Elamur Research⌕
☰
home / papers / 2026 / AROMA: Augmented Reasoning Over a Multimodal Architecture for Virtual Cell Genetic Perturbation Modeling

AROMA: Augmented Reasoning Over a Multimodal Architecture for Virtual Cell Genetic Perturbation Modeling

Paper · 2026 · Wang Zhenyu, Geyan Ye, Wei Liu et al.

Summary

AROMA tackles gene-level virtual cell modeling. Taking as input a cell line, the gene being perturbed and a gene of interest, the model classifies the latter as upregulated, downregulated or unchanged, and it also explains why. The authors argue that current approaches fall short in three ways. LLMs reason without biological constraints. Expression predictors give no rationale. Retrieval-based systems ignore regulatory topology.

AROMA addresses this by combining three sources of evidence:

  • retrieved textual gene descriptions and path evidence from two purpose-built knowledge graphs;
  • GNN-derived structural embeddings;
  • ESM-2 protein sequence features.

A cross-attention module fuses these to capture the dependency between the perturbation and the target. Training runs in two stages: multimodal supervised fine-tuning, then GRPO reinforcement learning on a new reasoning corpus, PerturbReason.

Across four cell lines, AROMA beats existing methods. It degrades only mildly in zero-shot evaluation on an unseen cell line and stays comparatively stable on rarely studied, weakly connected genes. The released graphs, dataset and model give the virtual cell community reusable infrastructure for interpretable perturbation prediction.

Key points

  • Resources:
    • Gene-KG merges STRING and CORUM, with over 18k nodes and 700k edges.
    • Path-KG is built from Gene Ontology and Reactome, with over 80k nodes and 400k edges.
    • PerturbReason contains more than 498k samples.
  • Architecture: retrieval-augmented LLM reasoning combined with GNN topology embeddings and ESM-2 protein features, fused through cross-attention. Training uses SFT followed by GRPO.
  • Evaluation: experiments cover the K562, HepG2, Jurkat and RPE1 cell lines, plus a zero-shot test on a held-out cell line.
  • Long-tail robustness (RPE1): on low-popularity genes, AROMA scores 0.71 versus 0.56 for a variant without retrieval and multimodal modules. On low-degree genes, it scores 0.69 versus 0.57.
  • GRPO sensitivity and availability: sampling 16 trajectories per query gave the best or tied-best Macro-F1 on all cell lines, so this setting is used in the main experiments. Weights are on Hugging Face (blazerye/AROMA) and code is on GitHub.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🏢TencentShenzhen · China163 works · 21,032 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksComputational Theory and Mathematics · 321 worksBiochemistry, Genetics and Molecular Biology · 263 worksMolecular Biology · 260 works

🧭 Topics 2Research subjects and languages the work is about.Reasoning · 1,685 worksVision & language · 811 works

Cite

Cite as Wang Zhenyu, Geyan Ye, Wei Liu and Man Tat Alexander Ng. 2026. AROMA: Augmented Reasoning Over a Multimodal Architecture for Virtual Cell Genetic Perturbation Modeling. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1594

@inproceedings{zhenyu2026aroma,
  title = {{AROMA: Augmented Reasoning Over a Multimodal Architecture for Virtual Cell Genetic Perturbation Modeling}},
  author = {Wang Zhenyu, Geyan Ye, Wei Liu and Man Tat Alexander Ng},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1594},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1594}
}

References

52 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abhinav K Adduri, Dhruv Gautam, Beatrice Bevilacqua, Alishba Imran, Rohan Shah, Mohsen Naghipourfar, Noam Teyssier, Rajesh Ilango, Sanjay Nagaraj, Mingze Dong, and 1 others. 2025. Predicting cellular responses to perturbation across diverse contexts with state. BioRxiv, pages 2025–06. find
  2. Maria Ahmad, Adam Wolberg, and Chadi I Kahwaji. 2018. Biochemistry, electron transport chain. find
  3. Leif C Andersson, Kenneth Nilsson, and Carl G Gahm-berg. 1979. K562a human erythroleukemic cell line. International journal of cancer, 23(2):143–147. find
  4. Viktoriia A Arzumanian, Olga I Kiseleva, and Ekaterina V Poverennaya. 2021. The curious case of the hepg2 cell line: 40 years of expertise. International journal of molecular sciences, 22(23):13135. find
  5. Michael Ashburner, Catherine A Ball, Judith A Blake, David Botstein, Heather Butler, J Michael Cherry, Allan P Davis, Kara Dolinski, Selina S Dwight, Janan T Eppig, and 1 others. 2000. Gene ontology: tool for the unification of biology. Nature genetics, 25(1):25–29. find
  6. Alan Bundy and Lincoln Wallen. 1984. Breadth-first search. In Catalogue of artificial intelligence tools, pages 13–13. Springer. find
  7. Charlotte Bunne, Yusuf Roohani, Yanay Rosen, Ankit Gupta, Xikun Zhang, Marcel Roed, Theo Alexandrov, Mohammed AlQuraishi, Patricia Brennan, Daniel B Burkhardt, and 1 others. 2024. How to build the virtual cell with artificial intelligence: Priorities and opportunities. Cell, 187(25):7045–7063. find
  8. Pierre Cardol, Marie Lapaille, Pierre Minet, Fabrice Franck, René F Matagne, and Claire Remacle. 2006. Nd3 and nd4l subunits of mitochondrial complex i, both nucleus encoded in chlamydomonas reinhardtii, are required for activity and assembly of the enzyme. Eukaryotic cell, 5(9):1460–1467. find
  9. Yiqun Chen and James Zou. 2024. Genept: a simple but effective foundation model for genes and cells built from chatgpt. bioRxiv, pages 2023–10. find
  10. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arX…
  11. Sarah A Comerford, Zhiguang Huang, Xinlin Du, Yun Wang, Ling Cai, Agnes K Witkiewicz, Holly Walters, Mohammed N Tantawy, Allie Fu, H Charles Manning, and 1 others. 2014. Acetate dependence of tumors. Cell, 159(7):1591–1602. find
  12. David Croft, Gavin Okelly, Guanming Wu, Robin Haw, Marc Gillespie, Lisa Matthews, Michael Caudy, Phani Garapati, Gopal Gopinath, Bijay Jassal, and 1 others. 2010. Reactome: a database of reactions, pathways and biological processes. Nucleic acids research, 39(suppl_1):D691–D697. find
  13. Haotian Cui, Chloe Wang, Hassaan Maan, Kuan Pang, Fengning Luo, Nan Duan, and Bo Wang. 2024. scgpt: toward building a foundation model for single-cell multi-omics using generative ai. Nature methods, 21(8):1470–1480. find
  14. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and 1 others. 2024. The llama 3 herd of models. arXiv e-prints, pages arXiv–2407.
  15. Ahmed Elnaggar, Michael Heinzinger, Christian Dallago, Ghalia Rehawi, Yu Wang, Llion Jones, Tom Gibbs, Tamas Feher, Christoph Angerer, Martin Steinegger, and 1 others. 2021. Prottrans: Toward understanding the language of life through self-supervised learning. IEEE transactions on pattern analysis a… find
  16. Adibvafa Fallahpour, Andrew Magnuson, Purav Gupta, Shihao Ma, Jack Naimer, Arnav Shah, Haonan Duan, Omar Ibrahim, Hani Goodarzi, Chris J Maddison, and 1 others. 2025. Bioreason: Incentivizing multimodal biological reasoning within a dna-llm model. arXiv preprint arXiv:2505.23579. — [arXiv:2505.23579](https://arxiv.org/abs/2505.23579) find
  17. Xue Gao, Shu-Hai Lin, Feng Ren, Jin-Tao Li, Jia-Jia Chen, Chuan-Bo Yao, Hong-Bin Yang, Shu-Xia Jiang, Guo-Quan Yan, Di Wang, and 1 others. 2016. Acetate functions as an epigenetic metabolite to promote lipid synthesis under hypoxia. Nature communications, 7(1):11960. find
  18. Louis Gioia, Azeem Siddique, Steven R Head, Daniel R Salomon, and Andrew I Su. 2018. A genome-wide survey of mutations in the jurkat cell line. BMC genomics, 19(1):334. find
  19. Madalina Giurgiu, Julian Reinhard, Barbara Brauner, Irmtraud Dunger-Kaltenbach, Gisela Fobo, Goar Frishman, Corinna Montrone, and Andreas Ruepp. 2019. Corum: the comprehensive resource of mammalian protein complexes2019. Nucleic acids research, 47(D1):D559–D563. find
  20. Google Cloud. 2025. Gemini 2.5 pro: Generative ai on vertex ai. https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/ gemini/2-5-pro?hl=zh-cn. Accessed: 2025-12-21. ↗
  21. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  22. Minsheng Hao, Jing Gong, Xin Zeng, Chiming Liu, Yucheng Guo, Xingyi Cheng, Taifeng Wang, Jianzhu Ma, Xuegong Zhang, and Le Song. 2024. Large-scale foundation model on single-cell transcriptomics. Nature methods, 21(8):1481–1491. find
  23. Ibarra, Sanjay R Srivatsan, Mohsen Naghipourfar, Riza M Daza, Beth Martin, and 1 others. 2023. Predicting cellular responses to complex perturbations in high-throughput screens. Molecular systems biology, 19(6):e11517. find
  24. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276.
  25. Hehuan Ma, Yatao Bian, Yu Rong, Wenbing Huang, Tingyang Xu, Weiyang Xie, Geyan Ye, and Junzhou Huang. 2022. Cross-dependent graph neural networks for molecular property prediction. Bioinformatics, 38(7):2003–2009. find
  26. Ana-Maria Istrate, Fausto Milletari, Fabrizio Castrotor-res, Jakub M Tomczak, Michaela Torkar, Donghui Li, and Theofanis Karaletsos. 2025. rbio1-training scientific reasoning llms with biological world models as soft verifiers. bioRxiv, pages 2025–08. find
  27. Christian von Mering, Martijn Huynen, Daniel Jaeggi, Steffen Schmidt, Peer Bork, and Berend Snel. 2003. String: a database of predicted functional associations between proteins. Nucleic acids research, 31(1):258–261. find
  28. Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM computing surveys, 55(12):1–38.
  29. John R Moffett, Narayanan Puthillathu, Ranjini Vengilote, Diane M Jaworski, and Aryan M Namboodiri. 2020. Acetate revisited: A key biomolecule at the nexus of metabolism, epigenetics and onco-genesispart 1: Acetyl-coa, acetogenesis and acyl-coa short-chain synthetases. Frontiers in Physiology, 11:58… find
  30. Jonathan R Karr, Jayodita C Sanghvi, Derek N Macklin, Miriam V Gutschow, Jared M Jacobs, Benjamin Bolival, Nacyra Assad-Garcia, John I Glass, and Markus W Covert. 2012. A whole-cell computational model predicts phenotype from genotype. Cell, 150(2):389–401. find
  31. Irem Kaymak, McLane J Watson, Brandon M Oswald, Shixin Ma, Benjamin K Johnson, Lisa M DeCamp, Batsirai M Mabvakure, Katarzyna M Luda, Eric H Ma, Kin Lau, and 1 others. 2024. Acly and acss2 link nutrient-dependent chromatin accessibility to cd8 t cell effector responses. Journal of Experimental Medic… find
  32. Yunsoo Kim, Jinge Wu, Yusuf Abdulle, and Honghan Wu. 2024. Medexqa: Medical question answering benchmark with multiple explanations. arXiv preprint arXiv:2406.06331. find
  33. Dominik Klein, Jonas Simon Fleck, Daniil Bobrovskiy, Lea Zimmermann, Sören Becker, Alessandro Palma, Leander Dony, Alejandro Tejada-Lapuerta, Guillaume Huguet, Hsiu-Chuan Lin, and 1 others. 2025. Cellflow enables generative single-cell phenotype modeling with flow matching. bioRxiv, pages 2025– 04. find
  34. Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, and 1 others. 2023. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637):1123–1130. find
  35. Rui Ling, Gong Chen, Xiang Tang, Na Liu, Yuepeng Zhou, and Deyu Chen. 2022. Acetyl-coa synthetase 2 (acss2): a review with a focus on metabolism and tumor development. Discover Oncology, 13(1):58. find
  36. Mohammad Lotfollahi, Anna Klimovskaia Susmelj, Carlo De Donno, Leon Hetzel, Yuge Ji, Ignacio L Osval A Montesinos-Lopez, Moises Chavira-Flores, Kismiantini, Leo Crespo-Herrera, Carolina Saint Piere, HuiHui Li, Roberto Fritsche-Neto, Khalid Al-Nowibet, Abelardo Montesinos-López, and José Crossa. 2024… find
  37. OpenAI. 2025. Openai o3 and o4-mini system card. https://openai.com/zh-Hant-HK/index/ o3-o4-mini-system-card/. Accessed: 2025-12-21. ↗
  38. OpenAI Team. 2025. Gpt-5 system card. https://openai.com/index/gpt-5-system-card/. Accessed: 2025-12-21. ↗
  39. Lawrence Phillips, Marc Boubnovski Martell, Aditya Misra, Josefa Lia Stoisser, Cesar A Prada-Medina, Rory Donovan-Maiye, and Kaspar Märtens. 2025. Synthpert: Enhancing llm biological reasoning via synthetic reasoning traces for cellular perturbation prediction. arXiv preprint arXiv:2509.25346. — [arXiv:2509.25346](https://arxiv.org/abs/2509.25346) find
  40. William A Pryor, Kendall N Houk, Christopher S Foote, Jon M Fukuto, Louis J Ignarro, Giuseppe L Squadrito, and Kelvin JA Davies. 2006. Free radical biology and medicine: it’s a gas, man! American Journal of Physiology-Regulatory, Integrative and Comparative Physiology, 291(3):R491–R511. find
  41. Yusuf Roohani, Kexin Huang, and Jure Leskovec. 2024. Predicting transcriptional outcomes of novel multigene perturbations with gears. Nature Biotechnology, 42(6):927–935. find
  42. Lokendra K Sharma, Jianxin Lu, and Yidong Bai. 2009. Mitochondrial respiratory complex i: structure, function and implication in human diseases. Current medicinal chemistry, 16(10):1266–1277. find
  43. Allison C Sharrow, Emily Megill, Amanda J Chen, Afifa Farooqi, Naveen Kumar Tangudu, Apoorva Uboveja, Stacy McGonigal, Nadine Hempel, Nathaniel W Snyder, Ronald J Buck-anovich, and 1 others. 2024. Acetate drives ovarian cancer quiescence via acss2-mediated acetyl-coa production. Molecular Metabolism… find
  44. Cosma Spalluto, David I Wilson, and Tom Hearn. 2013. Evidence for reciliation of rpe1 cells in late g1 phase, and ciliary localisation of cyclin b1. FEBS open bio, 3:334–340. find
  45. Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2017. Graph attention networks. arXiv preprint arXiv:1710.10903.
  46. Zhijian Wei, Runze Ma, Zichen Wang, Zhongmin Li, Shuotong Song, and Shuangjia Zheng. 2025. Vcworld: A biological world model for virtual cell simulation. arXiv preprint arXiv:2512.00306. — [arXiv:2512.00306](https://arxiv.org/abs/2512.00306) find
  47. Jacob White. 2020. Pubmed 2.0. Medical reference services quarterly, 39(4):382–387. find
  48. Menghua Wu, Russell Littman, Jacob Levine, Lin Qiu, Tommaso Biancalani, David Richmond, and Jan-Christian Huetter. 2025. Contextualizing biological perturbation experiments through language. arXiv preprint arXiv:2502.21290. — [arXiv:2502.21290](https://arxiv.org/abs/2502.21290) find
  49. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  50. Geyan Ye, Xibao Cai, Houtim Lai, Xing Wang, Junhong Huang, Longyue Wang, Wei Liu, and Xiangxiang Zeng. 2025. Drugassist: A large language model for molecule optimization. Briefings in Bioinformatics, 26(1):bbae693. find
  51. Chaohao Yuan, Songyou Li, Geyan Ye, Yikun Zhang, Long-Kai Huang, Wenbing Huang, Wei Liu, Jianhua Yao, and Yu Rong. 2025. Annotation-guided protein design with multi-level domain alignment. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1, pages 1855–1866. find
  52. Yikun Zhang, Geyan Ye, Chaohao Yuan, Bo Han, Long-Kai Huang, Jianhua Yao, Wei Liu, and Yu Rong. 2025. Atomas: Hierarchical adaptive alignment on molecule-text for unified molecule understanding and generation. In The Thirteenth International Conference on Learning Representations. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhenyu-2026-aroma-augmented-reasoning-over-multimodal/