Using agentic AI for data harmonization and access: lessons from NMDC

This presentation outlines how the National Microbiome Data Collaborative (NMDC) is leveraging agentic AI to improve metadata capture, data harmonization, and programmatic access to multi-omics microbiome data. It emphasizes that foundation models require well-labeled, curated, and linked data, and positions NMDC as an AI-ready platform integrating diverse omics datasets with structured metadata and APIs. The talk highlights challenges in applying community standards like MIxS and the prevalence of incomplete or inconsistent metadata in public repositories, motivating the integration of AI into the NMDC submission portal. Initial AI features include metadata suggestion using LLMs and deterministic lookups, evolving toward agentic interfaces capable of literature discovery, dataset linking, and automated spreadsheet-to-schema mapping. Recognizing new failure modes such as hallucination and over-extrapolation, the presentation advocates “harness engineering” using LinkML-based schemas, validation tools, and guardrails to ensure quality and reproducibility. The key takeaways are that agents can perform complex data wrangling with human oversight, NMDC can serve as a surface for BER AI applications, and the submission portal can function as a primary agentic data-ingest interface.