Visual DNA representation
DNA is rendered into structured multi-page images with nucleotide-level bounding boxes, linking one-dimensional genome coordinates to two-dimensional visual regions.
ICML 2026 · Seoul, South Korea
Rethinking Genomic Modeling Through Optical Character Recognition
OpticalDNA turns DNA sequences into structured visual documents and learns OCR-style genomic reasoning, enabling compact, reconstructible visual compression for ultra-long genomic contexts.
20×fewer effective tokens on 450k-base tasks
985×fewer activated parameters than large baselines
0.867average AUROC on DNALONGBench with MLP head
12.3 minrice whole-genome inference time
Conference poster
ICML 2026 Poster
Click the poster to inspect it at a larger size. The preview is embedded in this HTML file, so it remains visible even without a separate image asset.
View the original poster PDF inside this page
Core idea
Genomes are not just sentences. They can be read as documents.
Sequential models spend computation scanning long low-information backgrounds. OpticalDNA converts genomic sequences into coordinate-indexed pages, making genomic reading, grounding, retrieval, and completion natural document-understanding tasks.
Understanding-driven compression
A visual encoder produces compact, reconstructible tokens that preserve fine-grained sequence information while reducing the effective token budget for long inputs.
Region-aware reasoning
OCR-style supervision supports explicit localization, subsequence retrieval, ROI transcription, and masked-region completion through prompt-conditioned genomic tasks.
Method
OpticalDNA pipeline
The method reframes genomic modeling as OCR-style document understanding: render, prompt, encode, decode, and reuse the learned visual DNA encoder for downstream prediction.
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.
1. Render
Convert DNA sequences into fixed-resolution pages with base-level coordinate annotations.
Convert DNA sequences into fixed-resolution pages with base-level coordinate annotations.
2. Learn
Train an OCR-capable vision-language model on reading, grounding, retrieval, and completion prompts.
Train an OCR-capable vision-language model on reading, grounding, retrieval, and completion prompts.
3. Transfer
Reuse the compact visual DNA encoder with lightweight heads for long-range genomic prediction.
Reuse the compact visual DNA encoder with lightweight heads for long-range genomic prediction.
Pretraining tasks
Six OCR-inspired genomic primitives
Reading
Free-form DNA transcription from rendered pages.
Grounded reading
Transcribe DNA and locate each line or block.
ROI transcription
Read DNA text inside user-provided boxes.
Completion
Recover masked DNA regions from visual context.
Retrieval
Find all occurrences of a query subsequence.
Recognition
Predict chromosome-level document labels.
Results
Strong accuracy, efficiency, and interpretability
OpticalDNA is evaluated from canonical long-range regulatory modeling to cross-subspecies generalization and whole-genome phenotype prediction.
256K 30× fewer
Parameter-efficient adaptation
Linear probing reaches 0.852 average AUROC using only 256K trainable parameters, outperforming JanusDNA-MLP while requiring far fewer task-specific parameters.
0.858 robust
Layout robustness
Rendering perturbations do not break the visual formulation: wide+dense rendering reaches 0.858 AUROC on AS eQTL, above the original 0.813 layout.
0.595/0.745 transfer
Cross-species transfer
The HG38-pretrained OpticalDNA transfers strongly to rice, approaching the rice-pretrained model and clearly surpassing HG38-pretrained sequence baselines.
DNALONGBench summary: accuracy with lightweight adaptation
DNALONGBench. OpticalDNA achieves the best average AUROC on eQTL tasks while using a lightweight probing setup. Relative to JanusDNA-MLP, OpticalDNA-MLP improves the average AUROC by +3.2%; relative to JanusDNA without mid-attention, it improves by +9.6%. Average AUROC is computed over nine GTEx eQTL tissues.
| Model | Average AUROC ↑ | Trainable params | Relative note |
|---|---|---|---|
| JanusDNA w/o mid-Attn | 0.791 | 7.662M | — |
| JanusDNA-MLP w/o mid-Attn | 0.840 | 7.745M | strongest JanusDNA baseline |
| OpticalDNA Linear Probing | 0.852 | 256K | +1.4% vs. JanusDNA-MLP |
| OpticalDNA MLP | 0.867 | 1.3M–2.3M | +3.2% vs. JanusDNA-MLP |
RiceSubBench: in-domain to far-OOD generalization
RiceSubBench. OpticalDNA generalizes from in-domain japonica to near-, mid-, and far-OOD rice subspecies, with especially clear accuracy gains on rufipogon, barthii, and glaberrima. Accuracy / AUROC are shown below. Percentage badges report relative gains over the strongest non-OpticalDNA baseline for each metric; for glaberrima AUROC, OpticalDNA is slightly below LucaOne but still improves over Evo-2. OpticalDNA uses 409M parameters, compared with 7B for Evo-2 and 1.8B for LucaOne.
| Model | japonica In-domain | aus Near-OOD | rufipogon Mid-OOD | barthii Far-OOD | glaberrima Far-OOD |
|---|---|---|---|---|---|
| Evo-2 (7B) | 0.486 / 0.700 | 0.509 / 0.714 | 0.500 / 0.714 | 0.532 / 0.725 | 0.489 / 0.705 |
| LucaOne (1.8B) | 0.510 / 0.703 | 0.551 / 0.723 | 0.589 / 0.760 | 0.556 / 0.745 | 0.526 / 0.736 |
| OpticalDNA (409M) | 0.590 / 0.739+15.7% Acc+5.1% AUROC | 0.556 / 0.725+0.9% Acc+0.3% AUROC | 0.639 / 0.762+8.5% Acc+0.3% AUROC | 0.608 / 0.747+9.4% Acc+0.3% AUROC | 0.599 / 0.731+13.9% Acc+3.7% AUROC vs. Evo-2 |
Best accuracyon all 5 splits
Best AUROCon 4 of 5 splits
+13.9%accuracy gain on glaberrima vs. LucaOne
17.1×fewer parameters than Evo-2
RiceWGPB: whole-genome phenotype prediction
RMSE and inference time under ~400M-base rice genome inputs. The original paper reports OpticalDNA inference time as 12.3 minutes.
| Model | TGW RMSE ↓ | LRI-15SZ RMSE ↓ | Inference time ↓ | Speed note |
|---|---|---|---|---|
| Evo-2 (7B) | 3.056 | 9.617 | 5h 40m | OpticalDNA is 27.6× faster |
| LucaOne (1.8B) | 8.817 | 9.740 | 32.5 min | OpticalDNA is 2.6× faster |
| OpticalDNA (409M) | 2.952best | 9.531best | 12.3 minfastest | best RMSE on both traits |
Interpretability. Grad-CAM highlights region-level attributions and biologically meaningful splice signals, showing that OpticalDNA can localize relevant genomic evidence.
Visual compression remains stable across rendering resolutions
Average AUROC on nine eQTL tasks changes minimally while the visual-token compression ratio increases.
| Resolution | #Pages | #Tokens / page | Compression ratio | Average AUROC ↑ |
|---|---|---|---|---|
| 512 | 370 | 64 | 19.0 | 0.849 |
| 640 | 226 | 100 | 19.9 | 0.852best |
| 1024 | 84 | 256 | 20.9 | 0.851 |
| 1280 | 53 | 400 | 21.2 | 0.849 |
Rendering Robustness: OpticalDNA is not tied to a single visual layout
Key finding. OpticalDNA remains robust under realistic rendering perturbations, including text density, aspect ratio, starting position, and font changes. Dense and wide+dense layouts even improve over the original AS eQTL AUROC, suggesting that the OCR-style formulation captures genomic content rather than overfitting to a brittle page style.
0.858best AUROC under wide+dense rendering
0.839 / 0.835robust under sans / serif font changes
0.812–0.828stable under moderate downsampling
Fixed windowsmain degradation source, due to artificial fragment boundaries
Rendering perturbations
Figure 5. Common formatting changes preserve performance; wide+dense and font-variant renderings outperform the original layout, while rigid fixed-window layouts degrade performance.
Low-resolution robustness
Figure 6. OpticalDNA remains stable under moderate downsampling and degrades mainly at extreme low resolution, supporting the robustness of visual DNA tokens.
Cross-Species Transfer: visual genomic representations generalize beyond pretraining species
Key finding. HG38-pretrained OpticalDNA transfers to rice far better than HG38-pretrained sequence baselines and nearly matches rice-pretrained OpticalDNA, indicating that the gain comes from a transferable OCR-style visual genomic representation rather than only domain-matched pretraining.
HG38 → rice transfer
Figure 7. OpticalDNA-HG38 clearly surpasses HG38-pretrained sequence baselines and approaches the rice-pretrained OpticalDNA model.
+24.7%Accuracy gain over JanusDNA-HG38
+10.7%AUROC gain over JanusDNA-HG38
+32.5%Accuracy gain over Caduceus-HG38
+31.2%AUROC gain over Caduceus-HG38
Transfer summary
| Model | Accuracy ↑ | AUROC ↑ | Finding |
|---|---|---|---|
| Caduceus (HG38) | 0.449 | 0.568 | HG38-pretrained sequence baseline |
| JanusDNA (HG38) | 0.477 | 0.673 | stronger HG38-pretrained baseline |
| OpticalDNA (HG38) | 0.595+24.7%vs. JanusDNA | 0.745+10.7%vs. JanusDNA | +32.5% Acc and +31.2% AUROC vs. Caduceus |
| OpticalDNA (Rice) | 0.598+25.4%vs. JanusDNA | 0.741+10.1%vs. JanusDNA | +0.5% Acc and −0.5% AUROC vs. OpticalDNA-HG38 |
Takeaway. OpticalDNA-HG38 reaches 0.595 / 0.745 on rice, nearly matching the rice-pretrained reference (0.598 / 0.741) while delivering large gains over HG38-pretrained sequence baselines.
Takeaways
What OpticalDNA changes
A new genomic modeling paradigm
Instead of treating DNA as a flat token sequence, OpticalDNA models genomes as coordinate-indexed visual documents.
Practical long-context scaling
Compact visual tokens reduce the effective token count and make ultra-long sequence prediction more efficient.
Interpretable region access
The visual formulation makes genomic regions explicit, supporting localization, retrieval, and evidence inspection.
BibTeX
@inproceedings{xiang2026rethinking,
title = {Rethinking Genomic Modeling Through Optical Character Recognition},
author = {Xiang, Hongxin and Ma, Pengsen and Cao, Yunkang and Yu, Di and Chen, Haowen and Yang, Xinyu and Zeng, Xiangxiang},
booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
year = {2026},
url = {https://openreview.net/forum?id=nggzekChuU}
}
@misc{xiang2026rethinkinggenomicmodelingoptical,
title = {Rethinking Genomic Modeling Through Optical Character Recognition},
author = {Hongxin Xiang and Pengsen Ma and Yunkang Cao and Di Yu and Haowen Chen and Xinyu Yang and Xiangxiang Zeng},
year = {2026},
eprint = {2602.02014},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2602.02014}
}