JingjingZhai/maize-in-silico

Maize In-Silico Variant Windows (8 kb)

Reference sequence windows centered on annotated maize SNP positions, built for in-silico variant-effect modelling with DNA language models.

Each row is one SNP position paired with an 8192 bp window of reference genome sequence, with the reference allele at position 4096 (1-based) of the window. Alternate alleles are not encoded in seq — substitute the ALT base at index 4095 (0-based) to generate the mutated sequence.

Fields

field	type	description
chr	string	Chromosome (chr8 throughout this release)
pos	int32	1-based reference coordinate of the SNP
variant_type	string	Variant effect class (see below)
seq	string	8192 bp uppercase reference sequence, [pos-4095, pos+4096] inclusive

Class frequencies

variant_type	n
TE	125,402
intron_variant	99,392
missense_variant	97,922
downstream_gene_variant	97,037
upstream_gene_variant	97,016
3_prime_UTR_variant	95,221
synonymous_variant	94,641
5_prime_UTR_variant	93,725
stop_gained	85,570
intergenic_variant	73,024
splice_acceptor_variant	29,061
splice_donor_variant	28,369
start_lost	12,403
stop_lost	12,173
stop_retained_variant	4,903
total	1,045,859

Construction

• Reference genome: B73 NAM v5 (Zm-B73-REFERENCE-NAM-5.0). Sequence is uppercased; the non-soft-masked assembly was used.
• Variant effects: Ensembl VEP consequence calls, downsampled per consequence class.
• TE class: SNPs originally called intergenic_variant whose single-base position falls inside a transposable-element annotation (Zm-B73-REFERENCE-NAM-5.0.TE.gff3, EDTA) were relabelled TE. Overlap was evaluated on the SNP coordinate only, not the 8 kb window. Counted TE classes are the transposon superfamilies (Gypsy, Copia, unclassified LTR, helitron, hAT, PIF/Harbinger, CACTA, Mutator, Tc1/Mariner, L1 LINE, RTE LINE, LINE element); tandem/satellite features in the same GFF (knob, centromeric repeat, subtelomere, low complexity) were not counted as TE, a choice affecting <0.3% of the class. Of the relabelled SNPs, 117,055 fall in exactly one TE annotation and 8,347 fall in nested annotations.

Validation

• Every row is exactly 8192 bp; no row contains N.
• The base at window position 4096 matches VEP's REF_ALLELE for all 1,314,907 source records (0 mismatches).
• Windows were cross-checked byte-for-byte against samtools faidx.
• No variant lies within 4096 bp of a chromosome end, so no window required padding.

Usage

from datasets import load_dataset

ds = load_dataset("JingjingZhai/maize-in-silico", split="train")
row = ds[0]
assert len(row["seq"]) == 8192

# build the ALT sequence for a given alternate base
alt = "T"
mutated = row["seq"][:4095] + alt + row["seq"][4096:]

Coverage

This release covers chromosome 8 only.