The summary.
MUCH OF THE FIELD IS STRING ALGORITHMS AND STATISTICS AT SCALE
With enough biology to know what the question means. Files are enormous — stream rather than load, and index for regional access.
REFERENCE VERSIONS AND COORDINATE CONVENTIONS CAUSE THE MOST SILENT ERRORS
Formats differ in whether coordinates start at zero or one, and chromosome naming differs between sources. Mixing versions produces results that are wrong but not obviously so.
PIN EVERY TOOL VERSION, REFERENCE AND PARAMETER
Tool versions change results, and undocumented ones make an analysis irreproducible.
Never put a spreadsheet in the middle of a pipeline — the software alters values on import and has corrupted published datasets.
CORRECT FOR MULTIPLE TESTING, ALWAYS
Testing tens of thousands of genes guarantees false positives at conventional thresholds.
Batch effects can perfectly mimic the biological effect you are looking for — randomise processing at collection time.
SPLIT MACHINE LEARNING DATA BY SIMILARITY, NOT RANDOMLY
Biological sequences are related by evolution, so random splitting leaves near-identical examples on both sides.
TREAT GENOMIC DATA AS PERMANENTLY IDENTIFYING
It can be re-identified, it cannot be revoked, and it exposes relatives who never consented.
