MAGen develops and validates ML/AI tools for genomic translational research using multimodal data and shared, reproducible resources. This page summarizes the types of data MAGen works with, how data are handled across sites, and what tools and resources the consortium aims to make available.

Data used in MAGen

MAGen leverages existing datasets available at participating sites and supports integration of diverse data types to better predict how individuals with pathogenic (germline) variants manifest disease.

Common data modalities include (as available by site and use case):

• Genomic data (e.g., germline variants; sequencing- or array-derived data)
• Other omics (when available) such as transcriptomics, proteomics, metabolomics
• Clinical phenotypes derived from electronic health records (EHRs)
  (diagnoses, labs, medications, procedures, clinical notes-derived features where appropriate)
• Demographics and ancestry-related variables (used carefully, with attention to bias and interpretability)
• Environmental and lifestyle factors (where available)
• Social determinants of health (SDOH) and contextual measures
• Biomarkers and clinical measurements
• Wearables / biosensors (when available)

Because MAGen is a multi-site consortium, not every dataset exists at every site; MAGen’s approach is designed to support robust tool development and evaluation across heterogeneous data environments.

Cross-site data strategy: harmonize, then validate

MAGen’s data approach balances collaboration with local governance:

Shared definitions and data models
MAGen establishes common specifications (e.g., cohort definitions, phenotype algorithms, feature dictionaries, and metadata expectations) to support consistent implementation across sites.

Local control, local governance
Each site maintains stewardship of its own data and complies with institutional policies, IRB requirements, and applicable regulations. Where needed, analyses may be performed locally with shared code and standardized reporting.

Independent validation on distinct datasets
A core MAGen principle is that tools are evaluated on independent datasets at other sites, strengthening evidence for generalizability and reproducibility.

Data quality, documentation, and reproducibility

MAGen emphasizes practices that enable trustworthy ML/AI research:

• clear inclusion/exclusion criteria and cohort provenance
• transparent feature definitions and data processing steps
• versioning of datasets (where feasible), code, and model artifacts
• consistent evaluation metrics and reporting templates
• calibration and subgroup performance reporting when appropriate

Privacy and responsible use of data

MAGen integrates Ethical, Legal, and Social Implications (ELSI) and responsible AI considerations throughout its data lifecycle, including attention to:

• privacy and security protections
• appropriate data use and governance
• bias and representativeness of datasets
• transparency about limitations and intended use
• evaluation across subgroups to identify potential inequities

Tools & resources (what MAGen produces)

MAGen aims to generate tools and resources that can be reused and tested across institutions. Depending on the project, outputs may include:

ML/AI tool packages

Validated, portable implementations that may include:

• model specifications and parameters
• training and inference pipelines
• input/output schemas
• documentation of assumptions, intended use, and limitations
• evaluation scripts and test harnesses to support cross-site validation

Shared protocols and templates

• phenotype and cohort definition documentation
• data dictionaries / feature lists
• model reporting checklists (performance, calibration, subgroup analyses)
• cross-site validation playbooks and standard operating procedures

ELSI framework

• methods and tools to systematically identify and address risk
• mechanisms for incorporating community-based input into model design
• playbook for continuous monitoring

Publications and dissemination materials

• peer-reviewed manuscripts and preprints
• methodological reports and consortium summaries
• presentations and workshop materials (as available)