Repository navigation
TElocal
On by default (telocal.enabled: true); TElocal provides locus-level
TE quantification as a complement to TEcount's subfamily-level counts. Where
TEcount pools all copies of a TE subfamily into one count (e.g. all L1PA2
instances → one L1PA2 row), TElocal resolves TEs per genomic instance
using the same EM-based multi-mapper assignment, reporting counts per locus
(e.g. chr1:564318:564741(L1PA2:+):L1PA2:L1:LINE).
TElocal uses the same unsorted BAM as TEcount (results/star/{sample}_Aligned.out.bam), so no re-alignment is needed. Results go to
results/telocal/{sample}.cntTable.gz. Per-sample summary barplots
(gene-vs-TE assignment and TE class composition) are rendered in the MultiQC
report (results/telocal/qc/telocal_assignment_mqc.json and
telocal_te_class_mqc.json).
A sample-QC view mirrors the TEcount one (telocal.qc.*: enabled,
view-only filters, pca_transform, feature_class): the per-sample tables are
merged into a locus x sample counts matrix
(results/telocal/counts_matrix.tsv.gz) that drives a PCA scatter and a
sample-to-sample clustering heatmap in the report. The default transform is
log2 — locus-level matrices hold one row per TE instance (often millions),
which makes DESeq2's vst/rlog prohibitively slow; opt in deliberately if you
want them.
TElocal requires a pre-built .locInd file — a pickled index built from the
same TE GTF that TEcount uses. Build one with:
TElocal_indexer --afile TE.gtf --itype TE # ~30 GB RAM for human genomeOr download prebuilt indices from https://www.mghlab.org/software/telocal. Set the path in your config:
telocal:
enabled: true
locind: /path/to/TE_annotation.locIndWhen locind is empty, the workflow auto-builds the index from your TE GTF
into results/telocal/telocal.locInd (the .locInd suffix is mandatory —
TElocal rejects any --TE file whose path lacks it). It is kept by default;
outputs.keep_telocal_index: false deletes it once all TElocal runs finish
(a user-provided locind path is never touched). Auto-building trades disk
for convenience.
Auto-building uses telocal.indexer: fast (default) — a drop-in
reimplementation of TElocal_Toolkit's own index builder that fixes an
O(n²) bottleneck (a list-membership check that should have been a set),
verified against the original on real mouse rmsk data (3.7M TE instances):
identical output, ~1000x faster (a build that took ~21 h dropped to ~1 min).
Set telocal.indexer: legacy to fall back to TElocal_Toolkit's own
unmodified builder if ever needed.
Set telocal.enabled: false to skip locus-level quantification entirely.
TEcount and TElocal report at two different levels of the same hierarchy:
subfamily L1PA2 <- ONE TEcount row: all copies pooled
|
+-- L1PA2_dup1 chr1:1,000-7,200 <- one TElocal row each
+-- L1PA2_dup2 chr4:88,100-94,300
+-- L1PA2_dup3 chrX:2,410-8,600
So a subfamily that looks uniformly expressed in TEcount may turn out to be
one or two highly-expressed copies in TElocal, with the rest silent -- which
is the whole reason to run both. The chimera screens' te_id is the
insertion-level key, so it joins directly against TElocal rows; their
te_subfamily column joins against TEcount rows.
TElocal's cntTables key TE rows by
transcript_id:gene_id:family_id:class_id but never include genomic
coordinates. results/telocal/telocal_locations.bed maps every such key
to its coordinates as a standard BED6 (chrom, start, end, name, score, strand, 0-based; name is the TE key, score unused) — one pass over the
TE GTF, always built when telocal.enabled is true — so you can join
cntTable output against name and also load the file directly into
bedtools/IGV. Kept uncompressed (unlike most pipeline outputs) so IGV can
load it directly — plain gzip isn't the random-access-indexed format IGV
needs.