FIMP: Foundation Model-Informed Message Passing for Graph Neural Networks

TL;DR

FIMP introduces a novel message-passing framework for GNNs that leverages pretrained foundation models' self-attention mechanisms to improve performance on graph-based tasks across various domains.

Contribution

The paper proposes FIMP, a new GNN framework that repurposes foundation models' self-attention layers for effective cross-node message passing in graphs.

Findings

FIMP outperforms strong baselines on real-world graph datasets.

FIMP effectively leverages foundation models in zero-shot and finetuned settings.

The approach demonstrates versatility across image, biological, and brain activity data.

Abstract

Foundation models have achieved remarkable success across many domains, relying on pretraining over vast amounts of data. Graph-structured data often lacks the same scale as unstructured data, making the development of graph foundation models challenging. In this work, we propose Foundation-Informed Message Passing (FIMP), a Graph Neural Network (GNN) message-passing framework that leverages pretrained non-textual foundation models in graph-based tasks. We show that the self-attention layers of foundation models can effectively be repurposed on graphs to perform cross-node attention-based message-passing. Our model is evaluated on a real-world image network dataset and two biological applications (single-cell RNA sequencing data and fMRI brain activity recordings) in both finetuned and zero-shot settings. FIMP outperforms strong baselines, demonstrating that it can effectively leverage state-of-the-art foundation models in graph tasks.