Pith. sign in

REVIEW 4 major objections 6 minor 51 references

This paper advances a new claim: treating DNA not as a one-dimensional token stream but as multi-page rendered documents, read by an OCR-style vision-language model, produces genomic representations that compress sequences nearly 20-fold in

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 05:29 UTC pith:WU3AYYJB

load-bearing objection The OCR-style genomic model is a genuinely new application with competitive eQTL results and an interesting token-compression angle, but the headline parameter-efficiency claim is built on inconsistent accounting and the long-range mechanism (mean pooling over pages) undercuts the paper's core narrative. the 4 major comments →

arxiv 2602.02014 v2 pith:WU3AYYJB submitted 2026-02-02 cs.CV cs.AIcs.CLcs.LG

Rethinking Genomic Modeling Through Optical Character Recognition

Hongxin Xiang , Pengsen Ma , Yunkang Cao , Di Yu , Haowen Chen , Xinyu Yang , Xiangxiang Zeng This is my paper
classification cs.CV cs.AIcs.CLcs.LG
keywords genomic foundation modelsoptical character recognitionvision-language modelDNA documentvisual tokenslong-range DNA modelingsequence compressionwhole-genome phenotype prediction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that genomic sequence modeling need not read DNA as a 1D token stream. Instead, it renders a sequence as pages of monospaced text and applies OCR-style document understanding, learning compact visual tokens that preserve base-level content under a roughly 20x reduced token budget. It claims this approach beats recent sequence-based genomic foundation models on long-range eQTL prediction, out-of-distribution rice subspecies classification, and whole-genome phenotype regression, while tuning only 256K parameters and processing a 400M-base rice genome in about 12 minutes. A sympathetic reader would care because the result suggests that the dominant computational bottleneck of genomic foundation models — exhaustive token-by-token processing — may be avoidable without sacrificing biological signal.

Core claim

On the paper's own terms, OpticalDNA's central discovery is that genomic function survives a change in modality: rendering a 1D sequence as 2D pages and reading it with an OCR-capable vision-language model yields representations that outperform LLM-style DNA foundation models on long-range regulatory benchmarks, cross-subspecies splice-site classification, and whole-genome trait prediction. The claim is carried by a visual encoder that converts each page into approximately 100 fused document tokens through a single self-attention layer followed by mean-pooling over pages, combined with generative pretraining over six prompt-conditioned tasks — transcription, grounded transcription, ROI trans

What carries the argument

The load-bearing object is the 'DNA document': a multi-page rendering of a genomic sequence as monospaced text (about 1,800 bases per page) with nucleotide-level bounding boxes. The visual encoder partitions each page into 16x16 patches, downsamples by 16, projects the features to decoder width, and a one-layer, 20-head self-attention fusion module aggregates pages into L=100 document tokens. Those tokens condition an autoregressive OCR decoder pretrained on six prompted task families, so that the tokens must be compact yet reconstructible enough to support base-level reading, region grounding, and subsequence retrieval. The mean-reduction over pages is what turns a variable-length, arbitrar

Load-bearing premise

The load-bearing premise is that mean-pooling features across rendered pages preserves the long-range regulatory signal in a 450k-base (and 400M-base) sequence; if relevant biological dependencies span page boundaries, averaging over pages would erase exactly the signal the paper claims to capture.

What would settle it

A controlled eQTL-style task in which the causal variant and the regulated region are placed on different pages (separated by more than approximately 1,800 bases), where a model that attends jointly over both pages outperforms one that mean-pools per-page features. If OpticalDNA's page-pooled representation does not beat a single-page-averaging baseline on such a task, the long-range claim would be falsified. Concretely, shuffling the order of test pages should barely change predictions if cross-page information is unused.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, long-range genomic modeling no longer needs to scale with sequence length in tokens: effective token count is set by page count and image resolution, giving nearly 20x fewer tokens than base/k-mer baselines at comparable or better accuracy.
  • Whole-genome-scale phenotype prediction becomes practical: a 389.8M-base rice genome is scored in 12.3 minutes versus 5h40m for a 7B-parameter sequence model, with lower RMSE on thousand-grain weight and leaf rolling index.
  • OCR-style pretraining transfers across species and subspecies: the same frozen encoder handles in-domain japonica and far-OOD barthii/glaberrima splice-site classification better than Evo-2 (7B) and LucaOne (1.8B) without retraining on the target taxonomy.
  • The compression ratio has headroom: raising rendering resolution from 512 to 1280 pixels raises visual tokens per page from 64 to 400 and the compression ratio from 19.0 to 21.2, while eQTL AUROC stays between 0.849 and 0.852.
  • The learned visual representations localize biologically meaningful regions: Grad-CAM on the fused tokens concentrates around donor splice sites and ranks the containing page highest, providing a new form of interpretability for long-context genomic models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the reported efficiency gain is partly purchased by mean-pooling over pages, which discards cross-page interactions; if real regulatory effects routinely span more than one page (~1,800 bases), a cross-page attentive fusion would be a natural next step that the paper does not test.
  • Editorial inference: the document formulation suggests a testable extension — rendering DNA with a functional-element-aware layout, such as placing an enhancer and its target promoter on the same page, could inject biological priors into the visual tokenization and further improve grounding tasks.
  • Editorial inference: the six-task OCR pretraining suite defines a transferable visual-DNA reading objective that could be applied to non-model species without architectural change; whether the 12.3-minute whole-genome latency scales to a ~3.1G-base human genome remains untested.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes OpticalDNA, a vision-based genomic foundation model that renders DNA sequences as multi-page images and applies OCR-style prompted pretraining over six tasks (transcription, grounding, retrieval, completion, classification). It claims state-of-the-art results on DNALONGBENCH eQTL tasks, rice subspecies classification, and rice whole-genome phenotype prediction, while using nearly 20x fewer effective tokens than sequence baselines and reportedly surpassing models with up to 985x more activated parameters using only 256k trainable parameters. The appendices provide detailed rendering, task, and evaluation protocols, including a controlled 1D/2D CNN comparison and per-task metrics.

Significance. If validated, the OCR-based reformulation is a creative departure from 1D token-centric DNA language models and could offer a more efficient route to long-context genomic analysis. The eQTL comparison in Table 2 uses identical splits to JanusDNA, which is a methodological strength, and the rendering and task definitions are unusually detailed and reproducible. However, the headline efficiency claim rests on inconsistent parameter accounting, the central comparisons lack error bars, and several self-defined capability metrics (T4 span-8 EM, T5 detection at IoU=0.99, T6 accuracy) are reported in the appendices at levels that undercut the claim of fine-grained understanding. The core idea is promising, but the current evidence does not support the paper's strongest quantitative claims.

major comments (4)
  1. [Table 2, Section 4.2.1, and Abstract] The 'activated param' column lists OpticalDNA Linear Probing as 256K, but this is the trainable linear head, not the inference-time parameter count. The paper itself reports OpticalDNA (409M) in Table 3 and Section 4.2.2. The 985x ratio (252M/256K ≈ 984) compares the Expert Model's full parameters to OpticalDNA's probe only; the actual inference stack is ~409M, about 1.6x the Expert Model. This invalidates the abstract's 'up to 985x more activated parameters' claim and the 'parameter-efficient adaptation' narrative in Section 4.2.1. Token compression (~20x) is a separate claim and may survive, but the parameter accounting must be corrected and the claims re-quantified.
  2. [Section 4.1, Eq. (7)] Downstream features are extracted as per-page visual features and then mean-pooled across pages. The fusion module Fθ applies one self-attention layer over per-page tokens followed by a mean reduction over the page dimension. Any signal that spans pages is collapsed to an average; the model cannot represent cross-page long-range interactions. This directly undercuts the claim (Section 4.2.1) that OpticalDNA 'preserves long-range regulatory signals under ultra-long genomic contexts' for sequences up to 450k bases. Please provide evidence that the reported benchmarks do not require cross-page dependencies, or modify the architecture/claims to reflect this limitation.
  3. [Appendix I.3–I.5, Section 4.5] The self-defined pretraining tasks are evaluated in-distribution, and several results are near failure: T4 span-8 Exact Match is 0.0077/0.0013 (Table S26), T5 detection at IoU=0.99 is 0.0055/0.0623 (Table S24), and T6 accuracy is 0.062/0.104 (Table S27). These numbers appear in the appendices but are not discussed in the main text. The claim (Section 4.5) that OpticalDNA 'learns structured, region-aware representations' is only partially supported. Since these tasks are presented as a core contribution, these limitations should be acknowledged in the main text, or the paper should demonstrate that downstream quality does not depend on these capabilities.
  4. [Tables 2–4] No error bars, confidence intervals, or multiple-seed results are reported. The central eQTL comparison claims superiority over several baselines, but single AUROC values could be within noise. This is especially important for RiceWGPB (Table 4), where the dataset has only 102–130 samples and the TGW RMSE difference between OpticalDNA (2.952) and Evo-2 (3.056) is 0.104. Please report variance over at least three seeds or a statistical test to support the 'consistently outperforms' claim.
minor comments (6)
  1. [Section 4.1, Table 3] The model is labeled 'OpticalDNA (409M)' in Table 3 but '256K' in Table 2. Please define parameter reporting conventions (total inference-time parameters vs. trainable parameters) once and apply them consistently.
  2. [Table 7, Abstract] The compression ratio ranges from 19.0 to 21.2 across resolutions in Table 7, but the abstract says 'nearly 20x'. Clarify whether 20x refers to a specific resolution or an average.
  3. [Section 3.3, Eq. (5)] In Eq. (5), Y6 = Y, but Y is not defined. Define the label space for T6 explicitly.
  4. [Appendix C.2] The rendering parameters (font size 14, line spacing 1.6) are given, but the approximate per-page capacity of ~1800 nucleotides is not exact. State the exact per-page nucleotide capacity used in the experiments.
  5. [General] No code or data availability statement is provided. The appendices are detailed, but releasing code and trained checkpoints would substantially improve reproducibility.
  6. [Section 4.1 and Appendix G.2] For Evo-2, the fragment length is set to 70,000, and features are mean-pooled. This is a reasonable practical choice, but the paper should note that this departs from Evo-2's full long-context setting and may affect the comparison.

Circularity Check

0 steps flagged

No significant circularity: the central accuracy and efficiency claims are tested on external benchmarks with independent labels, not derived from the model's own pretraining objectives.

full rationale

The paper's headline results (DNALONGBENCH eQTL AUROC, rice subspecies OOD accuracy/AUROC, whole-genome phenotype RMSE) are evaluated on externally defined benchmarks with independent biological labels and published baselines, so the main comparisons are not self-referential. The OCR-style tasks T1-T6 are the pretraining objectives themselves, and the evaluations in Sections 4.3-4.4 sample held-out pages from the same pretraining corpora; these are in-distribution capability checks rather than independent evidence, but they are not load-bearing for the paper's central claims. The ablation compares OpticalDNA to its external DeepSeek-OCR backbone, and no load-bearing result rests on a self-citation chain. The '985× fewer activated parameters' claim in the abstract and Table 2 uses a probe-only count (256K) while the paper itself reports a 409M parameter encoder in Section 4.2.2; that is an inconsistent accounting or correctness concern, not a circular derivation. No equation in the paper reduces a predicted quantity to a fitted input by construction, so the circularity score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper contributes a new representation and training protocol, not new physical entities. The central load-bearing choices are the hand-set rendering geometry, the fixed fusion token count, the task-sampling mix, and three domain assumptions: glyph layout preserves order, natural-image OCR features transfer, and mean-pooled pages preserve long-range signal.

free parameters (3)
  • Rendering geometry (font size 14, line spacing 1.6) = font_size=14, line_spacing=1.6 (~1800 nt/page)
    Chosen by hand; controls the 2D neighborhood and the ~20x token compression ratio; not ablated in the paper.
  • Fused document token count L = 100
    Architectural choice for the fixed-length document representation after multi-page fusion; affects capacity and compression.
  • Task sampling distribution pi over T1-T6 = Stage1: (0.25,0.20,0.15,0.15,0.15,0.10); Stage2/rice: (0.17,...)
    Hand-set balancing that controls the pretraining mix of OCR-style tasks.
axioms (4)
  • domain assumption Rendered monospace glyph layout preserves all nucleotide order needed for genomic prediction.
    Section 3.2: sequences are written row-by-row without auxiliary markers; base order is encoded only through glyph positions, and line wrapping changes local spatial neighborhoods.
  • domain assumption DeepSeek-OCR's natural-image visual front-end transfers to synthetic DNA text pages.
    Section 3.4: SAM-Conv-CLIP-L is frozen; no ablation of alternative visual backbones is provided.
  • ad hoc to paper Mean pooling over pages retains enough long-range regulatory signal for the claimed 450k/400M-base results.
    Section 3.4 Eq. 7 and Section 4.1: cross-page interactions are collapsed to an average, so inter-page dependencies cannot be represented.
  • ad hoc to paper The six OCR-style prompt tasks are an adequate pretraining proxy for genomic understanding.
    Section 3.3 T1-T6: the tasks have only indirect relation to downstream regulatory labels; their sufficiency is assumed.

pith-pipeline@v1.3.0-alltime-deepseek · 34975 in / 14418 out tokens · 134658 ms · 2026-08-03T05:29:02.266100+00:00 · methodology

0 comments
read the original abstract

Recent genomic foundation models largely adopt large language model architectures that treat DNA as a one-dimensional token sequence. However, exhaustive sequential reading is structurally misaligned with sparse and discontinuous genomic semantics, leading to wasted computation on low-information background and preventing understanding-driven compression for long contexts. Here, we present OpticalDNA, a vision-based framework that reframes genomic modeling as Optical Character Recognition (OCR)-style document understanding. OpticalDNA renders DNA into structured visual layouts and trains an OCR-capable vision--language model with a visual DNA encoder and a document decoder, where the encoder produces compact, reconstructible visual tokens for high-fidelity compression. Building on this representation, OpticalDNA defines prompt-conditioned objectives over core genomic primitives-reading, region grounding, subsequence retrieval, and masked span completion-thereby learning layout-aware DNA representations that retain fine-grained genomic information under a reduced effective token budget. Across diverse genomic benchmarks, OpticalDNA consistently outperforms recent baselines; on sequences up to 450k bases, it achieves the best overall performance with nearly 20$\times$ fewer effective tokens, and surpasses models with up to 985$\times$ more activated parameters while tuning only 256k trainable parameters.

Figures

Figures reproduced from arXiv: 2602.02014 by Di Yu, Haowen Chen, Hongxin Xiang, Pengsen Ma, Xiangxiang Zeng, Xinyu Yang, Yunkang Cao.

Figure 1
Figure 1. Figure 1: From sequential reading to selective genomic scanning. (a) Sparse, discontinuous genomic signals make sequential modeling inefficient. (b) OCR-inspired genomic modeling enables efficient, reconstructible visual compression. (c) 2D CNNs outperform 1D CNNs in accuracy–efficiency trade-offs on eQTL prediction. (d) Vision tokens substantially reduce the effective token count. (Fig. 1c; Appendix A), motivating … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of OpticalDNA. (a) Render a 1D genomic sequence into a multi-page DNA document with bounding-box annotations. (b) Construct six OCR-style prompted genomic tasks. (c) Pretrain a visual encoder–document decoder under prompt supervision. Here c (p) k ∈ {A, C, G, T, N} is the rendered nucleotide, g (p) k ∈ [0, N) is the global nucleotide index in S, and r (p) k denotes the glyph bounding box in page p… view at source ↗
Figure 3
Figure 3. Figure 3: Ablation on DNALONGBENCH (AUROC). Green arrows and numbers denote the relative gain over DeepSeek-OCR. (+2.77%). These results indicate that our DNA-specific doc￾ument formulation and pretraining yield transferable rep￾resentations for long-range functional genomics prediction. Detailed per-tissue results are provided in Appendix H.1. Q2: DNA transcription. We randomly sample 1,000 val￾idation examples fro… view at source ↗
Figure 4
Figure 4. Figure 4: Grad-CAM visualization on multi-page fusion for a donor case (two pages). Purple boxes indicate donor splice sites; numbers denote page-level mean attribution [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

51 extracted references · 1 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Ashley, E. A. Towards precision medicine. Nature Reviews Genetics, 17 0 (9): 0 507--522, 2016

    2016

  3. [3]

    R., Grabska-Barwinska, A., Taylor, K

    Avsec, Z ., Agarwal, V., Visentin, D., Ledsam, J. R., Grabska-Barwinska, A., Taylor, K. R., Assael, Y., Jumper, J., Kohli, P., and Kelley, D. R. Effective gene expression prediction from sequence by integrating long-range interactions. Nature methods, 18 0 (10): 0 1196--1203, 2021

    2021

  4. [4]

    Predicting gene expression from dna sequence using deep learning models

    Barbadilla-Mart \' nez, L., Klaassen, N., van Steensel, B., and de Ridder, J. Predicting gene expression from dna sequence using deep learning models. Nature Reviews Genetics, pp.\ 1--15, 2025

    2025

  5. [5]

    Nougat: Neural optical understanding for academic documents

    Blecher, L., Cucurull, G., Scialom, T., and Stojnic, R. Nougat: Neural optical understanding for academic documents. arXiv preprint arXiv:2308.13418, 2023

    Pith/arXiv arXiv 2023

  6. [6]

    G., Ku, J., Poli, M., Brockman, G., Chang, D., Gonzalez, G

    Brixi, G., Durrant, M. G., Ku, J., Poli, M., Brockman, G., Chang, D., Gonzalez, G. A., King, S. H., Li, D. B., Merchant, A. T., et al. Genome modeling and design across all domains of life with evo 2. BioRxiv, pp.\ 2025--02, 2025

    2025

  7. [7]

    Browning, B. L. and Browning, S. R. Genotype imputation with millions of reference samples. The American Journal of Human Genetics, 98 0 (1): 0 116--126, 2016

    2016

  8. [8]

    Glyph: Scaling context windows via visual-text compression

    Cheng, J., Liu, Y., Zhang, X., Fei, Y., Hong, W., Lyu, R., Wang, W., Su, Z., Gu, X., Liu, X., et al. Glyph: Scaling context windows via visual-text compression. arXiv preprint arXiv:2510.17800, 2025 a

    arXiv 2025

  9. [9]

    Dnalongbench: a benchmark suite for long-range dna prediction tasks

    Cheng, W., Song, Z., Zhang, Y., Wang, S., Wang, D., Yang, M., Li, L., and Ma, J. Dnalongbench: a benchmark suite for long-range dna prediction tasks. Nature Communications, 16 0 (1): 0 10108, Nov 2025 b . ISSN 2041-1723. doi:10.1038/s41467-025-65077-4. URL https://doi.org/10.1038/s41467-025-65077-4

    work page doi:10.1038/s41467-025-65077-4 2025

  10. [10]

    Consortium, E. P. et al. An integrated encyclopedia of dna elements in the human genome. Nature, 489 0 (7414): 0 57, 2012

    2012

  11. [11]

    Consortium, G. R. et al. Genome reference consortium human build 38 patch release 13 (grch38. p13). National Library of Medicine, Bethesda, Maryland, 2019

    2019

  12. [12]

    H., Oteri, F., Dallago, C., Trop, E., de Almeida, B

    Dalla-Torre, H., Gonzalez, L., Mendoza-Revilla, J., Lopez Carranza, N., Grzywaczewski, A. H., Oteri, F., Dallago, C., Trop, E., de Almeida, B. P., Sirelkhatim, H., et al. Nucleotide transformer: building and evaluating robust foundation models for human genomics. Nature Methods, 22 0 (2): 0 287--297, 2025

    2025

  13. [13]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pp.\ 4171--4186, 2019

    2019

  14. [14]

    Janus DNA : A powerful bi-directional hybrid DNA foundation model

    Duan, Q., Huang, B., Song, Z., Lehmann, I., Gu, L., Eils, R., and Wild, B. Janus DNA : A powerful bi-directional hybrid DNA foundation model. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. URL https://openreview.net/forum?id=9PL1DIIB7e

    2025

  15. [15]

    Dolphin: Document image parsing via heterogeneous anchor prompting

    Feng, H., Wei, S., Fei, X., Shi, W., Han, Y., Liao, L., Lu, J., Wu, B., Liu, Q., Lin, C., Tang, J., Liu, H., and Huang, C. Dolphin: Document image parsing via heterogeneous anchor prompting. In Che, W., Nabende, J., Shutova, E., and Pilehvar, M. T. (eds.), Findings of the Association for Computational Linguistics: ACL 2025, pp.\ 21919--21936, Vienna, Aust...

    doi:10.18653/v1/2025.findings-acl.1130 2025

  16. [16]

    Benchmarking dna foundation models for genomic and genetic tasks

    Feng, H., Wu, L., Zhao, B., Huff, C., Zhang, J., Wu, J., Lin, L., Wei, P., and Wu, C. Benchmarking dna foundation models for genomic and genetic tasks. Nature communications, 16 0 (1): 0 10780, 2025 b

    2025

  17. [17]

    Generalized biological foundation model with unified nucleic acid and protein language

    He, Y., Fang, P., Shan, Y., Pan, Y., Wei, Y., Chen, Y., Chen, Y., Liu, Y., Zeng, Z., Zhou, Z., et al. Generalized biological foundation model with unified nucleic acid and protein language. Nature Machine Intelligence, pp.\ 1--12, 2025

    2025

  18. [18]

    Layoutlmv3: Pre-training for document ai with unified text and image masking

    Huang, Y., Lv, T., Cui, L., Lu, Y., and Wei, F. Layoutlmv3: Pre-training for document ai with unified text and image masking. In Proceedings of the 30th ACM international conference on multimedia, pp.\ 4083--4091, 2022

    2022

  19. [19]

    Ocr-free document understanding transformer

    Kim, G., Hong, T., Yim, M., Nam, J., Park, J., Yim, J., Hwang, W., Yun, S., Han, D., and Park, S. Ocr-free document understanding transformer. In European Conference on Computer Vision, pp.\ 498--517. Springer, 2022

    2022

  20. [20]

    R., Hu, H., Liu, F., Eisenschlos, J

    Lee, K., Joshi, M., Turc, I. R., Hu, H., Liu, F., Eisenschlos, J. M., Khandelwal, U., Shaw, P., Chang, M.-W., and Toutanova, K. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In International Conference on Machine Learning, pp.\ 18893--18912. PMLR, 2023

    2023

  21. [21]

    Diminishing return for increased mappability with longer sequencing reads: implications of the k-mer distributions in the human genome

    Li, W., Freudenberg, J., and Miramontes, P. Diminishing return for increased mappability with longer sequencing reads: implications of the k-mer distributions in the human genome. BMC bioinformatics, 15 0 (1): 0 2, 2014

    2014

  22. [22]

    Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model

    Liu, A., Feng, B., Wang, B., Wang, B., Liu, B., Zhao, C., Dengr, C., Ruan, C., Dai, D., Guo, D., et al. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Genome evolution and diversity of wild and cultivated rice species

    Long, W., He, Q., Wang, Y., Wang, Y., Wang, J., Yuan, Z., Wang, M., Chen, W., Luo, L., Luo, L., et al. Genome evolution and diversity of wild and cultivated rice species. Nature Communications, 15 0 (1): 0 9994, 2024

    2024

  24. [24]

    Layoutllm: Layout instruction tuning with large language models for document understanding

    Luo, C., Shen, Y., Zhu, Z., Zheng, Q., Yu, Z., and Yao, C. Layoutllm: Layout instruction tuning with large language models for document understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 15630--15640, 2024

    2024

  25. [25]

    The blast sequence analysis tool

    Madden, T. The blast sequence analysis tool. The NCBI handbook, 2 0 (5): 0 425--436, 2013

    2013

  26. [26]

    I., Teufel, F., Horlacher, M., Madsen, D., Pultz, D., Winther, O., and Boomsma, W

    Marin, F. I., Teufel, F., Horlacher, M., Madsen, D., Pultz, D., Winther, O., and Boomsma, W. Bend: Benchmarking dna language models on biologically meaningful tasks. arXiv preprint arXiv:2311.12570, 2023

    Pith/arXiv arXiv 2023

  27. [27]

    Hyenadna: Long-range genomic sequence modeling at single nucleotide resolution

    Nguyen, E., Poli, M., Faizi, M., Thomas, A., Wornow, M., Birch-Sykes, C., Massaroli, S., Patel, A., Rabideau, C., Bengio, Y., et al. Hyenadna: Long-range genomic sequence modeling at single nucleotide resolution. Advances in neural information processing systems, 36: 0 43177--43201, 2023 a

    2023

  28. [28]

    Hyenadna: Long-range genomic sequence modeling at single nucleotide resolution

    Nguyen, E., Poli, M., Faizi, M., Thomas, A., Wornow, M., Birch-Sykes, C., Massaroli, S., Patel, A., Rabideau, C., Bengio, Y., et al. Hyenadna: Long-range genomic sequence modeling at single nucleotide resolution. Advances in neural information processing systems, 36: 0 43177--43201, 2023 b

    2023

  29. [29]

    G., Kang, B., Katrekar, D., Li, D

    Nguyen, E., Poli, M., Durrant, M. G., Kang, B., Katrekar, D., Li, D. B., Bartie, L. J., Thomas, A. W., King, S. H., Brixi, G., et al. Sequence modeling and design from molecular to genome scale with evo. Science, 386 0 (6723): 0 eado9336, 2024

    2024

  30. [30]

    W., Wasserman, W

    Novakovsky, G., Dexter, N., Libbrecht, M. W., Wasserman, W. W., and Mostafavi, S. Obtaining genetics insights from deep learning via explainable artificial intelligence. Nature Reviews Genetics, 24 0 (2): 0 125--137, 2023

    2023

  31. [31]

    V., Mikheenko, A., Vollger, M

    Nurk, S., Koren, S., Rhie, A., Rautiainen, M., Bzikadze, A. V., Mikheenko, A., Vollger, M. R., Altemose, N., Uralsky, L., Gershman, A., et al. The complete sequence of a human genome. Science, 376 0 (6588): 0 44--53, 2022

    2022

  32. [32]

    Oliver, S. G. From dna sequence to biological function. Nature, 379 0 (6566): 0 597--600, 1996

    1996

  33. [33]

    T., et al

    Poplin, R., Chang, P.-C., Alexander, D., Schwartz, S., Colthurst, T., Ku, A., Newburger, D., Dijamco, J., Nguyen, N., Afshar, P. T., et al. A universal snp and small-indel variant caller using deep neural networks. Nature biotechnology, 36 0 (10): 0 983--987, 2018

    2018

  34. [34]

    and Xie, X

    Quang, D. and Xie, X. Danq: a hybrid convolutional and recurrent deep neural network for quantifying the function of dna sequences. Nucleic acids research, 44 0 (11): 0 e107--e107, 2016

    2016

  35. [35]

    Quinlan, A. R. and Hall, I. M. Bedtools: a flexible suite of utilities for comparing genomic features. Bioinformatics, 26 0 (6): 0 841--842, 2010

    2010

  36. [36]

    S., Huntley, M

    Rao, S. S., Huntley, M. H., Durand, N. C., Stamenova, E. K., Bochkov, I. D., Robinson, J. T., Sanborn, A. L., Machol, I., Omer, A. D., Lander, E. S., et al. A 3d map of the human genome at kilobase resolution reveals principles of chromatin looping. Cell, 159 0 (7): 0 1665--1680, 2014

    2014

  37. [37]

    E., Cargill, M., Bolk, S., Ireland, J., Sabeti, P

    Reich, D. E., Cargill, M., Bolk, S., Ireland, J., Sabeti, P. C., Richter, D. J., Lavery, T., Kouyoumjian, R., Farhadian, S. F., Ward, R., et al. Linkage disequilibrium in the human genome. Nature, 411 0 (6834): 0 199--204, 2001

    2001

  38. [38]

    M., and Poetsch, A

    Sanabria, M., Hirsch, J., Joubert, P. M., and Poetsch, A. R. Dna language model grover learns sequence context in the human genome. Nature Machine Intelligence, 6 0 (8): 0 911--923, 2024

    2024

  39. [39]

    Caduceus: Bi-directional equivariant long-range dna sequence modeling

    Schiff, Y., Kao, C.-H., Gokaslan, A., Dao, T., Gu, A., and Kuleshov, V. Caduceus: Bi-directional equivariant long-range dna sequence modeling. Proceedings of machine learning research, 235: 0 43632, 2024

    2024

  40. [40]

    A super pan-genomic landscape of rice

    Shang, L., Li, X., He, H., Yuan, Q., Song, Y., Wei, Z., Lin, H., Hu, M., Zhao, F., Zhang, C., et al. A super pan-genomic landscape of rice. Cell Research, 32 0 (10): 0 878--896, 2022

    2022

  41. [41]

    V., Lyu, P., Wan, X., Li, G., Peng, S., Wang, W., Wu, L., Shen, H., Zhou, Y., Tang, C., et al

    Team, H. V., Lyu, P., Wan, X., Li, G., Peng, S., Wang, W., Wu, L., Shen, H., Zhou, Y., Tang, C., et al. Hunyuanocr technical report. arXiv preprint arXiv:2511.19575, 2025

    arXiv 2025

  42. [42]

    N., Kaiser, ., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, ., and Polosukhin, I. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  43. [43]

    Deepseek-ocr: Contexts optical compression

    Wei, H., Sun, Y., and Li, Y. Deepseek-ocr: Contexts optical compression. arXiv preprint arXiv:2510.18234, 2025

    Pith/arXiv arXiv 2025

  44. [44]

    Layoutlm: Pre-training of text and layout for document image understanding

    Xu, Y., Li, M., Cui, L., Huang, S., Wei, F., and Zhou, M. Layoutlm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pp.\ 1192--1200, 2020

    2020

  45. [45]

    Integrating convolution and self-attention improves language model of human genome for interpreting non-coding regions at base-resolution

    Yang, M., Huang, L., Huang, H., Tang, H., Zhang, N., Yang, H., Wu, J., and Mu, F. Integrating convolution and self-attention improves language model of human genome for interpreting non-coding regions at base-resolution. Nucleic acids research, 50 0 (14): 0 e81--e81, 2022

    2022

  46. [46]

    Megabyte: Predicting million-byte sequences with multiscale transformers

    Yu, L., Simig, D., Flaherty, C., Aghajanyan, A., Zettlemoyer, L., and Lewis, M. Megabyte: Predicting million-byte sequences with multiscale transformers. Advances in Neural Information Processing Systems, 36: 0 78808--78823, 2023

    2023

  47. [47]

    A., Ainslie, J., Alberti, C., Ontanon, S., Pham, P., Ravula, A., Wang, Q., Yang, L., et al

    Zaheer, M., Guruganesh, G., Dubey, K. A., Ainslie, J., Alberti, C., Ontanon, S., Pham, P., Ravula, A., Wang, Q., Yang, L., et al. Big bird: Transformers for longer sequences. Advances in neural information processing systems, 33: 0 17283--17297, 2020

    2020

  48. [48]

    Cancer diagnosis with dna molecular computation

    Zhang, C., Zhao, Y., Xu, X., Xu, R., Li, H., Teng, X., Du, Y., Miao, Y., Lin, H.-c., and Han, D. Cancer diagnosis with dna molecular computation. Nature nanotechnology, 15 0 (8): 0 709--715, 2020

    2020

  49. [49]

    Rice functional genomics and breeding database (rfgb)-3k-rice snp and indel sub-database

    ZHANG, H., ZiChao, L., JiaYang, L., Jue, R., Hong, Y., JianLong, X., ZHANG, G., ZHANG, D., ShuaiShuai, T., ChaoChun, W., et al. Rice functional genomics and breeding database (rfgb)-3k-rice snp and indel sub-database. Chinese Science Bulletin, 60 0 (4): 0 367--371, 2015

    2015

  50. [50]

    Single-cell transcriptome atlas and chromatin accessibility landscape reveal differentiation trajectories in the rice root

    Zhang, T.-Q., Chen, Y., Liu, Y., Lin, W.-H., and Wang, J.-W. Single-cell transcriptome atlas and chromatin accessibility landscape reveal differentiation trajectories in the rice root. Nature communications, 12 0 (1): 0 2053, 2021

    2053

  51. [51]

    Dnabert-2: Efficient foundation model and benchmark for multi-species genome

    Zhou, Z., Ji, Y., Li, W., Dutta, P., Davuluri, R., and Liu, H. Dnabert-2: Efficient foundation model and benchmark for multi-species genome. arXiv preprint arXiv:2306.15006, 2023

    Pith/arXiv arXiv 2023