Summary
STARlong (--runMode alignReads) aborts with a fatal EXIT_CODE_BUG during splice-junction collapsing:
EXITING because of BUG: different annotation status for the same junction while collapsing junctions:1287415241 2508 0 1
...... FATAL ERROR, exiting
The crash is deterministic and data-triggered — re-running on the same input aborts on the same junction(s).
Version
STAR 2.7.11b (built from source at tag/commit b1edc12 "2.7.11b").
Where
source/OutSJ.cpp, Junction::collapseOneSJ() — the annot branch. The companion motif branch immediately above does not fire.
What's happening
The assertion fires when two junction entries with identical coordinates (same *start, same *gap) are collapsed but carry different annot values (0 = de-novo, 1 = annotated / sjdb). In my runs every crash was annot 0 vs 1, on a small set of distinct junctions, e.g.:
*start |
*gap |
annot1 |
annot2 |
| 1287415241 |
2508 |
0 |
1 |
| 346775474 |
6054 |
0 |
1 |
| 1907313072 |
2917 |
0 |
1 |
i.e. the same junction was detected both de-novo (annot=0) and via the annotated-junction (sjdb) path (annot=1), so by the time the global SJ table is sorted and collapsed the two entries disagree on annotation status, and collapseOneSJ exits instead of merging them.
Reproduction context
STARlong alignment of 10x Genomics 5′ single-cell RNA-seq reads (--soloType CB_UMI_Simple, --soloFeatures Gene GeneFull, --soloStrand Reverse).- GRCh38 (refdata-gex-GRCh38-2020-A) genome with a few custom transgene contigs appended; sjdb generated from the GENCODE GTF at genomeGenerate time.
- Permissive long-read params:
--seedSearchStartLmax 50 --seedPerReadNmax 100000 --seedPerWindowNmax 200 --alignTranscriptsPerReadNmax 100000 --alignTranscriptsPerWindowNmax 10000 --winAnchorMultimapNmax 200 --outFilterMismatchNmax 1000 --outFilterMismatchNoverReadLmax 0.1 --alignIntronMax 1000000, --outSAMtype None.
- Affects a subset of samples; others run through the same genome/params/pipeline fine. (Full command available on request.)
Note
collapseOneSJ's own doc comment says it should "choose max overhangs, motif, annot", and the lenient reconciliation (*(isj1P+annotP) = *(isjP+annotP)) is present but commented out, replaced by the fatal exit. Reconciling annot to the max (annotated wins) would let these runs complete — flagging in case the assert was instead intended to catch a different upstream inconsistency in how annot is assigned along the de-novo vs sjdb detection paths.
Summary
STARlong(--runMode alignReads) aborts with a fatalEXIT_CODE_BUGduring splice-junction collapsing:The crash is deterministic and data-triggered — re-running on the same input aborts on the same junction(s).
Version
STAR 2.7.11b (built from source at tag/commit
b1edc12"2.7.11b").Where
source/OutSJ.cpp,Junction::collapseOneSJ()— theannotbranch. The companionmotifbranch immediately above does not fire.What's happening
The assertion fires when two junction entries with identical coordinates (same
*start, same*gap) are collapsed but carry differentannotvalues (0= de-novo,1= annotated / sjdb). In my runs every crash wasannot 0 vs 1, on a small set of distinct junctions, e.g.:*start*gapi.e. the same junction was detected both de-novo (annot=0) and via the annotated-junction (sjdb) path (annot=1), so by the time the global SJ table is sorted and collapsed the two entries disagree on annotation status, and
collapseOneSJexits instead of merging them.Reproduction context
STARlongalignment of 10x Genomics 5′ single-cell RNA-seq reads (--soloType CB_UMI_Simple,--soloFeatures Gene GeneFull,--soloStrand Reverse).--seedSearchStartLmax 50 --seedPerReadNmax 100000 --seedPerWindowNmax 200 --alignTranscriptsPerReadNmax 100000 --alignTranscriptsPerWindowNmax 10000 --winAnchorMultimapNmax 200 --outFilterMismatchNmax 1000 --outFilterMismatchNoverReadLmax 0.1 --alignIntronMax 1000000,--outSAMtype None.Note
collapseOneSJ's own doc comment says it should "choose max overhangs, motif, annot", and the lenient reconciliation (*(isj1P+annotP) = *(isjP+annotP)) is present but commented out, replaced by the fatal exit. Reconcilingannotto the max (annotated wins) would let these runs complete — flagging in case the assert was instead intended to catch a different upstream inconsistency in howannotis assigned along the de-novo vs sjdb detection paths.