DNA language models are powerful predictors of genome-wide variant effects (PNAS 2023)
Proceedings of the National Academy of Sciences of the United States of America · 2023 · PNAS 120(44) e2311219120, published 26 October 2023; Europe PMC full text JATS XML of PMC10622914 · 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.
- Code availability for the GPN paper: Code availability "Code to reproduce all results, including instructions to load the pretrained model, is available at https://github.com/songlab-cal/gpn". No checkpoint name, revision or hash is given in the paper, and sequence logo predictions for Arabidopsis thaliana are offered as a UCSC Genome Browser session at https://genome.ucsc.edu/s/gbenegas/gpn-arabidopsis. (Data, Materials, and Software Availability; Abstract)
- Correlation between GPN scores and phastCons: Score correlation phastcons r = 0.13, P < 10^-300, over the 1001 Genomes variants both methods score. The paper does not say whether r is a Pearson or a Spearman correlation. (Results, 'Benchmarking using allele frequencies in 1001 genomes', sentence "In fact, GPN scores are only weakly correlated with phyloP (r = 0.22, P < 10-300) and phastCons (r = 0.13, P < 10-300).")
- Correlation between GPN scores and phyloP: Score correlation phylop r = 0.22, P < 10^-300, over the 1001 Genomes variants both methods score. The paper does not say whether r is a Pearson or a Spearman correlation. (Results, 'Benchmarking using allele frequencies in 1001 genomes', sentence "In fact, GPN scores are only weakly correlated with phyloP (r = 0.22, P < 10-300) and phastCons (r = 0.13, P < 10-300).")
- GPN architecture as published in the PNAS paper: Architecture Masked language model over single-nucleotide tokens, not k-mers or byte-pair encoding. A 512 bp window is one-hot encoded and processed by 25 convolutional blocks, each a dilated convolution followed by a feed-forward layer with residual connections and layer normalisation, at a fixed embedding dimension of 512 and kernel size 9, with the dilation increased exponentially and then cycled. A final layer outputs four nucleotide probabilities at each masked position. Implemented with the Hugging Face library. (Figure 1 caption; Materials and Methods, 'Pretraining'; SI Appendix Table S3)
- GPN pretraining corpus and held-out chromosomes: Training data Eight RefSeq-annotated Brassicales reference genomes, one per genus, filtered from NCBI Genome: Arabidopsis thaliana, Brassica rapa, Camelina sativa, Capsella rubella, Carica papaya, Eutrema salsugineum, Raphanus sativus and Tarenaya hassleriana. Arabidopsis thaliana chromosomes 4 and 5 were held out for validation and testing. Windows of 512 bp at a step of 256 bp were sampled with the reverse complement, drawn from exons with a small intronic flank, promoters 1,000 bp upstream of transcription start sites and an equivalent amount of random windows, with each genome subset to the number of windows from Arabidopsis. (SI Appendix Table S1; Materials and Methods, 'Pretraining')
- GPN training compute as published: Training compute Each of the three models was trained for 150,000 steps, taking approximately 4 days with 4 NVIDIA A100 80 GB GPUs. (Materials and Methods, 'Pretraining', sentence "We trained each model for 150,000 steps, taking approximately 4 d with 4 NVIDIA A100 80 GB GPUs.")
- How GPN scores a variant: Scoring definition The variant position is masked and the score is the log-likelihood ratio between the alternate and the reference allele, averaged between the forward and reverse strands. During training 15% of positions are masked; at variant effect prediction only the variant position is masked. (Figure 1 caption; Materials and Methods, 'Variant Effect Prediction')
- k-mer frequency baseline for gene region classification: Baseline range Classification from k-mer frequencies alone spans 8% to 70% per class at k = 3 and 15% to 67% at k = 6. The paper prints these only as ranges over classes, not per class. (Results, 'Unsupervised Clustering of Genomic Regions', sentence beginning "This level of classification accuracy cannot be achieved merely through k-mer frequencies"; SI Appendix Figure S1)
- Promoter motifs recovered by TF-MoDISco on GPN predictions: Motif analysis Approximately a hundred and sixty motifs were identified in regions 1,000 bp either side of transcription start sites, after filtering repeats and coding exons. Four are shown in Figure 3B, the first two with a significant match in PlantTFDB at a Tomtom q-value below 0.05. (Results, 'DNA Motifs Revealed by High-Confidence Model Predictions'; Figure 3 B caption; SI Appendix Figure S5; Materials and Methods, 'Motif Analysis', sentence beginning "An adaptation of TF-MoDISco was run", for the 1,000 bp promoter window and the repeat and coding exon filter)
- Share of variants that alignment-based scores cannot reach: Coverage advantage 14.2% of all 1001 Genomes variants could not be scored by phyloP or phastCons because the whole-genome alignment failed there; GPN scores them because it needs no alignment. Where the paper compares with phyloP and phastCons, those variants were excluded. (Results, 'Benchmarking using allele frequencies in 1001 genomes', sentence beginning "A notable advantage of GPN"; Materials and Methods, 'Variant Effect Prediction')
- The GPN paper evaluates Arabidopsis thaliana only: Evaluation scope Every evaluation in this paper is in Arabidopsis thaliana: held-out sequence modelling, gene region classification from embeddings, in silico mutagenesis, 1001 Genomes rare against common enrichment and AraGWAS hit enrichment. The paper reports no human variant effect results, and no human model is trained or scored in it. Its stated motivation is that a model trained on DNA sequence alone can be applied to understudied non-model organisms. (Abstract; Results; Discussion, paragraph beginning "Here, we present an unsupervised genome-wide variant effect predictor")
- The GPN paper's model comparisons are printed only as figures: Extraction gap Every comparison of GPN against phyloP, phastCons, abs(phyloP), LD Score and their LD-weighted forms is printed as a bar chart with no values in the text layer: Figures 5D, 6D and 6E of the article and Figures S6, S8, S9, S10, S11 and S12 of the SI Appendix. The same is true of the contingency tables drawn in Figures 5C and 6C and of the confusion matrix in Figure 2B. The published article contains no table at all; the SI Appendix contains three, of which only Table S2 holds results. The comparator values are therefore a gap, not a finding. (Figures 5 C, 5 D, 6 C, 6 D, 6 E and 2 B of the article; SI Appendix Figures S6, S8, S9, S10, S11 and S12)
- The only comparator figure the GPN paper prints for Figure 6D: Comparator bound "while less than 7.5-fold enrichment was observed for other methods (Fig. 6D)". The other methods are phyloP x LD, phastCons x LD, abs(phyloP) x LD and LD Score. Figure 6D prints no per-method value in its text layer, so none is recorded. (Results, 'Enrichment of GWAS hits in regions with low GPN scores', sentence beginning "In particular, by analyzing odds ratios"; Figure 6 D caption; Materials and Methods, 'Variant Effect Prediction', sentences beginning "We similarly defined phyloP x LD" and "We considered the baseline LD Score", where the four comparator methods are defined)
- Training of the single-species GPN model: Training setting Trained on the single Arabidopsis thaliana genome with a repeat weight of 0.1 and the same hyperparameters as the main model, except for only 12,000 steps with decaying learning rate, because the authors saw it would soon start overfitting. (Materials and Methods, 'Pretraining', sentence beginning "We also trained a separate model on the single genome of A. thaliana")
- Version history of the GPN paper: Version history bioRxiv 10.1101/2022.08.22.504706 has three versions: v1 posted 2022-08-23 and v2 posted 2023-04-12, both titled "DNA language models are powerful zero-shot predictors of non-coding variant effects" (v1) and "... of genome-wide variant effects" (v2), and v3 posted 2023-08-04 under the published title "DNA language models are powerful predictors of genome-wide variant effects". All three are licensed CC BY-NC-ND and all name 10.1073/pnas.2311219120 as the published article. (bioRxiv details API response for 10.1101/2022.08.22.504706, fields title, date, version, license and published)
- Where bioRxiv v3 and the published article print different numbers: Version difference The positional perplexity sentence differs. bioRxiv v3 reports medians over the whole test chromosome for four classes: splice acceptors 1.03, splice donors 1.02, start codons 1.10, stop codons 2.81, against intergenic and intronic 3.14, all Mann-Whitney p-values < 10^-239. The published article reports medians over a 1 Mb region of the test chromosome containing 264 genes and 471 transcripts, and adds the three coding frames: splice donors 1.02, splice acceptors 1.03, start codons 1.08, CDS2 2.24, CDS1 2.44, CDS3 2.79, stop codons 2.8, against intergenic and intronic 3.24, all Mann-Whitney P-values < 10^-17. This is a different measurement region, not a corrected value. Every other printed value this batch records appears verbatim in both versions, asserted by extract/parse_values.py. (bioRxiv v3 Results, sentence beginning "Across the test chromosome, model perplexities in splice acceptors"; published article Results, sentence beginning "Across a 1-Mb region in the test chromosome")
Models, benchmarks and other records
Configurations (5)
- GPN trained on eight Brassicales genomes, repeat loss weight 0 (PNAS 2023)
- GPN trained on eight Brassicales genomes, repeat loss weight 0.1 (PNAS 2023)
- GPN trained on eight Brassicales genomes, repeat loss weight 1 (PNAS 2023)
- GPN trained on the Arabidopsis thaliana genome alone (PNAS 2023)
- GPN x LD, GPN scores weighted by linkage disequilibrium (PNAS 2023)
Benchmarks (2)
Protocols (7)
- Gene region classification from averaged GPN embeddings, leave-one-chromosome-out
- GWAS hit enrichment in the lowest 1% of scores, AraGWAS permutation threshold
- Median held-out perplexity by position class in Chr5:3,500,000-4,500,000
- Rare versus common variant enrichment in the lowest 0.1% of scores, 1001 Genomes
- Share of simulated repeat variants scored below the first decile of missense variants
- Test perplexity on Arabidopsis thaliana chromosome 5, chromosome-wide
- Test perplexity on Arabidopsis thaliana chromosome 5, non-repeat regions
Datasets (4)
Dataset subsets (4)
Evaluations (15)
- Gene region classification from averaged GPN embeddings (PNAS 2023)
- GWAS hit enrichment for GPN x LD, AraGWAS (PNAS 2023)
- Median held-out perplexity by position class, Chr5:3,500,000-4,500,000 (PNAS 2023)
- Rare versus common variant enrichment for GPN, 1001 Genomes (PNAS 2023)
- Repeat variants below the missense first decile, GPN with repeat weight 0 (PNAS 2023)
- Repeat variants below the missense first decile, GPN with repeat weight 0.1 (PNAS 2023)
- Repeat variants below the missense first decile, GPN with repeat weight 1 (PNAS 2023)
- Test perplexity, chromosome-wide, GPN trained on Arabidopsis thaliana alone (PNAS 2023)
- Test perplexity, chromosome-wide, GPN with repeat weight 0 (PNAS 2023 Table S2)
- Test perplexity, chromosome-wide, GPN with repeat weight 0.1 (PNAS 2023 Table S2)
- Test perplexity, chromosome-wide, GPN with repeat weight 1 (PNAS 2023 Table S2)
- Test perplexity, non-repeat, GPN trained on Arabidopsis thaliana alone (PNAS 2023)
- Test perplexity, non-repeat, GPN with repeat weight 0 (PNAS 2023 Table S2)
- Test perplexity, non-repeat, GPN with repeat weight 0.1 (PNAS 2023 Table S2)
- Test perplexity, non-repeat, GPN with repeat weight 1 (PNAS 2023 Table S2)
Sources (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 |
|---|---|---|
| Article licence CC BY-NC-ND 4.0 (https://creativecommons.org/licenses/by-nc-nd/4.0/), stated in the article permissions Source metadata | DNA language models are powerful predictors of genome-wide variant effects (PNAS 2023) No field-specific location recorded Version: PNAS 120(44) e2311219120, published 26 October 2023; Europe PMC full text JATS XML of PMC10622914 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Format: JATS XML |
| Artifact format JATS XML Source metadata | DNA language models are powerful predictors of genome-wide variant effects (PNAS 2023) No field-specific location recorded Version: PNAS 120(44) e2311219120, published 26 October 2023; Europe PMC full text JATS XML of PMC10622914 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Format: JATS XML |
| Artifact SHA-256 8ff2184c07f5e6807af0b9d5103d669e1920e735ab5a4b1ea856c070077ae1e0 Source metadata | DNA language models are powerful predictors of genome-wide variant effects (PNAS 2023) No field-specific location recorded Version: PNAS 120(44) e2311219120, published 26 October 2023; Europe PMC full text JATS XML of PMC10622914 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Format: JATS XML |
| Doi 10.1073/pnas.2311219120 Source metadata | DNA language models are powerful predictors of genome-wide variant effects (PNAS 2023) No field-specific location recorded Version: PNAS 120(44) e2311219120, published 26 October 2023; Europe PMC full text JATS XML of PMC10622914 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Format: JATS XML |
| Licence CC BY-NC-ND 4.0 Source metadata | DNA language models are powerful predictors of genome-wide variant effects (PNAS 2023) No field-specific location recorded Version: PNAS 120(44) e2311219120, published 26 October 2023; Europe PMC full text JATS XML of PMC10622914 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Format: JATS XML |
Limitations2 values
| DNA language models are powerful predictors of genome-wide variant effects (PNAS 2023) No field-specific location recorded Version: PNAS 120(44) e2311219120, published 26 October 2023; Europe PMC full text JATS XML of PMC10622914 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Format: JATS XML |
| Media type application/xml Source metadata | DNA language models are powerful predictors of genome-wide variant effects (PNAS 2023) No field-specific location recorded Version: PNAS 120(44) e2311219120, published 26 October 2023; Europe PMC full text JATS XML of PMC10622914 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Format: JATS XML |
| Original url https://www.pnas.org/doi/10.1073/pnas.2311219120 Source metadata | DNA language models are powerful predictors of genome-wide variant effects (PNAS 2023) No field-specific location recorded Version: PNAS 120(44) e2311219120, published 26 October 2023; Europe PMC full text JATS XML of PMC10622914 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Format: JATS XML |
| Publication status peer_reviewed Source metadata | DNA language models are powerful predictors of genome-wide variant effects (PNAS 2023) No field-specific location recorded Version: PNAS 120(44) e2311219120, published 26 October 2023; Europe PMC full text JATS XML of PMC10622914 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Format: JATS XML |
| Retrieval note www.pnas.org returned HTTP 403 to a direct request for the article PDF and for the supplementary PDF, and the PMC PDF endpoint returned a proof-of-work interstitial rather than the file, so the Europe PMC full text XML is the pinned artifact. It was fetched twice on 2026-11-16 with identical bytes. Source metadata | DNA language models are powerful predictors of genome-wide variant effects (PNAS 2023) No field-specific location recorded Version: PNAS 120(44) e2311219120, published 26 October 2023; Europe PMC full text JATS XML of PMC10622914 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Format: JATS XML |
Sources and history
Release 2026-11-21-c8bea25e5abb · Record review: source checked
Peer-reviewed article introducing GPN, a masked DNA language model trained on eight Brassicales genomes and evaluated on Arabidopsis thaliana variants.
1 source record and release history
- DNA language models are powerful predictors of genome-wide variant effects (PNAS 2023) · Original source · PNAS 120(44) e2311219120, published 26 October 2023; Europe PMC full text JATS XML of PMC10622914
Technical metadata and extraction receipts
Stable ID: gpn-paper-20261116-source-pnas-2023
- areas
- dna-genomes
- contexts
- research
- article licence
- CC BY-NC-ND 4.0 (https://creativecommons.org/licenses/by-nc-nd/4.0/), stated in the article permissions
- artifact format
- JATS XML
- artifact sha256
- 8ff2184c07f5e6807af0b9d5103d669e1920e735ab5a4b1ea856c070077ae1e0
- doi
- 10.1073/pnas.2311219120
- licence
- CC BY-NC-ND 4.0
- limitations
- The published article prints no tables in its main text. Every comparison of GPN against phyloP, phastCons, abs(phyloP) and LD Score (Figures 5D, 6D, 6E and SI Appendix Figures S6, S8 to S12) is a bar chart with no values in the text layer, so those numbers are not in this batch.; The Europe PMC rendering of the publisher's XML, not the publisher's own file; it carries the article text, figure captions and metadata but not the typeset page.
- media type
- application/xml
- original url
- https://www.pnas.org/doi/10.1073/pnas.2311219120
- publication status
- peer_reviewed
- retrieval note
- www.pnas.org returned HTTP 403 to a direct request for the article PDF and for the supplementary PDF, and the PMC PDF endpoint returned a proof-of-work interstitial rather than the file, so the Europe PMC full text XML is the pinned artifact. It was fetched twice on 2026-11-16 with identical bytes.
- retrieved at
- 2026-11-16T00:29:00Z
- review
- artifact sha256: 8ff2184c07f5e6807af0b9d5103d669e1920e735ab5a4b1ea856c070077ae1e0; date: 2026-11-16; method: source-hash-verification; automated-source-review; method note: Extraction: the published article prints no tables in its main text, so Table S2 was parsed from the pdftotext -layout text layer of the pinned SI Appendix PDF (extract/si-appendix.layout.txt) by extract/parse_values.py with all three row labels and the header asserted, and every other value was taken from the article body by matching its whole printed sentence in the pinned Europe PMC JATS XML, so a changed sentence fails the parse.; note: Fetched twice on 2026-11-16 with identical bytes. The DOI, PMID 37883436, PMC10622914, volume 120 issue 44 and the 2023-10-26 publication date were read from the article metadata in the same file.; retrieval url: https://www.ebi.ac.uk/europepmc/webservices/rest/PMC10622914/fullTextXML; reviewer: claude; reviewer note: Claude (Opus 5) extraction agent. Extraction only: no independent review is claimed, and a different agent reviews this batch.
- url
- https://www.ebi.ac.uk/europepmc/webservices/rest/PMC10622914/fullTextXML
- venue
- Proceedings of the National Academy of Sciences of the United States of America
- version
- PNAS 120(44) e2311219120, published 26 October 2023; Europe PMC full text JATS XML of PMC10622914
- year
- 2023