cembra

CEMBRA OA pseudo-SNP SVD64 development archive

This private repository is a curated model-and-evidence export from the oa_tl research codebase. It scores paired 3,000 bp reference/alternate DNA windows for osteoarthritis (OA) variant prioritization using a 9,216-feature hashed DNA representation, a 64-dimensional encoder initialized by label-free TruncatedSVD pretraining, and a tanh/linear ranker.

Research archive only. This is a development model, not an externally confirmed, clinically validated, calibrated, or deployment-ready predictor. It must not be used for diagnosis, treatment, or patient-level decisions.

Why this model was selected

The repository contains many historical and experimental checkpoints. This export selects the compact pseudo-SNP SVD64 arm because it is the only model that passed its locked development gate and it remained the best native single arm in the later five-model NT/Borzoi comparison.

On the governed OA2025 PIP >= 0.30 development cohort (168 rows, 56 positives, 49 guarded components), its pooled out-of-fold result was:

The gain was modest and not uniform across folds. External confirmation was not opened or scored. See results/pseudosnp/AUDIT.md and results/pseudosnp/model_selection_gate.json for the exact claim boundary.

Important negative evidence

• The formal leakage-resistant NT-v2 comparison was a development NO-GO: pretrained-minus-random matched concordance was -0.1136 on 22 matched groups. No NT-v2 model lock was justified.
• The later fixed 1:1 NT+Borzoi fusion was also a NO-GO. Its AUROC/AP were 0.500638/0.356623, below pseudo-SNP SVD64 by 0.110651/0.077032.
• July 2026 NT-v2 metrics and eight approximately 374 MiB checkpoints are retained locally as historical artifacts but are superseded for model selection and intentionally not uploaded here.

Files

Weight format

Every fold file is a safe NumPy .npz archive containing only float32 arrays:

• encoder_weight: [9216, 64]
• encoder_bias: [64]
• scorer_weight: [64]
• scorer_bias: scalar

The raw fold score is

tanh(features @ encoder_weight + encoder_bias) @ scorer_weight + scorer_bias.

The uploaded artifacts are cross-validation fold models, not a single full-development refit. The convenience mean across the five fold scores in inference.py is useful for research inspection but was not the evaluated OOF quantity and is not a validated deployment ensemble.

Usage

from inference import PseudoSnpSVD64

model = PseudoSnpSVD64(".")
fold_scores = model.score_pair_by_fold(reference_3000bp, alternate_3000bp)
research_only_mean = model.score_pair(reference_3000bp, alternate_3000bp)

The two sequences must be exactly 3,000 bases long, contain the variant at zero-based index 1,500, and differ at that position only. Scores are raw ranking logits, not probabilities. Do not apply a clinical threshold.

To verify the package against the stored OOF evidence:

python verify_export.py

Provenance

• Source repository: [email protected]:zibin-zhao/cembra-bio.git
• Exported from commit: 76a214e44d83d7577d7d9bc7a67796240e970d25
• Source branch at export: codex/nt-borzoi-oof-fusion
• Selected run: 20260808T003427Z_multiscale_pseudosnp_svd_344bbf0cca5b
• Selected model schema: oa-tl-tanh-encoder-linear-triplet-ranker.v1
• Feature schema: oa-tl-multiscale-pseudosnp-hashed-dna.v1
• Original audited run-manifest SHA-256: ad72b7a4f5585dd6ebd3c9639d0ff48d7a69d4d124472e204c2b2edd8533925a

The 2026-08-22 fusion audit and runner existed as uncommitted worktree artifacts at export time; its own audit records result-affecting source and artifact hashes. It is included as later negative context, not as selected model provenance.

License

Proprietary, all rights reserved. The source repository license grants no reuse rights without written permission from the owner.