Yu Rong
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Research Interests
Yu Rong has a total of 217 co-authors. Here are the top co-authors by number of shared publications:
- Wenbing Huang
Renmin University of China
21 shared publications - Junzhou Huang
The University of Texas at Arlington
19 shared publications - Kangfei Zhao
Beijing Institute of Technology, Beijing, China
12 shared publications - Tingyang Xu
Tencent AI Lab
12 shared publications - Hong Cheng
The Chinese University of Hong Kong, Hong Kong, China
10 shared publications - Deli Zhao
DAMO Academy, Alibaba Group, Hangzhou, China and Hupan Lab, Hangzhou, China
9 shared publications
Latest Publications
- Open Access
- 6 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 8081-8085
Type-aware Prompting for Community Search over Heterogeneous Information Networks
- 6 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 8081-8085
Open AccessType-aware Prompting for Community Search over Heterogeneous Information Networks
- 24 September 2026
- Published by Springer Nature in Nature Cell Biology
- Vol. 28  (10) , 2179-2192
The dual-enhanced graph learning framework DePass allows paired data integration in single-cell and spatial multiomics
Recent sequencing advances have enabled abundant multi-omics data generation for both single-cell and spatial contexts. Integrating such multimodal data is critical for decoding cellular and tissue-level complexity. However, compared with single-modality profiling, multimodal data often exhibit higher levels of noise, and existing methods typically overlook this challenge during integration. Meanwhile, most current approaches are tailored to either single-cell or spatial data, limiting their applicability across data types. Here we present DePass, a scalable graph learning framework for paired data integration in both single-cell and spatial multi-omics. We propose a coupled enhancement-integration architecture that iteratively denoises data and improves integrated embeddings. We systematically benchmarked DePass across 6 modalities, 9 tissue types and 13 experimental platforms, demonstrating superior integration accuracy. In the in-house colorectal cancer data, DePass further uncovered immune niche substructure and spatial tumour heterogeneity at near single-cell resolution. These results establish DePass as a unified and generalizable solution for multi-omics integration across diverse biological contexts.
- 24 September 2026
- Published by Springer Nature in Nature Cell Biology
- Vol. 28  (10) , 2179-2192
The dual-enhanced graph learning framework DePass allows paired data integration in single-cell and spatial multiomics
Recent sequencing advances have enabled abundant multi-omics data generation for both single-cell and spatial contexts. Integrating such multimodal data is critical for decoding cellular and tissue-level complexity. However, compared with single-modality profiling, multimodal data often exhibit higher levels of noise, and existing methods typically overlook this challenge during integration. Meanwhile, most current approaches are tailored to either single-cell or spatial data, limiting their applicability across data types. Here we present DePass, a scalable graph learning framework for paired data integration in both single-cell and spatial multi-omics. We propose a coupled enhancement-integration architecture that iteratively denoises data and improves integrated embeddings. We systematically benchmarked DePass across 6 modalities, 9 tissue types and 13 experimental platforms, demonstrating superior integration accuracy. In the in-house colorectal cancer data, DePass further uncovered immune niche substructure and spatial tumour heterogeneity at near single-cell resolution. These results establish DePass as a unified and generalizable solution for multi-omics integration across diverse biological contexts.
- Top Cited PapersOpen Access
- 22 August 2026
- Published by Association for Computing Machinery (ACM) in ACM Computing Surveys
- Vol. 58  (15) , 1-35
A Survey of Graph Transformers: Architectures, Theories and Applications
Graph Transformers (GTs) have demonstrated a strong capability in modeling graph structures by addressing the intrinsic limitations of graph neural networks (GNNs), such as over-smoothing and over-squashing. Recent studies have proposed diverse architectures, enhanced explainability, and practical applications for Graph Transformers. In light of these rapid developments, we conduct a comprehensive review of Graph Transformers, covering aspects such as their architectures, theoretical foundations, and applications. In this survey, we first categorize the architecture of Graph Transformers according to their strategies for processing structural information, including graph tokenization, positional encoding, structure-aware attention, and model ensemble. Then, from the theoretical perspective, we examine the expressivity of Graph Transformers in various discussed architectures and contrast them with other advanced graph learning algorithms to discover their connections. For applications, we organize the literature around four graph organization forms, from relational, geometric, dynamic to heterogeneous. A Practical Guidance table then maps architectural components to these graph forms by adoption frequency, so practitioners can narrow down which design families to consider for a given input structure. Lastly, we will discuss the current challenges and prospective directions in Graph Transformers for potential future research.
- 22 August 2026
- Published by Association for Computing Machinery (ACM) in ACM Computing Surveys
- Vol. 58  (15) , 1-35
Top Cited PapersOpen AccessA Survey of Graph Transformers: Architectures, Theories and Applications
Graph Transformers (GTs) have demonstrated a strong capability in modeling graph structures by addressing the intrinsic limitations of graph neural networks (GNNs), such as over-smoothing and over-squashing. Recent studies have proposed diverse architectures, enhanced explainability, and practical applications for Graph Transformers. In light of these rapid developments, we conduct a comprehensive review of Graph Transformers, covering aspects such as their architectures, theoretical foundations, and applications. In this survey, we first categorize the architecture of Graph Transformers according to their strategies for processing structural information, including graph tokenization, positional encoding, structure-aware attention, and model ensemble. Then, from the theoretical perspective, we examine the expressivity of Graph Transformers in various discussed architectures and contrast them with other advanced graph learning algorithms to discover their connections. For applications, we organize the literature around four graph organization forms, from relational, geometric, dynamic to heterogeneous. A Practical Guidance table then maps architectural components to these graph forms by adoption frequency, so practitioners can narrow down which design families to consider for a given input structure. Lastly, we will discuss the current challenges and prospective directions in Graph Transformers for potential future research.
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 11351-11362
One Path to Model Them All: Learnable-Time Flow Matching for Crystal Structure and Energy Prediction
Crystals are cornerstone materials for semiconductors and renewable energy, yet their discovery is hindered by the prohibitive cost of Density Functional Theory (DFT). While geometric graph neural networks have advanced Crystal Structure Prediction (CSP) and energy estimation, existing methods treat these tasks as disparate problems with task-specific architectures, failing to exploit cross-task synergies. In this paper, we propose UniPath, a novel framework based on Learnable-Time Flow Matching that unifies crystal structure and energy prediction within a single probability path. Our key insight is modeling unstable crystals as an intermediate state along the probability path from a prior to the stable distribution. We introduce two pivotal mechanisms: (1) a learnable time parameter that dynamically identifies the intermediate state, and (2) an auxiliary velocity field that parameterizes tailored flows to synchronize the intermediate and unstable distributions. We theoretically prove that under optimal velocity fields, these paths converge to a consistent trajectory, enabling a unified backbone to maximize shared representation learning capabilities. Extensive evaluations on CSP and energy prediction benchmarks demonstrate that UniPath significantly outperforms task-specific baselines. Notably, UniPath achieves over 50% higher Match Rate on MPTS-52 and over 30% lower energy MAE on MPT-MPTS-52. Furthermore, UniPath generalizes effectively to small molecules, surpassing conventional models in molecular conformation generation. The source code is publicly available at https://github.com/GLAD-RUC/UniPath.
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 11351-11362
Open AccessOne Path to Model Them All: Learnable-Time Flow Matching for Crystal Structure and Energy Prediction
Crystals are cornerstone materials for semiconductors and renewable energy, yet their discovery is hindered by the prohibitive cost of Density Functional Theory (DFT). While geometric graph neural networks have advanced Crystal Structure Prediction (CSP) and energy estimation, existing methods treat these tasks as disparate problems with task-specific architectures, failing to exploit cross-task synergies. In this paper, we propose UniPath, a novel framework based on Learnable-Time Flow Matching that unifies crystal structure and energy prediction within a single probability path. Our key insight is modeling unstable crystals as an intermediate state along the probability path from a prior to the stable distribution. We introduce two pivotal mechanisms: (1) a learnable time parameter that dynamically identifies the intermediate state, and (2) an auxiliary velocity field that parameterizes tailored flows to synchronize the intermediate and unstable distributions. We theoretically prove that under optimal velocity fields, these paths converge to a consistent trajectory, enabling a unified backbone to maximize shared representation learning capabilities. Extensive evaluations on CSP and energy prediction benchmarks demonstrate that UniPath significantly outperforms task-specific baselines. Notably, UniPath achieves over 50% higher Match Rate on MPTS-52 and over 30% lower energy MAE on MPT-MPTS-52. Furthermore, UniPath generalizes effectively to small molecules, surpassing conventional models in molecular conformation generation. The source code is publicly available at https://github.com/GLAD-RUC/UniPath.
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 1039-1050
PLACE: Prompt Learning for Attributed Community Search in Large Graphs
Attributed Community Search (ACS) aims to identify communities in an attributed graph with structural cohesiveness and attribute homogeneity for given queries. While algorithmic approaches often suffer from structural inflexibility and attribute irrelevance, recent years have witnessed a boom in learning-based approaches that employ Graph Neural Networks (GNNs) to simultaneously model structural and attribute information. To improve GNN efficacy and efficiency, these approaches adopt various techniques to refine the input graph and queries, including query-dependent node pruning and feature enhancement. However, these refinements are either detached from the end-to-end optimization with the backbone model, or impose additional burdens on training resources, limiting their adaptability and practical usage in diverse graphs. In this paper, we propose PLACE(Prompt Learning for Attributed Community Search), a graph prompt learning framework for ACS. Inspired by prompt-tuning in Natural Language Processing (NLP), PLACE constructs a query-specific prompt graph consisting of learnable attribute tokens and virtual node tokens, and inserts it into the original graph to form a prompt-augmented graph. The learned prompt tokens act as a bridge that enriches the query context within the graph, enabling the GNN to more effectively capture patterns of structural cohesiveness and attribute similarity pertinent to each query. We employ an alternating training paradigm to jointly optimize the prompt parameters and the GNN, and design a divide-and-conquer strategy that scales PLACEto million-scale graphs. Extensive experiments on 9 real-world graphs across 3 types of ACS queries demonstrate that PLACE outperforms 6 baselines, achieving an average F1 improvement of 14% over the state-of-the-art.
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 1039-1050
Open AccessPLACE: Prompt Learning for Attributed Community Search in Large Graphs
Attributed Community Search (ACS) aims to identify communities in an attributed graph with structural cohesiveness and attribute homogeneity for given queries. While algorithmic approaches often suffer from structural inflexibility and attribute irrelevance, recent years have witnessed a boom in learning-based approaches that employ Graph Neural Networks (GNNs) to simultaneously model structural and attribute information. To improve GNN efficacy and efficiency, these approaches adopt various techniques to refine the input graph and queries, including query-dependent node pruning and feature enhancement. However, these refinements are either detached from the end-to-end optimization with the backbone model, or impose additional burdens on training resources, limiting their adaptability and practical usage in diverse graphs. In this paper, we propose PLACE(Prompt Learning for Attributed Community Search), a graph prompt learning framework for ACS. Inspired by prompt-tuning in Natural Language Processing (NLP), PLACE constructs a query-specific prompt graph consisting of learnable attribute tokens and virtual node tokens, and inserts it into the original graph to form a prompt-augmented graph. The learned prompt tokens act as a bridge that enriches the query context within the graph, enabling the GNN to more effectively capture patterns of structural cohesiveness and attribute similarity pertinent to each query. We employ an alternating training paradigm to jointly optimize the prompt parameters and the GNN, and design a divide-and-conquer strategy that scales PLACEto million-scale graphs. Extensive experiments on 9 real-world graphs across 3 types of ACS queries demonstrate that PLACE outperforms 6 baselines, achieving an average F1 improvement of 14% over the state-of-the-art.
- Open Access
- 27 March 2026
- Published by Springer Nature in Communications Biology
- Vol. 9  (1)
Accurate protein-protein interactions modeling through physics-informed geometric invariant learning
AlphaFold has set a new standard for predicting protein structures from primary sequences; however, it faces challenges with protein complexes across species, engineered proteins, and antigen-antibody interactions, where co-evolutionary signals may be sparse or missing. Herein, we present ProTact, a SE(3)-invariant geometric graph neural network that integrates physics-informed geometric complementarity and trigonometric constraints as inductive biases to enhance protein-protein contact predictions. ProTact is applicable to both experimental and predicted monomer structures and utilizes a modulated key point matching algorithm to approximate accurate docking poses. Experimental evaluations demonstrate that ProTact consistently outperforms state-of-the-art sequence-based and structure-based methods on benchmark datasets, achieving notable relative improvements of 31.63% in average top-10 precision (Precision@10) for CASP 13 and 14 targets and 31.94% for DIPS-Plus datasets on high-quality structures. While performance naturally declines on the more challenging unbound complexes due to large conformational changes, ProTact maintains a competitive edge over baselines. Moreover, when combined with AlphaFold3 as re-scoring functions, ProTact surpasses its default confidence scores, offering over 30.48% improvements in low-MSA contexts. We anticipate that the proposed framework will advance our understanding of protein interactions, functions, and design.
- 27 March 2026
- Published by Springer Nature in Communications Biology
- Vol. 9  (1)
Open AccessAccurate protein-protein interactions modeling through physics-informed geometric invariant learning
AlphaFold has set a new standard for predicting protein structures from primary sequences; however, it faces challenges with protein complexes across species, engineered proteins, and antigen-antibody interactions, where co-evolutionary signals may be sparse or missing. Herein, we present ProTact, a SE(3)-invariant geometric graph neural network that integrates physics-informed geometric complementarity and trigonometric constraints as inductive biases to enhance protein-protein contact predictions. ProTact is applicable to both experimental and predicted monomer structures and utilizes a modulated key point matching algorithm to approximate accurate docking poses. Experimental evaluations demonstrate that ProTact consistently outperforms state-of-the-art sequence-based and structure-based methods on benchmark datasets, achieving notable relative improvements of 31.63% in average top-10 precision (Precision@10) for CASP 13 and 14 targets and 31.94% for DIPS-Plus datasets on high-quality structures. While performance naturally declines on the more challenging unbound complexes due to large conformational changes, ProTact maintains a competitive edge over baselines. Moreover, when combined with AlphaFold3 as re-scoring functions, ProTact surpasses its default confidence scores, offering over 30.48% improvements in low-MSA contexts. We anticipate that the proposed framework will advance our understanding of protein interactions, functions, and design.
- Open Access
- 27 February 2026
- Published by MDPI AG in Preprints.org
A Comprehensive Survey of Multimodal LLMs for Scientific Discovery
Recent advances in artificial intelligence (AI), especially large language models, have accelerated the integration of multimodal data in scientific research. Given that scientific fields involve diverse data types, ranging from text and images to complex biological sequences, graphs, and structures, multimodal large language models (MLLMs) have emerged as powerful tools to bridge these modalities, enabling more comprehensive data analysis and intelligent decision-making. This work, S3-Bench, provides a comprehensive overview of recent advances in MLLMs, focusing on their diverse applications across science. We systematically review the progress of MLLMs in key scientific domains, including drug discovery, molecular & protein design, materials science, and genomics. The work highlights model architectures, domain-specific adaptations, benchmark datasets, and promising future directions. More importantly, we benchmarked open-source MLLMs on a range of critical molecular and protein property prediction tasks. Our work aims to serve as a valuable resource for both researchers and practitioners interested in the rapidly evolving landscape of multimodal AI for science.
- 27 February 2026
- Published by MDPI AG in Preprints.org
Open AccessA Comprehensive Survey of Multimodal LLMs for Scientific Discovery
Recent advances in artificial intelligence (AI), especially large language models, have accelerated the integration of multimodal data in scientific research. Given that scientific fields involve diverse data types, ranging from text and images to complex biological sequences, graphs, and structures, multimodal large language models (MLLMs) have emerged as powerful tools to bridge these modalities, enabling more comprehensive data analysis and intelligent decision-making. This work, S3-Bench, provides a comprehensive overview of recent advances in MLLMs, focusing on their diverse applications across science. We systematically review the progress of MLLMs in key scientific domains, including drug discovery, molecular & protein design, materials science, and genomics. The work highlights model architectures, domain-specific adaptations, benchmark datasets, and promising future directions. More importantly, we benchmarked open-source MLLMs on a range of critical molecular and protein property prediction tasks. Our work aims to serve as a valuable resource for both researchers and practitioners interested in the rapidly evolving landscape of multimodal AI for science.
- Open Access
- 21 February 2026
- Published by Association for Computing Machinery (ACM)
- p. 881-891
ParaFormer: A Generalized PageRank Graph Transformer for Graph Representation Learning
Graph Transformers (GTs) have emerged as a promising graph learning tool, leveraging their all-pair connected property to effectively capture global information. To address the over-smoothing problem in deep GNNs, global attention was initially introduced, eliminating the necessity for using deep GNNs. However, through empirical and theoretical analysis, we verify that the introduced global attention exhibits severe over-smoothing, causing node representations to become indistinguishable due to its inherent low-pass filtering. This effect is even stronger than that observed in GNNs. To mitigate this, we propose PageRank Transformer (ParaFormer), which features a PageRank-enhanced attention module designed to mimic the behavior of deep Transformers. We theoretically and empirically demonstrate that ParaFormer mitigates over-smoothing by functioning as an adaptive-pass filter. Experiments show that ParaFormer achieves consistent performance improvements across both node classification and graph classification tasks on 11 datasets ranging from thousands to millions of nodes, validating its efficacy. The supplementary material, including code and appendix, can be found in https://github.com/chaohaoyuan/ParaFormer.
- 21 February 2026
- Published by Association for Computing Machinery (ACM)
- p. 881-891
Open AccessParaFormer: A Generalized PageRank Graph Transformer for Graph Representation Learning
Graph Transformers (GTs) have emerged as a promising graph learning tool, leveraging their all-pair connected property to effectively capture global information. To address the over-smoothing problem in deep GNNs, global attention was initially introduced, eliminating the necessity for using deep GNNs. However, through empirical and theoretical analysis, we verify that the introduced global attention exhibits severe over-smoothing, causing node representations to become indistinguishable due to its inherent low-pass filtering. This effect is even stronger than that observed in GNNs. To mitigate this, we propose PageRank Transformer (ParaFormer), which features a PageRank-enhanced attention module designed to mimic the behavior of deep Transformers. We theoretically and empirically demonstrate that ParaFormer mitigates over-smoothing by functioning as an adaptive-pass filter. Experiments show that ParaFormer achieves consistent performance improvements across both node classification and graph classification tasks on 11 datasets ranging from thousands to millions of nodes, validating its efficacy. The supplementary material, including code and appendix, can be found in https://github.com/chaohaoyuan/ParaFormer.
- Top Cited Papers
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- p. 1-18
Lingshu: Generalist Foundation Model for Unified Multimodal Medical Understanding and Reasoning
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- p. 1-18
Top Cited PapersLingshu: Generalist Foundation Model for Unified Multimodal Medical Understanding and Reasoning
- Open Access
- 4 December 2025
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 3  (6) , 1-28
Can Large Language Models Be Query Optimizer for Relational Databases?
Query optimization is a complex planning and decision-making problem within the exponentially growing plan space in database management systems (DBMS). Traditional optimization techniques have been extensively studied over decades, leaving limited room for further improvement along this track. Recent developments of Large Language Models (LLMs) have demonstrated their potential in solving complex planning and decision-making problems, such as arithmetic and programmatic tasks. In this paper, we try to explore the potential of LLMs in handling query optimization and propose a tentative LLM-based query optimizer dubbed LLM-QO, established on PostgreSQL's execution engine. In LLM-QO, we formulate query optimization in an autoregressive fashion which directly generates the execution plan without explicit plan enumeration. To investigate the essential input of LLM-QO, we design a customized data recipe named QInstruct to collect the training data from various optimizers and serialize the database's meta data, queries and corresponding plans into a textual format. Based on QInstruct, we implement a two-stage fine-tuning pipeline, Query Instruction Tuning (QIT) and Query Direct Preference Optimization (QDPO), to empower the capability of general-purpose LLMs in handling query optimization. In our experiments, LLM-QO can generate valid and high-quality plans and consistently outperforms both traditional and learned optimizers on three query workloads. Our findings verify that LLMs can be derived as query optimizers where generalization, efficiency and adaptivity deserve further research efforts.
- 4 December 2025
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 3  (6) , 1-28
Open AccessCan Large Language Models Be Query Optimizer for Relational Databases?
Query optimization is a complex planning and decision-making problem within the exponentially growing plan space in database management systems (DBMS). Traditional optimization techniques have been extensively studied over decades, leaving limited room for further improvement along this track. Recent developments of Large Language Models (LLMs) have demonstrated their potential in solving complex planning and decision-making problems, such as arithmetic and programmatic tasks. In this paper, we try to explore the potential of LLMs in handling query optimization and propose a tentative LLM-based query optimizer dubbed LLM-QO, established on PostgreSQL's execution engine. In LLM-QO, we formulate query optimization in an autoregressive fashion which directly generates the execution plan without explicit plan enumeration. To investigate the essential input of LLM-QO, we design a customized data recipe named QInstruct to collect the training data from various optimizers and serialize the database's meta data, queries and corresponding plans into a textual format. Based on QInstruct, we implement a two-stage fine-tuning pipeline, Query Instruction Tuning (QIT) and Query Direct Preference Optimization (QDPO), to empower the capability of general-purpose LLMs in handling query optimization. In our experiments, LLM-QO can generate valid and high-quality plans and consistently outperforms both traditional and learned optimizers on three query workloads. Our findings verify that LLMs can be derived as query optimizers where generalization, efficiency and adaptivity deserve further research efforts.