CRISPRme+ reference data & precomputed indexes
A fast, reliable mirror of the reference resources CRISPRme+ needs, plus precomputed indexes β so users can skip the slow raw-data download, the multi-hour variant enrichment, and the index build entirely.
Newer indexes exist. CRISPRme+ 2.10.0 publishes its indexes, rebuilt by CRISPRitz 2.10.0 in index format 2 and named for their guide length and side (
20-NRR_3_hg38+hg38_HGDP1kGP), with each panel's Dictionaries once and the enriched-genome kits they were built from, inlucapinello/crisprme_plus_indices, which 2.10.0'sdownloadreads by default. The archives underindexes/andbuild-inputs/here are CRISPRme+ 2.8.1's (index format 1), kept unchanged for 2.8.1 and earlier, which cannot read index format 2; 2.10.0 still fetches one by its name, and says which index of 2.10.0 replaces it. The genome, annotations, PAMs, sample lists and VCFs here serve every release.
Pulled automatically by crisprme.py download (and the web Settings β Data
Manager). See the CRISPRme+ docs for the full data layout and quickstart:
https://github.com/pinellolab/crisprme-plus
Why this exists
A from-scratch, genome-wide CRISPRme+ run is dominated by data movement + enrichment, not the search itself:
- Downloading the 1000 Genomes VCFs from EBI ran at
1.4 MB/s (3 h); the UCSC genome download is similarly throttled. Hugging Face measured ~6.5Γ faster from the same machine. - The variant enrichment stage was ~93 % of the ~13.5 h wall-clock (single-threaded per chromosome).
Shipping precomputed indexes removes both the download and the enrichment/index-build for the common case, taking a genome-wide setup from many hours to minutes.
Layout
crisprme-data/
βββ indexes/ # PRECOMPUTED search indexes (genome_library tarballs)
β βββ NRR_3_hg38.tar.gz # reference-only (no variants), NRR (2.10.0's default PAM: its index is 20-NRR_3_hg38, in the repo above)
β βββ NRR_3_hg38+hg38_HGDP1kGP.tar.gz # β the recommended variant default's 2.8.1 archive
β βββ NRR_3_hg38+hg38_HPRC.tar.gz # HPRC pangenome panel (HPRC Release 2, graph v2.0)
β βββ NRR_3_hg38+hg38_mega.tar.gz # sites-only aggregate (worst-case fallback)
β βββ NRG_3_hg38{,+hg38_HGDP1kGP,+hg38_HPRC,+hg38_mega}.tar.gz # NRG PAM (the default before 2.10.0)
β βββ NGG_3_hg38{,+hg38_HGDP1kGP}.tar.gz # NGG PAM (cannot serve NAG off-targets)
β βββ NNN_3_hg38{,+hg38_HGDP1kGP}.tar.gz # PAM-LESS 3-nt flank β serves ANY 3-nt 3β² PAM
β βββ genotypes_hg38_{HGDP1kGP,HPRC}.tar.gz # per-sample genotype companions
β βββ SHA256SUMS # checksum per archive (sha256sum -c --ignore-missing, inside indexes/)
βββ build-inputs/ # PRE-ENRICHED kits (fast rebuild / new PAM: download --what build-inputs)
β βββ hg38_HGDP1kGP.tar.gz (+ .SHA256SUMS)
β βββ hg38_mega.tar.gz (+ .SHA256SUMS)
βββ vcfs/ # SOURCE VCFs β build any index from scratch / with a new PAM
β βββ hg38_{HGDP1kGP,HPRC,mega}/ # download --what vcf --dataset <name> (hg38_HPRC: graph v2.0, the input of 2.8.1's archives)
βββ genomes/hg38/ # hg38 FASTA mirror, one file per sequence (download --what genome unpacks it)
β βββ chr1.fa.gz β¦ chrY.fa.gz, chrM.fa.gz # gzipped; the alt, random and unplaced contigs too (455 files)
β βββ chr22.fa # chr22 also unpacked: complete-test and setup fetch it as is
βββ annotations/ # functional + cancer-driver tracks (see below)
βββ pams/ # PAM files, all installed by download --what pams (and --what all)
β βββ 20bp-NRR-SpCas9.txt # SpCas9 NRR: the default PAM since 2.10.0
β βββ 20bp-{NRG,NGG}-SpCas9.txt # SpCas9 NRG (the default before 2.10.0) and NGG
β βββ 20bp-NNN-NO-PAM.txt # 2.8.1's PAM-less file (the NNN_3 indexes above)
β βββ 20bp-NNNNNNNN-NO-PAM.txt # the general index's (PAM-less, 8-nt flank)
β βββ 22bp-NNGRRN-SaCas9.txt, {21,22}bp-NNGRRT-SaCas9.txt # SaCas9 (2.10.0's NNGRRN indexes serve NNGRRT, pruned)
β βββ {20,23}bp-{TTTN,TTTV}-Cas12a.txt # Cas12a (2.10.0's TTTN indexes serve TTTV, pruned)
βββ samplesIDs/ # per-dataset sample-id lists
βββ guides/ # sg1617.txt: the example guide CTAACAGTTGCTTTTATCACNNN (download does not fetch it)
βββ README.md
Precomputed indexes
NRR is the recommended PAM since CRISPRme+ 2.10.0 (the SpCas9 NGG, NAG, NGA and
NAA superset β the conservative choice for off-target screening; it was NRG, NGG +
NAG, before); NRR and NRG are published on all three variant panels, NGG and a PAM-less
index on HGDP + 1000 Genomes only. CRISPRme+ 2.10.0 and later search an NRR index for NGG,
NAG, NRG or NGA, and an NRG one for NGG or NAG, pruning it to that PAM by itself. All are scored with
CFD + CRISPR-Bulge. Every index here stores a 3-nt PAM region, which is
what any CRISPRme release reads for a 3-nt PAM, so any CRISPRme release can read
them; use the rebuilt NNN_3 archives with CRISPRme+ 2.8.1 or later (see below).
Variant panels (the axis that matters most)
| Index | Variants | Phasing | Notes |
|---|---|---|---|
NRR_3_hg38 |
none | β | reference-only |
NRR_3_hg38+hg38_HGDP1kGP |
HGDP + 1000 Genomes (4,091 samples = 925 HGDP + 3,166 1000G) | genotyped-phased (SHAPEIT5 re-phased) | the recommended variant default's 2.8.1 archive (2.10.0's is 20-NRR_3_hg38+hg38_HGDP1kGP, in lucapinello/crisprme_plus_indices) β CONFIRMED cis co-occurrence + named carriers |
NRR_3_hg38+hg38_HPRC |
HPRC pangenome (232 samples incl. CHM13; HPRC Release 2's Minigraph-Cactus graph v2.0, where 2.10.0's 20-NRR_3_hg38+hg38_HPRC is of v2.1) |
genotyped-phased | pangenome panel |
NRR_3_hg38+hg38_mega |
aggregate (1000G + HGDP + gnomAD + TOPMed + All of Us) | sites-only | worst-case fallback; a listed haplotype may not exist in any individual |
The two genotyped panels ship a genotypes_hg38_* companion that enables
--per-sample resolution (CONFIRMED cis + exact joint AF + named carriers).
The mega for CRISPRme+ 2.10.0 is rebuilt from the same merged VCFs (vcfs/hg38_mega) by
seq_script/merge_panels/mega_rebuild.sh, with CRISPRitz 2.10.0 (index format 2, every allele
of a split site coded): its registry and its indel frequencies hold exactly the records its
genome codes, those whose FILTER is PASS or . (the archives above, of 2.8.1, list every
record of the VCF, 2,442,989 SNVs outside PASS among them), one allele once (of an allele on
two records, the more frequent record's: 8 long deletions genome-wide), and its indel log carries
each indel's AF_max, so an indel row shows its frequency. Searched with CRISPRme+ 2.10.0, a
sites-only dataset's candidates in windows of 2^20 sequences or more (the immunoglobulin loci
above all) are withheld and listed in the dense-region files, and noted on the rows there
(--no-sites-only-density-guard analyses them all), and its
rows of two or more variants whose frequencies fail the bound (product under 1e-06, or the
rarest under 0.001) are ranked after the others; with the guard a search of it took 1.35 to 1.46
(chr22) and 1.53 to 1.73 (chr14) times the shipped mega's, at 4 mismatches and 1 + 1 bulges
(during 2.10.0's development), and genome-wide 1.21 times with the same code and 1.18 times 2.8.1's
own (the release's timing gate, the README guide population-level with NRR, 32 threads). The
rebuilt archives are published with the release, in the location of the 2.10.0 indexes; the ones
above stay for earlier releases.
Other PAMs
| PAM | Indexes | When to use |
|---|---|---|
NRG |
NRG_3_hg38, +hg38_HGDP1kGP, +hg38_HPRC, +hg38_mega |
the default before CRISPRme+ 2.10.0 (NAG/NGG); a narrower, faster screen than NRR |
NGG |
NGG_3_hg38, +hg38_HGDP1kGP |
canonical SpCas9 PAM only. NGG cannot serve NAG or NGA off-targets, so NRR is preferred for a safety screen |
NNN (PAM-less) |
NNN_3_hg38, +hg38_HGDP1kGP |
one index for ANY 3-nt 3β² PAM, selected at search time with --prune-pam. Larger (71 GB / 110 GB downloads, GB = 10βΉ bytes) but removes the per-PAM rebuild |
NRR_3_hg38+hg38_mega was replaced for CRISPRme+ 2.8.1. The archive uploaded on
2026-10-27 shipped an empty indel index and empty indel logs, so no indel off-target
was ever reported from it, and installing it next to NRG_3_hg38+hg38_mega emptied
that index's indel logs as well (an install replaces each Dictionaries/ folder).
If you downloaded it before the replacement, remove
genome_library/NRR_3_hg38+hg38_mega and genome_library/NRR_3_hg38+hg38_mega_INDELS
and download it again; that also restores NRG mega's logs. Both mega archives now
carry identical Dictionaries/.
NRR_3_hg38+hg38_HPRC was replaced too, for a smaller reason: its registry lacked
the stamp that makes a search on a phased panel default to per-sample analysis, so it
ran population-level unless --per-sample was passed (and, installed next to
NRG_3_hg38+hg38_HPRC, did the same to that one). Its index is unchanged; re-download
it the same way to get the per-sample default back.
NNN_3_hg38 and NNN_3_hg38+hg38_HGDP1kGP were rebuilt for CRISPRme+ 2.8.1, with
the genomic-N flank fix in CRISPRitz 2.9.1, and replaced here on 2026-10-30. The
archives uploaded on 2026-10-25 lack every target whose 3-nt flank holds an assembly
N. CRISPRme+ 2.8.1 never reports a target with an N inside its PAM, so a search
with a 3-nt PAM (NGG, NRG, NRR, or the unconstrained 20bp-NNN-NO-PAM) gives
the same report on the old and the rebuilt archives (for NRG verified genome-wide,
per-sample, byte-for-byte against NRG_3_hg38+hg38_HGDP1kGP). A shorter 3β² PAM
searched with --prune-pam, such as 20bp-NG-SpCas9NG, does change: the old archives
miss each site whose base just past the PAM is an N (at an assembly-gap edge, and in
the indel index, whose windows are separated by Ns), and the rebuilt ones report it,
as a dedicated index for that PAM does. If you downloaded either archive before the
replacement and search it that way, remove genome_library/NNN_3_hg38,
genome_library/NNN_3_hg38+hg38_HGDP1kGP and
genome_library/NNN_3_hg38+hg38_HGDP1kGP_INDELS, then download again.
Use the rebuilt archives with CRISPRme+ 2.8.1 or later. 2.8.0 keeps the targets with
an N in their PAM that the rebuild adds, and a search that reaches one fails in
post-analysis or, on the CRISPRitz 2.9.0 that the 2.8.0 image builds and with RNA
bulges only, reports it as a 0-mismatch, 0-bulge off-target whose sequence stops at the
N. If you searched them with 2.8.0, run that search again with 2.8.1.
Wider PAM-less flanks (flank > 3, e.g. NNNNNNNN_3) are not readable by a
CRISPRme release that predates --prune-pam, so they are published separately in
lucapinello/crisprme-data-general-index
and never land in this repo.
Annotations: dhs+encode_screenv4+gencode+cosmic.hg38.bed.gz (functional +
cancer, with COSMIC gated behind a licence attestation),
intogen_drivers.hg38.bed.gz (IntOGen cancer drivers, CC0, default-on), and
gencode.protein_coding.bed.gz (closest-gene).
Usage
The indexes are fetched automatically:
# one-shot: genome + annotations + PAMs + sample lists + the default reference index
# (CRISPRme+ 2.8.1: NRG_3_hg38, from here; 2.10.0: 20-NRR_3_hg38, from
# lucapinello/crisprme_plus_indices)
crisprme.py download --what all --path .
# then the recommended variant default's 2.8.1 archive (or any other index by name; under
# 2.10.0, 20-NRR_3_hg38+hg38_HGDP1kGP fetches its index of format 2)
crisprme.py download --what index --index-name NRR_3_hg38+hg38_HGDP1kGP --path .
Direct download of single files also works (crisprme.py download additionally unpacks and
installs them). Name the files: the whole repo is about 432 GB.
hf download lucapinello/crisprme-data indexes/NRR_3_hg38+hg38_HGDP1kGP.tar.gz \
indexes/SHA256SUMS --repo-type dataset --local-dir crisprme-data
License & citation
Reference data are redistributed under the terms of their original sources (1000 Genomes, HGDP, HPRC, gnomAD, GENCODE/ENCODE, IntOGen, UCSC hg38, etc.). CRISPRme+ is licensed under the MGB Open Access License 1.0 (MGBOA 1.0) β see https://github.com/pinellolab/crisprme-plus/blob/main/LICENSE.
If you use CRISPRme, please cite: Cancellieri S, et al. Human genetic diversity alters off-target outcomes of therapeutic gene editing. Nat Genet 55, 34β43 (2023). doi:10.1038/s41588-022-01257-y
- Downloads last month
- 1,121