The Evidence Layer for Scientific AI

 Long Analysis Plan Copied

Evidence-Backed Analysis Plan: Prefrontal Control of Hippocampal Map Reorganization During Sleep

The plan below builds a reproducible bioinformatics/coding workflow around the central finding that post-encoding sleep selectively reinstated familiar CA1 place-cell maps in a similar novel environment, and that this generalization required mPFC activity during sleep Primary mouse miniscope imaging study.

1. Data Ingestion & Preprocessing (Python 3.11)

• Ingest raw UCLA Miniscope V4.4 Ca2+ videos and run the Minian/CNMF-E pipeline (as used by the authors) to extract CA1 soma activity; register behavior videography via OpenCV-based tracking and Facemap Imaging and analysis pipeline details.
• Score sleep vs wake from video immobility (no EEG available — flag as measurement bias); segment sessions into F, N, retention (sleep/wake), and re-exposure (N', F') epochs.
• Build an anndata-backed trial store: per-mouse, per-trial rate maps, event times, and group labels (n=5 imaging mice; n=4 DCZ/vehicle; n=7 behavior) Sample size and exclusions.

2. Core Quantitative Replication

• Recompute F-N' cell-by-cell rate-map correlations (expected sleep 0.32±0.11 vs wake 0.10±0.07) and PV correlations; note the PV analysis did not reach significance (p=0.104) — treat as sensitivity check, not primary outcome Map similarity statistics.
• Train a Poisson maximum-likelihood decoder on F, test decoding error in N' (sleep 9.93±1.23 cm vs wake 11.23±0.64 cm, Mann-Whitney p=0.018); bootstrap confidence intervals given only 9 vs 8 trials from 5 mice Decoding accuracy comparison.
• Compute the Generalization Index across conditions; verify Kruskal-Wallis p=0.0002 and Dunn's post-hocs (DCZ vs sleep p=0.003; sleep vs awake p=0.002; DCZ vs awake p=0.644) mPFC inactivation Generalization Index.

Higher values indicate that N' maps shifted toward the familiar (F) representation. Sleep produces positive generalization; mPFC silencing during sleep collapses this toward the awake value, supporting prefrontal causal control Generalization Index graph data.

3. Behavioral & Mechanistic Layers

• Quantify N' exploration behavior: normalized speed (sleep 2.49±0.51 vs wake 1.72±0.66, Wilcoxon p=0.005), stops (18.5±11.38 vs 37.42±22.04, p=0.007), rearing (p=0.01) Behavioral familiarity statistics. Implement as a Python analysis module with Wilcoxon signed-rank tests (statsmodels/scipy).
• Reimplement the 10-neuron Hebbian recurrent network (5 familiar + 5 novel units, global inhibition) testing whether sleep-phase F-N co-activation suffices to bias N' responses toward the familiar population — a falsifiable mechanistic simulation Recurrent network model.

4. Critical Caveats to Encode in the Pipeline

• Small samples: 4-5 mice per neural comparison; bootstrap or exact tests are mandatory; the sleep-vs-wake behavioral comparison (12 trials, 7 mice) is underpowered for medium effects.
• Low novelty salience: F and N were highly congruent, deliberately favoring generalization — the effect may not extend to high-novelty environments Author-reported limitations.
• Sleep scoring without EEG: immobility-based scoring can misclassify quiet wake as sleep — a key confound to quantify with velocity-threshold sensitivity analyses.
• Male-only cohort: findings cannot be assumed to generalize to females.
• Falsification path: a preregistered replication fails if post-encoding sleep does not raise F-N' similarity over wake, or if mPFC inactivation does not reduce the Generalization Index Falsification criteria.