OpticalDNA prints genomes as OCR pages and beats Enformer with 20× fewer tokens

Rethinking Genomic Modeling Through Optical Character Recognition

Hongxin Xiang, Pengsen Ma, Yunkang Cao, Di Yu, Haowen Chen, Xinyu Yang, Xiangxiang Zeng

ICML 2026

cs.CV, cs.AI, cs.CL, cs.LG

2026-02-02

OpticalDNA treats DNA as OCR documents. On 450kb eQTL it hits 0.852 AUROC with a 256k probe, beating Enformer (0.681) with nearly 20× fewer tokens.

What problem this solves

Functional DNA is sparse. Regulatory variants, splice sites, and other useful intervals sit inside long stretches of low-information sequence. Genomic foundation models inherited the LLM habit of reading A/T/C/G as a one-dimensional token stream. Most of the compute then goes to background, and long-range jumps never become first-class interval operations.

Compression should be the point of long-context genomics, yet high-fidelity compression needs to know what to keep. HyenaDNA, Caduceus, and Nucleotide Transformer still charge a similar cost per base or k-mer. At 450kb eQTL inputs, NT-v2-500M and GENERator-1.2B cannot take the sequence natively in this setup, so they are chunked and linearly probed. Tasks biologists actually run (locate a variant, retrieve a subsequence, fill a gap) stay implicit in positional encodings.

Method

OpticalDNA, from Hunan University and Yuelushan Laboratory, recasts genomic modeling as OCR-style document understanding. DNA is rasterized in a monospace layout onto 640×640 pages (font size 14, line spacing 1.6, about 1,800 bases per page) and paginated when it overflows. Only A/C/G/T/N appear on the canvas. Every nucleotide keeps a pixel box, so sequence intervals and image regions map both ways. The claim is computational, not biological: DNA is still a 1D polymer, the 2D page is a way to skip instead of scan.

The visual encoder is DeepSeek-OCR's frozen SAM–Conv–CLIP-L front-end: 16×16 patches, a 16× convolutional token downsample, a projector, then a one-layer 20-head fusion that collapses all pages into 100 document tokens. The decoder is DeepSeek-3B MoE with about 570M activated parameters, tuned with LoRA. After pretraining the decoder is dropped. Downstream heads are a 256k linear probe or a 1.3M–2.3M MLP.

Six prompt families train the OCR loop: full-page transcription, grounded reading, ROI reading, masked span completion, subsequence localization, and chromosome classification. Human pretraining samples 2,048-bp windows from HG38 on 8×H100 for 227,600 then 190,000 steps (about 11 days). A rice model is trained for 150,000 steps on the Nipponbare T2T genome.

Results

The main human benchmark is DNALONGBENCH eQTL prediction on 450kb inputs across nine GTEx tissues (AUROC).

MethodTrainable paramsMean AUROC
Enformer252M0.681
HyenaDNA1.6M0.514
Caduceus-Ph7.7M0.750
NT-v2-500M (chunked)1.03K0.772
GENERator-1.2B (chunked)10.24K0.782
JanusDNA7.7M0.791
JanusDNA MLP7.7M0.840
OpticalDNA linear probe256K0.852
OpticalDNA MLP1.3M–2.3M0.867

Raising render resolution from 512 to 1280 grows visual tokens per page from 64 to 400 and the compression ratio from 19.0 to 21.2; mean AUROC stays in 0.849–0.852. On whole blood the MLP variant reaches 0.927 versus JanusDNA MLP at 0.821; on thyroid, 0.876 versus 0.793.

A 409M rice encoder is trained separately. On RiceSubBench splice-site classification, japonica accuracy/AUROC is 0.590/0.739 and far-OOD O. glaberrima is 0.599/0.731, ahead of Evo-2 7B (0.486/0.700 and 0.489/0.705) and LucaOne 1.8B (0.510/0.703 and 0.526/0.736). On whole-genome RiceWGPB (400M bases), thousand-grain weight and leaf-rolling RMSE are 2.952 and 9.531 in 12.3 minutes, against Evo-2 at 3.056/9.617 in 5h40m.

Against an unmodified DeepSeek-OCR backbone, mean eQTL AUROC moves from 0.823 to 0.867. Full-page transcription exact match is 79.6% on HG38 and 74.9% on rice, versus about 1.2% for DeepSeek-OCR. With a matched 10,000-token budget and the same Transformer, the 2D document representation scores 0.626 AUROC against 0.558 for 1D sequence tokens. A lightweight TinyNet-E 2D CNN hits 0.786; the strongest 1D CNN variant in that sweep only reaches 0.649.

Why it matters

For long-range regulation and whole-genome traits, the usual tax is paying full attention on background bases. OpticalDNA is a reproducible detour: print the sequence without loss, compress it visually by about 20×, and adapt with a 256k probe. On 450kb human eQTL that probe already clears a line of sequence foundation models.

This is a change of representation, not a new biological prior. Downstream runs do not even keep the OCR decoder; the working object is the visual encoder plus a small head. Teams already predicting on ultra-long DNA who can tolerate a rendering step have something to try. Short-sequence classification and base-level generation still fit sequence models better.

Limitations

Chromosome classification on 2,048-bp windows is effectively unsolved: 0.062 accuracy on HG38 and 0.104 on rice. The authors say local motifs recur across chromosomes, so short context is not distinctive. Transcription exact match falls to 42.3% near single-page capacity (1,792–2,048 bp). The abstract's "985× fewer activated parameters" compares the 256k probe with Enformer's 252M; the frozen visual encoder is still about 409M. Appendix FLOPs per base are 177.76 MFLOPs for OpticalDNA, above JanusDNA at 17.95 and Caduceus at 27.27, and fine-tuning throughput is 0.12 M bases/s versus Caduceus at 0.52. The savings are in effective tokens and wall-clock against 7B sequence models, not a universal FLOP win.

RiceWGPB has 130 and 102 accessions. Thousand-grain-weight RMSE of 2.952 versus 3.056 is a thin gap on a tiny set. NT-v2 and GENERator are chunked because they cannot take 450kb, so that comparison is asymmetric. Fixed-window layouts drop adipose-subcutaneous AUROC from 0.813 to 0.758–0.774; the representation is sensitive to whether local sequence stays visually continuous. Pretraining is not cheap: about 11 days on 8×H100 for the human model.

Terms

Source

What people are saying

Related papers

All paper explainers