The data, and what we do to it

The single most important distinction for understanding this project's data. Every archaic individual can exist as either (or both), and they are not interchangeable.

BAM
Binary Alignment Map

The actual sequencing reads — every short DNA fragment, placed ("aligned") on the reference genome, carrying per-base quality, mapping quality, and the ancient-DNA damage still visible at read ends. Large (GBs). From a BAM you can re-derive coverage, assess damage, and (re-)call genotypes any way you like. The raw-ish evidence.

VCF
Variant Call Format

The genotype conclusions drawn from a BAM: at chosen positions, "this individual is A/G", plus depth/quality annotations. Small (MBs). You cannot recover the underlying reads or the damage signal from a VCF — the read-level information is gone. A derived summary.

The archaic genomes are published — as BAMs aligned to hg19/GRCh37 — in two public archives: MPI-EVA (Max Planck, ftp.eva.mpg.de) and EBI-ENA (ftp.sra.ebi.ac.uk). Lily's pipeline started from these BAMs and called its own genotypes. Sources marked "confirm" are being recovered in Stage 0 of the rebuild.

Modern comparison: the Simons Genome Diversity Project (SGDP), 15 individuals, downloaded as whole-genome VCFs on hg19 (thesis: a 34.4 M-SNP whole-genome panel; pca/modern/sgdp.wg on the cluster — verified hg19, contains the pigmentation SNPs by rsID).

Coverage — exact, current

Per-sample mean depth over the panel, computed exactly from the committed *_depth.txt (full record + caveats: coverage_current.tsv). Read these as caveated: they're measured at the panel's current (hg38) mis-located coordinates (bug A2), so they are realized depths at the wrong positions, not true per-locus coverage; Denisova 3's 0.5× is artifactual (bug A1 — its true coverage is ~30×); and a max far above the mean (Vindija 87, Scladina) is a pileup artifact that motivates the max-depth cap. True per-locus coverage comes after the rebuild (Q6). 10 of the 14 samples are under ~3.2× — which is why hard-calling is the wrong tool and we move to genotype likelihoods (ANGSD).

The whole pipeline is one chain. Colour marks the level each step operates at — the BAM→VCF boundary (red→blue) is where "photographs" become "written findings" and damage/depth information is fixed in place.

What "contig" means (it comes up a lot): a contig ("contiguous sequence") is one continuous stretch of the reference the reads align to — in a human reference, essentially each chromosome, plus small unplaced/decoy pieces. Every BAM and VCF opens with a header line per contig, @SQ SN:<name> LN:<length>, and the SN value is its label. "Contig naming" is whether that label is bare (1) or prefixed (chr1). Two files can describe the same genome yet be unusable together if one says 1 and the other chr1 — the software matches by name, finds nothing, and silently returns empty. That one-sample mismatch is bug A1 (Denisova 3's BAM was chr-named against a bare-named reference and panel), which is why the "rename contigs" step above harmonizes every sample to bare names up front.

Left: what Lily's original pipeline did (and where the bugs entered). Right: what the rebuild does differently, with a QC gate at each stage. Full rebuild plan: rebuild_from_raw.md (incl. the per-genome reuse-vs-rerun matrix — the 5 high-coverage genomes are reused from published VCFs; only 3 low-coverage genomes are targeted-re-acquired). The justified sample subset (maximize reliable pigmentation SNPs, not genome count) is in papers/SAMPLE_INCLUSION.md. What each published VCF's QC actually was: papers/PIPELINE_QC_BY_PAPER.md.

They differ on nearly every axis that matters — which is why one uniform recipe never fit them.

Format / source

single BAMs (ENA accessions) vs per-chromosome BAMs (Vindija, MPI-EVA); different filenames and conventions.

Coverage

~4 high-coverage (~16–30×: Vindija 33.19, Chagyrskaya 8, Denisova 5, Denisova 25) vs the rest very low (0.1–3×). This drives the missingness story.

Lab treatment

some UDG-treated (damage largely removed), some not — changes how much read-end deamination there is, and how aggressively we must handle it.

Type

Neanderthal, Denisovan, and the Denisova 11 F1 hybrid. Note Denisova 5 is a Neanderthal despite the cave name.

Contig naming

Denisova 3's BAM was chr-prefixed while the others were bare — bug A1.

Data at all

El Sidrón (Sid1253) has none → 14 analyzed, not 15.

Current cluster state

the full sorted BAMs are gone (only wrong-region filtered_*.bam + already-called VCFs remain) — which is why the rebuild re-downloads the raw reads.

A dated record of what we've actually done to the data, so the pipeline's real state is always legible. Newest at the bottom.