Xiaorui Su
Xiaorui Su

Xiaorui Su

Postdoctoral Fellow · Zitnik Lab, Harvard Medical School

Building AI that holds up in the clinic, with reasoning that stays faithful to the evidence.

About

I'm a Postdoctoral Fellow at Harvard Medical School, where I work in the Zitnik Lab with Prof. Marinka Zitnik. Previously, I completed my Ph.D. at the University of Chinese Academy of Sciences and my B.E. at Xiamen University, and I was a visiting Ph.D. student at the University of Illinois Chicago, working with Prof. Philip S. Yu.

I build AI agents for medicine — systems that reason over biomedical knowledge and verify their own evidence, so their outputs are ones clinicians and scientists can trust. My work spans medical reasoning, drug and therapeutic outcome prediction, and clinical decision making, connecting large biomedical knowledge bases with models that can explain and justify what they recommend.

My work has been published in leading venues across machine learning and biomedicine, including Nature Biomedical Engineering, Nature Communications, npj Digital Medicine, Communications Biology, ICML, ICLR, AAAI, and IEEE Transactions on Knowledge and Data Engineering.

📢 I am actively looking for faculty and industry positions starting Fall 2027. Please kindly reach out to me for any opportunities. Thanks!

Research Interests

News

  • Aug 2026Cancer-gene identification via single-cell contrastive learning accepted (in principle) at Nature Communications.
  • Aug 2026Invited talk, Towards Agentic AI in Biomedical Reasoning Problems, at the Joint Statistical Meetings (JSM), USA.
  • Jul 2026Activity Cliff-Informed Contrastive Learning accepted (in principle) at Nature Communications.
  • Jan 2026KnowGuard — knowledge-driven abstention for multi-round clinical reasoning — accepted at ICLR 2026.
  • Dec 2025Co-organizing CURE-Bench at NeurIPS 2025, the first benchmark on AI reasoning for therapeutics.
  • Jul 2025RNA m6A modification paper published in Communications Biology.
  • May 2025MedTok accepted at ICML 2025.
  • Mar 2025Madrigal preprint on drug-combination outcomes released.
  • Jan 2025KGARevion accepted at ICLR 2025.
  • Jan 2025Cancer-gene discovery paper published in Nature Biomedical Engineering.
  • Jul 2024Received the Chinese Academy of Sciences President Award (Excellent Prize); joined the Zitnik Lab at Harvard Medical School.
  • Feb 2024Dual-Channel DDI (TIGER) accepted at AAAI 2024.
  • 2023KG2ECapsule published in IEEE TKDE.

Selected Publications

† equal contribution · ‡ corresponding author · full list on Google Scholar.

Cancer-gene single-cell model overview
Leveraging Cell-type Specificity with Contextualized Contrastive Learning for Cancer Gene Identification by Single-cell Sequencing
A contextualized contrastive-learning model that leverages cell-type specificity in single-cell sequencing to identify cancer genes.
Ying Chang†, Xiaorui Su†, Yue Yang, Guodong Li, Ziwen Cui, Dongxu Li, Hengchuang Yin, Pengwei Hu, Lun Hu‡
Nature Communications, 2026 (accepted in principle)
ACANet molecular property model
Activity Cliff-Informed Contrastive Learning for Molecular Property Prediction
A contrastive-learning framework that leverages activity-cliff structure to learn molecular representations for more accurate property prediction.
Chao Cui, Xiaorui Su, Zaixi Zhang, Alejandro Velez Arce, et al.
Nature Communications, 2026 (in press)
Preprint
KnowGuard clinical reasoning agent
KnowGuard: Knowledge-Driven Abstention for Multi-Round Clinical Reasoning
A clinical reasoning agent that retrieves multi-modal knowledge to decide when to keep investigating versus when to safely abstain across multi-round patient interactions.
Xilin Dang, Kexin Chen, Xiaorui Su, Ayush Noori, Iñaki Arango, Lucas Vittor, Xinyi Long, Yuyang Du, Marinka Zitnik, Pheng Ann Heng
International Conference on Learning Representations (ICLR), 2026
arXiv OpenReview
MedTok architecture
MedTok: Multimodal Medical Code Tokenizer
A tokenizer that fuses the text descriptions and ontology structure of medical codes, improving EHR foundation models on prediction and drug recommendation.
Xiaorui Su, Shvat Messica, Yepeng Huang, Ruth Johnson, Lukas Fesser, Shanghua Gao, Faryad Sahneh, Marinka Zitnik‡
International Conference on Machine Learning (ICML), 2025
Poster
RNA m6A prediction model
Interpretability-Guided RNA N6-Methyladenosine Modification Site Prediction with Invertible Neural Networks
An invertible-neural-network model that predicts RNA m6A modification sites and reveals interpretable sequence and structure binding regions.
Guodong Li, Xiaorui Su, Yue Yang, Dongxu Li, Ziwen Cui, Xun Deng, Pengwei Hu, Lun Hu‡
Communications Biology, 2025
Paper
KGARevion agent overview
KGARevion: An AI Agent for Knowledge-Intensive Biomedical QA
An LLM agent that generates candidate knowledge-graph triplets, verifies them against a curated KG, and reasons over the filtered evidence to answer biomedical questions.
Xiaorui Su, Yibo Wang, Shanghua Gao, Xiaolong Liu, Valentina Giunchiglia, Djork-Arné Clevert, Marinka Zitnik‡
International Conference on Learning Representations (ICLR), 2025
Poster Video · Discover AI (92K views)
TxAgent tool-use loop
TxAgent: An AI Agent for Therapeutic Reasoning Across a Universe of Tools
A treatment-reasoning agent that calls 200+ biomedical tools to deliver evidence-grounded therapy recommendations across thousands of personalized scenarios.
Shanghua Gao, Richard Zhu, Zhenglun Kong, Ayush Noori, Xiaorui Su, Curtis Ginder, Theodoros Tsiligkaridis, Marinka Zitnik‡
arXiv:2503.10970, 2025
arXiv
Madrigal model overview
Multimodal AI Predicts Clinical Outcomes of Drug Combinations from Preclinical Data
Madrigal — a multimodal model with attention-bottleneck fusion, robust to missing modalities, that predicts clinical outcomes of drug combinations from preclinical data.
Yepeng Huang, Xiaorui Su, Varun Ullanat, Intae Moon, Ivy Liang, et al., Marinka Zitnik‡
arXiv:2503.02781, 2025
arXiv
Cancer-gene model overview
Interpretable Identification of Cancer Genes across Biological Networks via Transformer-Powered Graph Representation Learning
An interpretable Transformer-graph model over biological networks that identifies cancer genes and their mechanisms, validated in clinical cohorts.
Xiaorui Su, Pengwei Hu, Dongxu Li, Bowei Zhao, Zhaomeng Niu, Thomas Herget, Philip S. Yu, Lun Hu‡
Nature Biomedical Engineering, 2025
Paper
Dual-Channel DDI model
Dual-Channel Learning Framework for Drug-Drug Interaction Prediction via Relation-Aware Heterogeneous Graph Transformer
A relation-aware heterogeneous graph Transformer that predicts drug–drug interactions through dual-channel representation learning.
Xiaorui Su, Pengwei Hu, Zhuhong You, Lun Hu‡, Philip S. Yu
AAAI Conference on Artificial Intelligence, 2024
Paper Code
KG2ECapsule model
Biomedical Knowledge Graph Embedding with Capsule Network for Multi-Label Drug-Drug Interaction Prediction
KG2ECapsule — a capsule-network knowledge-graph embedding that generates and scores biomedical triplets for multi-label drug–drug interaction prediction.
Xiaorui Su, Zhuhong You, Deshuang Huang, Lei Wang, Leon Wong, Boya Ji, Bowei Zhao
IEEE Transactions on Knowledge and Data Engineering (TKDE), 2023
Paper

Talks & Service

Education