Subgraph Neural Networks

SubGNN is a general framework for subgraph representation learning. SubGNN learns meaningful representations for subgraphs and supports prediction of any subgraph properties.

We present SubGNN, a general method for subgraph representation learning. It addresses a fundamental gap in current graph neural network (GNN) methods that are not yet optimized for subgraph-level predictions.

Our method implements in a neural message passing scheme three distinct channels to each capture a key property of subgraphs: neighborhood, structure, and position. We have generated four synthetic datasets highlighting a specific subgraph property. By performing an ablation study over the channels, we demonstrate that the performance of individual channels aligns closely with their inductive biases.

SubGNN outperforms baseline methods across both synthetic and real-world datasets. Along with SubGNN, we have released eight new datasets representing a diverse collection of tasks to pave the way for innovating subgraph neural networks in the future.


Subgraph Neural Networks
Emily Alsentzer, Samuel G. Finlayson, Michelle M. Li, Marinka Zitnik
NeurIPS 2020 [arXiv] [poster]

  title={Subgraph Neural Networks},
  author={Alsentzer, Emily and Finlayson, Samuel G and Li, Michelle M and Zitnik, Marinka},
  booktitle={Proceedings of Neural Information Processing Systems, NeurIPS},


Encoding subgraphs for GNNs is not well-studied or commonly used. Rather, current GNN methods are optimized for node-, edge-, and graph-level predictions, but not yet at the subgraph-level.

Representation learning for subgraphs presents unique challenges.

  • Subgraphs require that we make joint predictions over structures of varying sizes. They do not necessarily cluster together, and can even be composed of multiple disparate components that are far apart from each other in the graph.
  • Subgraphs contain rich higher-order connectivity patterns, both internally and externally with the rest of the graph. The challenge is to inject information about border and external subgraph structure into the GNN’s neural message passing.
  • Subgraphs can be localized or distributed throughout the graph. We must effectively learn about subgraph positions within the underlying graph.

The following figure depicts a simple base graph and five subgraphs, each with different structures. For instance, subgraphs S2, S3, and S5 comprise of single connected components in the graph, whereas subgraphs S1 and S4 each form two isolated components. Colors indicate subgraph labels. The right panel illustrates an example of internal connectivity versus border structure

SubGNN framework

SubGNN takes as input a base graph and subgraph information to train embeddings for each subgraph. Each channel’s output embeddings are then concatenated to generate one final subgraph embedding.

The figure above depicts SubGNN’s architecture. The left panel illustrates how property-specific messages are propagated from anchor node patches to subgraph components. The right panel shows the three channels designed to capture a distinct subgraph property.


We present four new synthetic datasets and four novel real-world social and biological datasets to stimulate subgraph representation learning research.

Synthetic datasets

Each synthetic dataset challenges the ability of our methods to capture:

  • DENSITY: Internal structure of subgraph topology.
  • CUT RATIO: Border structure of subgraph topology.
  • CORENESS: Border structure and position of subgraph topology.
  • COMPONENT: Internal and external position of subgraph topology.

Real-world datasets

  • PPI-BP is a molecular biology dataset, where the subgraphs are a group of genes and their labels are the genes’ collective cellular function. The base graph is a human protein-protein interaction network.
  • UDN-METAB is a clinical dataset, where the subgraphs are a collection of phenotypes associated with a rare monogenic disease and their labels are the subcategory of the metabolic disorder most consistent with those phenotypes. The base graph is a knowledge graph containing phenotype and genotype information about rare diseases.
  • UDN-NEURO is similar to UDN-METAB but for one or more neurological disorders (multilabel classification), and shares the same base graph.
  • EM is a social dataset, where the subgraphs are the workout history of a user and their labels are the gender of the user. The base graph is a social fitness network from Endomondo.


Source code is available in the GitHub repository.


Latest News

Apr 2024:   Biomedical AI Agents

Mar 2024:   Efficient ML Seminar Series

We started a Harvard University Efficient ML Seminar Series. Congrats to Jonathan for spearheading this initiative. Harvard Magazine covered the first meeting focusing on LLMs.

Mar 2024:   UniTS - Unified Time Series Model

UniTS is a unified time series model that can process classification, forecasting, anomaly detection and imputation tasks within a single model with no task-specific modules. UniTS has zero-shot, few-shot, and prompt learning capabilities. Project website.

Mar 2024:   Weintraub Graduate Student Award

Michelle receives the 2024 Harold M. Weintraub Graduate Student Award. The award recognizes exceptional achievement in graduate studies in biological sciences. News Story. Congratulations!

Mar 2024:   PocketGen - Generating Full-Atom Ligand-Binding Protein Pockets

PocketGen is a deep generative model that generates residue sequence and full-atom structure of protein pockets, maximizing binding to ligands. Project website.

Feb 2024:   SPECTRA - Generalizability of Molecular AI

Feb 2024:   Kaneb Fellowship Award

The lab receives the John and Virginia Kaneb Fellowship Award at Harvard Medical School to enhance research progress in the lab.

Feb 2024:   NSF CAREER Award

The lab receives the NSF CAREER Award for our research in geometric deep learning to facilitate algorithmic and scientific advances in therapeutics.

Feb 2024:   Dean’s Innovation Award in AI

Jan 2024:   AI's Prospects in Nature Machine Intelligence

We discussed AI’s 2024 prospects with Nature Machine Intelligence, covering LLM progress, multimodal AI, multi-task agents, and how to bridge the digital divide across communities and world regions.

Jan 2024:   Combinatorial Therapeutic Perturbations

New paper introducing PDGrapher for combinatorial prediction of chemical and genetic perturbations using causally-inspired neural networks.

Nov 2023:   Next Generation of Therapeutics Commons

Oct 2023:   Structure-Based Drug Design

Geometric deep learning has emerged as a valuable tool for structure-based drug design, to generate and refine biomolecules by leveraging detailed three-dimensional geometric and molecular interaction information.

Oct 2023:   Graph AI in Medicine

Graph AI models in medicine integrate diverse data modalities through pre-training, facilitate interactive feedback loops, and foster human-AI collaboration, paving the way to clinically meaningful predictions.

Sep 2023:   New papers accepted at NeurIPS

Sep 2023:   Future Directions in Network Biology

Excited to share our perspectives on current and future directions in network biology.

Aug 2023:   Scientific Discovery in the Age of AI

Jul 2023:   PINNACLE - Contextual AI protein model

PINNACLE is a contextual AI model for protein understanding that dynamically adjusts its outputs based on biological contexts in which it operates. Project website.

Jun 2023:   Our Group is Joining the Kempner Institute

Excited to join Kempner’s inaugural cohort of associate faculty to advance Kempner’s mission of studying the intersection of natural and artificial intelligence.

Jun 2023:   Welcoming a New Postdoctoral Fellow

An enthusiastic welcome to Shanghua Gao who is joining our group as a postdoctoral research fellow.

Zitnik Lab  ·  Artificial Intelligence in Medicine and Science  ·  Harvard  ·  Department of Biomedical Informatics