Relational Deep Learning on Biological Networks

33 concepts26 papers1 proposal

Traditional deep learning architectures excel at processing data structured in Euclidean space. Convolutional neural networks expect regular grids of pixels, while transformers and recurrent networks are designed for sequential tokens. However, biological systems rarely fit into these rigid geometries. A cell's functional state, a protein's behavior, or a drug's mechanism of action is fundamentally defined by a web of interactions—such as protein-protein interaction (PPI) networks, metabolic pathways, and gene regulatory circuits (Davide Bacciu et al.).

When researchers force these non-Euclidean structures into flat tables or fixed-length vectors, they discard the underlying topology. This loss of context limits a model's ability to capture the intricate, systemic dependencies that govern living organisms (Jianxin Wang et al.). To accurately model biology, deep learning must shift from analyzing isolated entities to learning directly from relational networks.

The Mechanism: Neighborhood Aggregation

Graph Neural Networks (GNNs) solve the geometric mismatch by introducing a relational inductive bias directly into the network architecture. Instead of treating biological entities in isolation, GNNs represent them as nodes connected by edges within a graph (Dongmin Bang et al.).

The core mechanism that drives GNNs is neighborhood aggregation (or message passing). To update the latent representation (embedding) of a specific biological entity, the network collects and aggregates feature vectors from its immediate neighbors (Kisan Thapa et al.).

[Neighbor Node A] \
[Neighbor Node B]  --> [Aggregation Function] --> [Updated Target Node Embedding]
[Neighbor Node C] /

By stacking multiple layers of these operations, information propagates across the network. A single node progressively incorporates broader topological context:

  • Layer 1: Captures immediate local interactions (e.g., direct physical binding between two proteins).
  • Layer 2: Captures secondary relationships (e.g., downstream enzymatic targets).
  • Higher Layers: Capture complex, non-linear pathways across the entire system (Kerr Ding et al.).

Depending on the biological context, this mechanism can be adapted. Graph Attention Networks (GATs) dynamically weight the importance of different interactions, while graph autoencoders map these complex networks into low-dimensional latent spaces while preserving the global topology (Jian-Xin Wang 0001 et al.; Chang Guo et al.).

Mirroring the Hierarchy of Life

The primary reason GNNs show remarkable efficacy in biological domains is structural alignment. As Wei Jiang et al. observe, the architecture of a GNN naturally mirrors the hierarchical organization of life itself, which spans from molecular interactions to cellular networks and organ-level systems.

This structural alignment makes GNNs an ideal framework for integrating prior biological knowledge and multi-omics data. Rather than forcing a model to learn complex biological rules from scratch, researchers can encode established biological facts—such as the semantic hierarchies of the Gene Ontology (GO) or known pathway-pathway interactions—directly into the graph structure (Xingyi Cheng et al.; Jianxin Wang 0001 et al.).

This "biologically-informed" design shifts deep learning away from uninterpretable "black boxes" toward models with bio-centric interpretability, allowing clinical researchers to trace predictions back to specific network pathways (Marie Zufferey et al.).

Empirical Impact Across the Biological Hierarchy

The power of relational biological modeling is evident across several distinct biomedical scales:

Protein and Molecular Interactions

By explicitly modeling the relational context between proteins rather than relying solely on individual sequence features, GNNs achieve superior accuracy in predicting protein-protein interactions (Hang Zhou et al.). This is particularly true for hyperbolic GNNs, which use non-Euclidean geometry to naturally capture the hierarchical, tree-like structures common in evolutionary and metabolic pathways.

Single-Cell Transcriptomics

In single-cell analysis, graph-based architectures leverage cellular interaction networks to outperform traditional neural networks in tasks like spatial domain identification and modality prediction (Chang Su et al.; Jiayuan Ding et al.). The graph structure allows the model to preserve cellular heterogeneity that is often smoothed out by standard matrix factorization techniques.

Drug Discovery and Repurposing

Multi-modal knowledge graphs serve as a unifying structure where drugs, diseases, genes, and tissue types are represented together. GNNs navigate these heterogeneous graphs to identify novel therapeutic targets and optimize drug combinations for complex conditions like cancer and neurodegenerative diseases (Jingwen Hui et al.; Yawei Ma et al.). By learning the topological patterns of successful treatments, the models can predict how a novel compound will perturb a disease network.

Key Foundational Work

Bioinformatics and medicine in the era of deep learning (2018) — contributed 2 concept elements to this direction.
PIKE-R2P: Protein-protein interaction network-based knowledge embedding with graph neural network for single-cell RNA to protein prediction (2021) — contributed 2 concept elements to this direction.
KG4SL: knowledge graph neural network for synthetic lethality prediction in human cancers (2021) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_graph_neural_networks_for_biological_net_f74a86df,
  title = {graph neural networks for biological networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-neural-networks-for-biological-networks/rdr_f74a86df8e1d41d99610f1cc8ea2db27},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_graph_neural_networks_for_biological_net_f74a86df_v1,
  title = {graph neural networks for biological networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-neural-networks-for-biological-networks/rdr_f74a86df8e1d41d99610f1cc8ea2db27/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake