CoChat
Sign Up
Cited by 157Genomics and Phylogenetic StudiesGenetic Associations and EpidemiologyGenomics and Rare Diseases

DNA language models are powerful predictors of genome-wide variant effects

Open access
Benegas, Gonzalo, Batra, Sanjit Singh, Song, Yun S.·2023·Proceedings of the National Academy of Sciences
DOI 10.1073/pnas.2311219120PMC PMC10622914

Published version of a preprint

Open this paper in CoChat — free account

Abstract

The expanding catalog of genome-wide association studies (GWAS) provides biological insights across a variety of species, but identifying the causal variants behind these associations remains a significant challenge. Experimental validation is both labor-intensive and costly, highlighting the need for accurate, scalable computational methods to predict the effects of genetic variants across the entire genome. Inspired by recent progress in natural language processing, unsupervised pretraining on large protein sequence databases has proven successful in extracting complex information related to proteins. These models showcase their ability to learn variant effects in coding regions using an unsupervised approach. Expanding on this idea, we here introduce the Genomic Pre-trained Network (GPN), a model designed to learn genome-wide variant effects through unsupervised pretraining on genomic DNA sequences. Our model also successfully learns gene structure and DNA motifs without any supervision. To demonstrate its utility, we train GPN on unaligned reference genomes of Arabidopsis thaliana and seven related species within the Brassicales order and evaluate its ability to predict the functional impact of genetic variants in A. thaliana by utilizing allele frequencies from the 1001 Genomes Project and a comprehensive database of GWAS. Notably, GPN outperforms predictors based on popular conservation scores such as phyloP and phastCons. Our predictions for A. thaliana can be visualized as sequence logos in the UCSC Genome Browser ( https://genome.ucsc.edu/s/gbenegas/gpn-arabidopsis ). We provide code ( https://github.com/songlab-cal/gpn ) to train GPN for any given species using its DNA sequence alone, enabling unsupervised prediction of variant effects across the entire genome.

AI snapshot

Written from the paper’s full text.

TL;DR

The authors introduce GPN, a multispecies DNA language model trained by self-supervision on reference genomes, and show it can score genome-wide variant effects without functional genomics data. In Arabidopsis thaliana, GPN achieves state-of-the-art unsupervised variant effect prediction and improves GWAS hit enrichment when combined with linkage disequilibrium.

Key findings

  • GPN embeddings distinguish genomic regions without supervision: logistic regression on averaged embeddings reached 96% accuracy for coding sequences and 51% for ncRNA, far above k-mer baselines (k=3: 8%–70%; k=6: 15%–67%).
  • In a 1-Mb test region, GPN perplexities were lower at constrained sites—splice donors median 1.02, splice acceptors 1.03, start codons 1.08, CDS2 2.24, CDS1 2.44, CDS3 2.79, stop codons 2.8—than in intergenic/intronic regions (median 3.24; Mann–Whitney P < 10⁻¹⁷).
  • Across over 10 million SNPs from the 1001 Genomes Project, the lowest 0.1% of GPN scores showed 5.5-fold enrichment in rare variants; GPN outperformed phyloP and phastCons, despite weak correlations with phyloP (r=0.22) and phastCons (r=0.13), and could score 14.2% of variants that phyloP/phastCons could not.
  • For AraGWAS variants, the lower 1% of GPN × LD scores were 10.3-fold enriched in GWAS hits versus the upper 99%, while other methods showed less than 7.5-fold enrichment; with the full variant set (n=510,462), GPN × LD achieved even higher odds ratios.

Methods

  • GPN was pretrained on unaligned reference genomes from A. thaliana and seven related Brassicales species (8 genomes total), holding out A. thaliana chromosomes 4 and 5 for validation and testing. Training used 512-bp windows with 256-bp step, reverse-complement augmentation, and masked language modeling with 15% of nucleotides masked; the model was a convolutional neural network with 25 convolutional blocks, 512-dimensional embeddings, trained for 150,000 steps on 4 NVIDIA A100 80 GB GPUs.
  • Variant effect scores were computed as the log-likelihood ratio between alternate and reference alleles, averaged over forward and reverse strands. Analyses included in silico mutagenesis, over 10 million naturally occurring SNPs from the 1001 Genomes Project, and around half a million AraGWAS variants across 462 phenotypes.
  • GPN × LD was defined as a stratified LD score: a weighted sum of GPN scores using squared Pearson correlations (r²) within 100 kb, with Fisher’s exact test used for odds ratios and P-values.

Limitations

  • The paper states that repetitive elements remain challenging: down-weighting repeats only partially mitigates issues, and the current proof-of-concept model is considerably smaller—by 200 times—than the largest published protein language model. It also notes that incorporating multiple sequence alignments would be bottlenecked by noncoding alignment quality, and that the current method averages forward/reverse strand outputs rather than using reverse-complement equivariance.
  • The paper states that performance relative to phyloP and phastCons varies by variant class and threshold, and that GPN’s advantage is strongest at the extreme tail of the score distribution.
  • Evident from the text: evaluation is limited to A. thaliana and related Brassicales; no human or broader species validation is presented, and the single-species model performed substantially worse, indicating dependence on multispecies training data.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1073/pnas.2311219120

Cites

65 of 65 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 13 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    GPN classification accuracy cannot be achieved merely through k-mer frequencies.

    “This level of classification accuracy cannot be achieved merely through k-mer frequencies (k=3: 8% to 70%; k=6: 15% to 67%; see SI Appendix, Fig. S1).”
    Quote in the paperSubstantial claim
  • Findingown data

    GPN × LD effectively distinguishes GWAS hits from nonhits in the example locus.

    “With this approach, GPN × LD effectively distinguishes GWAS hits from nonhits in this example locus (Fig. 6 A, Bottom).”
    Quote in the paperSubstantial claim
  • Findingown data

    Putative functional SNPs, defined as the lowest 0.1% of GPN scores, exhibit a 5.5-fold enrichment in rare variants.

    “Putative functional SNPs, defined as the lowest 0.1% of GPN scores, exhibit a 5.5-fold enrichment in rare variants (Fig. 5C); see SI Appendix, Fig. S8 for different allele frequency thresholds.”
    Quote in the paperSubstantial claim
  • Findingown data

    The highest confusion in classifying genomic regions using GPN embeddings was between intergenic regions and ncRNAs.

    “As summarized in Fig. 2B, the highest confusion was observed between intergenic regions and ncRNAs; this may be partly explained by errors in ncRNA annotation, which is especially challenging given their low expression levels and poor conservation (33).”
    Quote in the paperNot a substantial claim
  • Findingown data

    GPN presents a unified approach capable of accurate unsupervised prediction of genome-wide variant effects.

    “GPN presents a unified approach capable of accurate unsupervised prediction of genome-wide variant effects.”
    Quote in the paperSubstantial claim
  • May combine two claims. 14
  • Classified as probably not a substantial claim. 22

  • May combine two claims.

    Findingown data

    GPN achieves significant odds ratios when computed only within particular variant classes, but its performance relative to phyloP and phastCons varies.

    “GPN also achieves significant odds ratios when computed only within particular variant classes, but its performance relative to phyloP and phastCons varies (SI Appendix, Fig. S9).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Moderate down-weighting results in a similar improvement in perplexity for nonrepetitive regions without sacrificing genome-wide perplexity as much as full down-weighting.

    “Compared to full down-weighting, moderate down-weighting results in a similar improvement in perplexity for nonrepetitive regions without sacrificing genome-wide perplexity as much.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Most variants in the 1001 Genomes Project have a neutral GPN score, with a heavy tail of putative functional variants with negative GPN scores.

    “While most variants have a neutral GPN score, there is a heavy tail of putative functional variants with negative GPN scores (Fig. 5A).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    SNPs with the lower 1% of GPN × LD scores are 10.3-fold enriched in GWAS hits compared to the upper 99% of GPN × LD scores, while less than 7.5-fold enrichment was observed for other methods.

    “In particular, by analyzing odds ratios (Fig. 6C), we found that SNPs with the lower 1% of GPN × LD scores are 10.3-fold enriched in GWAS hits compared to the upper 99% of GPN × LD scores, while less than 7.5-fold enrichment was observed for other methods (Fig. 6D); see SI Appendix, Fig. S11 for different thresholds.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    A slightly higher odds ratio is achieved by the GPN model trained with an intermediate loss weight on repeats.

    “On a separate note, a slightly higher odds ratio is achieved by the GPN model trained with an intermediate loss weight on repeats (SI Appendix, Fig. S6C).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    A logistic regression classifier trained on averaged GPN embeddings achieved highest accuracy on CDS (96%) and lowest on ncRNA (51%).

    “To quantify GPN’s ability to distinguish genomic regions, we trained a logistic regression classifier using the averaged embeddings as features, achieving the highest accuracy on CDS (96%) and the lowest on ncRNA (51%), the least frequent class.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    GPN is considerably smaller-by 200 times-than the largest published protein language model.

    “Our current proof-of-concept model is considerably smaller—by 200 times—than the largest published protein language model (49).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [49] Z. Lin, Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379, 1123– 1130 (2023). 36927031 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    GPN embeddings can distinguish different repeat families to some extent.

    “We also note that, to some extent, GPN embeddings can distinguish different repeat families (SI Appendix, Fig. S2).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In models with 0.0 and 0.1 down-weighting, respectively, 8% and 9% of repeat variants are ranked before the first decile of missense variants, compared to 27% in the model without any down-weighting.

    “More precisely, in models with 0.0 and 0.1 down-weighting, respectively, 8% and 9% of repeat variants are ranked before the first decile of missense variants. These represent a substantial decrease compared to the 27% observed in the model without any down-weighting (SI Appendix, Fig. S6B, Fisher’s exact test P<10−300).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    GPN does not rely on whole-genome alignments, which can often have lower quality in noncoding regions.

    “Furthermore, GPN does not rely on whole-genome alignments, which can often have a lower quality in noncoding regions.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    GPN scores are only weakly correlated with phyloP (r=0.22, P<10−300) and phastCons (r=0.13, P<10−300).

    “In fact, GPN scores are only weakly correlated with phyloP (r=0.22,P<10−300) and phastCons (r=0.13,P<10−300).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    GPN is genome-wide, can be trained on genomic sequence alone, and is cell-type and mechanism agnostic.

    “Similar to phyloP and phastCons, GPN is genome-wide, can be trained on genomic sequence alone, and is cell-type and mechanism agnostic (46).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [46] P. F. Sullivan, Leveraging base pair mammalian constraint to understand genetic variation and human disease. Science 380, abn293 (2023). PMC10259825 37104612

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    Perplexity in CDS2 is significantly smaller than that in CDS1, which is significantly smaller than that in CDS3.

    “Perplexity in CDS2 is significantly smaller than that in CDS1, which in turn is significantly smaller than that in CDS3 (all Mann–Whitney P-values <10−300), consistent with their different expected levels of constraint (18).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [18] K. S. Pollard, M. J. Hubisz, K. R. Rosenbloom, A. Siepel, Detection of nonneutral substitution rates on mammalian phylogenies. Genome Res. 20, 110– 121 (2010). 19858363 PMC2798823 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    The separate model trained on the single genome of A. thaliana obtained a test perplexity of 3.13 (3.17 on nonrepeat regions).

    “This model obtained a higher test perplexity of 3.13 (3.17 on nonrepeat regions).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    The separate model trained on the single genome of A. thaliana would soon start overfitting.

    “as we noticed it would soon start overfitting”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    In our initial experiments, convolutional models converged faster than transformer models.

    “In our initial experiments, we noticed that convolutional models converged faster than transformer models.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationnone given

    phyloP and phastCons only consider nucleotide frequencies at a specific site, whereas GPN can learn from joint nucleotide distributions across all similar contexts appearing in the genome.

    “The key distinction is that while phyloP and phastCons only consider nucleotide frequencies at a specific site, GPN can learn from joint nucleotide distributions across all similar contexts appearing in the genome.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    The lower perplexity in nonrepetitive regions upon down-weighting repeats can be attributed to the model allocating fewer parameters exclusively to repetitive elements.

    “The lower perplexity in nonrepetitive regions upon down-weighting repeats can be attributed to the model allocating fewer parameters exclusively to repetitive elements.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    GPN can be readily integrated into GWAS fine-mapping and polygenic risk scores.

    “Our results lay the foundation for developing state-of-the-art genome-wide variant effect predictors for any species using genomic sequence alone, which can be readily integrated into GWAS fine-mapping and polygenic risk scores.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationcitation

    The confusion between intergenic regions and ncRNAs may be partly explained by errors in ncRNA annotation.

    “As summarized in Fig. 2B, the highest confusion was observed between intergenic regions and ncRNAs; this may be partly explained by errors in ncRNA annotation, which is especially challenging given their low expression levels and poor conservation (33).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [33] Z. Lu, Identification and characterization of novel lncRNAs in Arabidopsis thaliana. Biochem. Biophys. Res. Commun. 488, 348– 354 (2017). 28499866

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Since GPN is trained only on DNA sequence, it can be readily applied to understudied nonmodel organisms even in the absence of extensive functional genomics data, while still providing state-of-the-art unsupervised variant effect prediction genome-wide.

    “Since GPN is trained only on DNA sequence, it can be readily applied to understudied nonmodel organisms even in the absence of extensive functional genomics data, while still providing state-of-the-art unsupervised variant effect prediction genome-wide.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Larger models and more extensive training data can enhance performance.

    “Mounting evidence suggests that larger models and more extensive training data can enhance performance (48).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [48] J. Kaplan, Scaling laws for neural language models. arXiv [Preprint] (2020). http://arxiv.org/abs/2001.08361 (Accessed 3 July 2023).

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    A. thaliana chromosomes 4 and 5 were held out for validation and testing, respectively.

    “We held out A. thaliana chromosomes 4 and 5 for validation and testing, respectively.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    The linear complexity of convolution simplifies inference or fine-tuning on longer sequences such as entire chromosomes.

    “The linear complexity of convolution also simplifies inference or fine-tuning on longer sequences such as entire chromosomes, which in the case of transformers might require chunking (with some overlap) and aggregating the results.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    Variants with different consequences in different transcripts were excluded, and only variant types with at least 1,000 variants were compared.

    “We compared scores for variant types with at least 1,000 variants, and we excluded variants with different consequences in different transcripts.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    When comparing to phyloP and phastCons, variants where these scores are undefined were excluded.

    “When comparing to phyloP and phastCons, we excluded variants where these scores are undefined (due to the lack of whole-genome alignment).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Convolutional models have shown similarly good performance as transformer architectures in natural language and protein modeling.

    “While language model pretraining successes were first showcased by transformer architectures, convolutional models have shown similarly good performance in natural language (55) and protein modeling (56).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [55] Y. Tay, “Are pretrained convolutions better than pretrained transformers?” in Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), C. Zong, F. Xia, W. Li, R. Navigli, Eds. (Association for Computational Linguistics, Online, 2021), pp. 4349–4359.

      Not checked in the free run.

    • cites [56] K. K. Yang, A. X. Lu, N. Fusi, “Convolutions are competitive with transformers for protein sequence pretraining” in ICLR2022 Machine Learning for Drug Discovery (2022). 38428432

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisereasoning

    The 1001 Genomes genotype matrix is binary because all accessions are homozygous, as Arabidopsis is predominantly selfing.

    “This genotype matrix is binary, since all the accessions are homozygous, as Arabidopsis is predominantly selfing.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Arabidopsis is predominantly selfing.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    The association P-value is invariant to recoding of reference and alternate alleles.

    “Since the association P-value is invariant to recoding of reference and alternate alleles, we took the absolute value of GPN scores.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    Using single-nucleotide tokens makes interpretation easier, particularly for unsupervised variant effect prediction.

    “While a thorough benchmark of different tokenization strategies is lacking, using single-nucleotide tokens makes interpretation easier, in particular for unsupervised variant effect prediction.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Windows in each chromosome were predicted by a model trained on the remaining chromosomes.

    “Windows in each chromosome were predicted by a model trained on the remaining chromosomes.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationown data

    The locality of convolutions may be a good inductive bias for modeling DNA sequences at this scale.

    “The locality of convolutions may be a good inductive bias for modeling DNA sequences at this scale.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Windows with ambiguous annotation were excluded from the analysis.

    “Windows with ambiguous annotation (e.g., 50% CDS and 50% intron) were excluded from the analysis.”
    Quote in the paperNot a substantial claim
33 of 33

  • The paper relies on this without support.

    The GPN score, defined as the log-likelihood ratio between the alternate and reference allele, is a valid pathogenicity or functionality score for single-nucleotide polymorphisms.

    “GPN can be employed to calculate a pathogenicity or functionality score for any single-nucleotide polymorphism (SNP) in the genome using the log-likelihood ratio between the alternate and reference allele (GPN score, Fig. 1).”

  • The paper relies on this without support.

    Rare variants are more likely to be deleterious or functional than common variants.

    “Putative functional SNPs, defined as the lowest 0.1% of GPN scores, exhibit a 5.5-fold enrichment in rare variants (Fig. 5C); see SI Appendix, Fig. S8 for different allele frequency thresholds.”

  • The paper relies on this without support.

    Arabidopsis is predominantly selfing.

    “as Arabidopsis is predominantly selfing”
3 of 3

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • The limited separation of 3′ UTR regions from other elements suggests that the model has not fully learned to distinguish this type of element, or that many of these regions might be mis-annotated.

    “The limited separation of 3′ UTR regions from other elements suggests that the model has not fully learned to distinguish this type of element, or as previously suggested, that many of these regions might be mis-annotated37.”

    From Nucleotide Transformer: building and evaluating robust foundation models for human genomics (2025) · cites this paper as [37]

    The citation is attached to the suggestion that many 3′ UTR regions might be mis-annotated. In Results, “Unsupervised Clustering of Genomic Regions” (page unavailable), the paper suggests annotation errors for ncRNAs, not 3′ UTRs. Searches for 3′ UTR annotation and classification found no passage in our copy addressing that suggestion. (LLM, conf=0.72)

    Experimental — accuracy not yet measured

Cites

Flagged first, then by citations

Research works printed but not linked: 3

  1. S. Mo “Multi-modal self-supervised pre-training for large-scale genome data” in NeurIPS 2021 AI for Science Workshop (2021).
  2. H. Zhou A. Shrikumar A. Kundaje “Towards a better understanding of reverse-complement equivariance for deep learning models in genomics” in Machine Learning in Computational Biology D. A. Knowles S. Mostafavi S.-I. Lee Eds. (PMLR 2022) pp. 1–33.
  3. E.W. Sayers Genome. NCBI. https://www.ncbi.nlm.nih.gov/data-hub/genome. Accessed 2 June 2023.

Cited by

Sorted by citations

Related

Authors
Benegas, G., Batra, S. S., & Song, Y. S.
Affiliations
University of California, Berkeley (US)
Corresponding author
Song, Yun S.
ORCID
Venue
Proceedings of the National Academy of Sciences
Publisher
National Academy of Sciences
Published
2023-10-26
Volume
120
Issue
44
Pages
e2311219120-e2311219120
ISSN
  • 0027-8424 (ISSN-L)
  • 1091-6490
DOI
10.1073/pnas.2311219120
PMC
PMC10622914
OpenAlex
W4387966979
Cited by
157
References
65
Citations by year
2026: 52 · 2025: 61 · 2024: 35 · 2023: 8 · 2022: 1
FWCI
13.90
Type
journal article
Language
en
Open access
Licence
Locations
  • Proceedings of the National Academy of Sciences (publishedVersion, cc-by-nc-nd)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
Funders
  • HHS | NIH | National Institute of General Medical Sciences (R35-GM134922)
  • Koret Foundation (Koret-UC Berkeley-Tel Aviv University Initiative in Computational Biology and Bioinformatics)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Genomics and Phylogenetic Studies
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Genetics › Genetic Associations and Epidemiology
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Genetics › Genomics and Rare Diseases
Keywords
DNA language models, variant effect prediction, genome-wide association study, causal variant identification, Arabidopsis thaliana, Brassicales, conservation scores, sequence logos
MeSH terms
DNA, Genome, Arabidopsis, Genomics, Genome-Wide Association Study

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-10.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-10.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review