Biological Reasoning-Informed Regression for Interpretable Regulatory DNA Activity Prediction

Bing Su; Chuan Cao; Jiwei Zhu; Yi Duan; Ying Ba; Zhao Yang

arxiv: 2606.08147 · v1 · pith:4KKQNSS5new · submitted 2026-06-06 · 🧬 q-bio.GN · cs.LG

Biological Reasoning-Informed Regression for Interpretable Regulatory DNA Activity Prediction

Yi Duan , Zhao Yang , Jiwei Zhu , Ying Ba , Chuan Cao , Bing Su This is my paper

Pith reviewed 2026-06-27 18:59 UTC · model grok-4.3

classification 🧬 q-bio.GN cs.LG

keywords regulatory DNAenhancer predictionlarge language modelsinterpretable regressioncis-regulatory elementsmechanistic reasoningCRE-ReasonBenchgene expression prediction

0 comments

The pith

R3LM teaches language models to reason over structured biological data before regressing regulatory DNA activity.

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

The paper presents R3LM as a way to make language models predict how DNA sequences control gene expression by first giving them structured regulatory information and explicit mechanistic reasoning traces. It creates a new dataset called CRE-ReasonBench that pairs sequences and activity scores with step-by-step biological explanations, then trains models in two stages: reasoning first, then numerical regression. This setup is intended to fix the black-box limitation of direct sequence-to-score models and the weak results from feeding raw DNA to language models. If the approach works, predictions become both more accurate across cell types and accompanied by explanations that biologists can inspect and use for designing regulatory elements.

Core claim

By structuring DNA regulatory information into a biologically grounded format and building CRE-ReasonBench to link sequences, activity scores, and mechanistic reasoning traces, two-stage training first teaches language models to reason over the structured knowledge and then to perform regression, yielding state-of-the-art enhancer activity prediction across three cell types while generating interpretable explanations that outperform both raw-sequence language models and specialized DNA models.

What carries the argument

The R3LM framework, which uses a biologically grounded data format for regulatory sequences together with reasoning traces from the CRE-ReasonBench dataset and applies two-stage training to enable reasoning-informed regression in language models.

If this is right

Enhancer activity prediction reaches higher accuracy across three cell types than either raw-sequence language models or existing specialized DNA models.
Predictions are accompanied by mechanistic reasoning traces that make the outputs interpretable.
The trained model can function as an interpretable reward model for assisting in cis-regulatory element design.
Language models can be adapted to biological regression tasks through added reasoning supervision rather than raw sequence input alone.

Where Pith is reading between the lines

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

The same combination of structured biological formatting and reasoning traces could be applied to other sequence-to-function tasks such as promoter strength or variant effect prediction.
If the performance gain comes mainly from the reasoning stage, general language models might reduce reliance on task-specific DNA architectures when similar datasets exist.
Explanations generated by the model could be checked against experimental assays to iteratively improve the quality of the reasoning traces in future datasets.

Load-bearing premise

The assumption that supplying a structured biological format plus mechanistic reasoning traces will let language models reach higher regression accuracy and produce accurate explanations after two-stage training.

What would settle it

After two-stage training on CRE-ReasonBench, the resulting model fails to exceed baseline performance on enhancer activity prediction in the three tested cell types or produces mechanistic explanations that do not align with known regulatory biology on held-out sequences.

Figures

Figures reproduced from arXiv: 2606.08147 by Bing Su, Chuan Cao, Jiwei Zhu, Yi Duan, Ying Ba, Zhao Yang.

**Figure 2.** Figure 2: Stage-1 generator diagnostics and CREReasonBench statistics across three cell types. (a) Distribution of discrete activity level prediction error Δ = ˆℓ − ℓ for the stage-1 reasoning generator. (b) Motif count distribution in the RCC prompts. (c) Distribution of rationale tokens length. (d) Motif coverage of rationales, i.e., the fraction of RCC prompt motifs explicitly referenced in the generated ratio… view at source ↗

read the original abstract

DNA cis-regulatory elements (CREs) such as enhancers control gene expression levels. Accurately predicting regulatory activity from DNA sequences is valuable but challenging, as it requires understanding complex biological regulatory processes. Existing methods typically regress activity scores from sequences in a black-box manner, limiting both interpretability and regression performance. Meanwhile, large language models (LLMs) benefit from explicit reasoning processes, yet directly applying LLMs to raw DNA sequences performs poorly. In this paper, we bridge this gap by introducing R3LM, a framework that teaches LLMs reasoning-informed regression on regulatory DNA through structured biological knowledge. Specifically, we design a biologically grounded data format that structures DNA's regulatory information for improved LLM understanding, and construct CRE-ReasonBench, the first dataset that associates DNA sequences and activity scores with mechanistic reasoning traces. Through two-stage training that first teaches LLMs reasoning over structured biological information then performs regression, R3LM achieves state-of-the-art performance on enhancer prediction across three cell types, outperforming both LLMs with raw sequence input and specialized DNA models while providing interpretable mechanistic explanations. We expect R3LM as an interpretable reward model that can effectively assist biologists in CRE design. Code is available at https://github.com/DuanYi516/R3LM.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper introduces a new reasoning-trace dataset and two-stage LLM training for interpretable DNA regulatory activity prediction, but lacks visible metrics to support its SOTA claims.

read the letter

This paper's main contribution is the CRE-ReasonBench dataset, which pairs regulatory DNA sequences with activity scores and mechanistic reasoning traces, along with a two-stage training procedure for LLMs that first focuses on reasoning over structured biological information before doing regression. That combination is not in the prior work they cite.

The paper does a good job identifying the limitations of black-box regression on DNA sequences and of feeding raw sequences to LLMs. Structuring the input with biological knowledge and requiring explicit reasoning steps is a logical step toward interpretability. The idea that this could serve as a reward model for CRE design makes sense for practical use in biology.

Where it falls short is the lack of supporting details in the abstract. It claims state-of-the-art performance on enhancer prediction across three cell types but does not report any metrics, baselines, or significance tests. This makes it impossible to judge whether the gains are real or how large they are. The construction of the reasoning traces in the dataset is also not described, which leaves open questions about reproducibility and potential biases in how the biological knowledge was encoded.

The two-stage training is presented as key, but without ablations showing the contribution of the reasoning stage versus just the structured format, it's hard to know what drives any improvement. The assumption that biologically grounded reasoning will improve both accuracy and explanation quality is plausible but needs empirical backing that isn't visible here.

This work is aimed at researchers in regulatory genomics who are exploring LLM-based methods for sequence-to-function prediction. Someone looking for new benchmarks or ways to add interpretability to DNA models would find it relevant. It deserves a serious referee because the core idea is distinct from existing approaches and could be valuable if the results hold.

I recommend sending it to peer review. The full paper likely contains the missing experimental details, and referees can assess whether the dataset and training procedure deliver on the claims.

Referee Report

1 major / 0 minor

Summary. The paper introduces R3LM, a framework that teaches LLMs reasoning-informed regression for predicting regulatory DNA (enhancer) activity. It uses a biologically grounded structured data format for DNA sequences and constructs CRE-ReasonBench, a dataset pairing sequences and activity scores with mechanistic reasoning traces. A two-stage training process first teaches reasoning over the structured information then performs regression; the abstract claims this yields state-of-the-art performance across three cell types, outperforming both raw-sequence LLMs and specialized DNA models while supplying interpretable mechanistic explanations usable as a reward model for CRE design.

Significance. If the central performance and interpretability claims hold with rigorous evidence, the work would be a meaningful contribution to regulatory genomics by moving beyond black-box sequence regression toward biologically grounded LLM reasoning, while also releasing a new reasoning-annotated benchmark dataset.

major comments (1)

[Abstract] Abstract: the claim that R3LM 'achieves state-of-the-art performance on enhancer prediction across three cell types, outperforming both LLMs with raw sequence input and specialized DNA models' is load-bearing for the paper's central contribution, yet the abstract supplies no quantitative metrics, baselines, statistical tests, or experimental details to support it.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their review. We address the single major comment below and agree that the abstract requires strengthening with quantitative details.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that R3LM 'achieves state-of-the-art performance on enhancer prediction across three cell types, outperforming both LLMs with raw sequence input and specialized DNA models' is load-bearing for the paper's central contribution, yet the abstract supplies no quantitative metrics, baselines, statistical tests, or experimental details to support it.

Authors: We agree that the abstract would be strengthened by including concrete quantitative support for the performance claims. In the revised manuscript we will update the abstract to report the key Pearson correlation values achieved by R3LM on the three cell types, the specific baselines (both raw-sequence LLMs and specialized DNA models) against which it was compared, and a brief statement of the evaluation protocol and statistical significance. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper presents an empirical ML framework: construction of CRE-ReasonBench dataset with reasoning traces, a biologically structured data format, and two-stage LLM training for regression. No equations, fitted parameters renamed as predictions, self-definitional loops, or load-bearing self-citations appear in the provided text. Central claims rest on experimental SOTA results rather than any reduction to inputs by construction. This is a standard non-circular empirical training setup.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the effectiveness of the biologically grounded data format and the utility of mechanistic reasoning traces for LLM training; these are introduced in the paper rather than derived from prior benchmarks.

free parameters (1)

two-stage training hyperparameters
Standard LLM fine-tuning choices whose specific values are not stated in the abstract.

axioms (1)

domain assumption Explicit reasoning processes improve LLM performance on complex tasks
Invoked in the abstract as the motivation for the two-stage approach.

pith-pipeline@v0.9.1-grok · 5765 in / 1180 out tokens · 24537 ms · 2026-06-27T18:59:55.872730+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

44 extracted references · 7 canonical work pages

[1]

Žiga Avsec, Vikram Agarwal, Daniel Visentin, Joseph R Ledsam, Agnieszka Grabska-Barwinska, Kyle R Taylor, Yannis Assael, John Jumper, Pushmeet Kohli, and David R Kelley. 2021. Effective gene expression prediction from sequence by integrating long-range interactions.Nature methods18, 10 (2021), 1196–1203

2021
[2]

Žiga Avsec, Natasha Latysheva, Jun Cheng, Guido Novati, Kyle R Taylor, Tom Ward, Clare Bycroft, Lauren Nicolaisen, Eirini Arvaniti, Joshua Pan, et al. 2026. Advancing regulatory variant effect prediction with AlphaGenome.Nature649, 8099 (2026), 1206–1218

2026
[3]

Ying Ba, Tianyu Zhang, Yalong Bai, Wenyi Mo, Tao Liang, Bing Su, and Ji- Rong Wen. 2025. Enhancing Reward Models for High-quality Image Generation: Beyond Text-Image Alignment. arXiv:2507.19002 [cs.CV] https://arxiv.org/abs/ 2507.19002

arXiv 2025
[4]

Timothy L Bailey, James Johnson, Charles E Grant, and William S Noble. 2015. The MEME suite.Nucleic acids research43, W1 (2015), W39–W49

2015
[5]

Xingyu Chen, Shihao Ma, Runsheng Lin, Jiecong Lin, and Bo Wang. 2025. Ctrl- DNA: Controllable Cell-Type-Specific Regulatory DNA Design via Constrained RL. arXiv:2505.20578 [cs.LG] https://arxiv.org/abs/2505.20578

arXiv 2025
[6]

Cheng-Han Chiang, Hung yi Lee, and Michal Lukasik. 2025. TRACT: Regression- Aware Fine-tuning Meets Chain-of-Thought Reasoning for LLM-as-a-Judge. arXiv:2503.04381 [cs.CL] https://arxiv.org/abs/2503.04381

arXiv 2025
[7]

de Almeida, Hassan Sirelkhatim, Guillaume Richard, Marcin Skwark, Karim Beguir, Marie Lopez, and Thomas Pierrot

Hugo Dalla-Torre, Liam Gonzalez, Javier Mendoza-Revilla, Nicolas Lopez Car- ranza, Adam Henryk Grzywaczewski, Francesco Oteri, Christian Dallago, Evan Trop, Bernardo P. de Almeida, Hassan Sirelkhatim, Guillaume Richard, Marcin Skwark, Karim Beguir, Marie Lopez, and Thomas Pierrot. 2025. Nucleotide Trans- former: building and evaluating robust foundation m...

work page doi:10.1038/s41592-024-02523-z 2025
[8]

Bernardo P. de Almeida, Guillaume Richard, Hugo Dalla-Torre, Christopher Blum, Lorenz Hexemer, Priyanka Pandey, Stefan Laurent, Chandana Rajesh, Marie Lopez, Alexandre Laterre, Maren Lang, Uğur Şahin, Karim Beguir, and Thomas Pierrot. 2025. A multimodal conversational agent for DNA, RNA and protein tasks.Nature Machine Intelligence7, 6 (01 Jun 2025), 928–...

work page doi:10.1038/s42256- 2025
[9]

Bernardo P de Almeida, Christoph Schaub, Michaela Pagani, Stefano Secchia, Eileen EM Furlong, and Alexander Stark. 2024. Targeted design of synthetic enhancers for selected tissues in the Drosophila embryo.Nature626, 7997 (2024), 207–211

2024
[10]

Carl G de Boer, Eeshit Dhaval Vaishnav, Ronen Sadeh, Esteban Luis Abeyta, Nir Friedman, and Aviv Regev. 2020. Deciphering eukaryotic gene-regulatory logic with 100 million random promoters.Nature biotechnology38, 1 (2020), 56–65

2020
[11]

Seppe De Winter, Vasileios Konstantakos, and Stein Aerts. 2025. Modelling and design of transcriptional enhancers.Nature Reviews Bioengineering3 (2025), 374–389

2025
[12]

Kseniia Dudnyk, Donghong Cai, Chenlai Shi, Jian Xu, and Jian Zhou. 2024. Sequence basis of transcription initiation in the human genome.Science384, 6694 (2024), eadj0116

2024
[13]

Maddison, and Bo Wang

Adibvafa Fallahpour, Andrew Magnuson, Purav Gupta, Shihao Ma, Jack Naimer, Arnav Shah, Haonan Duan, Omar Ibrahim, Hani Goodarzi, Chris J. Maddison, and Bo Wang. 2025. BioReason: Incentivizing Multimodal Biological Reasoning within a DNA-LLM Model. arXiv:2505.23579 [cs.LG] https://arxiv.org/abs/2505.23579

arXiv 2025
[14]

Tianshun Gao and Jiang Qian. 2020. EnhancerAtlas 2.0: an updated resource with enhancer annotation in 586 tissue/cell types across nine species.Nucleic acids research48, D1 (2020), D58–D64

2020
[15]

Ilias Georgakopoulos-Soares, Chengyu Deng, Vikram Agarwal, Candace SY Chan, Jingjing Zhao, Fumitaka Inoue, and Nadav Ahituv. 2023. Transcription factor binding site orientation and order are major drivers of gene regulatory activity. Nature communications14, 1 (2023), 2333

2023
[16]

Laura H Goetz and Nicholas J Schork. 2018. Personalized medicine: motivation, challenges, and progress.Fertility and sterility109, 6 (2018), 952–963

2018
[17]

Sager J Gosai, Rodrigo I Castro, Natalia Fuentes, John C Butts, Kousuke Mouri, Michael Alasoadura, Susan Kales, Thanh Thanh L Nguyen, Ramil R Noche, Arya S Rao, et al. 2024. Machine-guided design of cell-type-targeting cis-regulatory elements.Nature634, 8036 (2024), 1211–1220

2024
[18]

Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, ...

work page doi:10.1038/s41586-025-09422-z 2026
[19]

Jiaxin Guo, Zewen Chi, Li Dong, Qingxiu Dong, Xun Wu, Shaohan Huang, and Furu Wei. 2025. Reward Reasoning Model. arXiv:2505.14674 [cs.CL] https: //arxiv.org/abs/2505.14674

arXiv 2025
[20]

Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2023. Large Language Models are Zero-Shot Reasoners. arXiv:2205.11916 [cs.CL] https://arxiv.org/abs/2205.11916

Pith/arXiv arXiv 2023
[21]

Avantika Lal, David Garfield, Tommaso Biancalani, and Gokcen Eraslan. 2024. De- signing realistic regulatory DNA with autoregressive language models.Genome Research34, 9 (2024), 1411–1420

2024
[22]

Zehui Li, Vallijah Subasri, Yifei Shen, Dongsheng Li, Wentao Gu, Guy-Bart Stan, Yiren Zhao, and Caihua Shan. 2026. Omni-DNA: a Genomic model supporting sequence understanding, long-context, and textual annotation.Advances in Neural Information Processing Systems38 (2026), 133044–133072

2026
[23]

Bo Liu, Xingchao Liu, Xiaojie Jin, Peter Stone, and Qiang Liu. 2024. Conflict- Averse Gradient Descent for Multi-task Learning. arXiv:2110.14048 [cs.LG] https://arxiv.org/abs/2110.14048

arXiv 2024
[24]

Aaron J Lorenz, Shiaoman Chao, Franco G Asoro, Elliot L Heffner, Takeshi Hayashi, Hiroyoshi Iwata, Kevin P Smith, Mark E Sorrells, and Jean-Luc Jannink
[25]

Genomic selection in plant breeding: knowledge and prospects.Advances in agronomy110 (2011), 77–123

2011
[26]

Damla Ovek Baydar, Ieva Rauluseviciute, Dina R Aronsen, Romain Blanc-Mathieu, Ine Bonthuis, Herman de Beukelaer, Katalin Ferenc, Alice Jegou, Vipin Ku- mar, Roza Berhanu Lemma, Jérémy Lucas, Mathis Pochon, Chang M Yun, Vivekanandan Ramalingam, Salil Sanjay Deshpande, Aman Patel, Georgi K Marinov, Austin T Wang, Alejandro Aguirre, Jaime A Castro-Mondragon,...

work page doi:10.1093/nar/gkaf1209 2025
[27]

Tomas Pachano, Víctor Sánchez-Gaya, Thais Ealo, Maria Mariner-Faulí, Tore Bleckwehl, {Helena G.} Asenjo, Patricia Respuela, Sara Cruz-Molina, María {Muñoz-San Martín}, Endika Haro, {Wilfred F.J.} {van IJcken}, David Landeira, and Alvaro Rada-Iglesias. 2021. Orphan CpG islands amplify poised enhancer regulatory activity and determine target gene responsive...

work page doi:10.1038/s41588-021-00888-x 2021
[28]

Aniketh Janardhan Reddy, Xinyang Geng, Michael Herschl, Sathvik Kolli, Aviral Kumar, Patrick Hsu, Sergey Levine, and Nilah Ioannidis. 2024. Designing cell- type-specific promoter sequences using conservative model-based optimization. Advances in Neural Information Processing Systems37 (2024), 93033–93059

2024
[29]

Aniketh Janardhan Reddy, Michael H Herschl, Xinyang Geng, Sathvik Kolli, Amy X Lu, Aviral Kumar, Patrick D Hsu, Sergey Levine, and Nilah M Ioannidis
[30]

2023–02 pages

Strategies for effectively modelling promoter-driven gene expression using transfer learning. 2023–02 pages. doi:10.1101/2023.02.24.529941

work page doi:10.1101/2023.02.24.529941 2023
[31]

Eilon Sharon, Yael Kalma, Ayala Sharp, Tali Raveh-Sadka, Michal Levo, Danny Zeevi, Leeat Keren, Zohar Yakhini, Adina Weinberger, and Eran Segal. 2012. Infer- ring gene regulatory logic from high-throughput measurements of thousands of systematically designed promoters.Nature biotechnology30, 6 (2012), 521–530

2012
[32]

Ibrahim I Taskiran, Katina I Spanier, Hannah Dickmänken, Niklas Kempynck, Alexandra Pančíková, Eren Can Ekşi, Gert Hulselmans, Joy N Ismail, Koen Theu- nis, Roel Vandepoel, et al. 2024. Cell-type-directed design of synthetic enhancers. Nature626, 7997 (2024), 212–220

2024
[33]

Eeshit Dhaval Vaishnav, Carl G de Boer, Jennifer Molinet, Moran Yassour, Lin Fan, Xian Adiconis, Dawn A Thompson, Joshua Z Levin, Francisco A Cubillos, and Aviv Regev. 2022. The evolution, evolvability and engineering of gene regulatory DNA.Nature603, 7901 (2022), 455–463

2022
[34]

Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-Consistency Improves Chain of Thought Reasoning in Language Models. arXiv:2203.11171 [cs.CL] https://arxiv.org/abs/2203.11171

Pith/arXiv arXiv 2023
[35]

Yajie Wang, Pu Xue, Mingfeng Cao, Tianhao Yu, Stephan T Lane, and Huimin Zhao. 2021. Directed evolution: methodologies and applications.Chemical reviews 121, 20 (2021), 12384–12444

2021
[36]

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv:2201.11903 [cs.CL] https: //arxiv.org/abs/2201.11903

Pith/arXiv arXiv 2023
[37]

Anouck Wijgaerts, Christine Wittevrongel, Chantal Thys, Timothy Devos, Kathe- lijne Peerlinck, Marloes R Tijssen, Chris Van Geet, and Kathleen Freson. 2017. The transcription factor GATA1 regulates NBEAL2 expression through a long- distance enhancer.Haematologica102, 4 (2017), 695

2017
[38]

An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, ...

Pith/arXiv arXiv 2025
[39]

Zhao Yang, Bing Su, Chuan Cao, and Ji-Rong Wen. 2025. Regulatory DNA Sequence Design with Reinforcement Learning. InInternational Conference on Learning Representations, Y. Yue, A. Garg, N. Peng, F. Sha, and R. Yu (Eds.), Vol. 2025. 73499–73524. https://proceedings.iclr.cc/paper_files/paper/2025/file/ b65b7e0d910fbb9f05b2c1d241e7ade1-Paper-Conference.pdf

2025
[40]

Zhao Yang, Jiwei Zhu, and Bing Su. 2025. SPACE: Your Genomic Profile Predictor is a Powerful DNA Foundation Model. InForty-second International Conference on Machine Learning

2025
[41]

Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, and Zheyan Luo
[42]

doi: 10.18653/v1/2024.acl-demos.38

LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Yixin Cao, Yang Feng, and Deyi Xiong (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 400–410. doi:10.18653/v1/2024.acl-demos.38

work page doi:10.18653/v1/2024.acl-demos.38 2024
[43]

Zhihan Zhou, Yanrong Ji, Weijian Li, Pratik Dutta, Ramana Davuluri, and Han Liu
[44]

Step 1–Step 7

DNABERT-2: Efficient Foundation Model and Benchmark For Multi-Species Genome. arXiv:2306.15006 [q-bio.GN] https://arxiv.org/abs/2306.15006 A Training Details Data split.All training/validation/test samples follow the official regLM split. Stage-I supervised fine-tuning.For each cell type, we curate 1,000 instruction-following training instances generated ...

arXiv 2026

[1] [1]

Žiga Avsec, Vikram Agarwal, Daniel Visentin, Joseph R Ledsam, Agnieszka Grabska-Barwinska, Kyle R Taylor, Yannis Assael, John Jumper, Pushmeet Kohli, and David R Kelley. 2021. Effective gene expression prediction from sequence by integrating long-range interactions.Nature methods18, 10 (2021), 1196–1203

2021

[2] [2]

Žiga Avsec, Natasha Latysheva, Jun Cheng, Guido Novati, Kyle R Taylor, Tom Ward, Clare Bycroft, Lauren Nicolaisen, Eirini Arvaniti, Joshua Pan, et al. 2026. Advancing regulatory variant effect prediction with AlphaGenome.Nature649, 8099 (2026), 1206–1218

2026

[3] [3]

Ying Ba, Tianyu Zhang, Yalong Bai, Wenyi Mo, Tao Liang, Bing Su, and Ji- Rong Wen. 2025. Enhancing Reward Models for High-quality Image Generation: Beyond Text-Image Alignment. arXiv:2507.19002 [cs.CV] https://arxiv.org/abs/ 2507.19002

arXiv 2025

[4] [4]

Timothy L Bailey, James Johnson, Charles E Grant, and William S Noble. 2015. The MEME suite.Nucleic acids research43, W1 (2015), W39–W49

2015

[5] [5]

Xingyu Chen, Shihao Ma, Runsheng Lin, Jiecong Lin, and Bo Wang. 2025. Ctrl- DNA: Controllable Cell-Type-Specific Regulatory DNA Design via Constrained RL. arXiv:2505.20578 [cs.LG] https://arxiv.org/abs/2505.20578

arXiv 2025

[6] [6]

Cheng-Han Chiang, Hung yi Lee, and Michal Lukasik. 2025. TRACT: Regression- Aware Fine-tuning Meets Chain-of-Thought Reasoning for LLM-as-a-Judge. arXiv:2503.04381 [cs.CL] https://arxiv.org/abs/2503.04381

arXiv 2025

[7] [7]

de Almeida, Hassan Sirelkhatim, Guillaume Richard, Marcin Skwark, Karim Beguir, Marie Lopez, and Thomas Pierrot

Hugo Dalla-Torre, Liam Gonzalez, Javier Mendoza-Revilla, Nicolas Lopez Car- ranza, Adam Henryk Grzywaczewski, Francesco Oteri, Christian Dallago, Evan Trop, Bernardo P. de Almeida, Hassan Sirelkhatim, Guillaume Richard, Marcin Skwark, Karim Beguir, Marie Lopez, and Thomas Pierrot. 2025. Nucleotide Trans- former: building and evaluating robust foundation m...

work page doi:10.1038/s41592-024-02523-z 2025

[8] [8]

Bernardo P. de Almeida, Guillaume Richard, Hugo Dalla-Torre, Christopher Blum, Lorenz Hexemer, Priyanka Pandey, Stefan Laurent, Chandana Rajesh, Marie Lopez, Alexandre Laterre, Maren Lang, Uğur Şahin, Karim Beguir, and Thomas Pierrot. 2025. A multimodal conversational agent for DNA, RNA and protein tasks.Nature Machine Intelligence7, 6 (01 Jun 2025), 928–...

work page doi:10.1038/s42256- 2025

[9] [9]

Bernardo P de Almeida, Christoph Schaub, Michaela Pagani, Stefano Secchia, Eileen EM Furlong, and Alexander Stark. 2024. Targeted design of synthetic enhancers for selected tissues in the Drosophila embryo.Nature626, 7997 (2024), 207–211

2024

[10] [10]

Carl G de Boer, Eeshit Dhaval Vaishnav, Ronen Sadeh, Esteban Luis Abeyta, Nir Friedman, and Aviv Regev. 2020. Deciphering eukaryotic gene-regulatory logic with 100 million random promoters.Nature biotechnology38, 1 (2020), 56–65

2020

[11] [11]

Seppe De Winter, Vasileios Konstantakos, and Stein Aerts. 2025. Modelling and design of transcriptional enhancers.Nature Reviews Bioengineering3 (2025), 374–389

2025

[12] [12]

Kseniia Dudnyk, Donghong Cai, Chenlai Shi, Jian Xu, and Jian Zhou. 2024. Sequence basis of transcription initiation in the human genome.Science384, 6694 (2024), eadj0116

2024

[13] [13]

Maddison, and Bo Wang

Adibvafa Fallahpour, Andrew Magnuson, Purav Gupta, Shihao Ma, Jack Naimer, Arnav Shah, Haonan Duan, Omar Ibrahim, Hani Goodarzi, Chris J. Maddison, and Bo Wang. 2025. BioReason: Incentivizing Multimodal Biological Reasoning within a DNA-LLM Model. arXiv:2505.23579 [cs.LG] https://arxiv.org/abs/2505.23579

arXiv 2025

[14] [14]

Tianshun Gao and Jiang Qian. 2020. EnhancerAtlas 2.0: an updated resource with enhancer annotation in 586 tissue/cell types across nine species.Nucleic acids research48, D1 (2020), D58–D64

2020

[15] [15]

Ilias Georgakopoulos-Soares, Chengyu Deng, Vikram Agarwal, Candace SY Chan, Jingjing Zhao, Fumitaka Inoue, and Nadav Ahituv. 2023. Transcription factor binding site orientation and order are major drivers of gene regulatory activity. Nature communications14, 1 (2023), 2333

2023

[16] [16]

Laura H Goetz and Nicholas J Schork. 2018. Personalized medicine: motivation, challenges, and progress.Fertility and sterility109, 6 (2018), 952–963

2018

[17] [17]

Sager J Gosai, Rodrigo I Castro, Natalia Fuentes, John C Butts, Kousuke Mouri, Michael Alasoadura, Susan Kales, Thanh Thanh L Nguyen, Ramil R Noche, Arya S Rao, et al. 2024. Machine-guided design of cell-type-targeting cis-regulatory elements.Nature634, 8036 (2024), 1211–1220

2024

[18] [18]

Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, ...

work page doi:10.1038/s41586-025-09422-z 2026

[19] [19]

Jiaxin Guo, Zewen Chi, Li Dong, Qingxiu Dong, Xun Wu, Shaohan Huang, and Furu Wei. 2025. Reward Reasoning Model. arXiv:2505.14674 [cs.CL] https: //arxiv.org/abs/2505.14674

arXiv 2025

[20] [20]

Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2023. Large Language Models are Zero-Shot Reasoners. arXiv:2205.11916 [cs.CL] https://arxiv.org/abs/2205.11916

Pith/arXiv arXiv 2023

[21] [21]

Avantika Lal, David Garfield, Tommaso Biancalani, and Gokcen Eraslan. 2024. De- signing realistic regulatory DNA with autoregressive language models.Genome Research34, 9 (2024), 1411–1420

2024

[22] [22]

Zehui Li, Vallijah Subasri, Yifei Shen, Dongsheng Li, Wentao Gu, Guy-Bart Stan, Yiren Zhao, and Caihua Shan. 2026. Omni-DNA: a Genomic model supporting sequence understanding, long-context, and textual annotation.Advances in Neural Information Processing Systems38 (2026), 133044–133072

2026

[23] [23]

Bo Liu, Xingchao Liu, Xiaojie Jin, Peter Stone, and Qiang Liu. 2024. Conflict- Averse Gradient Descent for Multi-task Learning. arXiv:2110.14048 [cs.LG] https://arxiv.org/abs/2110.14048

arXiv 2024

[24] [24]

Aaron J Lorenz, Shiaoman Chao, Franco G Asoro, Elliot L Heffner, Takeshi Hayashi, Hiroyoshi Iwata, Kevin P Smith, Mark E Sorrells, and Jean-Luc Jannink

[25] [25]

Genomic selection in plant breeding: knowledge and prospects.Advances in agronomy110 (2011), 77–123

2011

[26] [26]

Damla Ovek Baydar, Ieva Rauluseviciute, Dina R Aronsen, Romain Blanc-Mathieu, Ine Bonthuis, Herman de Beukelaer, Katalin Ferenc, Alice Jegou, Vipin Ku- mar, Roza Berhanu Lemma, Jérémy Lucas, Mathis Pochon, Chang M Yun, Vivekanandan Ramalingam, Salil Sanjay Deshpande, Aman Patel, Georgi K Marinov, Austin T Wang, Alejandro Aguirre, Jaime A Castro-Mondragon,...

work page doi:10.1093/nar/gkaf1209 2025

[27] [27]

Tomas Pachano, Víctor Sánchez-Gaya, Thais Ealo, Maria Mariner-Faulí, Tore Bleckwehl, {Helena G.} Asenjo, Patricia Respuela, Sara Cruz-Molina, María {Muñoz-San Martín}, Endika Haro, {Wilfred F.J.} {van IJcken}, David Landeira, and Alvaro Rada-Iglesias. 2021. Orphan CpG islands amplify poised enhancer regulatory activity and determine target gene responsive...

work page doi:10.1038/s41588-021-00888-x 2021

[28] [28]

Aniketh Janardhan Reddy, Xinyang Geng, Michael Herschl, Sathvik Kolli, Aviral Kumar, Patrick Hsu, Sergey Levine, and Nilah Ioannidis. 2024. Designing cell- type-specific promoter sequences using conservative model-based optimization. Advances in Neural Information Processing Systems37 (2024), 93033–93059

2024

[29] [29]

Aniketh Janardhan Reddy, Michael H Herschl, Xinyang Geng, Sathvik Kolli, Amy X Lu, Aviral Kumar, Patrick D Hsu, Sergey Levine, and Nilah M Ioannidis

[30] [30]

2023–02 pages

Strategies for effectively modelling promoter-driven gene expression using transfer learning. 2023–02 pages. doi:10.1101/2023.02.24.529941

work page doi:10.1101/2023.02.24.529941 2023

[31] [31]

Eilon Sharon, Yael Kalma, Ayala Sharp, Tali Raveh-Sadka, Michal Levo, Danny Zeevi, Leeat Keren, Zohar Yakhini, Adina Weinberger, and Eran Segal. 2012. Infer- ring gene regulatory logic from high-throughput measurements of thousands of systematically designed promoters.Nature biotechnology30, 6 (2012), 521–530

2012

[32] [32]

Ibrahim I Taskiran, Katina I Spanier, Hannah Dickmänken, Niklas Kempynck, Alexandra Pančíková, Eren Can Ekşi, Gert Hulselmans, Joy N Ismail, Koen Theu- nis, Roel Vandepoel, et al. 2024. Cell-type-directed design of synthetic enhancers. Nature626, 7997 (2024), 212–220

2024

[33] [33]

Eeshit Dhaval Vaishnav, Carl G de Boer, Jennifer Molinet, Moran Yassour, Lin Fan, Xian Adiconis, Dawn A Thompson, Joshua Z Levin, Francisco A Cubillos, and Aviv Regev. 2022. The evolution, evolvability and engineering of gene regulatory DNA.Nature603, 7901 (2022), 455–463

2022

[34] [34]

Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-Consistency Improves Chain of Thought Reasoning in Language Models. arXiv:2203.11171 [cs.CL] https://arxiv.org/abs/2203.11171

Pith/arXiv arXiv 2023

[35] [35]

Yajie Wang, Pu Xue, Mingfeng Cao, Tianhao Yu, Stephan T Lane, and Huimin Zhao. 2021. Directed evolution: methodologies and applications.Chemical reviews 121, 20 (2021), 12384–12444

2021

[36] [36]

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv:2201.11903 [cs.CL] https: //arxiv.org/abs/2201.11903

Pith/arXiv arXiv 2023

[37] [37]

Anouck Wijgaerts, Christine Wittevrongel, Chantal Thys, Timothy Devos, Kathe- lijne Peerlinck, Marloes R Tijssen, Chris Van Geet, and Kathleen Freson. 2017. The transcription factor GATA1 regulates NBEAL2 expression through a long- distance enhancer.Haematologica102, 4 (2017), 695

2017

[38] [38]

An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, ...

Pith/arXiv arXiv 2025

[39] [39]

Zhao Yang, Bing Su, Chuan Cao, and Ji-Rong Wen. 2025. Regulatory DNA Sequence Design with Reinforcement Learning. InInternational Conference on Learning Representations, Y. Yue, A. Garg, N. Peng, F. Sha, and R. Yu (Eds.), Vol. 2025. 73499–73524. https://proceedings.iclr.cc/paper_files/paper/2025/file/ b65b7e0d910fbb9f05b2c1d241e7ade1-Paper-Conference.pdf

2025

[40] [40]

Zhao Yang, Jiwei Zhu, and Bing Su. 2025. SPACE: Your Genomic Profile Predictor is a Powerful DNA Foundation Model. InForty-second International Conference on Machine Learning

2025

[41] [41]

Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, and Zheyan Luo

[42] [42]

doi: 10.18653/v1/2024.acl-demos.38

LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Yixin Cao, Yang Feng, and Deyi Xiong (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 400–410. doi:10.18653/v1/2024.acl-demos.38

work page doi:10.18653/v1/2024.acl-demos.38 2024

[43] [43]

Zhihan Zhou, Yanrong Ji, Weijian Li, Pratik Dutta, Ramana Davuluri, and Han Liu

[44] [44]

Step 1–Step 7

DNABERT-2: Efficient Foundation Model and Benchmark For Multi-Species Genome. arXiv:2306.15006 [q-bio.GN] https://arxiv.org/abs/2306.15006 A Training Details Data split.All training/validation/test samples follow the official regLM split. Stage-I supervised fine-tuning.For each cell type, we curate 1,000 instruction-following training instances generated ...

arXiv 2026