Skip to content

Confidence Tiers

Ali Altintas edited this page Aug 28, 2026 · 2 revisions

Confidence Tiers

results/chimera_junction/chimera_evidence.tsv.gz gives every (gene_id, te_id) pair a confidence_tier from 1 to 4. This page is what those numbers mean, what was deliberately left out of them, and how to use them to cut a candidate list down to something defensible.

For where this table sits among the pipeline's other outputs, see Interpreting Your Results.

The ladder

A tier is the highest rung a pair reaches — the rungs are not cumulative requirements, and a pair does not need the ones below it. Each rung answers a different way the call could be wrong, and they are ordered by how well they actually separate signal from artifact:

  strongest
      ^
      |                    ,------------------------------.
      |                    |  4   BOTH SCREENS AGREE      |
      |                    `------------------------------'
      |                  the junction screen (chimeric reads) and the
      |                  assembly screen (transcript structure) are blind
      |                  to opposite things -- see Chimera Detection. Both
      |                  implicating the SAME insertion is notable -- but
      |                  read the measured caveat below before relying on
      |                  it: this rung is the least established of the four.
      |
      |              ,--------------------------------------.
      |              |  3   A CANONICAL SPLICE MOTIF        |
      |              `--------------------------------------'
      |            STAR reported GT/AG, GC/AG or AT/AC at the break.
      |            Real introns are ~100% canonical; template switching,
      |            ligation and PCR chimeras carry no motif at all.
      |            The single biggest jump on this ladder.
      |
      |        ,----------------------------------------------------.
      |        |  2   SEEN IN MORE THAN ONE SAMPLE                  |
      |        `----------------------------------------------------'
      |      Not a one-library accident. Weaker than it looks, though:
      |      a template switch driven by local sequence happens in EVERY
      |      library, so replication does not rule the main artifact out.
      |
      |  ,--------------------------------------------------------------.
      |  |  1   A GENE-TE CALL EXISTS                                   |
      |  `--------------------------------------------------------------'
      |  The floor. One breakpoint hit a gene, the other hit a TE. Check
      |  `direction_ambiguous` before trusting even the direction.
  weakest

Everything below tier 1 — the raw all_events.tsv.gz catalog — is mostly neither gene-TE nor real, and carries no confidence claim at all.

What is deliberately not on the ladder

Two things look like support and are not. Both are still reported as columns; neither promotes a pair:

   +----------------------+   why it is not evidence
   |  read depth          |   The metric artifacts inflate MOST. A hot PCR
   |  (junction_reads)    |   chimera is often the single deepest event in
   +----------------------+   a run. Useful only to rank WITHIN a tier.

   +----------------------+   Measured on a real 4-sample mouse run
   |  telocal_active      |   (19,503 gene-TE junction events):
   |  (TE locus expressed)|
   +----------------------+       present on 91% of junctions
                                  -> discriminates almost nothing

                                  canonical rate when the TE locus is
                                    EXPRESSED (count > 10) :  6.7%
                                    QUIET     (count <= 10) : 10.2%
                                  -> anti-correlated with the one signal
                                     that separates real splicing from
                                     artifacts

                              More reads at a locus means more chances for
                              template switching. Expression is CONTEXT --
                              worth knowing, not a reason to believe.

telocal_active was tier 5 in the first version of this table, ranked above canonical, on the reasoning that independent evidence the TE is transcribed corroborates the chimera. The measurement above is why it was removed. The reasoning was plausible, so the numbers are recorded here rather than just the conclusion.

Tier 4 is under review

Measured on a real 4-sample mouse run, screen agreement did not behave like corroboration:

measured
genes flagged by both screens 3,124
...also agreeing on the TE insertion 127 (4.1%)
median TEs nominated per gene, each screen 1
canonical rate among both pairs 7.8% (background: 8.6%)
both pairs seen in >1 sample 6 of 129 (4.7%)

Two things follow. Screen agreement showed no enrichment for splice motifs — 10 observed against 11.1 expected by chance. And since each screen nominates about one TE per gene, agreeing on the same insertion ~4% of the time is close to what independent picking produces when a gene contains a few dozen TE insertions.

The likely reason is that this join is at (gene, TE insertion) locus level, not event level. Both screens implicating the same locus does not mean they saw the same molecule: the junction screen found a non-colinear break there, the assembly screen built a transcript there, and those can be different things.

The measurement that settles it is the number of TE insertions inside the genes both screens flag — if that is ~25, 4.1% is chance. Until then, treat tier 4 as "two methods pointed at the same locus" rather than "one event confirmed twice", and note that a tier-3 pair with a splice motif and replicate support may well be the better candidate.

Why canonical outranks reproducibility

These two swapped places for a specific reason. Ask what each one rules out:

   artifact mode                      caught by
   ---------------------------------  --------------------------------
   template switching during RT       canonical  (no splice motif)
     - sequence-driven                NOT replication -- the same
     - recurs in every library          local sequence does it again
                                        in every library

   ligation / PCR chimera             canonical  (no splice motif)
     - largely random partners        partly replication

   one-off library accident           replication

The dominant artifact mode in this screen recurs across libraries, so seeing a junction in every sample does not rule it out. A splice motif does.

Reading a row

  gene_id             te_id           found_by  tier  canonical  samples
  ------------------  --------------  --------  ----  ---------  -------
  ENSMUSG00000098284  L1_Mus3_dup469  both      4     yes        1
  ENSMUSG00000041995  MT2B1_dup5641   both      4     no         3
  ENSMUSG00000012345  B1_Mus1_dup112  junction  3     yes        2
  ENSMUSG00000067890  IAPLTR1_dup77   junction  1     no         1
  • Row 1 is the best kind of call: both screens and a splice motif. That it appears in only one sample matters far less than either of those.
  • Row 2 is tier 4 on screen agreement alone, with no motif and one sample. On the measurement above that is the weakest kind of tier 4 — open it in IGV before believing the breakpoint.
  • Row 3 is junction-only with a motif. A real candidate; the assembly screen being silent says nothing, since it cannot see non-colinear structures.
  • Row 4 has nothing going for it. It is in the file because the screen is annotate-only, not because it is a finding.

Filtering to a defensible list

The tier alone is a coarse cut. In practice you want tier 4, plus tier 3 with a motif:

# both screens, or a splice motif
zcat results/chimera_junction/chimera_evidence.tsv.gz \
  | awk -F'\t' 'NR==1 || $7>=3'

# the strictest useful set: both screens AND a motif
zcat results/chimera_junction/chimera_evidence.tsv.gz \
  | awk -F'\t' 'NR==1 || ($6=="both" && $11=="yes")'

That last set is small — on a real 4-sample run there were 129 both pairs in total, and only ~9% of gene-TE junctions carry a motif — and it is the set worth spending validation effort on. The other tens of thousands of rows are a screen, not a result: both screens are annotate-only by design, writing out everything they see so you can apply your own cutoff rather than having one applied silently upstream.

Add gene symbols with results/reference/gene_id_to_name.tsv.gz (see Output Layout) before handing the list to anyone.

Related

  • Interpreting Your Results — how this table relates to the pipeline's other outputs, and which are independent
  • Chimera Detection — what each screen sees, what it is blind to, and every column in the underlying tables
  • TElocal — the locus-level TE quantification behind telocal_active

Clone this wiki locally