luoyiming1991/spliceai-mane-v1.5-d500-m1-snv

SpliceAI MANE Select v1.5 GRCh38 SNV scores

Masked SpliceAI 1.3.1 scores for all three possible non-reference single- nucleotide variants at every A/C/G/T position within the union of MANE Select v1.5 transcript spans on the 24 GRCh38 primary chromosomes.

The predictions were computed with a maximum distance of 500 nucleotides (D=500) and masking enabled (M=1). The release contains 3,408,398,835 SNV records in 24 bgzip-compressed, tabix-indexed VCF files. The compressed VCFs total 34.32 GB (31.96 GiB).

This is an independent research dataset. It is not affiliated with or endorsed by Illumina. “SpliceAI” is used descriptively to identify compatibility with the SpliceAI score format.

Why this dataset was generated

The original GRCh38 precomputed SpliceAI scores used an older transcript set and genome-build liftover. Martin-Geary et al. reported transcript-selection and liftover-related discrepancies in that resource (doi:10.1101/2025.08.27.25334471). The Broad SpliceAI Lookup service provides updated calculations but is designed for small, rate-limited queries rather than genome-wide annotation. Ensembl provides MANE-based SNV scores with a smaller reported distance. This dataset provides a bulk, reproducible MANE Select v1.5 resource with D=500 and M=1. A wider search distance permits SpliceAI to report predicted splice-site changes farther from the queried variant; see Pitsava et al. (doi:10.1016/j.gim.2025.101574).

Dataset contents

data/
  spliceai-mane-v1.5-d500-m1.snv.chr1.vcf.gz
  spliceai-mane-v1.5-d500-m1.snv.chr1.vcf.gz.tbi
  ...
  spliceai-mane-v1.5-d500-m1.snv.chrY.vcf.gz
  spliceai-mane-v1.5-d500-m1.snv.chrY.vcf.gz.tbi
metadata/
  release-manifest.json
  broad-validation-summary.json
README.md
USAGE.md
LICENSE.md
NOTICE.md
CITATION.cff
SHA256SUMS

Each VCF record has one alternate allele and a SpliceAI INFO value:

ALLELE|SYMBOL|DS_AG|DS_AL|DS_DG|DS_DL|DP_AG|DP_AL|DP_DG|DP_DL

DS fields are delta scores for acceptor gain, acceptor loss, donor gain, and donor loss. DP fields are the corresponding signed positions relative to the variant. Scores are stored to two decimal places. A delta position associated with a displayed score of 0.00 should not be interpreted as evidence of an effect.

No score threshold was applied. Users can select thresholds appropriate to their intended research application.

Genome and transcript scope

• Genome assembly: GRCh38/hg38
• Contigs: 1–22, X, and Y
• Transcript set: MANE Select v1.5
• Primary-assembly MANE Select transcripts: 19,299
• Variant class: SNVs only
• Search distance: 500 nucleotides
• Masking: enabled (M=1)
• SpliceAI version: 1.3.1, five-model ensemble

The release excludes 64 MANE Select v1.5 transcripts located only on GRCh38 patch or alternate sequences. It does not contain MANE Plus Clinical-only transcripts, indels, mitochondrial variants, alternate loci, or patches.

Validation

The completed production release was audited across all 3,419 source shards:

• expected and observed records: 3,408,398,835;
• missing SpliceAI annotations: 0;
• all 24 primary chromosomes present;
• all chromosome VCFs passed BGZF, tabix, count, and checksum checks;
• deterministic production validation against the official SpliceAI 1.3.1 implementation passed;
• 120 chromosome-balanced production SNVs were compared with the Broad SpliceAI Lookup API at GRCh38, D=500, and M=1;
• comparisons performed: 120/120;
• matching MANE Select responses: 120/120;
• delta-position differences: 0;
• maximum absolute score difference: 0.005.

The Broad service reports three decimal places while this VCF stores two, so a maximum difference of 0.005 is consistent with two-decimal quantization.

Checksums and the machine-readable release manifest are included. The public VCF headers and manifests contain no execution-host paths.

Usage

See USAGE.md for download, checksum verification, tabix lookup, VCF annotation, chromosome naming, and interpretation examples.

Limitations and responsible use

SpliceAI scores are computational predictions, not measurements of RNA splicing and not classifications of pathogenicity. They should be considered with transcript relevance, phenotype, population frequency, other evidence, and—when appropriate—RNA or other functional studies.

This dataset is research material, not a validated clinical diagnostic device. It must not be the sole basis for diagnosis, treatment, or other patient-care decisions. Users are responsible for validation appropriate to their setting and for compliance with applicable professional, institutional, and regulatory requirements.

Because this resource is limited to MANE Select transcripts, clinically important effects on alternative transcripts may be absent. Masking suppresses some predictions that coincide with annotated splice gains or unannotated splice losses. GRCh38 coordinates must not be used directly with another assembly.

License

The dataset is distributed under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0). Commercial use is not permitted under this license.

The source software used to produce the data has separate GPL-3.0-or-later terms. The SpliceAI trained models and upstream materials retain their own terms and attribution. No model files, reference genome sequence, MANE source annotation, patient data, or Broad API response cache is included here.

Citation

Please cite this dataset, the generation software, and the original SpliceAI publication:

• Luo Y. SpliceAI MANE Select v1.5 GRCh38 SNV scores (D=500, masked). Dataset, version 1.0.0, 2026.
• Luo Y. Batched inference for SpliceAI. Version 0.1.0rc4. https://github.com/yimingluo-md/batched-inference-for-spliceai
• Jaganathan K, et al. Predicting Splicing from Primary Sequence with Deep Learning. Cell. 2019;176(3):535–548.e24. doi:10.1016/j.cell.2018.12.015
• Morales J, et al. A joint NCBI and EMBL-EBI transcript set for clinical genomics and research. Nature. 2022;604:310–315. doi:10.1038/s41586-022-04558-8

Reproducibility

Generation software: https://github.com/yimingluo-md/batched-inference-for-spliceai

Production source revision: a145b3c6aeee8fe020033a8a6a0e71816772e73a

The public manifest records the reference, MANE annotation, runner, container, and source-code SHA-256 identifiers used for production.

AI-assisted development disclosure

The generation software and release preparation were developed with assistance from OpenAI Codex (GPT-5.6 Sol) and Anthropic Claude Code (Claude Opus 5). AI-generated suggestions were reviewed and tested before inclusion. These tools are not project authors, and their use does not imply endorsement by OpenAI or Anthropic.