Rethinking Genomic Modeling Through Optical Character Recognition

AI summary

Overview

• Research area: Computer vision applied to genomics — specifically, reframing DNA sequence modeling as OCR-style visual document understanding (category: Computer Vision, arXiv:2602.02014v2).
• Technical level: Intermediate. The core idea is intuitive, but the full method assumes familiarity with genomic foundation models, vision–language architectures, and sequence benchmark protocols.
• Scope: The paper introduces OpticalDNA, a vision-based DNA foundation model that renders nucleotide sequences into multi-page visual documents and trains an OCR-capable vision–language model to read, ground, retrieve, and complete them.

What This Paper Is About

Existing genomic foundation models borrow large language model architectures and treat DNA as a one-dimensional token sequence that must be read exhaustively. The authors argue this is structurally mismatched to genomic semantics, which are sparse and discontinuous, so most computation is spent on low-information background rather than functional regions, and long contexts cannot be compressed in an understanding-driven way. The goal of the paper is to replace sequential token modeling with OCR-style document understanding: render DNA into structured 2D page layouts and train a vision–language model to interpret them region by region.

Key Contributions

1. A new framing of genomic modeling. The authors reformulate genomic sequence modeling as an OCR-style document understanding problem, which they describe as a new paradigm for genome-scale representation learning. A controlled comparison reported in the paper finds that 2D CNN backbones show a clearly better accuracy–efficiency trade-off than 1D CNN baselines on eQTL prediction.
2. OpticalDNA, a vision-based DNA foundation model. The framework renders a 1D sequence into a multi-page DNA document with nucleotide-level bounding boxes, then pretrains a visual encoder (SAM–Conv–CLIP-L front-end with a learned projector and a multi-page fusion module) together with an autoregressive document decoder based on DeepSeek-3B (Mixture-of-Experts, 570M activated parameters).
3. Six OCR-inspired genomic tasks. The paper defines prompt families T1–T6 spanning four primitives — recognition/reading, grounding, retrieval, and completion — including free-form transcription, transcription with spatial grounding, ROI transcription, masked-region completion, query-driven subsequence localization, and chromosome-level document classification.
4. Efficiency and accuracy claims across benchmarks. On sequences up to 450k bases, OpticalDNA is reported to achieve the best overall performance with nearly 20 times fewer effective tokens, surpassing models with up to 985 times more activated parameters while tuning only 256k trainable parameters.

Main Findings

• eQTL prediction on DNALONGBENCH: With a 256K-parameter linear probe, OpticalDNA reaches 0.852 average AUROC across nine GTEx tissues, compared with HyenaDNA (0.514, +65.8% relative), Caduceus-Ph (0.750, +13.6%), JanusDNA w/o mid-Attn (0.791, +7.7%), NT-v2-500M (0.772, +10.4%), and GENERator-1.2B (0.782, +9.0%). It exceeds the expert model (Enformer, 252M activated parameters at inference, 0.681 average AUROC).
• Lightweight MLP head improves results further: Replacing the linear probe with an MLP head of 1.3M–2.3M parameters raises average AUROC from 0.852 to 0.867 (+1.8% relative) and gives the best result on 5/9 tissues, including WB (0.927 vs. 0.821 for JanusDNA MLP w/o mid-Attn) and Thyroid (0.876 vs. 0.793).
• Parameter efficiency on adaptation: OpticalDNA outperforms JanusDNA on all tissues while using approximately 30 times fewer trainable parameters for adaptation (256K vs. approximately 7.7M).
• Rice subspecies generalization (RiceSubBench): With 409M parameters versus Evo-2 (7B) and LucaOne (1.8B), OpticalDNA achieves the best accuracy on all splits, including in-domain japonica (0.590). Gains grow under stronger distribution shift: +8.49% on rufipogon (0.639 vs. 0.589), +9.35% on barthii (0.608 vs. 0.556), and +13.88% on glaberrima (0.599 vs. 0.526). It achieves the best AUROC on 4/5 subspecies splits.
• Whole-genome phenotype prediction (RiceWGPB): On approximately 400M-base rice genome inputs, OpticalDNA attains the lowest RMSE for thousand-grain weight (2.952 vs. 3.056 for Evo-2 and 8.817 for LucaOne) and for LRI-15SZ (9.531 vs. 9.617 and 9.740). On a representative 389.8M-base genome (NH158), inference time is 12.3 minutes versus 5h40m for Evo-2 and 32.5m for LucaOne.
• Ablation against the backbone: Compared with DeepSeek-OCR under the same protocol, OpticalDNA improves eQTL AUROC on all nine tissues, with a +5.37% relative gain on the overall average; the largest gains are on Thyroid (+16.86%) and SNSES (+14.30%), with MS (+7.09%), AS (+3.37%), NT (+3.14%), and SSELL (+2.77%) also improving.
• DNA transcription fidelity: On 1,000 validation examples each from HG38 and rice under random tail truncation of up to 90%, OpticalDNA reaches near-perfect exact match at short prefixes (97.3 on HG38 and 98.5 on rice at 10% GT prefix) and remains strong at full length (79.6 and 74.9 at 100%), while DeepSeek-OCR stays near zero EM throughout. Character similarity stays in the 90.6–100.0 range.
• Region-of-interest grounding: On task T2 under a strict IoU = 0.99 criterion, OpticalDNA-HG38 attains line-count match 0.9871, joint accuracy 0.9820, and page-level strict success 0.8763; OpticalDNA-Rice attains 0.9991, 0.9692, and 0.8084. Both achieve pixel-level localization error of ℓ∞ ≤ 0.3941 px.
• Compression robustness: As rendering resolution increases from 512 to 1280, visual tokens per page rise from 64 to 400 and compression ratio rises from 19.0 to 21.2, while average AUROC stays nearly unchanged (0.849–0.852).
• Rendering robustness: On the Adipose Subcutaneous eQTL task, OpticalDNA remains robust under common formatting changes, with several variants matching or exceeding the original layout (0.813 AUROC), including dense rendering, wide+dense rendering, and font variations. The paper notes the main degradation appears in rigid fixed-layout settings; the provided text is truncated at this point.
• Interpretability: Grad-CAM on the multi-page fusion representations shows localized attributions over short contiguous regions rather than diffuse base-level attention. Activations concentrate around annotated donor splice sites, and the corresponding page receives higher overall attribution.
• Contamination analysis: The paper reports negligible train–test leakage, with full-sequence overlap rates of 0.279% for eQTL and 0.489% for RiceSubBench, and similarly low rates for RiceWGPB.

Methodology in Plain English

The authors stop treating DNA as a sentence and start treating it as a document. A genomic sequence is rasterized as monospace-like text onto fixed-resolution 640×640 pages, written row by row, using only the characters A/C/G/T/N, with font size 14 and line spacing 1.6 — roughly 1,800 nucleotides per page. Longer sequences simply continue onto more pages, so a sequence becomes a multi-page document.

Every rendered nucleotide gets a bounding box that records its character, its global index in the sequence, and its pixel coordinates. These boxes let the system map any sequence interval to a region on a page, and any region back to an interval. This is what makes region-based operations — "read this part," "find this subsequence," "fill in this masked area" — directly expressible.

The model has two parts. A visual encoder converts page images into compact visual tokens: pages are split into 16×16 patches, encoded by a SAM–Conv–CLIP-L front-end with fixed downsampling along the token axis, projected to the decoder width, then aggregated across pages by a multi-page fusion module (one attention layer, 20 heads, mean reduction over the page dimension) into a fixed-length sequence of 100 document tokens. A document decoder, based on DeepSeek-3B, then generates answers autoregressively from those tokens plus a task prompt.

Pretraining uses six prompt families organized as two-turn multimodal conversations, where the user turn carries the prompt and the page images and the assistant turn carries the target. Training is standard teacher-forced autoregressive loss computed only over the assistant span. Tasks are sampled from a categorical distribution to balance them, and robustness is added by optionally truncating supervision tails, randomizing sampled spans, and varying the query length for subsequence localization.

For efficiency, the visual front-end is frozen, the decoder is fine-tuned with LoRA, the multi-page fusion module gets full-parameter updates, and the projector is tuned either with LoRA or fully. Pretraining runs on 8× H100 GPUs on HG38 using 2048-bp windows and the same sampling protocol as JanusDNA, in two stages (227,600 steps over about 8 days, then 190,000 projector steps over about 3 days). A separate rice model is pretrained on the Nipponbare (O. sativa japonica) T2T-NIP genome from RiceSuperPIRdb with a 2048-bp window and 1920-bp overlap for 150,000 steps (about 5 days). After pretraining, the encoder is reused as a general representation extractor with a lightweight linear or MLP head for downstream predictions.

Why This Matters

The paper argues that exhaustive sequential reading is structurally wasteful for genomes, because functional signal is sparse and embedded in long low-information background. If region-aware visual processing can deliver better accuracy at far lower token and parameter cost, it changes how genome-scale models are built and how long contexts are handled.

Real-world applications:

• Disease diagnosis and precision medicine: Genomic sequence modeling is described as foundational to disease diagnosis and precision medicine, where variant localization and interval-level reasoning matter directly.
• Drug discovery: The paper cites drug discovery as an application area for genotype-to-phenotype modeling.
• Agricultural genomics: The rice benchmarks target cross-subspecies generalization and organism-level traits such as thousand-grain weight and leaf rolling index, which are relevant to crop breeding.
• Scalable genome-wide analysis: The 12.3-minute inference on a 389.8M-base genome addresses the practical cost of running genome-scale prediction.

Industry relevance: Companies and labs that train or serve genomic foundation models face steep compute and memory costs that scale poorly with sequence length. OpticalDNA's claim of competitive accuracy with nearly 20 times fewer effective tokens and 256k trainable parameters on adaptation points to cheaper fine-tuning and cheaper inference, plus explicit coordinate outputs (bounding boxes) that integrate more naturally with interval-based genomic tooling than implicit positional embeddings.

Future Directions

• Scaling the decoder and encoder beyond the current configuration. The paper uses a DeepSeek-3B decoder with 570M activated parameters and a 100-token document representation; whether benefits persist at larger scales is not established here.
• Extending the OCR-style task set and species coverage. The six tasks cover reading, grounding, retrieval, and completion; additional genomic primitives and broader multi-species pretraining are natural extensions given the rice and human setups.
• Pushing the compression frontier. AUROC was nearly flat as resolution rose from 512 to 1280 with compression ratios of 19.0 to 21.2; how far compression can go before accuracy degrades is left open, along with the token-bottleneck and end-to-end cost comparisons the paper points to in its appendices.
• Understanding and mitigating rendering sensitivity. The paper reports robustness across density, aspect ratio, layout mode, starting position, and font, but notes degradation in rigid fixed-layout settings — the truncated text leaves the full failure analysis and its implications unstated in the provided content.

Target Audience

This paper is most useful to machine learning researchers working on genomic foundation models, vision–language model practitioners interested in cross-domain applications of document understanding, and computational biologists or bioinformaticians who need long-context, region-aware genome analysis at practical cost. Teams building or serving genome-scale models in industry and clinical genomics settings will find the efficiency and inference-time results particularly relevant.

Authors’ 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.