JingjingZhai/full_conservation_within_andropogoneae

Full Conservation Within Andropogoneae

This dataset contains 8,192 bp Sorghum bicolor v3 sequence windows for the Evolutionary_constraint split from kuleshov-group/cross-species-single-nucleotide-annotation.

Columns:

• sequence: 8,192 bp DNA sequence window.
• label: original binary conservation label.

The source coordinates match:

/workdir/jz963/genomes/Sorghum_bicolor_v3.1.1/assembly/Sbicolor_454_v3.0.1.fa

(Phytozome Sorghum bicolor v3.0.1 assembly, v3.1.1 annotation.)

Window definition:

chrom:pos-4095 through chrom:pos+4096, so the original center base is at position 4096 using 1-based sequence indexing.

Splits

The split is held-out chromosome, not random:

split	chromosomes	rows
train	Chr01-Chr09	858,086
validation	Chr10	38,060

There is no overlap between them: 0 shared (chrom, pos), and 0 identical 8,192 bp sequences (verified by exhaustive comparison, not sampling).

Labels are exactly balanced (448,073 per class over both splits). Negatives are sampled close to their matched positives - on Chr01 every positive has a negative within 8,192 bp, median distance 283 bp - so a positive and its matched negative have heavily overlapping windows and differ chiefly at the center base. This pairing is always within a chromosome, so it never crosses the train/validation boundary.

Genomic-class annotation (annotation config)

annotation/train_8192_annotated.tsv and annotation/valid_8192_annotated.tsv are aligned row-for-row with train.tsv and valid.tsv - same order, same number of rows, no join required. They can be concatenated column-wise directly. The same rows and order also apply to the 512 / 1,024 / 4,096 bp variants of this dataset, which share identical coordinates.

Columns:

• chrom, pos: 1-based coordinate of the center base on the v3.0.1 assembly (equal to sequence index 4096, 1-based).
• label: repeated from the main files, for verification.
• finalType: genomic class, one of CDS, utr5, utr3, promoter, intron, intergenic.
• is_TE: overlap with a RepeatMasker feature. Computed independently of finalType, which it does not override.

Classes were assigned from Sbicolor_454_v3.1.1.gene_exons.gff3 with GenomicFeatures, using promoters of 1,000 bp upstream of the transcript start (strand-aware) and this precedence, later overriding earlier:

intergenic < intron < promoter < utr5 < utr3 < CDS

is_TE comes from Sbicolor_454_v3.1.1.repeatmasked_assembly_v3.0.1.gff3. annotation/annotate_snps.R is the exact script used.

Class composition:

finalType	train	validation	% label=1
CDS	593,553	27,969	50.5
intron	145,970	6,471	52.2
intergenic	44,637	1,136	25.2
promoter	31,249	1,065	37.9
utr3	27,056	797	68.7
utr5	15,621	622	70.8

Note that finalType is not matched between labels: UTRs are strongly enriched for conserved sites (utr5 71%, utr3 69% label=1) while intergenic is depleted (25%). Stratified analyses should account for this skew. TE overlap, by contrast, is matched exactly 1:1 between the two labels within each split.