AI tools for exploiting macromolecular structure knowledge graph

Macromolecular structure data provides valuable insights into the functional mechanisms of biological processes. Recent advances in experimental structure determination techniques and structure prediction have accumulated an abundance of structure models, revolutionising life-science research. The structure data can only benefit the broader scientific community if it is enriched by adding structural and functional annotations to provide biological context through analysis using various scientific software. PDBe Knowledge Base (PDBe-KB; be-kb.org) is an open, collaborative consortium that supports the use of 3D-structure data by collating structural and functional annotations for PDB data from 34 partner resources worldwide and representing this rich data as a PDBe knowledge graph (PDBKG) using the robust Neo4j graph database. The PDBKG complements the minimal annotations available in the PDB, is distributed via the FTP area, and powers the PDBe API and web portal. The PDBKG has allowed an efficient mechanism to aggregate all available structure data. Over the past five years after its establishment, PDBe-KB has successfully supported basic and translational research in life sciences, as evidenced by the more than 300K unique users each year with over 110 citations. To fully exploit the capabilities of the PDBKG, users must have expertise in Neo4j graph database and maintain infrastructure or depend on the more than 100 API calls developed by the PDBe team. These limitations restrict the use of this powerful resource, resulting in and slowing down the exploitation of macromolecular structure data to accelerate fundamental and translational research. This problem will be exacerbated when the expected advances in experimental techniques to study conformational changes or macromolecules inside cells in their natural context or AI methods to predict large multi-component complexes and their conformations materialise in the coming years. New technologies to democratise access and visualisation of 3D-structure models and associated annotations are essential to address the problem. In this project, we aim to prototype new ways to democratise access to the KG by supporting the use of natural language queries and contextualised visual storyboards to enable discoveries using 3D-structure data by: Developing a fine-tuned LLM model to translate scientific questions posed using natural language into cypher queries to interrogate the PDBKG stored in a Neo4j database and generate contextualised visual storyboards for displaying results using Mol* 3D-view specification capabilities and other interactive visualisation components. Implement a web-based prototype system integrating the fine-tuned LLM and visual storyboards to facilitate iterative user interaction and exploit the macromolecular structure data in the PDBKG. The prototype system that enables natural language queries and contextualised interactive visual storyboards would provide a highly efficient system to exploit enriched structure data for the over 800 million models available from multiple structure resources such as PDB and AlphaFold Protein structure database and ESMFold. The capability to translate natural language queries into cypher queries will also lower the barrier to in-house use of the PDBKG and accelerate innovation and efficient use of limited resources to accelerate PDBe API development. The new system will realise the promise of structural biology by driving hypothesis generation and experiment design to study macromolecular function and facilitate basic and translational research, e.g., in drug and protein design, supporting the design of better therapeutics and the green economy, for example, by aiding in the design of more effective enzymes for breaking down plastic pollution.