A Phylogenetic Approach to Genomic Language Modeling
2025 · preprint version in PMC · Preprint
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.
- b2-phylogpn-2025: status: Review status superseded (Table 1, row '3’ UTR', column 'PhyloGPN')
- ClinVar 3-prime UTR variant classification: attributes.benchmark_research.primary_sources: Benchmark research: primary sources ["phylogpn-2025"] (data/entities/sources.jsonl, records phylogpn-2025 and evidence-expansion-p2-phylogpn-2025-807f3a26cbfa, both with artifact_sha256 807f3a26cbfa9b5ce238d92164bd523302c67d1c5794b08273c51cca1acd4224)
- ClinVar 3-prime UTR variant classification: source_ids: Sources ["phylogpn-2025"] (data/entities/sources.jsonl, records phylogpn-2025 and evidence-expansion-p2-phylogpn-2025-807f3a26cbfa, both with artifact_sha256 807f3a26cbfa9b5ce238d92164bd523302c67d1c5794b08273c51cca1acd4224)
- evidence-expansion-p2-phylogpn-2025-807f3a26cbfa: status: Review status superseded (Both source records carry artifact_sha256 807f3a26cbfa9b5ce238d92164bd523302c67d1c5794b08273c51cca1acd4224)
- Licence of the pinned PhyloGPN preprint artifact: Licence CC BY-NC-ND 4.0 (Permissions statement of the PMC deposit of arXiv:2503.03773v1)
- PhyloGPN access: Access Model weights at https://huggingface.co/songlab/phylogpn and the code for the likelihood-based evaluation at https://github.com/songlab-cal/gpn. Neither is pinned to a revision in the paper. (Footnotes, Code and model availability)
- PhyloGPN architecture: Architecture Reverse-complement-equivariant dilated convolutional network adapted from CARP, itself adapted from ByteNet, with 40 residual blocks, a 481 bp receptive field and about 83M stored parameters. Reverse-complement weight tying leaves about half of the weights and biases free, plus the layer-normalisation parameters. (Methods, Architecture and Training (paragraph 1))
- PhyloGPN embeddings: Embeddings A 960-dimensional embedding of any 481 bp DNA sequence, taken from one named layer. Embeddings from other layers were not tried. (Methods, Embeddings (paragraph 1))
- PhyloGPN inference inputs: Inference inputs A single DNA sequence. The whole-genome alignment and the phylogeny enter the training loss only and are not required to make a prediction. (Abstract; Discussion and Conclusion (paragraph 1))
- PhyloGPN reported limitation: Reported limitation The authors attribute PhyloGPN coming up short of GPN-MSA on coding variant pathogenicity to its training alignment holding many mammals but fewer distantly related species, and state they did not harmonise the more diverse vertebrate alignment with their mammalian one. (Discussion and Conclusion (paragraph 4))
- PhyloGPN training compute: Training compute Eighteen epochs on four NVIDIA A100 GPUs with AdamW, a fixed learning rate of 1e-5 and no weight decay; 12 alignment blocks of about 10 kb per batch, with autosomal positions sampled four times per epoch, X chromosome positions three times and Y chromosome positions once. (Methods, Architecture and Training (paragraphs 2-3))
- PhyloGPN training data: Training data Length-481 windows of the human reference genome GRCh38, each paired with the nucleotides aligned to its central position in a whole-genome alignment of 447 placental mammal genomes from the Zoonomia Consortium, and with the minimum spanning tree of a species-level phylogeny for the species aligned at that position. Positions outside the primary assembly were excluded. (Methods, Training Data (paragraphs 1-2); Methods, Architecture and Training (paragraph 2))
- PhyloGPN training objective: Training objective A stable upper bound of a phylogenetic F81 likelihood for the aligned column at the central position, conditioned on the state of the human reference nucleotide so the model cannot simply copy it. (Methods, Loss Function (Eqs. 3 to 5))
- phylogpn-2025: attributes.evidence_concerns: Evidence concerns One scoped entry naming phylogpn-20261116-result-t2-chromatin-gena-lm: Table 2's GENA-LM Chromatin Accessibility cell, printed 0.84, is above every GENA-LM value BEND prints in its main results table, so it is withheld from automatic comparison. The printed value is kept. (Table 2, row 'GENA-LM', column 'Chromatin Accessibility (AUROC)')
- phylogpn-2025: attributes.licence: Licence CC BY-NC-ND 4.0 (Permissions statement of the PMC deposit of arXiv:2503.03773v1, in the pinned full text XML)
- phylogpn-2025: attributes.version_note: Version note Published as a peer-reviewed conference paper: Albors C., Li J.C., Benegas G., Ye C., Song Y.S., A Phylogenetic Approach to Genomic Language Modeling, Research in Computational Molecular Biology (RECOMB 2025), Lecture Notes in Computer Science (LNBI) volume 15647, pages 99-117, first online 25 April 2025, doi 10.1007/978-3-031-90252-9_7. The pinned artifact remains the preprint, so publication_status stays 'preprint'; the version of record is paywalled and was not compared with the preprint. (Springer chapter page citation metadata (series, volume, citation_firstpage, citation_lastpage, First Online); Crossref 10.1007/978-3-031-90252-9_7; Europe PMC core record for PMCID:PMC11908359)
- phylogpn-2025: links: Links One link: supersedes evidence-expansion-p2-phylogpn-2025-807f3a26cbfa (Both source records carry artifact_sha256 807f3a26cbfa9b5ce238d92164bd523302c67d1c5794b08273c51cca1acd4224 and the same Europe PMC fullTextXML artifact URL)
- phylogpn-2025: status: Review status source_checked (Pinned full text XML re-retrieved 2026-11-16; SHA-256 807f3a26cbfa9b5ce238d92164bd523302c67d1c5794b08273c51cca1acd4224)
- Reported AUROC for PhyloGPN: Printed value 0.94 (Table 1, 3-prime UTR row, PhyloGPN AUROC column)
- Version of record of the PhyloGPN paper: Version note Published as a peer-reviewed conference paper: Albors C., Li J.C., Benegas G., Ye C., Song Y.S., A Phylogenetic Approach to Genomic Language Modeling, Research in Computational Molecular Biology (RECOMB 2025), Lecture Notes in Computer Science (LNBI) volume 15647, pages 99-117, first online 25 April 2025, doi 10.1007/978-3-031-90252-9_7. The pinned artifact remains the preprint, whose publication_status is correctly 'preprint'; the version of record is paywalled and was not compared with the preprint. (Springer chapter page header and abstract; songlab-cal/gpn README model-families table and BibTeX entry albors2025phylogenetic)
Models, benchmarks and other records
Configurations (14)
- Caduceus-PS 131k, 256 model dimensions, 16 layers
- Caduceus, largest reverse-complement-equivariant model, frozen embeddings
- DNABERT, best model of the family on each BEND task
- GENA-LM, best model of the family on each BEND task
- GPN-MSA, allele likelihood ratios on variant-effect tasks
- GPN-MSA, concatenated 128 bp window embeddings
- GPN-MSA, published precomputed hg38 likelihood ratios
- HyenaDNA medium 160k, likelihood-ratio scoring
- HyenaDNA, best model of the family on each BEND task
- Nucleotide Transformer v2 500M multi-species, likelihood-ratio scoring
- Nucleotide Transformer, best model of the family on each BEND task
- PhyloGPN
- PhyloGPN (481 bp receptive field, about 83M parameters)
- PhyloGPN-X (pooled embeddings, 6,001 bp effective receptive field)
Tasks (1)
Protocols (1)
Datasets (2)
Dataset subsets (12)
- ClinVar 3’ UTR variants, at least one review star
- ClinVar 5’ UTR variants, at least one review star
- ClinVar Downstream of Transcript variants, at least one review star
- ClinVar Intronic variants, at least one review star
- ClinVar Missense variants, at least one review star
- ClinVar Noncoding variants, at least one review star
- ClinVar Splice Acceptor variants, at least one review star
- ClinVar Splice Donor variants, at least one review star
- ClinVar Start Codon variants, at least one review star
- ClinVar Stop Lost variants, at least one review star
- ClinVar Synonymous variants, at least one review star
- ClinVar Upstream of Transcript variants, at least one review star
Evaluations (125)
- Augustus on BEND Gene Finding
- Basset on BEND Chromatin Accessibility
- Basset on BEND CpG Methylation
- Basset on BEND Histone Modification
- Caduceus on BEND Chromatin Accessibility
- Caduceus on BEND CpG Methylation
- Caduceus on BEND Disease VEP
- Caduceus on BEND Enhancer Annotation
- Caduceus on BEND Expression VEP
- Caduceus on BEND Gene Finding
- Caduceus on BEND Histone Modification
- Caduceus on ClinVar 3’ UTR variants, likelihood-ratio scoring
- Caduceus on ClinVar 5’ UTR variants, likelihood-ratio scoring
- Caduceus on ClinVar Downstream of Transcript variants, likelihood-ratio scoring
- Caduceus on ClinVar Intronic variants, likelihood-ratio scoring
- Caduceus on ClinVar Missense variants, likelihood-ratio scoring
- Caduceus on ClinVar Noncoding variants, likelihood-ratio scoring
- Caduceus on ClinVar Splice Acceptor variants, likelihood-ratio scoring
- Caduceus on ClinVar Splice Donor variants, likelihood-ratio scoring
- Caduceus on ClinVar Start Codon variants, likelihood-ratio scoring
- Caduceus on ClinVar Stop Lost variants, likelihood-ratio scoring
- Caduceus on ClinVar Synonymous variants, likelihood-ratio scoring
- Caduceus on ClinVar Upstream of Transcript variants, likelihood-ratio scoring
- DeepSEA on BEND Disease VEP
- DeepSEA on BEND Expression VEP
Sources (2)
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.
25 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
| Artifact retrieved at 2026-09-16T10:38:57.558218+00:00 Source metadata | A Phylogenetic Approach to Genomic Language Modeling No field-specific location recorded Version: preprint version in PMC | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Artifact SHA-256 807f3a26cbfa9b5ce238d92164bd523302c67d1c5794b08273c51cca1acd4224 Source metadata | A Phylogenetic Approach to Genomic Language Modeling No field-specific location recorded Version: preprint version in PMC | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Artifact url https://www.ebi.ac.uk/europepmc/webservices/rest/PMC11908359/fullTextXML Source metadata | A Phylogenetic Approach to Genomic Language Modeling No field-specific location recorded Version: preprint version in PMC | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Missing metadata: doi: reason unextracted Source metadata | A Phylogenetic Approach to Genomic Language Modeling No field-specific location recorded Version: preprint version in PMC | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Publication status preprint Source metadata | A Phylogenetic Approach to Genomic Language Modeling No field-specific location recorded Version: preprint version in PMC | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Retrieved at 2026-09-15T23:33:26Z Source metadata | A Phylogenetic Approach to Genomic Language Modeling No field-specific location recorded Version: preprint version in PMC | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Review: artifact sha256 807f3a26cbfa9b5ce238d92164bd523302c67d1c5794b08273c51cca1acd4224 Source metadata | A Phylogenetic Approach to Genomic Language Modeling No field-specific location recorded Version: preprint version in PMC | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Review: date 2026-11-16 Source metadata | A Phylogenetic Approach to Genomic Language Modeling No field-specific location recorded Version: preprint version in PMC | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Review: method note Independent review of the PhyloGPN batch. The artifact was re-retrieved from the recorded artifact_url on 2026-11-16 and its SHA-256 matched the stored value byte for byte. Its two tables were re-read cell by cell, its Methods, Results, Discussion, permissions statement and footnotes were read for the descriptive claims, and the Europe PMC core record for PMCID:PMC11908359 was read for the licence and the published-version metadata. Source metadata | A Phylogenetic Approach to Genomic Language Modeling No field-specific location recorded Version: preprint version in PMC | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
Review method3 values
| A Phylogenetic Approach to Genomic Language Modeling No field-specific location recorded Version: preprint version in PMC | catalogued No individual claim review recorded Source has a recorded evidence concern. Consult its source page before using the claim. Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
Sources and history
Release 2026-11-21-adc15377fab2 · Record review: source checked
Primary paper retained with its original identifier. Metadata inherited from the literature collection; individual result checks are separate.
Release history
No supporting source is linked yet.
Supersedes A Phylogenetic Approach to Genomic Language Modeling
Download this release (gzip)Technical metadata and extraction receipts
Stable ID: phylogpn-2025
- areas
- dna-genomes
- url
- https://pmc.ncbi.nlm.nih.gov/articles/PMC11908359/
- version
- preprint version in PMC
- retrieved at
- 2026-09-15T23:33:26Z
- publication status
- preprint
- year
- 2025
- artifact sha256
- 807f3a26cbfa9b5ce238d92164bd523302c67d1c5794b08273c51cca1acd4224
- artifact url
- https://www.ebi.ac.uk/europepmc/webservices/rest/PMC11908359/fullTextXML
- artifact retrieved at
- 2026-09-16T10:38:57.558218+00:00
- scope decision
- included
- missing metadata
- doi: reason: unextracted
- licence
- CC BY-NC-ND 4.0
- version note
- Published as a peer-reviewed conference paper: Albors C., Li J.C., Benegas G., Ye C., Song Y.S., A Phylogenetic Approach to Genomic Language Modeling, Research in Computational Molecular Biology (RECOMB 2025), Lecture Notes in Computer Science (LNBI) volume 15647, pages 99-117, first online 25 April 2025, doi 10.1007/978-3-031-90252-9_7. The pinned artifact remains the preprint, so publication_status stays 'preprint'; the version of record is paywalled and was not compared with the preprint.
- review
- method: automated-source-review; independent-cell-check; source-hash-verification; method note: Independent review of the PhyloGPN batch. The artifact was re-retrieved from the recorded artifact_url on 2026-11-16 and its SHA-256 matched the stored value byte for byte. Its two tables were re-read cell by cell, its Methods, Results, Discussion, permissions statement and footnotes were read for the descriptive claims, and the Europe PMC core record for PMCID:PMC11908359 was read for the licence and the published-version metadata.; reviewer: claude; reviewer note: Claude review agent (independent of the extractor); no human review claimed. Nothing was reproduced and no value was recomputed.; date: 2026-11-16; artifact sha256: 807f3a26cbfa9b5ce238d92164bd523302c67d1c5794b08273c51cca1acd4224; retrieval url: https://www.ebi.ac.uk/europepmc/webservices/rest/PMC11908359/fullTextXML; note: Reviewed with data/omics/phylogpn-20261116 and recorded in docs/reviews/2026-11-16-phylogpn.md. publication_status stays 'preprint': the pinned artifact is the preprint, and the version of record is named in version_note and pinned by phylogpn-20261116-source-springer-chapter.
- evidence concerns
- applies to: phylogpn-20261116-result-t2-chromatin-gena-lm; artifact sha256: 807f3a26cbfa9b5ce238d92164bd523302c67d1c5794b08273c51cca1acd4224; message: Table 2's GENA-LM Chromatin Accessibility cell prints 0.84. The caption's rule is 'We report the best result among models in each family' and Section 4.2 says the row is reproduced from Marin et al. (BEND), but BEND's main results table prints 0.76 for GENA-LM BERT and 0.82 for GENA-LM BigBird in both the ICLR 2024 proceedings PDF pinned as evidence-expansion-bend-final-f709b6be and the arXiv rendition arXiv:2311.12570 retrieved on 2026-11-16. No subset of the GENA-LM family reproduces all seven printed cells of the row, while the DNABERT, HyenaDNA and Nucleotide Transformer rows each have one that does, so the printed value is not the family best this paper says it reports and its origin is unknown. The printed value is kept as published; this entry withholds that one result from automatic comparison. The row's other six cells and the rest of the source are not covered.; review method: independent-cell-check; review date: 2026-11-16; source id: phylogpn-2025; source locator: Table 2, row 'GENA-LM', column 'Chromatin Accessibility (AUROC)'
Related records
- supersedes: A Phylogenetic Approach to Genomic Language Modeling
- subject: phylogpn-2025: attributes.evidence_concerns
- subject: phylogpn-2025: attributes.licence
- subject: phylogpn-2025: status
- subject: phylogpn-2025: links
- subject: phylogpn-2025: attributes.version_note
- subject: Licence of the pinned PhyloGPN preprint artifact
- subject: Version of record of the PhyloGPN paper
- companion of: Springer chapter page for the RECOMB 2025 version of record
- companion of: Supplement 1 of the PhyloGPN preprint (arXiv:2503.03773v1)