Visual experience shapes population representations

Population geometry, linear decoding and shared PCA views of visual cortical responses.

· Figure notebook · Analysis protocol ·

Population geometry

D. Dark rearing collapses population orientation representation in dorsal stream higher visual areas. Lateral dorsal stream recovers with visual experience, while medial dorsal stream does not.	H. Stimulus contrast controls spread of population orientation representation. Medial dorsal stream fails to recover this dark-rearing-impaired mechanism with visual experience while lateral dorsal stream recovers.

NR / DR (normally / dark-reared), baseline / chronic (P90 / P120 = P90+30 days of visual experience). V1: Primary visual cortex; AM-PM: Medial dorsal stream; A-AL-RL: Lateral dorsal stream; LM: Ventral stream.

Reproduce the figures

uv sync --locked
uv run jupyter lab plotting/figures.ipynb

Run the notebook to export B, C, D, F, G and H, plus the B supplements and D-radius barplots listed below. Each panel is saved independently as PNG, PDF, editable SVG and outlined SVG in plotting/exports/.

The included plotting/source_data/ contains only the 19 numerical inputs needed to draw these panels. No recordings or R installation are needed to reproduce the figures. For a noninteractive export:

uv run python -m plotting

Analyze recordings

Place MATLAB v7.3/HDF5 recordings in this structure:

recordings/
├── darkReared_baseline/    Mouse_session.mat
├── normReared_baseline/    Mouse_session.mat
├── darkReared_chronic/     Mouse_session.mat
└── normReared_chronic/     Mouse_session.mat

The filename prefix before the first underscore identifies the mouse; use the same prefix across timepoints. Files are discovered automatically, in the condition order above and then filename order. That order defines deterministic session IDs and sampling seeds.

Each recording supplies Res/CaDeconCorrected (time × trial × neuron in MATLAB), Res/ax (seconds relative to stimulus onset), info/Stim/log (trial × 2: direction in degrees, contrast), and ABAroi/region (one atlas ID per neuron). The design has eight directions, eight contrast levels and 15 trials per direction–contrast cell. Atlas IDs 1–7 map to AM, PM, A, AL, RL, LM and V1; zero or a missing region field is unassigned. Missing behavior, anatomical coordinate and calcium fields do not prevent deconvolved analysis. See preprocess/README.md.

For the statistical analyses, install R and nparLD 2.3.3 with its dependencies. Use the source package because an older R binary repository may offer an older version:

Rscript -e 'install.packages("nparLD", repos="https://cloud.r-project.org", type="source"); stopifnot(as.character(packageVersion("nparLD")) == "2.3.3")'

Then run the complete pipeline:

uv run python run.py --input /path/to/recordings

This prepares the signals, computes geometry, LDA and PCA, runs mouse-level inference, and exports all panels to results/spikes/panels/. Preparation also works independently:

uv run python -m preprocess.prepare --input /path/to/recordings
uv run python run.py --stage geometry
uv run python run.py --stage decoding
uv run python run.py --stage pca
uv run python run.py --stage statistics
uv run python run.py --stage supplement
uv run python run.py --stage plot

Preparation stage (preprocess.prepare) does not replace existing data unless --overwrite is supplied. LDA reuses checkpoints only when their inputs and code match. To use raw calcium instead, pass --signal calcium; its input field is Res/CaSigCorrected, and its data/results are written separately. Only deconvolved results are included here.

Numerical protocol

• 10 neurons, 1,000 shared draws and five fixed stratified folds per session–area population. Neurons must have finite values across the recording. Activity is used without baseline subtraction, variance normalization or additional scaling.
• B/C/D/G/H average 0 ≤ t < 0.5 s. F takes the peak over native frames in 0 ≤ t ≤ 0.5 s, after averaging draws and sessions within each mouse.
• B/C/D pool the top two contrasts. F compares bottom two versus top two. G uses four adjacent pairs and displays 4→3, 3→2, 2→1. H pools bottom four versus top four, preserving all eight directions until after PCA.
• LDA uses analytic Ledoit–Wolf shrinkage of pooled training residuals and class-frequency priors. Held-out trials do not enter fitting. The estimator differs from sklearn LDA's classwise shrinkage="auto".
• Sessions are averaged within mouse; mice receive equal weight. Intervals use 5,000 mouse-bootstrap resamples. Draws and folds are technical repeats.
• Within-area alignment uses training centroids and rotates held-out centroids without scaling. Subsequent cross-area registration and PCA use group means and are descriptive. Ordinary cosine, distance and radius retain finite-trial noise; F also includes peak-selection bias.

Supplementary panels

The notebook exports each supplement separately:

Export name	Content
B_angular, B_euclidean, B_cross_contrast	Angular cosine, endpoint Euclidean distance, and cross-contrast cosine
B_angular_distance	Full-angle Euclidean distance, including individual mouse values
orientation_contrasts	Cosine and distance contrasts across areas
cosine_contrast, euclidean_contrast	Compact 150 × 100 pt contrast panels
B_cosine_endpoints, B_euclidean_endpoints	B curves with NR–DR tests at 180° and 90°
B_cosine_contrast, B_euclidean_contrast	B curves annotated with tests of the within-mouse contrast
radius_bars	D's full-space squared radius across areas, on a shared scale (150 × 100 pt)
radius_V1, radius_AM_PM, radius_A_AL_RL, radius_LM	Individual D-radius barplots with area-specific scales (55 × 100 pt)

Orientation contrasts subtract angles within each mouse and timepoint: C(180°) − C(90°) for cosine, d(90°) − d(180°) for distance. Pointwise 95% intervals bootstrap these mouse differences, preserving their covariance. Baseline and chronic retain every eligible mouse independently; longitudinal pairing is not required.

Statistics

Statistical tests compare mice, with multiple recording sessions averaged within each mouse. Nonparametric factorial ANOVA-type tests (nparLD) assess rearing, visual area, timepoint and the panel-specific factor, accounting for repeated measurements and retaining incomplete records. NR–DR post-hoc comparisons use two-sided exact Wilcoxon rank-sum tests. Holm correction covers four main effects per panel and, separately, two timepoints per area/endpoint (six timepoint × transition comparisons for G). Supplementary orientation-contrast tests additionally report a stricter six-comparison sensitivity correction. Intervals are pointwise 95% mouse-bootstrap intervals; PCA views are descriptive.

Organization

Folder	Analyses
preprocess/	MATLAB loading, response preparation, eligibility, neuron sampling and trial folds
geometry/	B and its supplements, D's radius inset, F's peak radius and inference
decoding/	Ledoit–Wolf LDA for C and G
dimensionality-reduction/	Alignment and shared PCA for D/H; imported as dimensionality_reduction
plotting/	Panel renderers, source data, README previews and the final notebook
tests/	Numerical checks, MATLAB ingestion and full pipeline/export tests

Prepared data defaults to data/<signal>/, results to results/<signal>/. Set VEP_DATA_DIR or VEP_RESULTS_DIR before starting Python to change those locations. For another results folder, use uv run python -m plotting --source /path/to/results --output /path/to/panels, or set RESULTS in the notebook and restart its kernel.

Raw recordings, intermediate arrays, caches, environments and generated exports are excluded from Git. The only bundled analysis data are the figure inputs in source_data; D/H preview images support this README. Notebook outputs are cleared to keep the repository small.