Bridge2AI Recommendations for AI-Ready Genomic Data

Rapid advancements in technology have led to an increased use of artificial intelligence (AI) technologies in medicine and bioinformatics research. In anticipation of this, the National Institutes of Health (NIH) assembled the Bridge to Artificial Intelligence (Bridge2AI) consortium to coordinate development of AI-ready datasets that can be leveraged by AI models to address grand challenges in human health and disease. The widespread availability of genome sequencing technologies for biomedical research presents a key data type for informing AI models, necessitating that genomics data sets are AI-ready. To this end, the Genomic Information Standards Team (GIST) of the Bridge2AI Standards Working Group has documented a set of recommendations for maintaining AI-ready genomics datasets. In this report, we describe recommendations for the collection, storage, identification, and proper use of genomics datasets to enable them to be considered AI-ready and thus drive new insights in medicine through AI and machine learning applications.

💡 Research Summary

The paper “Bridge2AI Recommendations for AI-Ready Genomic Data” presents a comprehensive set of guidelines developed by the Genomic Information Standards Team (GIST) within the NIH’s Bridge to Artificial Intelligence (Bridge2AI) consortium. The central objective is to define the standards and practices necessary to create genomic sequencing datasets that are “AI-ready”—meaning they are explainable, reusable, and computationally accessible for robust artificial intelligence and machine learning applications in biomedicine.

The authors begin by establishing the motivation: the explosive growth of genomic data coupled with the rise of AI presents a tremendous opportunity for discovery, but only if the data is prepared with the rigor required for automated, trustworthy analysis. They anchor the concept of AI-readiness in the principles of being FAIR (Findable, Accessible, Interoperable, Reusable), fully reliable, robustly defined, and computationally accessible.

The core of the paper is a detailed, prescriptive breakdown of the essential metadata that must accompany genomic datasets. This metadata is categorized into several critical domains, each presented in a corresponding table with “Must” and “Should” requirement levels. First, Sample Origin and Preparation metadata must include details like sample storage conditions (e.g., flash-frozen, FFPE), collection date/location, biospecimen type, clinical diagnosis, and pathological state. For human samples, sex assigned at birth, genetic ancestry, and phenotypes are recommended. Second, Sequencing Preparation and Process metadata must capture library preparation methods, the use of unique molecular identifiers (UMIs), sequencing platform and instrument model, run dates, and locations. For targeted sequencing, a BED file defining the covered regions is required.

Third, and crucial for reproducibility, are the Genomic Sequencing Processing and Procedure specifications. This includes the analyses performed, the exact bioinformatics workflows and software versions used, all parameters and algorithms (including any modifications), and the specific reference genome version and source. The use of standard sequence digests (e.g., GA4GH’s sha512t24u) for identifiers is recommended. Fourth, comprehensive Quality Control metrics, such as read mapping quality, mean depth of coverage, percent of target bases covered sufficiently, error rates, and GC content, are necessary to assess data reliability.

Finally, the paper provides clear recommendations for Data Storage Formats. It mandates that aligned read data be stored in CRAM files and that variant calls be stored as gVCF files compliant with the GA4GH Variant Call Format (minimum version 4.2, with 4.3 preferred). Most significantly, it strongly advocates for representing variation data using the GA4GH Variation Representation Specification (VRS), which provides a semantically precise, computable framework essential for cross-dataset integration by AI systems.

To illustrate the practical necessity of these guidelines, the authors include a use case demonstrating how missing metadata on sample source (saliva vs. blood) can lead an AI model to identify spurious associations based on technical artifacts rather than true biology. In conclusion, while focused primarily on small-variant detection studies, these recommendations establish a foundational framework for creating genomic data that is truly prepared for the demands and opportunities of the AI era, thereby enabling more trustworthy and impactful biomedical discoveries.