Collaborative Research: III: Medium: Contextualized and Multimodal Foundation Models for Graph Data in Scientific Discovery

Recent progress in deep learning has demonstrated the potential of foundation models built on massive datasets, particularly in scientific discovery. In the sciences, data ranging from particles, to molecules, to cells, to brain activity can be represented by nodes on a graph or as signals on a graph substrate. Therefore, successful AI foundation models in scientific discovery are required to possess the capability of handling such graph-structured data and integrating with other types of data such as text, images, and tabular data. The proposed model can be used as a general substrate to help scientists predict and understand a variety of data that are expressed in graphs, such as molecules, proteins, and connectome. Existing methods for building graph foundation models for scientific discovery are in general severely limited in that they: 1) do not consider contexts in which the vertices of a graph can themselves be complex structures such as molecular graphs; 2) do not incorporate multimodal information in the form of knowledge graphs and text; 3) have limited forms of message passing in the form of local averaging; and 4) are not versatile and have limited performance gains due to diversity of downstream tasks and graph data distributions. This team of researchers will address these issues by developing a general foundation model framework for data represented as a graph in scientific domains by systematically addressing these key limitations. The framework incorporates novel approaches of multi-level graph neural networks, graph signal processing, multimodal graph learning, graph-specific fine-tuning, and in-context learning. By capturing human scientific knowledge and express the complexity of the natural world, our framework has the potential to dramatically transform machine learning models in scientific discovery and will allow us to tackle a wide range of complex scientific tasks, even with scarce supervision labels. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.