BERIL: Integrated AI Infrastructure for Genome-to-Phenotype Research Using the BER Data Lakehouse DRAFT

The BER Data Lakehouse seeks to provide consistently labeled, harmonized datasets designed for AI-driven reasoning and inference across multiple heterogeneous data sources. The DOE Systems Biology Knowledgebase (KBase), Joint Genome Institute (JGI), Environmental Molecular Sciences Laboratory (EMSL), National Microbiome Data Collaborative (NMDC), the cross-BER data integration initiative BERtron, and other partners have contributed to the design of the Lakehouse infrastructure and data modeling. Additionally, recent efforts by these groups have demonstrated the value of specialized AI reasoning agents, such as the KBase narrative agent, which automates genome assembly, annotation, and manuscript preparation, and a literature-based curation agent that can instantly update genes in SwissProt. BERtron and NMDC have built platforms that leverage AI and ontologies to facilitate dynamic scientific discovery.

Building on these efforts, this project will develop an extensible, self‑updating AI ecosystem (BERIL, the BER Integrative Layer) that leverages the BER Data Lakehouse to orchestrate specialized reasoning agents and accelerate genome‑to‑phenotype discovery. BERIL’s unified agentic AI infrastructure will include a central orchestration agent (generalist agent) working collaboratively with multiple specialized agents. The generalist agent coordinates overall activities and user interaction and facilitates literature searches, tool execution and Data Lakehouse queries. It also manages reasoner traces, collecting detailed records of reasoning processes and engaging in rationale-driven refinement loops. This dual-layer feedback approach combines curiosity-driven interactions (reasoner traces) and formal benchmarking, enabling continuous improvement and close alignment with domain expert knowledge. The integrated agent coordination framework will employ a hybrid approach, combining curiosity-driven exploration and structured modular extensions via the model context protocol (MCP). Relevant analytical tools developed by BER user facilities and programs will be integrated—for example, the JGI Metabolomics team is developing an MCP for their Web of Microbes exometabolome data. Other specialized groups could develop MCPs for interpretation of their data and to describe their capabilities for experimental validation of predictions. These tools will dynamically interact, exchanging insights and maintaining adaptive mappings between domain-specific ontologies and knowledge graphs.

The specialized experiment design agent (EDA) will convert predictions to actionable laboratory tasks and metabolic pathway optimization by evaluating genome-wide predictions, strategically identifying validation experiments that reduce uncertainty while minimizing resource utilization. The EDA will generate ranked candidate experiments via MCP endpoints that describe experiment capabilities, cost, run time, etc. of the experiment providers, enabling interaction with the automated lab controller agents being planned by EMSL and ANL. Domain expert team members will define and regularly update a gold standard evaluation benchmark set of gene function assertions, genome-to-phenotype assertions and experiment designs. A random sample of new agent predictions against this benchmark set will be manually evaluated by domain experts, providing feedback to improve automated agent reasoning and develop mitigation strategies to identify and close any gaps. Researchers will measure success of this approach via metrics such as reduction in experimental design and validation costs and timeframes and uptake across BER user facilities.

As a pilot demonstration of BERIL, the team will select genes, pathways, and microbial strain targets that support DOE BER biomanufacturing and bioproduct development objectives, ensuring that the focus aligns strategically with critical BER research priorities and maximizes impact.