Datasetpublic

Taykhoom/mfass

MFASS Splicing Variant Effects This dataset packages 28,972 single-nucleotide variants from the Multiplexed Functional Assay of Splicing (MFASS) as one compact benchmark table. Each row contains the exact 170 bp transcript-oriented assay sequence pair, native exon-inclusion measurements, assay-relative geometry, and canonical GRCh38 locus. Of the 28,972 rows, 27,733 are evaluable and 1,050 are labeled splice-disrupting variants. Row identity: pair_id is the unique row key.… See the full description on the dataset page: https://huggingface.co/datasets/Taykhoom/mfass.

sourceHugging Faceotherupdated 28d agoView on Hugging Face

0likes88downloads

Dataset Card

MFASS Splicing Variant Effects

This dataset packages 28,972 single-nucleotide variants from the Multiplexed Functional Assay of Splicing (MFASS) as one compact benchmark table.

Each row contains the exact 170 bp transcript-oriented assay sequence pair, native exon-inclusion measurements, assay-relative geometry, and canonical GRCh38 locus. Of the 28,972 rows, 27,733 are evaluable and 1,050 are labeled splice-disrupting variants.

Row identity: pair_id is the unique row key. sequence is deliberately not unique: different tested variants in the same exon share the same reference construct. The table contains 2,199 distinct reference sequence values. Never deduplicate these rows by sequence.

Loading

python

from datasets import load_dataset

dataset = load_dataset("Taykhoom/mfass", split="test")

The Hugging Face split and the first row-level column, split, are both test for consistency with DART-Eval datasets that also contain train rows.

Sequence and target contract

sequence carries the canonical hg38 reference allele and alt_sequence carries the alternate allele. Both are the exact 170 bp constructs represented in transcript orientation; therefore, on strand == "-", the base in the sequence is the complement of the forward-genome ref or alt.

One source variant has a genuine hg19-to-hg38 reference swap. Its sequence order, inclusion measurements, and signed effects are reversed here so every row consistently means hg38 reference to alternate. Missing measurements and labels remain null.

The splice-disrupting-variant label is:

$$ \Delta\mathrm{PSI} \le -0.50 $$

where delta_psi is alternate minus reference exon inclusion.

Columns

Column	Description
`split`	Dataset partition. Always `test`; retained for DART-Eval schema consistency.
`source_index`	Zero-based row index in the pinned canonical MFASS source table.
`pair_id`	Stable unique identifier for the tested variant pair.
`component_id`	Ensembl exon identifier; use for grouped splits or bootstrap resampling.
`sequence`	Exact 170 bp transcript-oriented assay construct carrying the canonical hg38 reference allele.
`alt_sequence`	Matched 170 bp construct carrying the canonical hg38 alternate allele.
`label`	Nullable splice-disrupting label: `delta_psi <= -0.50`.
`delta_psi`	Alternate minus reference exon inclusion; the main quantitative target.
`delta_psi_rep1`	Replicate-1 alternate-minus-reference effect.
`delta_psi_rep2`	Replicate-2 alternate-minus-reference effect.
`ref_inclusion`	Exon inclusion measured for `sequence`.
`alt_inclusion`	Exon inclusion measured for `alt_sequence`.
`chrom`	GRCh38 chromosome using UCSC `chr` names.
`position`	1-based GRCh38 SNV position.
`ref`	Forward-genome GRCh38 reference allele.
`alt`	Forward-genome alternate allele.
`strand`	Corrected transcript orientation relative to GRCh38.
`variant_offset`	Zero-based focal-variant index within both 170 bp sequences.
`exon_start`	Zero-based exon start within the assay construct.
`exon_end`	Exclusive exon end within the assay construct.
`region`	Variant location: `upstream_intron`, `exon`, or `downstream_intron`.
`splice_site_offset`	Signed transcript-oriented offset from the relevant splice boundary. Upstream intron values are negative, downstream intron values are positive, and exonic values point inward from the nearest boundary.
`assay_hg38_alignment`	Whole-construct relation to the lifted GRCh38 interval: `exact`, `substitution`, or `insertion`.

Genomic context

No arbitrary fixed genomic window is stored. Use chrom, 1-based position, ref, and alt to extract whatever context length a model requires from the pinned GRCh38 no-alt assembly GCA_000001405.15.

The complete lifted assay interval was validated against each construct. Across 2,198 exons, 2,195 match exactly, two contain one assembly substitution, and one contains one additional GRCh38 base. This becomes 28,930 exact, 32 substitution, and 10 insertion variant rows.

Published baseline reproduction

Method	Rows	AUROC	Average precision
Pangolin	27,733	0.8882	0.4208
SpliceAI	27,733	0.8193	0.3208
SpliceTransformer	27,733	0.7857	0.3174
MMSplice	27,733	0.7582	0.2558
SPANR	27,663	0.7479	0.2279

The exhaustive source, score, and alignment provenance is intentionally omitted from Hugging Face. Stable processing documentation and the internal reproducibility manifest are available at:

—Processing repository: https://github.com/TaykhoomDalal/MFASS-Processing/tree/main
—Processing guide: https://github.com/TaykhoomDalal/MFASS-Processing/blob/main/README.md
—Root manifest: https://github.com/TaykhoomDalal/MFASS-Processing/blob/main/manifest.json

The Hugging Face repository is intentionally limited to this card and the compact Parquet. The 92-column full table, 8-column metrics table, root manifest, and processing code remain in the processing repository.

Limitations

—This release starts from the authors' processed MFASS table; it does not reproduce raw-read processing.
—Published baseline scores are verified and re-evaluated, not regenerated by rerunning the five models.
—The assay strings are synthetic transcript-oriented minigene constructs, not arbitrary forward-genome windows. Two exons contain one assembly substitution and one exon contains one additional GRCh38 base.
—Labels and dPSI are null for 1,239 rows. SPANR has 70 additional missing scores among otherwise evaluable rows.

Sources and citation

Rockie Chong et al. A multiplexed assay for exon recognition reveals that an unappreciated fraction of rare genetic variants cause large-effect splicing disruptions. Molecular Cell 73, 183-194.e8 (2019). https://doi.org/10.1016/j.molcel.2018.10.037

Published baseline scores:

https://github.com/brhanufen/spliceconsensus

License

Artifact-level license status is NOASSERTION; the license: other metadata does not grant a new license. The pinned MFASS repository and GEO record do not state explicit redistribution terms for the processed measurements. Users must review the original MFASS, SpliceConsensus/source-model, NCBI reference, and publication terms. See the processing `NOTICE.md` for authoritative links.

Permission for third-party redistribution of the processed MFASS measurements and derivatives in this repackaged form remains unresolved; obtain clarification or permission from the MFASS rights holders before redistribution.