Where predictive information resides in endogenous A-to-I RNA editing

This repository contains a study of where predictive information for endogenous A-to-I RNA editing resides. Starting from the published AdarEdit human inverted-Alu benchmark, the project traverses three biological information scales, a separate within-substrate dependence problem, and four classes of model-sufficiency tests under pair-disjoint development evaluation. The study retains supported, limited, unsupported, superseded, and data-limited branches.

Whole-duplex sequence and predicted structure provide a strong reference. Genomic organization adds between-substrate information about editing burden; correctly centered local representations improve use of the target neighborhood. Class-balanced pair-first learning reduces the influence of site-rich substrates within each class. Current context proxies have not supplied a consistent site-level correction in the retained comparisons. Neither substrate realization nor an underlying cellular state was measured directly.

The frozen predictor is an unchanged 50:50 equal-logit ensemble of two independently trained models: a pair-first whole-duplex-only graph model and a second pair-first graph model internally augmented with a target/helix-partner local encoder. A separately preregistered unified end-to-end architecture was developed before test access and rejected on development data alone.

Locked canonical result

The canonical test was evaluated once after the complete pipeline was frozen. The table below compares the reproduced supplied-checkpoint AdarEdit baseline with the frozen selected predictor on the original published task and split.

Context	AdarEdit AUROC	Selected predictor AUROC	AdarEdit AUPRC	Selected predictor AUPRC	AdarEdit F1	Selected predictor F1
Artery	0.9238	0.9477	0.9262	0.9485	0.8671	0.8869
Brain	0.9301	0.9508	0.9286	0.9504	0.8689	0.8902
Liver	0.9092	0.9414	0.9242	0.9500	0.8541	0.8812
Skeletal muscle	0.8688	0.9144	0.8757	0.9207	0.8139	0.8541
Combined	0.9331	0.9491	0.9356	0.9513	0.8674	0.8852

Mean gains across the five reported benchmark endpoints were +0.0277 AUROC, +0.0261 AUPRC, and +0.0253 F1. Every endpoint had a positive pair-clustered 95% interval for all three differences. In a post-test descriptive analysis at matched reproduced-baseline recall, the selected predictor avoided 126, 119, 158, 68, and 148 false positives, respectively, with positive clustered intervals throughout.

These held-out results are the authoritative performance claim. The lower absolute grouped out-of-fold scores in development are an internal research instrument used to compare candidates under identical folds; they do not replace or redefine the published benchmark.

What the campaign established

• Pair-first expanded high/low supervision improved all five development endpoints under the same-fold internal comparison.
• Recovering all valid high/low labels and correcting pair weighting made separable contributions in the muscle same-fold ablation.
• Inter-arm geometry and low-capacity pair models predicted substrate editing burden, but its tested site-level integrations remained limited.
• A graph model augmented with an aligned target/helix-partner local view added complementary information. An otherwise identical model centered 96 nt away underperformed, supporting correct target centering beyond added CNN capacity. This control does not isolate predicted partner identity.
• Fixed prediction-level integration outperformed the preregistered unified joint model at the available effective pair-level sample size.
• Tested eCLIP/PrismNet/RBPNet × bulk GTEx representations, current bulk H3K36me3 overlap, broad site-level annotations, frozen RNA foundation-model embeddings, and several richer-task formulations did not provide sufficiently robust incremental information for promotion. These are statements about the tested measurements and models, not claims that the corresponding biology is absent.

The independent source-code audit found that early residual screens reused pooled first-stage OOF scores in pair-disjoint second-stage fits. They are exploratory, not fully nested evaluations. The supplied Bio-aware comparison also changes features and architecture together and cannot isolate capacity. These limits, corrected sampler/comparator descriptions, and the unaffected controlled neural and final-test evidence are documented in reports/29_independent_submission_audit.md.

The full decision trail is indexed in docs/CAMPAIGN_INDEX.md. The scientific synthesis is in reports/FINAL_REPORT.md.

Manuscript

The PNAS Research Report manuscript is available as paper/main.pdf, with the complete campaign, controls, and fold-level evidence organized in paper/supplementary.pdf. Their LaTeX sources, official-template provenance, verified BibTeX, claim-to-artifact audit, and programmatically generated figure and table sources are retained in paper/. Author, affiliation, contribution, funding, and correspondence fields remain explicit placeholders pending author approval. The independent submission audit also identifies the residual-screen evidence gap and the unresolved PNAS AI-assisted-graphics policy question. Biological illustrations now use credited DBCLS/TogoTV and Servier SVG artwork. Historical GPT Image assets remain excluded from the active build; see archive provenance. The licensed-illustration revision records source selection, figure changes, and verification. These PDFs are revised author-review drafts, not certified submission-ready files. The sequential reader-comprehension walkthrough records where an adjacent-field reader could lose the argument, the resulting revisions, and the second-pass checks. It is an editorial simulation, not a human-participant comprehension study. The final scientific-writing pass and multilingual semantic audit record the implemented readability revisions, Hebrew/German/Japanese/Spanish meaning checks, quantitative preservation, and remaining author-completion items.

make paper-check

This regenerates every paper figure and SI table, compiles both PDFs, verifies the bibliography and PNAS field limits, and checks quantitative and claim traceability.

Repository map

Path	Contents
src/adarpro/, code/	ADARPro analysis/model code and pinned AdarEdit model code
experiments/	Append-only experiment registry, configurations, and retained run artifacts
models/	Frozen selected checkpoints, summaries, and development model artifacts
results/	Reproduced AdarEdit tables and the one-use locked-test result
data/	Canonical benchmark inputs and retained small derived tables
figures/	Reproducible figures and quantitative source tables
reports/	Campaign/editorial reports 00–34 and final synthesis
provenance/	Input hashes, freeze manifests, audit chain, inventory, and exclusions
paper/	PNAS manuscript, SI, verified bibliography, and paper source tables
docs/	Campaign, data, reproducibility, and upstream documentation

Reproduce and verify

The exact checkpoint-compatible environment is defined in environment.yml; the lightweight package/test dependencies are in pyproject.toml.

conda env create -f environment.yml
conda activate adaredit
python -m pip install -e '.[dev,tables]'
make verify

make verify checks the global pair split, result identities, registry integrity, tests, lint, and the curated release manifest. See docs/REPRODUCIBILITY.md for the complete paths from retained artifacts to result tables, figures, and paper.

Data and provenance

Canonical human benchmark files and the supplied AdarEdit checkpoints are retained. External resources and four large rebuildable feature matrices are not stored in Git; their source identities, SHA-256 hashes, roles, and exclusion reasons are recorded in provenance/source_workspace_inventory.csv, provenance/excluded_artifacts.csv, and docs/DATA.md. No negative or terminated result was omitted because it was unfavorable.

License

The repository retains the GNU General Public License v3 supplied with the AdarEdit codebase; see LICENSE. External datasets and pretrained weights remain governed by their original providers' terms.