article page 1422) states that mice were *"introduced to the same novel conspecific once daily over five consecutive days, with trials separated by 24-h intervals,"* and does not specify the duration of each daily social interaction session (e.g., 5 min or 10 min). Furthermore, in **Extended Data Fig. 7**, the sample sizes change between the MALDI glycomics panels and the behavioural panels (`WT-shScr` / `WT-shPGM3` changes from $n = 3$ per group in panel b to $n = 6$ and $n = 5$ in panel d; `WT-GlcN` / `WT-water` changes from $n = 4$ and $n = 3$ in panel f to $n = 3$ and $n = 5$ in panel g), which should be explained given the statement in the Methods that no animals or data points were excluded (and noting that $n = 3$ in panel g is very small for a behavioural assay).

6. **Validation of PGM3 Knockdown Efficiency and Mechanistic Context for Rapid (2-Week) Behavioural Rescue Dissociated from Neuropathology (Pages 7–11, 23–27; Figs. 5–6, Extended Data Figs. 6–9):**
   - **Knockdown validation:** While MALDI–MSI shows reduced N-glycan abundance in 5xFAD and PS19 brains 2 weeks after `shPGM3` lentiviral delivery (Fig. 5d–e, Extended Data Fig. 6b–c), no direct RT–qPCR, western blot, or immunofluorescence data are provided to quantify the efficiency and regional distribution of *Pgm3* mRNA or PGM3 protein knockdown across the brain parenchyma. Direct target engagement validation is particularly important because WT mice injected with the same `shPGM3` lentivirus showed no statistically significant reduction in individual N-glycans (Extended Data Fig. 7b).
   - **Dissociation from neuropathology and glycan turnover:** A 2-week intervention with `shPGM3`, `NGI-1`, or oral `glucosamine` in 8-month-old 5xFAD and PS19 mice produced bidirectional changes in social recognition memory without any detectable change in $\text{A}\beta$ plaque burden (6E10), phospho-tau (Thr231; Supplementary Fig. 14), or $\text{GFAP}^+$ reactive astrocytosis (Extended Data Figs. 8–9). Moreover, in **Extended Data Fig. 8**, no quantitative summary graphs or biological replicate numbers ($n$ animals) are reported, and in **Extended Data Fig. 9b,d,f**, the y-axes for reactive astrocytes and $\beta$-amyloid plaques are plotted only as `"Relative level (%)"` with control groups normalized to `100%` (and near-zero variance in several control bars) rather than reporting physical units (e.g., plaque count per $\text{mm}^2$ or percentage of area occupied). Providing unnormalized histological metrics, discussing how a 2-week window achieves large bulk N-glycan reductions in Fig. 5d–g (given that the 24-h chase experiment in Fig. 4e showed minimal N-glycan turnover over 24 hours) to rapidly modulate synaptic function independently of established dense-core plaques and gliosis, and acknowledging previously characterized preclinical small-molecule PGM3 inhibitors (such as FR054, utilized in vivo in ref. 47) when stating on page 11 that *"no blood–brain barrier-permeable inhibitors of PGM3 currently exist"* would strengthen the mechanistic and translational interpretation.

---

#### Minor Corrections and Typos

1. **Page 9 (Article Page 1418), Fig. 5b,d,e,i,j:**
   - In **Fig. 5b**, `"Lentivirus"` is misspelled as `"Lentie"`.
   - In **Fig. 5d,e**, the scrambled control is mislabelled as `"5xFAD-shSr"` and `"shSr"` instead of `"5xFAD-shScr"` and `"shScr"`, and in **Fig. 5e**, the violin plot x-axis label for the knockdown group reads `"sgPGM3"` (CRISPR single-guide RNA notation) instead of `"shPGM3"`.
   - In **Fig. 5d–g**, the y-axes of the violin plots lack axis titles (e.g., `"Relative abundance"`).
   - In **Fig. 5i,j**, genotype capitalization in the legends is inconsistent (`"5xFAD-shScr"` vs. `"5XFAD-shPGM3"`, `"5XFAD-Con"`, and `"5XFAD-NGI1"`).
2. **Page 9 (Article Page 1418), right column, line 6 vs. Page 14 (Article Page 1423):** `"Supplementary Fig. 17"` is cited on page 1418 for the 10-year EHR survival analysis adjusting for age, sex, and demographics, but `"Supplementary Fig. 17"` is also cited in the Methods (article page 1423) for the STT3A and ALG13 antibody pre-adsorption assay. Please verify and update the supplementary figure numbering.
3. **Page 10 (Article Page 1419), Fig. 6g,h,i:**
   - In **Fig. 6g**, the main text on page 7 states *"Across all three groups (AD, ADRD and MCI), approximately 8% of patients were identified as glucosamine users ... (Fig. 6g)"*, whereas Fig. 6g only plots two bars (`ADRD` and `MCI`) and lacks a `%` label on the broken x-axis.
   - In **Fig. 6h**, the y-axis has duplicate `0.6` tick labels (`0.5, 0.6, 0.6, 0.7, 0.8, 0.9, 1.0`).
   - In **Fig. 6i**, the *P*-value annotation uses a colon (`"P : 0.252"`) instead of an equals sign (`"P = 0.252"`).
4. **Pages 23–25, Extended Data Figs. 6 and 7:**
   - In **Extended Data Fig. 6c**, the violin plot x-axis label reads `"PS19-shSr"` instead of `"PS19-shScr"`.
   - In **Extended Data Fig. 7a**, `"Lentivirus"` is misspelled as `"Lentie"`.
   - In **Extended Data Fig. 7b**, the MALDI image label reads `"WT-shSr"` instead of `"WT-shScr"`, and the right violin plot x-axis labels are mislabelled as `"PS19-shSr"` and `"PS19-shPGM3"` instead of `"WT-shScr"` and `"WT-shPGM3"`.
   - In **Extended Data Fig. 7f**, both violin plot y-axes are labelled `"Relative abundance"`, whereas the caption on page 25 describes them as `"Relative and absolute intensities"`.
   - In **Extended Data Fig. 7d,g**, the legend states *"Data are mean ± s.e.m.; P values are indicated (two-tailed t-test); ns, not significant"*, but no *P*-values or `ns` annotations appear on the plots, and the 4-trial longitudinal social memory curves in panels **d** and **g** should be analysed via two-way repeated-measures ANOVA rather than a two-tailed *t*-test.

### [Experimental Methods, Mass Spectrometry Pipelines & Statistical Framework] (Pages: 12-14)
#### Potential Mistakes and Improvements

1. **Ambiguities in the MALDI–MSI Delipidation Workflow, Sequential Sectioning, and Signal Normalization (`MALDI imaging for spatial metabolomics, lipidomics, glycomics and hyaluronan` & `Stable isotope tracing and ion mobility mass spectrometry`, pages 1421–1422)**
   - **Xylene delipidation on fresh-frozen vs. FFPE human tissue:** On page 1421, the Methods state: *"Following metabolite and lipid imaging, slides were washed with ice-cold 100% methanol for 5 min to remove NEDC and fixed overnight in 10% neutral-buffered formalin. For human brain FFPE sections, lipid removal was performed using three sequential 1-h xylene washes before enzymatic treatment."* However, in the Results (page 1411) and Fig. 1a, the three-step 1-h xylene wash protocol was developed specifically to remove residual interfering lipids from **fresh-frozen** human brain sections during the sequential metabolomics/lipidomics/glycomics workflow after formalin fixation, whereas clinical FFPE sections (Fig. 2a) were used solely for single-omic spatial glycomics. Furthermore, sectioning thickness and standard deparaffinization/rehydration steps prior to antigen retrieval for archival FFPE human blocks are not distinguished from the post-methanol/formalin delipidation of fresh-frozen sections.
   - **Sequential processing vs. distinct cryosectioning temperatures and unreferenced glycogen imaging:** The Methods state that *"The sectioning temperature was −20 °C for metabolomics, lipidomics and glycomics runs, and −15 °C for hyaluronan runs"* and immediately thereafter state that *"Each slide was sequentially processed for spatial metabolomics, lipidomics, glycomics, glycogen and hyaluronan imaging."* If hyaluronan runs required sections cut at −15 °C (or adjacent sections, as also depicted in Extended Data Fig. 2b for CHCA/PNGase F), hyaluronan imaging could not be performed sequentially on the same −20 °C tissue section after PNGase F digestion and CHCA matrix application. Moreover, *"glycogen"* imaging is listed in this workflow, but no glycogen enzymatic digestion (e.g., isoamylase) or imaging data appear anywhere in the manuscript.
   - **Missing sprayer, TIMS, liquid diet, and normalization parameters:** For PNGase F and hyaluronidase application, the HTX sprayer temperature (`45 °C`) and velocity (`900 mm min−1`) are provided, but the flow rate, number of passes, track spacing, and gas pressure are omitted; for CHCA matrix application, no sprayer parameters are provided. Similarly, under `Stable isotope tracing and ion mobility mass spectrometry` (page 1422), the composition and preparation of the $^{13}\text{C}_6$-glucose liquid diet (and unlabelled acclimation/chase liquid diet) are omitted, and no trapped ion mobility spectrometry (TIMS) parameters ($1/K_0$ range, ramp time, or accumulation time) are provided for the Bruker timsTOF fleX. Finally, per-pixel Total Ion Current (TIC) normalization is only mentioned for the Waters Synapt G2-Si HDI workflow (page 1422), leaving it unclear how ion intensities were normalized in SCiLS Lab for the Bruker timsTOF fleX runs (note that Extended Data Fig. 3a states *"normalized to TIC"*, whereas Extended Data Fig. 5c reports *"absolute intensities"* and Extended Data Fig. 7f reports *"Relative and absolute intensities"*).

2. **Uncorrected Multiple Comparisons and Randomization Contradictions (`Quantification and statistical analysis`, page 1423)**
   - **Multiple-testing corrections:** Across the parallel multi-isotopologue comparisons (Fig. 4d,e; Extended Data Fig. 4a–c), multi-trial behavioural repeated-measures ANOVAs (Figs. 5i,j and 6e), and 96-well glycosylation RT–PCR arrays, the Methods do not state whether post-hoc multiple-comparison adjustments (such as Šídák, Tukey, or Benjamini–Hochberg false discovery rate corrections) were applied.
   - **Internal contradictions regarding randomization:** The statements *"Animals were assigned to experimental groups based on genotype and age. No further randomization was applied to group assignment"* and *"The order of sample analysis was not randomized"* conflict with the interventional experiments—in which animals of identical genotype and age were divided into active treatment versus vehicle/control cohorts (shPGM3 vs. shScr, NGI-1 vs. vehicle, glucosamine vs. water)—as well as with earlier statements on page 1422 (`LC–MS analysis of UDP-sugar precursors`: *"Samples were randomized and analysed in a blinded manner"*; `Behavioural testing by the social memory assay`: *"Experimental mice were randomly assigned to testing order"*).

3. **Glycoproteomics Alkylation Chemistry, Search Modification Mismatch, and Missing Quantification Details (`Glycoproteomics analysis`, page 1422)**
   - **Alkylation reagent quantity and modification mass:** The protocol states that membrane proteins dissolved in urea lysis buffer containing `100 mM DTT` were *"alkylated with IAA (55 mg) in the dark for 45 min at room temperature"*, without stating the reaction volume or molar concentration (likely intended as `55 mM` iodoacetamide, although 55 mM iodoacetamide is stoichiometrically insufficient to quench 100 mM DTT unless the sample was diluted beforehand). Furthermore, iodoacetamide adds a **carbamidomethyl** group (+57.021 Da) to cysteine residues, whereas the Byonic search parameters specify *"variable modifications for ... carboxymethylation (C)"* (+58.005 Da, the adduct of iodoacetic acid). Please clarify whether iodoacetamide (carbamidomethylation) or iodoacetic acid (carboxymethylation) was used and specify the exact molar concentration.
   - **Missing database and quantification criteria:** The reference proteome databases (e.g., UniProt species/release), protein- and peptide-level false discovery rate (FDR) filtering thresholds, and the MS1 precursor-intensity or spectral-counting method used to calculate relative glycopeptide abundances (Supplementary Figs. 7b and 8b) are omitted.

4. **Missing Digital PCR Protocol and Incomplete RT–PCR Details (`RT–PCR and digital PCR for glycosylation enzymes`, page 1422)**
   - Although both the subsection heading (`RT–PCR and digital PCR for glycosylation enzymes`) and the Results text (page 1414) state that digital PCR was used alongside RT–PCR to quantify glycosylation enzyme mRNA levels, the Methods text provides no instrumentation, reagents, partitioning workflow, or primer/probe details for digital PCR.
   - For the RT–PCR workflow, the text states that after isopropanol RNA precipitation at −20 °C, *"DNA was denatured using phenol"*, which is chemically non-standard for removing genomic DNA from precipitated RNA. In addition, the text states that cDNA was synthesized using the *"iScript One-Step RT–PCR kit with reverse transcriptase (Bio-Rad)"* with incubations at 25 °C for 5 min, 46 °C for 20 min, and 95 °C for 1 min; these are the thermal cycling conditions for the two-step **iScript cDNA Synthesis Kit**, and the subsequent qPCR master mix, amplification cycling protocol, and reference housekeeping genes used to normalize the 96-well Bio-Rad glycosylation arrays are not specified.

5. **Orphaned Primary Neuron Culture Protocol (`Primary neuron culture`, page 1423)**
   - Page 1423 includes a standalone subsection detailing the isolation and culture of E17–E18 C57BL/6J embryonic mouse primary neurons in eight-well chamber slides. However, no experiments or figures in the main manuscript or Extended Data report data from cultured primary neurons. If primary neurons were used exclusively for a supplementary experiment (such as the antibody pre-adsorption validation in Supplementary Fig. 17), this connection should be explicitly stated; otherwise, this subsection appears to be an orphaned protocol and should be removed.

---

#### Minor Corrections and Typos

* **Page 1421 (`Human and mouse tissue sources and preparation`):** In *"with a 14-h light/10-h dark with temperature (18–23 °C)"*, the word *"cycle"* is missing after *"10-h dark"*.

* **Page 1421 (`MALDI imaging for spatial metabolomics, lipidomics, glycomics and hyaluronan`):** In *"Hyaluronidase (Sigma-Aldrich H1136, 60 U per slide in 0.2 mM Tris–HCl pH 7.5 with 0.5 mM NaCl)"*, please verify whether *"0.2 mM"* is a typographical error for the spray volume (**0.2 ml**, matching *"PNGase F (20 µg per slide in 0.2 ml)"* in the preceding sentence) or a higher buffer molarity (such as 20 mM Tris–HCl), as 0.2 mM Tris has negligible buffering capacity.

* **Page 1422 (`Glycoproteomics analysis`):** Ultracentrifugation speeds are reported only in revolutions per minute (*"45,000 rpm for 15 min"* and *"45,000 rpm for 45 min"*) without specifying the rotor model or relative centrifugal force ($\times g$).

* **Page 1422 (`Immunofluorescence and immunohistochemistry analysis`) vs. Page 1423 (`Antibody pre-adsorption assay`):** The anti-STT3A antibody (catalog number `12034-1-AP`) is listed as `Thermo Scientific` on page 1422 (`STT3A (1:100,12034-1-AP; Thermo Scientific)`, also missing a space after the comma) but as `Proteintech` on page 1423 (`anti-STT3A (1:100, 12034-1-AP; Proteintech)`). Catalog number 12034-1-AP is a Proteintech product; please harmonize the supplier name and ensure consistent capitalization of `ALG13` vs. `Alg13` across both subsections.

### [Discussion, Broader Clinical Implications & Study Declarations] (Pages: 11-12, 14-16)
#### Potential Mistakes and Improvements

1. **Internal Contradiction Regarding Clinical Outcomes in the MCI Cohort (Discussion, Page 1420):**
   In the Discussion (page 1420, right column), the authors state: *"Our data suggest that glucosamine does not clinically worsen outcomes in individuals with MCI, but is associated with worse survival in patients with AD/ADRD."* This statement is a potential concern because it directly contradicts the Results section (page 1418 and Fig. 6j), where the authors report a statistically significant 25% increase ($P < 0.001$) in the rate of transition from MCI to AD/ADRD among glucosamine users and conclude that *"glucosamine supplementation is associated with accelerated disease progression or worsened disease phenotype."* While glucosamine was not associated with reduced 10-year all-cause survival in the MCI cohort (Fig. 6i, $P = 0.252$), accelerated conversion from MCI to clinical dementia represents a worsened clinical outcome. The text in the Discussion should be revised to accurately distinguish between all-cause mortality and MCI-to-AD/ADRD disease progression.

2. **Tempering Population-Level Extrapolations and Reconciling Prior Epidemiological Literature (Discussion, Pages 1420–1421):**
   In the Discussion (page 1421, left column), the authors extrapolate the ~8% glucosamine usage prevalence from the single-system UF Health EHR cohort to the entire US AD/ADRD population, concluding that *"potentially over 1 million patients may be unknowingly exacerbating their disease progression through glucosamine supplementation"* and that the data *"make a compelling argument that glucosamine is a negative driver of neurodegeneration."* While the convergence of preclinical and clinical observations is notable, a substantive concern is that the Discussion does not contextualize these extrapolations against prior epidemiological literature or preclinical scope boundaries:
   - **Reconciliation with Prior Epidemiological Literature:** Previous large-scale prospective cohort analyses (for example, in the UK Biobank) have reported neutral or protective associations between habitual glucosamine use and incident dementia in the general ageing population. Discussing how disease-stage-specific vulnerability (established MCI/ADRD versus healthy baseline cohorts) or methodological differences (such as unmeasured confounding by osteoarthritis indication or mobility in single-system EHR data) account for these divergent epidemiological findings would strengthen the interpretation.
   - **Preclinical Scope:** The mouse intervention studies were conducted over a relatively short 2-week window exclusively in 8- to 9-month-old female mice, and intracerebroventricular administration of shPGM3 or NGI-1—as well as oral glucosamine—altered social recognition memory without modifying amyloid-$\beta$ plaque burden, tauopathy, or GFAP astrocytosis. Explicitly discussing these observational and preclinical limitations is necessary to appropriately contextualize the public health claims.

3. **Inconsistent Author Initials and Omissions in Author Contributions (Author Contributions, Page 1425):**
   Comparing the masthead author list (page 1410) against the Author Contributions statement (page 1425) reveals several discrepancies that affect authorship attribution:
   - **Harrison A. Clarke (`H.A.C.`)** is listed as the sixth author on the masthead, but his initials do not appear anywhere in the Author Contributions section (`H.C.` is listed only once, which corresponds to co-author **Hongyu Chen**).
   - **Terrymar Medina (`T.M.`)** is referred to as `T.M.M.` under Investigation.
   - **Li Chen (`L.C.`)**: Under Investigation, both `C.L.` (which does not match any author on the masthead) and `L.C.` are listed.
   - **Craig W. Vander Kooi (`C.W.V.K.`)** is listed correctly as `C.W.V.K.` under Investigation and Writing—review and editing, but is truncated to `C.W.V.` under Supervision.

4. **Bibliographic Error in Reference 30 (References, Page 1424):**
   In Reference 30 (*"Harrison, A. C. et al. Spatial mapping of the brain metabolome lipidome and glycome. Nat. Commun. about:sanitized (2025)."*), the first author's first and middle names (**Harrison A. Clarke**) have been inverted and used as the surname (`Harrison, A. C.` instead of `Clarke, H. A.`). In addition, the cited DOI (`10.1101/2023.07.22.550155`) is the bioRxiv preprint identifier rather than the published *Nature Communications* article details (*Nat. Commun.* **16**, 4373 (2025); `https://doi.org/10.1038/s41467-025-59487-7`).

#### Minor Corrections and Typos

- **Page 1424 (Acknowledgements):** Missing period after `S.Q` in *"Z.L. and S.Q are supported by the MBI Gator NeuroScholar Program."*

- **Page 1425 (Peer review information):** Remove the extraneous article *"the"* before the reviewer's name in *"Nature Metabolism thanks the Yasuhiko Kizuka and the other, anonymous, reviewer(s)..."*