Research map / Vision-language and multimodal / Human and embodied modeling
Rethinking Genomic Modeling Through Optical Character Recognition
Hongxin Xiang, Pengsen Ma, Yunkang Cao, Di Yu, Haowen Chen, Xinyu Yang, xiangxiang Zeng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace linear probe with lightweight MLP head | OpticalDNA · DNALONGBENCH eQTL average | AUROC: 0.852 -> 0.867 | Improved |
| OpticalDNA pretraining on top of DeepSeek-OCR backbone (ablation) | DeepSeek-OCR · DNALONGBENCH eQTL average | AUROC: +5.37% relative | Improved |
| OpticalDNA pretraining vs DeepSeek-OCR on Thyroid tissue (ablation) | DeepSeek-OCR · eQTL Thyroid | AUROC: +16.86% | Improved |
| OpticalDNA pretraining vs DeepSeek-OCR on SNSES tissue (ablation) | DeepSeek-OCR · eQTL SNSES | AUROC: +14.30% | Improved |
| OpticalDNA DNA-specific pretraining for DNA transcription at full length (HG38) (ablation) | DeepSeek-OCR · HG38 transcription, 100% GT prefix | Exact Match (%): 1.2 -> 79.6 | Improved |
| OpticalDNA DNA-specific pretraining for DNA transcription at 10% prefix (HG38) (ablation) | DeepSeek-OCR · HG38 transcription, 10% GT prefix | Exact Match (%): 8.6 -> 97.3 | Improved |
| OpticalDNA pretraining for rice transcription at full length (ablation) | DeepSeek-OCR · Rice transcription, 100% GT prefix | Exact Match (%): 1.2 -> 74.9 | Improved |
| increase rendering resolution from 512 to 1280 (more tokens per page) (ablation) | OpticalDNA · eQTL average | AUROC: 0.849 -> 0.849 | No clear effect |
| render DNA as 2D image document instead of 1D sequence at 10,000 tokens (ablation) | shared Transformer backbone · eQTL 9 tissues | AUROC: 0.558 -> 0.626 | Improved |
| use 2D CNN (TinyNet-E) instead of best 1D CNN (CNN1D-small) (ablation) | CNN backbones · eQTL 9 tissues | AUROC: 0.649 -> 0.786 | Improved |
| use 2D TinyNet-E instead of 1D CNN family for training time (ablation) | CNN backbones · eQTL | Epoch time (s): 386.6-636.8 -> 66.9 | Improved |
| scale 1D CNN from base to xl (ablation) | 1D CNN · eQTL 9 tissues | AUROC: 0.643 -> 0.578 | Hurt |
Closest papers by meaning
- Model Decides How to Tokenize: Adaptive DNA Sequence Tokenization with MxDNA (NeurIPS 2024 · similarity 0.85)
- VQDNA: Unleashing the Power of Vector Quantization for Multi-Species Genomic Sequence Modeling (ICML 2024 · similarity 0.84)
- GeneZip: Region-Aware Compression for Long Context DNA Modeling (NeurIPS 2026 · similarity 0.84)
- BEND: Benchmarking DNA Language Models on Biologically Meaningful Tasks (ICLR 2024 · similarity 0.84)
- What DNA Foundation Models Learn Beyond Sequence Composition (NeurIPS 2026 · similarity 0.84)
- Revisiting Convolution Architecture in the Realm of DNA Foundation Models (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
