Study design & sample counts

Notes: Hi-C groups: adult WT (N=3), old WT (N=3), S6‑KO (N=3); CK‑p25 Hi‑C/RNA from published dataset subsets (Step1/Step2) integrated; shSIRT6 RNA-seq (in vitro) small-N. All extracted from the preprint text and methods Study design & sample details

Key reported quantitative outcomes

These counts come directly from the paper's Hi‑C annotation steps (insulation-based TAD calling and Chromosight loop calling). Differences: Old WT shows more loops and stronger TAD prominence; S6‑KO shows fewer loops and reduced TAD boundary strength TAD/loop counts from preprint

Compartment remodeling — reported fractions

Interpretation: global compartment identity is largely stable (A↔B flips extremely rare), but magnitude/intracompartment changes (wA/sA, wB/sB) affect ≈1–2% of 50‑kb bins. The Ighv cluster on chr12 is a recurrent wB→sB transition in all models Compartment change fractions

Synthesis: proposed model (visual)

Paper claim: aging moves neurons to a 'primed' state (increased local contacts, stronger TADs/loops, modest B‑compartment strengthening) whereas S6‑KO and CK‑p25 progress along a different axis toward architectural collapse (weakened TADs/loops, shorter loops, A‑B mixing increases for S6‑KO) — supported by PCA of compartment E1s and multiple structural metrics Model: aging primed vs pathology collapse

Critical appraisal — strengths and limitations (evidence-linked)

• Strength — matched neuronal Hi‑C from FANS-sorted NeuN+ nuclei: isolates neuron-specific 3D landscapes, reducing cell‑type confounds (Neuron-specific Hi-C).
• Strength — multilevel architectural metrics: authors use P(s), DLR, ICF, saddle plots, insulation/TAD strength, Chromosight loop calling, and loop-aggregate tests — giving convergent evidence for different scales of change (Multiscale Hi-C metrics).
• Strength — integration with CK‑p25 and transcriptomes: consistent convergence of architectural erosion with DNA-damage burden and REST/Polycomb TF network signal strengthens biological interpretation (Integration with CK-p25 and TF analyses).
• Limitation — small sample sizes for some comparisons: Hi‑C N=3 per in vivo group is typical but low; RNA replicates (shSIRT6 n=2 after QC exclusion) are underpowered for robust DEG calling and increase risk of false positives/overfitting (Replication and RNA sample sizes).
• Limitation — DNA damage association is correlative across models: S6‑KO and CK‑p25 have known DNA damage increases, and authors compare γH2AX stratified CK‑p25 neurons, but their own Hi‑C did not quantify DNA damage per nucleus, so damage–architecture causality remains inferential (Correlative DNA damage link).
• Limitation — species/model generalizability: mouse S6‑KO and CK‑p25 are models that capture aspects of AD pathology but differ from human sporadic AD; authors state species limitation explicitly (Species and sorting limitations).

Reproducibility & methods transparency

1. Data/code: authors provide GitHub repository (https://github.com/SIRT6/Sirt6_hic_paper) and plan GEO deposit; computational tools and versions are enumerated in Methods — strong reproducibility signal for Hi‑C pipelines (Data & code availability).
2. Hi‑C depth: libraries sequenced 150–350M PE reads; mapping MAPQ>30, ICE normalization and downsampling to equal contacts — appropriate QC steps for comparative Hi‑C (Hi-C sequencing depth & normalization).

Blindspots and possible alternative explanations

• Cell composition & sorting: NeuN+ selection enriches for neuronal nuclei but may miss neuronal subtypes differentially vulnerable to aging/pathology; CK‑p25 Step stratification used γH2AX to infer damage burden — differences in sorting strategy can drive apparent architecture differences between datasets (Sorting strategy confounds).
• Technical biases in loop/TAD calling: loop callers and insulation methods have parameter sensitivity; reported shifts in mean loop size (~95–105 kb) and cohesin-occupancy inferences depend on derivative analysis of Pc(s) and require cautious interpretation (Loop size estimation caveat).
• Translational uncertainty: Ighv locus changes (wB→sB) are intriguing but functional consequences in neurons (which silence Igh loci) are uncertain; enrichment may reflect regional genome features rather than neuronal functional rewiring (Ighv compartment change caution).

Falsifiability — what would change conclusions?

1. If higher‑replicate Hi‑C (>5 per group) and per‑nucleus γH2AX/Hi‑C co‑measurements show aging and S6‑KO/CK‑p25 have indistinguishable TAD/loop metrics after controlling damage burden, that would refute their divergence claim.
2. If cohesin/CTCF ChIP (or CUT&Tag) in the same nuclei shows no reduction in architectural proteins where loops/TADs are lost, the mechanistic link to architectural erosion would be weakened.
3. If perturbation to restore SIRT6 in S6‑KO rescues loop/TAD integrity in vivo, that would support causality; failure to rescue would suggest SIRT6-loss is correlative or secondary.

All three potential disproofs are experimentally tractable and would strengthen or reverse the study's model.

Concrete recommendations and high‑value follow-ups

• Co-assay γH2AX (or COMET) per nucleus combined with low‑input Hi‑C (or single‑cell Hi‑C/Hi‑CAR) to directly couple DNA‑damage burden and 3D architecture — would move the damage–architecture link from correlation to causation (Need for per-nucleus damage measurement).
• Perform cohesin (RAD21/SMC1) and CTCF CUT&Tag in matched nuclei to validate cohesin occupancy inferences from Pc(s) derivative analyses.
• Increase RNA-seq biological replicates (in vivo neuronal bulk or single-nucleus) and use matched nuclei to link specific 3D changes to gene regulation with higher statistical power.
• Test rescue: SIRT6 re-expression (AAV) in S6‑KO neurons in vivo, assess Hi‑C/TAD/loop restoration, and monitor DNA damage — critical test of causality.

Technical reproducibility checklist (from methods)

• Hi‑C enzyme: DpnII; crosslinking, biotin fill-in and ligation; sonication to <1000 bp; streptavidin pulldown; Illumina adapters; 150–350M PE reads/library.
• Processing: distiller (pairtools), bwa mapping, MAPQ>30, downsampling via cooltools.sample, ICE balancing via cooler balance, matrices 1 kb–1 Mb.
• Compartment calling: cooltools eigs_cis(n_eigs=8), saddle plots with cooltools.saddle(), differential PC1 via limma (lmFit/eBayes) with FDR <0.05 and |ΔE1|>0.3.
• TADs: insulation score (cooltools) at 10 kb, window 50 kb; differential boundaries = coordinate shift >10 kb.
• Loops: Chromosight (10 kb) min 35 kb max 3.5 Mb, Pearson threshold 0.45; differential loops = top/bottom deciles; permutation testing (10,000) for loop strength p‑values.

Having these settings makes reanalysis straightforward; code repo referenced in methods contains scripts to replicate.

Author review links

Open BGPT author reviews for paper authors:

(Other full-author buttons available on request.)