Genomic foundation models: task-first evidence guide

Research applicability

Accept as a bounded educational model-selection guide, not as a clinically validated protocol or a benchmark-result endorsement. The linked author guide gives substantive task distinctions, model-selection context and evidence categories; the forum explicitly identifies it as the author’s site.

Getting started

Editorial, untested guidance: read the task decision tree, specify whether your question concerns calling, effect scoring or classification, then consult the underlying primary studies before designing a local evaluation. Do not use the guide alone for patient decisions.

Maintenance

Live source retrieval is not a maintenance guarantee.

License

Not independently audited; check source and dependency terms before reuse.

Workflow & human oversight

Resource-level automation is unspecified. Researchers define the task, inspect sources/code and validate outputs; this review did not execute a workflow.

Workflow level is not an evidence score; higher is not better evidence.

Approach

Tim Richardson’s guide distinguishes variant calling, post-calling variant-effect scoring and custom DNA-sequence classification, and organises genomic foundation-model choices around the biological task and the kind of supporting evidence.

Limitations

Partial-documentation review: forum announcement and guide introduction, decision tree, model overview and opening efficiency discussion were read. The guide is primary evidence of its own tutorial content, not primary validation of the cited models. Downstream papers, benchmark data, cloud costs, code and clinical claims were not audited or reproduced. No model superiority, numeric outcome or clinical validity is endorsed. Not a validated clinical protocol.

Reproducibility

Documentation retrieval and source hashes verified; no software, benchmark or outcome reproduced.

Adoption requirements

Reading requires genomics and model-evaluation literacy. Any subsequent model experiment needs task-appropriate held-out data, checkpoint-specific license review, suitable compute and institutional patient-data governance; these are editorial prerequisites, not a tested recipe.