Skip to content

CLAUDE.md

Latest commit

 

History

History
135 lines (105 loc) · 6.23 KB

CLAUDE.md

File metadata and controls

135 lines (105 loc) · 6.23 KB
Raw

XORI: cross-orientation suppression in macaque V1

Two-photon calcium imaging of cross-orientation suppression across cortical depth in macaque V1 layer 2/3. ~4,800 ROIs, 28 fields of view, 140-518 um. First-author manuscript in final-push phase.

Bair Lab (UW Neurobiology and Biophysics) in collaboration with Allen Institute and WaNPRC.

Vault landmark

~/Library/Mobile Documents/iCloud~md~obsidian/Documents/Dreamless_Machine/03-Research/BAIR-Lab/

  • Plan.md. Manuscript state, figure inventory, key numbers, next 3 steps. Source of truth for project status.
  • Notes.md. Working notes.
  • BAIR-Lab.md. Hub.

Separate, unrelated project

V1 weight signatures lives at ~/v1-weight-signatures/. That is mechanistic interpretability work on pretrained vision nets vs V1 receptive fields. It is NOT this paper, and it is NOT what "xori" refers to. Different paper, different authorship, different scope, different agent. Do not conflate. Do not write to ~/v1-weight-signatures/ from this workspace.

Repo layout

~/xori/
├── raw_data/                  # ROI-level inputs, tuning curves
│   ├── bm_data/               # site_depth, roi_osi, roi_stat, roi_hw_orth, roi_lhi, roi_sf
│   └── tc_data/               # per-site tuning-curve responses
├── metric_data/               # cross-ori metrics S, R, etc, per ROI per site
│   ├── all_roi/               # primary
│   ├── cull_roi/              # high-SNR subsets
│   └── r_cull_roi/            # low-SNR subsets (QC)
├── depth_data/                # depth analysis outputs (plots, ROI maps)
├── data_baseline,_halfwidth,_lhi,_osi,_size,_spatial/
│                              # per-covariate analysis outputs
├── scripts/                   # core analysis
│   ├── m_calc/                # metric calculation from tuning curves
│   ├── d_calc/                # depth correlation analysis
│   └── bm_calc/               # covariate-metric relationships
├── stat/                      # Suite2p stat files (ROI spatial masks)
├── supplementary_analysis/    # extended stats (mixed-effects, partial corr, mediation)
└── paper/                     # manuscript + figures
    ├── manuscript.tex
    ├── references.bib
    ├── manuscript_stats.json  # canonical numbers (do NOT regenerate casually)
    ├── make_figures.py        # generates fig1, fig3..fig9
    ├── compute_stats.py       # prints LaTeX-formatted stats to stdout
    └── figures/               # PDF + PNG outputs

Quick commands

All commands assume cwd = repo root unless noted.

# activate env (Python 3.12 venv with numpy/pandas/matplotlib/scipy/statsmodels)
source .venv/bin/activate

# regenerate all figures (fig1, fig3..fig9). Figure 2 not yet implemented.
python paper/make_figures.py

# print canonical stats (LaTeX-formatted, suitable for tables)
python paper/compute_stats.py

# build manuscript PDF (requires MacTeX or BasicTeX, currently NOT installed)
cd paper && latexmk -pdf -interaction=nonstopmode -halt-on-error -outdir=build manuscript.tex

Site038 caveat (data, not exclusion)

Site038 is the shallowest site at 140 µm. All 134 ROIs have SNR_g < 0.5. The SNR-thresholded cull script (scripts/m_calc/cull_metric.py:240) correctly logs "no ROIs meet threshold (skipped)" and writes no files for site038 under any of metric_data/cull_roi/thr_cull/{above_0_5,above_1_0,above_1_5}/. This is documented behavior, not data corruption or analytic exclusion.

Site038 IS retained in metric_data/all_roi/, the percentile culls (per_cull/{top_70,top_80,top_90}/), the reverse culls, and all depth and covariate analyses. Do not exclude site038 from any analysis without explicit ask.

Setup blockers

  • LaTeX is not installed. No latexmk/pdflatex/bibtex on PATH. Install MacTeX or BasicTeX before claiming manuscript-ready. Until then, paper/manuscript.pdf is whatever the upstream repo committed.

Key metrics (from README + manuscript)

Code name Paper name Description
M_S_ratio S, P observed/predicted plaid response ratio. Primary.
M_C R, C Pearson r between observed and predicted curves.
M_S (legacy) signed difference observed minus predicted.
M_S_norm (legacy) signed diff normalized to baseline F.
M_X secondary additional cross-ori metric.
SNR_g/p filter grating/plaid SNR.

S < 1 = suppression, S > 1 = facilitation. log2(P) used for site-level stats.

Headline numbers

  • S vs depth: r = -0.768, [-0.908, -0.617], p = 1.83e-6, n=28 sites (log2-transformed).
  • R vs depth: r = +0.798, [+0.675, +0.891], p = 3.67e-7.
  • Mixed-effects S~Depth at single-ROI level: coef = -9.30, p = 7.9e-10, n=4785.
  • SF mediates ~40%, bandwidth ~21% of the S-depth effect.

Final push (current focus)

  1. Figure 2 (model schematic). Only missing figure. Methods §Modeling specifies Gabor pair (sf=1.0, SD_par=0.4, SD_orth=0.3) + Maxwell-difference temporal filter (s1=20ms, s2=40ms) + space-time separable filter outputs.
  2. Integrate fig9_depth_profile into manuscript.tex. Open question: replace fig3D depth-binned panel, or stand alone? Resolve before committing.
  3. Internal review pass. Verify every \ref resolves, every figure has \includegraphics, every numerical claim matches manuscript_stats.json, every reference in references.bib is cited.

Operator trigger

Post-commit hook at .git/hooks/post-commit calls ~/.local/lib/tri/bin/operator-trigger.sh. Operator refreshes BAIR-Lab/{Plan, Notes,BAIR-Lab}.md after each commit unless the subject contains [no-operator] or [operator] (the latter is set when Operator itself commits, to prevent loops).

Voice and tri usage

Inherited from ~/.claude/CLAUDE.md and ~/.claude/VOICE.md. For manuscript writing and Methods/Results edits, use Register 2 (academic), modeled on Bair's JNeurosci style. For figure captions: declarative, quantification-first, no "interestingly" or "notably."