Ye et al. 2026 GPN-Star, Nature Supplementary Information
Nature · 2026 · Supplementary Information PDF (MOESM1) of the version of record, 57 pages · Peer-reviewed
What this source contributed
Records in this release that cite this source. A record can cite several sources.
Results
Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.
Claims
Statements this source supports, each with where in the source it was checked.
- architecture_hyperparameters: GPN-Star: Architecture hyperparameters GPN-Star is an encoder-only masked language model whose encoder blocks apply sequence-wise self-attention, phylogeny-informed cross-attention and a feed-forward network. Supplementary Table 10 gives the three sizes as 25M, 85M and 200M parameters with 8, 12 and 16 encoder blocks, hidden dimension 512, 768 and 1024, intermediate size 2048, 3072 and 4096, 4, 6 and 8 self-attention heads, the same number of cross-attention heads, attention head dimension 64 and dropout rate 0.1 throughout. (Supplementary Information, Model Architecture (PDF pages 4-7) and Supplementary Table 10 (PDF page 27))
- benchmark_dataset_provenance: gpn-star-20261116-benchmark-gpn-star-vep: Benchmark dataset provenance The OMIM benchmark uses the processed Mendelian traits dataset from TraitGym and the non-coding fine-mapping benchmark uses the processed complex traits dataset from TraitGym, both with nine controls matched per positive on chromosome, molecular consequence and distance to the transcription start site, and for complex traits also on LD score and minor allele frequency. HGMD was processed in the same way as the TraitGym Mendelian traits dataset. Coding fine-mapped variants come from the same fine-mapping source but are matched on minor allele frequency only. The paper does not name a TraitGym revision. (Supplementary Information, Evaluation Benchmarks, Homo sapiens, Datasets (PDF pages 11-12))
- differences_from_gpn_msa: GPN-Star: Differences from gpn msa GPN-Star differs from GPN-MSA in three ways the authors name: it masks and predicts nucleotides across many species rather than the human genome alone; it encodes phylogenetic relationships explicitly through its attention modules; and it accepts alignments of any composition and size, so it does not need the manual species curation that GPN-MSA required, such as excluding species closely related to human. (Supplementary Information, 'Key differences between GPN-MSA and GPN-Star' (PDF page 2))
- mutation_rate_calibration: GPN-Star: Mutation rate calibration GPN-Star raw scores are calibrated to remove the effect of local mutation rate variation. The paper reports that the calibration substantially reduced correlations with Roulette mutation rate estimates while improving performance on most downstream benchmarks, and that the uncalibrated raw scores were used only for the head-to-head comparison with GPN-MSA in Supplementary Fig. 20. (Results, genome-wide evolutionary constraints (PDF page 7); Supplementary Information, Mutation Rate Calibration (PDF page 10) and Supplementary Fig. 20 caption)
- printed_values_scope: gpn-star-20261116-benchmark-gpn-star-vep: Printed values scope Of this paper's evaluations, four print numbers: Supplementary Table 3 (ablation studies), the bar labels of Fig. 4b, the bar labels of Fig. 5a for five non-human species, and the rare variant association gene counts in the Results. The model comparisons in Figs. 2 and 3, Fig. 4a,c,d, Fig. 5b-d, Extended Data Figs. 1-10 and Supplementary Figs. 1-27 are plots whose values are not printed, so the paper's headline comparisons against CADD, GPN-MSA, Evo 2, Nucleotide Transformer, ESM-1b, Enformer, Borzoi, AlphaGenome, AlphaMissense, PrimateAI-3D, PromoterAI, SpeciesLM, GPN-Promoter, PhyloP and PhastCons cannot be extracted as values from this source. One statistic inside a panel is an exception to this and was missed at extraction: Fig. 3e prints r = 0.47 and P = 0.0072 for the correlation between trait polygenicity and the enrichment difference between GPN-Star (P) and (M). It is a property of the models' relationship to each other rather than a score against a ground truth, so no result is recorded for it. (Figs. 2-5 and their captions; Extended Data Figs. 1-10; Supplementary Information, Supplementary Figs. 1-27)
- prioritized_variant_consequences: GPN-Star: Prioritized variant consequences Among the top 0.1% most constrained common variants by GPN-Star (P), missense variants are the largest class at 33% and 170-fold enriched, followed by variants in distal enhancer-like regions at 26% and 2.4-fold enriched and their flanks at 8% and 0.35-fold depleted. Compared with GPN-Star (M), GPN-Star (P) is enriched 1.14-fold in missense variants; compared with GPN-Star (V) it is enriched 1.52-fold in missense and 3.89-fold in synonymous variants and depleted in non-exonic classes such as intergenic. (Results, relevant evolutionary timescales (PDF page 5); Supplementary Information, relevant evolutionary timescales (PDF page 3) with Supplementary Tables 5-8 (PDF pages 24-26))
- released_checkpoints: GPN-Star: Released checkpoints The authors' Hugging Face collection listed eight model repositories on 2026-11-16: songlab/gpn-star-hg38-v100-200m, songlab/gpn-star-hg38-m447-200m, songlab/gpn-star-hg38-p243-200m, songlab/gpn-star-mm39-v35-85m, songlab/gpn-star-galGal6-v77-85m, songlab/gpn-star-dm6-i124-85m, songlab/gpn-star-ce11-n135-25m and songlab/gpn-star-tair10-b18-25m, alongside the training alignments, the benchmark datasets and the precomputed score set songlab/gpn-star-scores. None of the repositories was gated or private. Supplementary Table 1 describes 19 training setups, so not every trained model is released, and the paper does not map checkpoint names to repository ids. (items[] of the collection API response; Supplementary Information, Supplementary Table 1 (PDF page 21))
- training_data: GPN-Star: Training data The three human models are trained on whole-genome alignments at three timescales: multiz100way (100 vertebrates), cactus447way restricted to 447 mammals, and cactus447way restricted to 243 primates, all on hg38. Chromosomes 1-20, X and Y are training data, chromosome 21 is validation and chromosome 22 is a held-out test chromosome. Five further models are trained for Mus musculus (mm39, multiz35way), Gallus gallus (galGal6, multiz77way), Drosophila melanogaster (dm6, multiz124way), Caenorhabditis elegans (ce11, multiz135way) and Arabidopsis thaliana (tair10, multiz18way). (Supplementary Information, Supplementary Table 1 (PDF page 21) and Training Data (PDF pages 7-8))
- version: GPN-Star: Version The paper names no model release version. Its code is the gpn package at v0.8 (github.com/songlab-cal/gpn), archived at doi:10.5281/zenodo.21501177 and available under the MIT licence. Model checkpoints are named by assembly, alignment and parameter count (Supplementary Table 1); the three human models are called GPN-Star (V), (M) and (P) for the vertebrate, mammal and primate timescales. (Reporting Summary, Software and code, 'Data analysis' (PDF page 23); Code availability (PDF page 10); Results (PDF page 3); Supplementary Information, Supplementary Table 1 (PDF page 21))
Models, benchmarks and other records
Methods (1)
Configurations (39)
- CADD (GPN-Star Fig. 4b)
- GPN-MSA (GPN-Star Fig. 4b)
- GPN-Star (M) (GPN-Star Fig. 4b)
- GPN-Star (P) (GPN-Star Fig. 4b)
- GPN-Star (V) (GPN-Star Fig. 4b)
- GPN-Star ablation 'GPN-MSA (exclude 10 closest species)' (Supplementary Table 3)
- GPN-Star ablation 'GPN-MSA (no species curation)' (Supplementary Table 3)
- GPN-Star ablation 'GPN-Star default' (Supplementary Table 3)
- GPN-Star ablation 'MSA Transformer-like architecture' (Supplementary Table 3)
- GPN-Star ablation 'no evol. distance encoding' (Supplementary Table 3)
- GPN-Star ablation 'no pooling into clades' (Supplementary Table 3)
- GPN-Star ablation 'no species tree (no evol. distances, no clades)' (Supplementary Table 3)
- GPN-Star ablation 'only use one target species (human)' (Supplementary Table 3)
- GPN-Star ce11-25m
- GPN-Star ce11-85m
- GPN-Star dm6-25m
- GPN-Star dm6-85m
- GPN-Star galGal6-25m
- GPN-Star galGal6-85m
- GPN-Star hg38-mammal-25m
- GPN-Star hg38-mammal-85m
- GPN-Star hg38-mammal-200m
- GPN-Star hg38-primate-25m
- GPN-Star hg38-primate-85m
- GPN-Star hg38-primate-200m
Benchmarks (1)
Protocols (8)
- ClinVar pathogenic vs benign missense, AUPRC (GPN-Star Supplementary Table 3)
- COSMIC frequent vs gnomAD common missense, AUPRC (GPN-Star Supplementary Table 3)
- gnomAD rare versus common enrichment in the deleterious tail, odds ratio (GPN-Star Fig. 4b)
- HGMD pathogenic vs gnomAD common regulatory variants, AUPRC (GPN-Star Supplementary Table 3)
- OMIM pathogenic vs gnomAD common regulatory variants, AUPRC (GPN-Star Supplementary Table 3)
- ProteinGym human deep mutational scanning, mean Spearman (GPN-Star Supplementary Table 3)
- UK Biobank fine-mapped coding variants, AUPRC (GPN-Star Supplementary Table 3)
- UK Biobank fine-mapped non-coding variants, AUPRC (GPN-Star Supplementary Table 3)
Datasets (6)
- ClinVar pathogenic vs benign missense variants (GPN-Star paper)
- gnomAD rare versus common variants, consequence-balanced (GPN-Star paper)
- HGMD pathogenic vs gnomAD common regulatory variants, nine matched controls (GPN-Star paper)
- ProteinGym v0.1 human deep mutational scanning assays in genomic coordinates (GPN-Star paper)
- UK Biobank fine-mapped coding variants, MAF-matched controls (GPN-Star paper)
- UK Biobank whole-exome sequencing, 34 quantitative traits (GPN-Star paper)
Evaluations (68)
- CADD on gnomAD rare versus common enrichment (GPN-Star Fig. 4b)
- GPN-MSA (exclude 10 closest species) on ClinVar (GPN-Star Supplementary Table 3)
- GPN-MSA (exclude 10 closest species) on COSMIC (GPN-Star Supplementary Table 3)
- GPN-MSA (exclude 10 closest species) on Finemapped coding (GPN-Star Supplementary Table 3)
- GPN-MSA (exclude 10 closest species) on Finemapped non-coding (GPN-Star Supplementary Table 3)
- GPN-MSA (exclude 10 closest species) on HGMD (GPN-Star Supplementary Table 3)
- GPN-MSA (exclude 10 closest species) on OMIM (GPN-Star Supplementary Table 3)
- GPN-MSA (exclude 10 closest species) on ProteinGym (GPN-Star Supplementary Table 3)
- GPN-MSA (no species curation) on ClinVar (GPN-Star Supplementary Table 3)
- GPN-MSA (no species curation) on COSMIC (GPN-Star Supplementary Table 3)
- GPN-MSA (no species curation) on Finemapped coding (GPN-Star Supplementary Table 3)
- GPN-MSA (no species curation) on Finemapped non-coding (GPN-Star Supplementary Table 3)
- GPN-MSA (no species curation) on HGMD (GPN-Star Supplementary Table 3)
- GPN-MSA (no species curation) on OMIM (GPN-Star Supplementary Table 3)
- GPN-MSA (no species curation) on ProteinGym (GPN-Star Supplementary Table 3)
- GPN-MSA on gnomAD rare versus common enrichment (GPN-Star Fig. 4b)
- GPN-Star (M) on gnomAD rare versus common enrichment (GPN-Star Fig. 4b)
- GPN-Star (P) on gnomAD rare versus common enrichment (GPN-Star Fig. 4b)
- GPN-Star (V) on gnomAD rare versus common enrichment (GPN-Star Fig. 4b)
- GPN-Star default on ClinVar (GPN-Star Supplementary Table 3)
- GPN-Star default on COSMIC (GPN-Star Supplementary Table 3)
- GPN-Star default on Finemapped coding (GPN-Star Supplementary Table 3)
- GPN-Star default on Finemapped non-coding (GPN-Star Supplementary Table 3)
- GPN-Star default on HGMD (GPN-Star Supplementary Table 3)
- GPN-Star default on OMIM (GPN-Star Supplementary Table 3)
Evidence
Source checking verifies the cited claim or transcription. It does not establish independent reproduction.
Evidence table
Inspect claims, sources and review details
Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.
One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.
31 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
| Artifact SHA-256 3f7e8f71710bb63ec5ff032e697b6ef85f0d6d74411a31865fed73a757c7fc9e Source metadata | Ye et al. 2026 GPN-Star, Nature Supplementary Information No field-specific location recorded Version: Supplementary Information PDF (MOESM1) of the version of record, 57 pages | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the supplementary PDF as downloaded; a second download in the same session reproduced the hash. |
| Artifact url https://static-content.springer.com/esm/art%3A10.1038%2Fs41586-026-11005-5/MediaObjects/41586_2026_11005_MOESM1_ESM.pdf Source metadata | Ye et al. 2026 GPN-Star, Nature Supplementary Information No field-specific location recorded Version: Supplementary Information PDF (MOESM1) of the version of record, 57 pages | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the supplementary PDF as downloaded; a second download in the same session reproduced the hash. |
| Doi 10.1038/s41586-026-11005-5 Source metadata | Ye et al. 2026 GPN-Star, Nature Supplementary Information No field-specific location recorded Version: Supplementary Information PDF (MOESM1) of the version of record, 57 pages | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the supplementary PDF as downloaded; a second download in the same session reproduced the hash. |
| Hash scope SHA-256 of the supplementary PDF as downloaded; a second download in the same session reproduced the hash. Source metadata | Ye et al. 2026 GPN-Star, Nature Supplementary Information No field-specific location recorded Version: Supplementary Information PDF (MOESM1) of the version of record, 57 pages | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the supplementary PDF as downloaded; a second download in the same session reproduced the hash. |
| Licence CC BY 4.0 (Open Access) Source metadata | Ye et al. 2026 GPN-Star, Nature Supplementary Information No field-specific location recorded Version: Supplementary Information PDF (MOESM1) of the version of record, 57 pages | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the supplementary PDF as downloaded; a second download in the same session reproduced the hash. |
Limitations2 values
| Ye et al. 2026 GPN-Star, Nature Supplementary Information No field-specific location recorded Version: Supplementary Information PDF (MOESM1) of the version of record, 57 pages | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the supplementary PDF as downloaded; a second download in the same session reproduced the hash. |
| Media type application/pdf Source metadata | Ye et al. 2026 GPN-Star, Nature Supplementary Information No field-specific location recorded Version: Supplementary Information PDF (MOESM1) of the version of record, 57 pages | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the supplementary PDF as downloaded; a second download in the same session reproduced the hash. |
| Publication status peer_reviewed Source metadata | Ye et al. 2026 GPN-Star, Nature Supplementary Information No field-specific location recorded Version: Supplementary Information PDF (MOESM1) of the version of record, 57 pages | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the supplementary PDF as downloaded; a second download in the same session reproduced the hash. |
| Retrieved at 2026-11-16T00:12:33Z Source metadata | Ye et al. 2026 GPN-Star, Nature Supplementary Information No field-specific location recorded Version: Supplementary Information PDF (MOESM1) of the version of record, 57 pages | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the supplementary PDF as downloaded; a second download in the same session reproduced the hash. |
| Review: artifact sha256 3f7e8f71710bb63ec5ff032e697b6ef85f0d6d74411a31865fed73a757c7fc9e Source metadata | Ye et al. 2026 GPN-Star, Nature Supplementary Information No field-specific location recorded Version: Supplementary Information PDF (MOESM1) of the version of record, 57 pages | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the supplementary PDF as downloaded; a second download in the same session reproduced the hash. |
Sources and history
Release 2026-11-21-adc15377fab2 · Record review: source checked
Supplementary Information of the GPN-Star paper: supplementary methods, 27 supplementary figures and 11 supplementary tables.
Technical metadata and extraction receipts
Stable ID: gpn-star-20261116-source-nature-si
- areas
- dna-genomes
- contexts
- research; clinical_research
- artifact sha256
- 3f7e8f71710bb63ec5ff032e697b6ef85f0d6d74411a31865fed73a757c7fc9e
- artifact url
- https://static-content.springer.com/esm/art%3A10.1038%2Fs41586-026-11005-5/MediaObjects/41586_2026_11005_MOESM1_ESM.pdf
- doi
- 10.1038/s41586-026-11005-5
- hash scope
- SHA-256 of the supplementary PDF as downloaded; a second download in the same session reproduced the hash.
- licence
- CC BY 4.0 (Open Access)
- limitations
- Supplementary Figs. 1-27 print no numerical values.; Supplementary Tables 5-8 (variant consequence distributions) were not extracted as results; see notes.md.
- media type
- application/pdf
- publication status
- peer_reviewed
- retrieved at
- 2026-11-16T00:12:33Z
- review
- artifact sha256: 3f7e8f71710bb63ec5ff032e697b6ef85f0d6d74411a31865fed73a757c7fc9e; method: source-hash-verification; method note: Independent review pass of 2026-11-16. All four pinned artifacts were re-retrieved. The Nature article PDF, the Supplementary Information PDF and the Hugging Face collection API JSON each hashed to the value on their source record. The bioRxiv PDF did not and cannot: two downloads minutes apart gave b3e4d729d3363c16fb6b6cb99e4e18ffadb01b48614dba910480fb96f42b3d73 and 47f87f72baf93793177d3e40f70f43f03ac5d3050bbc5244f655e92041c7db4a, differing from each other in 34 bytes, all of them download-time stamps (XMP ModifyDate and MetadataDate, the xmpMM DocumentID and InstanceID, the document /ModDate and the second element of the trailer /ID); the page count, the CreationDate of 2025-09-21 and the pdftotext -layout text layer are identical across downloads. No value in this batch comes from the preprint. npm run batch -- check-values reproduced the extractor's figures exactly: 12 in_cell, 44 in_row, 16 locator_not_parsed, no not_in_row, no wrong_column and no overclaimed_in_row (value-check.tsv). All 72 results were then re-read by review/check_values_independent.py, written for this pass and importing nothing from extract/build.py. Supplementary Table 3 is printed rotated, so that script reads the page's word bounding boxes rather than a text dump: it recovers the seven column bands from the header words' y extents, pairing each 'Finemapped' header with whichever of 'coding' and 'non-coding' overlaps it, clusters the body words into eight row bands by x centre, and places every numeric token in the cell whose column centre is nearest, rejecting any token that is more than 45 per cent of a column pitch from its nearest column or less than 25 per cent of a pitch clear of the runner-up. Column order is therefore a measured property of the page and not an assumption, which is the check extract/build.py could not make: it asserted the column order positionally. Results are matched to cells through each source_locator's text, not through record IDs. Fig. 4b's 12 bar labels were required to be the first number printed after their series label on the same line, and the four rare variant counts were matched against the two sentences that print them, with the capture position fixing which count belongs to DeepRVAT with GPN-Star annotations and which to the original method. Result: 68 in cell, 4 in sentence, 0 failures (review/value-check-independent.tsv). Two negative controls confirm the checker is a cell check and not a row check: exchanging two ClinVar values between rows and moving a Fig. 4b label onto another series both fail (review/control-swap.tsv), and shifting a value into a neighbouring column of its own row fails and names the column that actually holds it, including across the Finemapped coding / OMIM and Finemapped coding / Finemapped non-coding boundaries (review/control-shift.tsv). Independently of both scripts, all 56 cells of Supplementary Table 3 and all 12 Fig. 4b labels were transcribed by hand from the pages rendered at 190 dpi with pdftoppm, into review/manual-transcription.tsv, and compared by review/cmp_manual.py: 68 values, 0 problems. That hand pass is what settles the column assignment the repository checker could only place at row level. The 12 descriptive claims were read against the pinned PDFs and the collection JSON.; note: Re-downloaded 2026-11-16 and matched the pinned hash byte for byte. 57 pages; Supplementary Table 3 is on PDF page 23 as the locators say.; retrieval url: https://static-content.springer.com/esm/art%3A10.1038%2Fs41586-026-11005-5/MediaObjects/41586_2026_11005_MOESM1_ESM.pdf; reviewed at: 2026-11-16T01:30:00Z; reviewer: claude; reviewer note: Claude review agent (independent of the extractor); no human review claimed. Values were checked against the pinned artifacts; nothing was reproduced and no value was recomputed.
- source locator
- Supplementary Table 1 (PDF page 21); Supplementary Table 3 (PDF page 23); Supplementary Tables 5-8 (PDF pages 24-26); Supplementary Table 10 (PDF page 27); Evaluation Benchmarks (PDF pages 11-14)
- url
- https://doi.org/10.1038/s41586-026-11005-5
- venue
- Nature
- version
- Supplementary Information PDF (MOESM1) of the version of record, 57 pages
- year
- 2026