Indel mutation types in 3 different classification systems with examples

This Zenodo record holds two large tables that map individual insertions and deletions simultaneously into three indel classification systems:

• 476-type (Koh_476), the fine-grained classification of Koh et al.
• 89-type (Koh_89), the reduced classification derived from it
• 83-type (COSMIC_83), the COSMIC ID83 classification

They are called Rosetta stone tables by analogy with the historical Rosetta stone, which carried one decree in three scripts and so let scholars read Egyptian hieroglyphs by comparing them with a script they already knew. Each row here is one real indel written out in all three classification systems at once. A reader who knows only the widely used 83-type scheme can therefore read across to the 476-type and 89-type categories, and vice versa.

These tables are the underlying evidence for Supplementary Table S1 of the accompanying paper. Supplementary Table S1 itself is a compact, human-readable Excel rendering with selected examples per category. These files are the full invididual indel occurrences behind it.

Files

rosetta_stone_full.csv.zip

One row per indel occurrence, 23,216,617 rows and 18 columns, about 2.7 GB uncompressed. The same physical mutation appears once for each sample in which it was called, so the row counts also indicate how common each category is.

Columns:

Insertions and deletions are left-justified, that is, shifted as far to the left as the reference sequence allows, before classification. See justify_indel() in the R package mSigSpectra.

rosetta_stone_full_cap9.csv.zip

The same table restricted to the categories that the 476-type and 89-type schemes actually distinguish, 15,985,113 rows, about 1.9 GB uncompressed. The columns are identical.

This is the file from which the examples in Supplementary Table S1 are drawn.

The 7,231,504 rows that were removed are single-base indels in repeat tracts of 10 or more units. This was not strictly necessary for this table, but both input data sets do not have indel calls for most longer repeat tracts, so we exclude those that were called. A repeat count of exactly 9 is relabelled R(9,) to match an open-interval convention that we adopted to avoid having to add additional indel categories for future data sets. However, future users of the categorization code should be aware of the limitations in the current data sets versus future data sets, which may have more reliable data for indels in longer poly-nucleotide tracts.

Provenance

The tables were built from annotated indel VCFs for two cohorts, HMF (Hartwig Medical Foundation) and PCAWG, specifically the 300 largest annotated indel VCF files from each cohort, 600 files in total. The input VCFs cannot be redistributed, because their data use agreements do not permit release of individual-level variant calls. The tables deposited here are derived summaries of indel sequence context and classification. They contain no sample identifiers, no genomic coordinates, and no other individual-level information.

On paper submission we will release links to the code that produced these tables.

Citation

These files accompany:

Liu et al., 2026, A unified repertoire of indel mutational signatures in 6,975 tumors

Manuscript in preparation, not yet submitted. 

Please cite both the paper and this Zenodo record.