xxl0001/mtedf-rice-vep

MTEDF Rice Variant-Effect Prediction

This public dataset contains the rice (Oryza sativa ssp. japonica) MSU7/IRGSP-1.0 reference genome, population SNVs with exact per-site allele numbers, conservation tracks, and genome-wide nucleotide log-likelihood ratios (LLRs) from multiple DNA language models. It also includes the Top4 Q25 ensemble and the analysis-ready merged variant table used by the accompanying rice variant-effect prediction benchmark.

The selected rice COD PlantCAD model is available separately at xxl0001/COD-PlantCAD-Rice.

Snapshot

Item	Value
Repository size	approximately 45.94 GB (42.79 GiB)
Repository data files	22, including this Dataset Card
Input SNVs	21,830,339
Analysis-ready scored SNVs	21,830,339
Reference assembly	MSU7 / IRGSP-1.0
Nuclear chromosomes	Chr1-Chr12
Raw model LLR matrices	12
Merged score columns	12 model LLRs, Top4 Q25, phyloP, phastCons

Quick start

For most analyses, download the two row-aligned Parquet files. The input table provides exact per-site AN; the merged table provides all cached scores.

import numpy as np
import pandas as pd
from huggingface_hub import hf_hub_download

repo_id = "xxl0001/mtedf-rice-vep"

variants_path = hf_hub_download(
    repo_id=repo_id,
    repo_type="dataset",
    filename="input_variants/variants.parquet",
)
scores_path = hf_hub_download(
    repo_id=repo_id,
    repo_type="dataset",
    filename="unmask_llr/result_all.parquet",
)

keys = ["chrom", "pos", "ref", "alt"]
variants = pd.read_parquet(
    variants_path,
    columns=[*keys, "AC", "AN", "AF"],
)
scores = pd.read_parquet(
    scores_path,
    columns=[
        *keys,
        "is_folded",
        "COD_FT_step12k_llr",
        "GPN_llr",
        "Top4_Elite_Q25",
        "phyloP",
        "phastCons",
    ],
)

assert variants[keys].equals(scores[keys])

# Exact-AN minor allele statistics.
ac = variants["AC"].to_numpy(np.int64)
an = variants["AN"].to_numpy(np.int64)
mac = np.minimum(ac, an - ac)
maf = mac / an

# result_all was initially oriented using AF > 0.5. Correct only rows where
# that decision differs from the exact-AN minor-allele direction.
exact_fold = (2 * ac) > an
fold_correction = scores["is_folded"].to_numpy(bool) ^ exact_fold
cod = scores["COD_FT_step12k_llr"].to_numpy(np.float32, copy=True)
cod[fold_correction] *= -1.0

Download the complete repository with the Hugging Face CLI:

hf download xxl0001/mtedf-rice-vep \
  --repo-type dataset \
  --local-dir ./mtedf-rice-vep

The complete snapshot is large. Prefer hf_hub_download or an include pattern when only one table or one model matrix is required.

Repository structure

mtedf-rice-vep/
|-- reference_genome/
|   |-- osa1_r7.asm.chrs.fa
|   `-- osa1_r7.asm.chrs.fa.fai
|-- input_variants/
|   `-- variants.parquet
|-- conservation_scores/
|   |-- Osj_PhyloP.bw
|   `-- Osj_PhastCons.bw
`-- unmask_llr/
    |-- result_all.parquet
    |-- gpn_llrs.pkl
    |-- plantcaduceus_l{20,24,28,32}_llrs.pkl
    |-- plantcad2_{small,medium,large}_llrs.pkl
    |-- ntv3_650m_pre_llrs.pkl
    |-- plantbimoe_llrs.pkl
    |-- cod_plantcad_llrs.pkl
    |-- cod_plantcad_step12k_llrs.pkl
    `-- Top4_Elite_Q25_rice_512bp/
        |-- Top4_Elite_Q25_llrs_rank0.pkl
        |-- Top4_Elite_Q25_meta.pkl
        `-- Top4_Elite_Q25_variant_scores.parquet

Input variants

input_variants/variants.parquet contains 21,830,339 SNVs and these columns:

Column	Meaning
chrom	Chromosome identifier, stored as 1-12
pos	One-based genomic position
ref	Reference allele
alt	Alternative allele
AC	ALT allele count
AN	Exact callable allele number for that site
AF	ALT allele frequency
consequence	Available consequence label; unknown where unavailable

Use MAC = min(AC, AN - AC) and MAF = MAC / AN. Do not infer a constant sample size when exact AN is available.

Primary table: result_all.parquet

unmask_llr/result_all.parquet is the recommended score entry point. Its 21,830,339 rows are positionally and key-wise aligned with input_variants/variants.parquet.

Variant and orientation columns

• chrom, pos, ref, alt, consequence
• AC_original, AF_original
• is_folded, AC_minor, MAF

The cached is_folded, AC_minor, and MAF fields reflect the earlier AF-based orientation. For the final exact-AN benchmark, compute MAC/MAF from the input table and apply the fold_correction shown in Quick start. Only 1,320 rows require this correction, but it is necessary for exact reproducibility.

Model score columns

• GPN_llr
• PlantCaduceus_l20_llr, PlantCaduceus_l24_llr, PlantCaduceus_l28_llr, PlantCaduceus_l32_llr
• PlantCAD2_Small_llr, PlantCAD2_Medium_llr, PlantCAD2_Large_llr
• NTv3_650M_llr
• PlantBiMoE_llr
• COD_PLANTCAD_llr (zero-shot COD PlantCAD)
• COD_FT_step12k_llr (rice fine-tuned COD PlantCAD)
• Top4_Elite_Q25
• phyloP, phastCons

The analysis used the lower 0.1% tail as stronger predicted effects after minor-allele orientation. phyloP and phastCons are site-level conservation scores and are not sign-flipped.

Raw genome-wide LLR matrices

The raw PKL files store REF-to-candidate-base scores in nucleotide order A, C, G, T:

LLR(base) = log P(base | reference context)
          - log P(reference base | reference context)

The reference-base column is zero up to numerical precision. More negative ALT values indicate that the alternative allele is less supported than the reference allele in that context.

The PKL payloads use segment IDs such as Chr1_0 and contain an LLR array plus segment metadata. Model window sizes are:

Models	Window
GPN	512 bp
PlantCaduceus l20/l24/l28/l32	512 bp
COD PlantCAD zero-shot and rice fine-tuned	512 bp
PlantCAD2 Small/Medium/Large	8192 bp
NTv3 650M pretrained	8192 bp
PlantBiMoE	8192 bp

Python pickle can execute code during loading. Only unpickle files obtained from a trusted, verified snapshot, and load one model at a time because each matrix is approximately 3 GB.

Top4 Q25 ensemble

The selected ensemble combines:

• COD PlantCAD zero-shot
• PlantCaduceus l32
• PlantCaduceus l28
• PlantCAD2 Large

The PlantCAD2 8192 bp source is normalized to 512 bp segments, then the 0.25 quantile is computed across the four model LLRs at each genomic position and candidate nucleotide. The directory contains:

• the full-genome 512 bp LLR matrix;
• generation and validation metadata;
• keyed scores for all 21,830,339 input SNVs.

In the exact-AN singleton benchmark (MAC = 1 versus MAF >= 0.05), the Top4 Q25 odds ratio is 11.204943925.

Reference genome

reference_genome/osa1_r7.asm.chrs.fa contains the 12 nuclear chromosomes of the MSU Rice Genome Annotation Project Release 7 / IRGSP-1.0 assembly. FASTA headers are Chr1 through Chr12; the Parquet tables store chromosome values without the Chr prefix. The .fai file provides byte offsets for indexed access.

Conservation tracks

Osj_PhyloP.bw and Osj_PhastCons.bw are indexed BigWig tracks used to attach site-level conservation scores. The keyed values used by the benchmark are also included directly in result_all.parquet.

Reproducibility notes

• Parquet positions are one-based; PKL segment starts are zero-based.
• Raw LLR nucleotide order is always A, C, G, T.
• Use exact per-site AN from input_variants/variants.parquet.
• Join or validate tables by chrom, pos, ref, and alt.
• Use the exact-AN orientation correction before benchmarking allele-specific score columns.
• Do not sign-flip phyloP or phastCons.
• Keep model window construction fixed when comparing raw matrices.

Associated code

The accompanying analysis notebook and genome-wide inference, ensemble, and distillation scripts are maintained in the MTEDF GitHub experiment repository under experiments/rice_vep.

License and attribution

This repository combines project-generated model outputs with third-party reference, population-variation, and conservation resources. No single permissive license is asserted for every component; the Dataset Card therefore uses license: other.

Users must review and comply with the original providers' licenses, citation requirements, and redistribution terms for the MSU7/IRGSP-1.0 reference, population variant source, phyloP/phastCons tracks, and each upstream model. When using COD PlantCAD predictions, cite the associated COD PlantCAD release and the upstream model publications as appropriate.