Skip to content

Interpreting Your Results

Ali Altintas edited this page Aug 28, 2026 · 5 revisions

Interpreting Your Results

A finished run produces a lot of sections that all mention TEs — TEcount, TElocal, two chimera screens, two cross-evidence joins — and it is easy to read them as several confirmations of the same finding when most of them are not independent. This page is the map: what each layer is, which ones can corroborate each other, and what order to read them in.

The MultiQC report opens with a Start here section that says the same thing for your run specifically (which layers actually ran, and whether a cross-check was available).

The two questions

The pipeline answers two separate questions off one alignment:

  1. What is expressed? — genes and TEs, quantified by TEcount and (optionally) TElocal.
  2. Where are genes and TEs fused into one transcript? — the two chimera screens.

Almost every misreading comes from treating an answer to question 1 as evidence for question 2, or vice versa. A TE subfamily being highly expressed is not evidence that it forms chimeras; a chimeric junction is not evidence that the TE is expressed at any meaningful level.

What is independent evidence, and what isn't

                    results/star/{sample}_Aligned.out.bam
                    (ONE alignment -- everything below shares it)
                                     |
        +----------------------------+----------------------------+
        |                            |                            |
   EXPRESSION                   EVIDENCE  A                  EVIDENCE  B
        |                            |                            |
   +----+----+          chimera.junction screen      chimera.assembly screen
   |         |          (STAR chimeric reads)        (StringTie structure)
TEcount   TElocal                    |                            |
 (sub-     (per                      |                            |
 family)  insertion)                 +------------+---------------+
   |         |                                    |
   +----+----+                          candidates_with_junction
        |                                    _evidence.tsv.gz
   same reads, same EM,                          |
   two resolutions --              found by BOTH = strongest call
   NOT confirmation                  the pipeline makes
   of each other

Read the diagram as three claims:

  • TEcount and TElocal are one layer, not two. Same reads, same EM multi-mapper assignment, different granularity. TElocal tells you which copy drives a TEcount subfamily signal. It never independently confirms it. See TElocal.
  • The two chimera screens are genuinely independent. They fail in different directions, which is the whole point — see the next section.
  • star.two_pass is a knob, not a result. It improves junction discovery for everything downstream, so it changes the numbers in both screens. It is never itself an answer. See Strandedness and STAR 2-pass.

Why two chimera screens

They are blind to opposite things, so the union is bigger than either and the intersection is stronger than either:

  a gene-TE chimera spliced through an ORDINARY canonical intron
  ---------------------------------------------------------------
     gene exon                              TE
     ========]-----------GT......AG-----------[==========
              \___ one clean, linear splice ___/

     junction screen : INVISIBLE  (STAR aligns this fine -- it is
                                   not a chimeric read at all)
     assembly screen : found      (StringTie builds the transcript)


  a gene-TE fusion with a NON-COLINEAR breakpoint
  ---------------------------------------------------------------
     gene exon                    TE on another chromosome / strand
     ========]                    x  [==========
              \__ read cannot be explained as one alignment __/

     junction screen : found      (this IS STAR's Chimeric.out.junction)
     assembly screen : INVISIBLE  (no assembler builds this structure)

So: absence in one screen is not evidence against a candidate. Presence in both is the strongest evidence the pipeline produces, and it is exactly what results/chimera_assembly/candidates_with_junction_evidence.tsv.gz records (and what the transcript-evidence screen's What to look at section marks with a star).

The confidence ladder

From weakest to strongest, for a gene-TE chimera candidate:

Tier What you have How much to trust it
— A row in all_events.tsv.gz Nothing yet — this is the raw catalog, mostly non-gene-TE and mostly artifact
1 A gene-TE call (te-gene-chimeras.tsv.gz) The floor; direction may still be ambiguous
2 Seen in multiple samples Reproducible, so not a one-library accident — but a sequence-driven template switch recurs across libraries too, so this is weaker than it looks
3 canonical: yes A recognised splice motif. Real introns are ~100% canonical; template-switching, ligation and PCR chimeras carry none
4 Found by both screens Two methods with opposite blind spots agree

Tier 3 is the single biggest jump, and it is where most triage should start. See Confidence Tiers for the diagram, the reasoning behind the ordering, and how to filter down to a defensible list.

The full ladder, including the two signals deliberately left out of it and the measurements behind that decision, is on its own page: Confidence Tiers. The short version: read depth and telocal_active (whether the TE locus is expressed) are reported but never promote a pair — on a real run TE expression covered 91% of gene-TE junctions and was anti-correlated with the splice motif.

A caution about tier 3 in practice: on real data only a small percentage of all chimeric junctions carry any motif (~5% in a real mouse run). That is expected and is not a sign the run failed. What matters is enrichment — but enrichment measured against the right background.

Compare within a donor group, not against other. The donor side alone moves the canonical rate a great deal, so gene_to_te vs other credits the class for its donor being a gene at all. Measured on a real run:

Donor Class Canonical vs its own donor group
gene gene_to_te 10.0% 1.19×
gene gene_to_gene 8.5% (background)
gene gene_to_other 8.5% 1.00×
te te_to_gene 7.2% 1.72×
te te_to_te 4.2% (background)
te te_to_other 2.6% 0.63×
other other_to_gene 5.4% 1.34×
other other 4.0% (background)
other other_to_te 3.5% 0.87×

Both true gene↔TE directions beat their own donor group, and the two classes with a TE but no gene partner (te_to_other, other_to_te) fall below theirs. That pattern is the evidence the classification is capturing something real. The margin is modest — well under 2× — not several-fold; the 2.5× you get by comparing gene_to_te against other is mostly the donor confounder, not signal.

Note also that read depth is deliberately not on this ladder. Depth is the metric most inflated by artifacts: a hot PCR chimera can easily be the deepest event in a run. Both What to look at sections rank depth last, on purpose.

What to open, in order

  1. QC first. FastQC, STAR alignment rates, then the Strandedness check table — it flags any sample whose sample sheet disagrees with what RSeQC infers. A wrong strandedness silently distorts every count downstream, so confirm it before trusting anything else. See Strandedness and STAR 2-pass.
  2. Sample QC. The PCA and sample-distance heatmap. If replicates do not group, fix that before interpreting any individual feature.
  3. Expression. TEcount for which subfamilies move; TElocal for which copy is responsible. See TEcounts Sample-QC.
  4. Gene-TE chimeras: read evidence → What to look at. The ranked gene-TE candidates, not the raw catalog.
  5. Gene-TE chimeras: transcript evidence → What to look at. Prefer candidates the read-evidence screen also found.
  6. Differential results. results/tetranscripts/ — only when the sample sheet has a condition column. These are not rendered in the report. See Automatic Differential Analysis.

Common misreadings

"The same TE shows up in TEcount, TElocal and the chimera screen — that's three lines of evidence." It is one expression measurement at two resolutions, plus one structural observation. Two, at most.

"Only 5% of junctions are canonical, so the screen is broken." No — the absolute rate is low by nature. See the enrichment caution above.

"gene_to_te is 2.5× the other rate, so it's strongly enriched." Most of that ratio is the donor side, not the gene-TE relationship. Compare it to gene_to_gene instead — see the table above.

"This candidate has 900 reads, it must be real." Depth is the weakest signal here. Check canonical and replicate support first.

"The TE is expressed, so the chimera is real." No — see the table above. An expressed TE locus is the common case, and its junctions are slightly less likely to carry a splice motif.

"The assembly screen didn't find it, so it's not real." The screens have opposite blind spots by construction. Absence in one says nothing.

"direction: gene_to_te, so the gene is upstream." Check direction_ambiguous first. TEs sit inside gene bodies constantly, so a breakpoint often hits both a gene and a TE; when it does, the direction was decided by branch order, not by evidence.

Where each file lives

See Output Layout for the full tree and Chimera Detection for what each chimera column means and which per-sample file is which.

Clone this wiki locally