pevo-msa-grch38-19way (dataset Hub)

EN: Training data, thesis tables, and reproduction artifacts for primate MSA variant-effect modeling.
中文： 灵长类 MSA 变异效应建模的训练数据与论文复现数据（不含模型权重）。

Results-status note (2026-07-29). The multi-seed metrics reported below are retained as historical registry records. They use an earlier scoring protocol and cohort convention, and are not comparable to the corrected strict-v2 results used for the final thesis conclusions. Do not use the values below for a current 10-way-versus-19-way claim; use the final thesis and reproducibility materials with their frozen corrected-score manifests instead.

Code & experiment map	github.com/jasperyeoh/pevo-msa-primate-genomics — see EXPERIMENT_MAP.md
Model checkpoints	jasperyeoh2/pevo-msa-mlm-19way
Full file tree	docs/HUGGINGFACE_DATASET_LAYOUT.md (this repo)
Upload audit	docs/UPLOAD_AUDIT.md on GitHub

---

Dataset vs model — which Hub?

  GitHub (code + small CSV)          THIS REPO (data)              MODEL REPO (weights)
  ─────────────────────────          ────────────────              ────────────────────
  scripts, configs, thesis CSV  →    Zarr, MAF, thesis_repro  +    base_models/*/checkpoint-best
                                     windows, conservation          configs/, inventory/

---

Quick download / 快速下载

pip install -U huggingface_hub

# A) Paper reproduction only (~hundreds of MB)
hf download jasperyeoh2/pevo-msa-grch38-19way --repo-type dataset --local-dir data_hf \
  --include 'thesis_repro/**' 'docs/**'

# B) Training corpus (Zarr per-chromosome tar — extract after download)
hf download jasperyeoh2/pevo-msa-grch38-19way --repo-type dataset --local-dir data_hf \
  --include 'new_dataset_27feb/msaasr_dataset_10primates_full/**'

# C) Full snapshot (100GB+)
hf download jasperyeoh2/pevo-msa-grch38-19way --repo-type dataset --local-dir data_hf

---

thesis_repro/ — paper tables & scores

Restore into a Git clone (paths expected by thesis_manu_v2_6/scripts/):

git clone https://github.com/jasperyeoh/pevo-msa-primate-genomics.git
cd pevo-msa-primate-genomics
hf download jasperyeoh2/pevo-msa-grch38-19way --repo-type dataset --local-dir ../data_hf \
  --include 'thesis_repro/**'

mkdir -p phylo_msa1/outputs/vep_all_base_models
cp ../data_hf/thesis_repro/phylo_msa1_outputs/VEP_*.csv phylo_msa1/outputs/
cp ../data_hf/thesis_repro/phylo_msa1_outputs/vep_all_base_models/*.parquet \
   phylo_msa1/outputs/vep_all_base_models/

cp -a ../data_hf/thesis_repro/dna_foundation_benchmark_results_final \
      downstream_tasks/dna_foundation_benchmark/results_final

mkdir -p thesis_manu_v2_6/artifacts
cp ../data_hf/thesis_repro/thesis_manu_v2_6_artifacts/* thesis_manu_v2_6/artifacts/

Tip: Many summary CSVs are also on GitHub now — git clone alone is enough for Tier-A LaTeX regeneration (see EXPERIMENT_MAP.md).

thesis_repro/ path	Used for
phylo_msa1_outputs/VEP_*.csv	VEP leaderboards, strict-v2 tables (tab:vep_*, tab:strictv2_*)
phylo_msa1_outputs/vep_all_base_models/*.parquet	Per-checkpoint VEP vectors, ablation registry
dna_foundation_benchmark_results_final/	Feng AUROC tables/figures (tab:feng_*, fig:feng_*)
thesis_manu_v2_6_artifacts/	Matched-width strict-v2 bootstrap (May 2026)

---

Historical multi-seed matched-width registry (Jun 2026; earlier protocol)

Historical only. The three-arm values in this section predate the corrected target-masked, strand-averaged strict-v2 scorer. They are preserved for provenance, not as a corrected matched-width comparison.

3 seeds × 3 arms (full_19way, matched_7way, human_only) trained on AutoDL (RTX 4090), evaluated on strict-v2 cohort (n=2000).

Artifact	Path on this Hub
VEP per-variant parquets (9 runs)	phylo_msa1/outputs/multi_seed/vep_scores/*_scores.parquet
Bootstrap / cross-seed CSVs	phylo_msa1/outputs/multi_seed/vep_scores/multi_seed_*.csv
Analysis report	phylo_msa1/outputs/multi_seed/vep_scores/multi_seed_analysis_report.md
Training + VEP logs	phylo_msa1/outputs/multi_seed/logs/
Config snapshots	phylo_msa1/outputs/multi_seed/configs/train_*.json
Checkpoint summaries	phylo_msa1/outputs/multi_seed/training_summaries/
Provenance index	phylo_msa1/outputs/multi_seed/MULTI_SEED_MANIFEST.md

Checkpoints (~75 MB each) live on the model Hub:
jasperyeoh2/pevo-msa-mlm-19way → base_models/multi_seed_{arm}_seed{seed}/checkpoint-best/

Historical cross-seed AUROC (mean ± std): full_19way 0.7925 ± 0.0051 · matched_7way 0.7068 ± 0.0042 · human_only 0.5085 ± 0.0020

# Download scores + logs only
hf download jasperyeoh2/pevo-msa-grch38-19way --repo-type dataset --local-dir data_hf \
  --include 'phylo_msa1/outputs/multi_seed/**'

# Download one checkpoint for re-scoring
hf download jasperyeoh2/pevo-msa-mlm-19way --repo-type model --local-dir models_hf \
  --include 'base_models/multi_seed_full_19way_seed42/**'

Reproduce: GitHub phylo_msa1/scripts/multi_seed/ (run_multi_seed_matched_width.sh, eval_multi_seed_vep.sh).

---

Main data prefixes / 主要数据目录

Prefix	Contents	Size
thesis_repro/	Paper CSV + VEP parquets + Feng summaries	small
phylo_msa1/outputs/multi_seed/	Multi-seed ablation VEP scores, logs, configs (Jun 2026)	~3 MB
new_dataset_27feb/msaasr_dataset_10primates_full/bundles/	24× msaasr.zarr.bundle_*.tar	~132 GB total
new_dataset_27feb/gpnmsa_1024bp_pipeline/	1024bp window parquets, VEP ablation (VCFs may be chunked)	mixed
window_workspace/neutral_model_10primates/	merged_10primates_training.maf	~23 GB
conservation_pipeline/scores/	phastCons / phyloP .wig per chromosome	moderate
Root train_*.parquet, windows_10primates_*.parquet	Multi-resolution window panels	moderate each
msaasr.zarr/ (legacy)	Old chunk tree at repo root	prefer bundles

---

Thesis figure/table index (data side)

LaTeX label	Data on this Hub
tab:vep_*, tab:strictv2_*, tab:vep_leaderboard_full	thesis_repro/phylo_msa1_outputs/VEP_*.csv (+ parquets)
fig:feng_*, tab:feng_*	thesis_repro/dna_foundation_benchmark_results_final/
fig:gnomad-tail-*	CSV on GitHub; score parquets cluster-only
Retrain from scratch	bundles/*.tar → extract msaasr.zarr/

Checkpoints for scoring: model Hub base_models/full_training_*/checkpoint-best/.
Full mapping: GitHub EXPERIMENT_MAP.md.

---

MSAASR Zarr restore / 训练语料恢复

1. Download new_dataset_27feb/msaasr_dataset_10primates_full/bundles/msaasr.zarr.bundle_*.tar
2. Read README_MSAASR_BUNDLES.md in that folder
3. Extract to new_dataset_27feb/msaasr_dataset_10primates_full/msaasr.zarr/
4. Do not rely on root msaasr.zarr/ chunk trees (millions of files; slow)

---

Regenerate thesis LaTeX (minimal)

cd pevo-msa-primate-genomics/thesis_manu_v2_6/scripts
python3 gen_strictv2_ci_artifacts.py
python3 gen_feng_combined_seq_and_l7_auroc.py
cd .. && latexmk -pdf main.tex

---

Citation & license

Cite the thesis/preprint and pin this dataset commit from History.
License: MIT (dataset card). gnomAD / ClinVar / EPO data remain under their original terms.