Split the annotation into separate axes, and widen what it reports

[emattei]

Why

location answered two questions at once, so a guide could not be genic and
exonic and in a promoter, which is the ordinary case for a first exon. And
the TSS set was one site per gene taken from the reference transcript, so a
guide at an alternative promoter read FALSE.

What changed

location                genic | intergenic          protein-coding only
feature                 exon | intron | ''          reference transcript
other_gene_types        semicolon-separated types of overlapping genes of
other_genes             other types, and their ids
promoter_enst/ensg/gene every transcript whose promoter window the
                        protospacer sits in
distance_to_gene_start  signed, from the gene's annotated 5' end
distance_from           midpoint (default) | target_site

The calls stay protein-coding. The store keeps one gene row for every other
type so a lncRNA overlap can be reported without it deciding where the guide
sits; rows() defaults to the annotated types so nothing widens by accident.
A store built before this lacks gene_type and is refused by name.

cut_site is renamed target_site throughout, and distances now measure
from the protospacer midpoint: this is CRISPRi, and dCas9 occupies the whole
protospacer rather than acting at one base. --distance-from target_site
restores the nuclease convention, and cds_position keeps it regardless.

Checked on the hiPSC-to-EC library, 43,778 guides

Faithful where it should be:

genic/intergenic vs the old call     0 rows differ
feature vs the old exon/intron       0 rows differ
promoter  FALSE -> TRUE          1,567
promoter  TRUE  -> FALSE             0

Gained only, never lost. It also finds 6,262 rows overlapping a gene of
another type, 5,899 of them lncRNA, which the old store could not see at all.

In the IGVF export, 284 elements change class, all of them into promoter:
115 from exon, 160 from intron, 9 from distal element. The nine were
checked individually: in 9 of 9 the reference transcript's TSS is not the
gene's annotated 5' end, inset by 69 bp to 72 kb, so they were measured from
the wrong place before rather than reclassified by accident.

This changes what would be submitted: 3,419 guides, 7.8% of the library,
move to promoter.

🤖 Generated with Claude Code