Input files

Input files

To estimate frequencies for one pool, kMate needs two things: the pooled reads, and a panel describing your founders.

pooled reads  ─┐
               ├──▶  kmate run  ──▶  allele frequencies
panel matrices ┘

---

Input reads

• Paired FASTQ from one pool (R1.fq.gz, R2.fq.gz). A single-end file works too.
• Adapter- and quality-trimmed.
• Deduplicated, if your library is PCR-based. PCR duplicates inflate k-mer counts.
• One pool per run. Run kMate once per sample.

Coverage: a few × is enough for usable estimates. Below ~1× the result is dominated by sampling noise. Check coverage and drop failed libraries before estimating, not after.

---

The panel

The panel is built once per founder-haplotype set and reused for every pool. It is four files per chromosome:

file	what it holds
kmer_pa_<CHR>.kmer_pa.npz	which founder haplotypes carry each k-mer
var_pa_<CHR>.var_pa.npz	which founder haplotypes carry the ALT at each variant
var_pa_<CHR>.var_called.npz	which founder haplotypes have a genotype call at each variant
var_pa_<CHR>.meta.npz	the variants themselves: chromosome, position, REF, ALT

kmer_pa also has a small .meta.npz beside it.

You build these yourself from your founders. See Building a panel. It is a one-time job per founder set; every pool you sequence afterwards reuses the same matrices.

The matrices are large (tens of GB for a few hundred founders) and are not stored in the git repository. The 231-founder Arabidopsis thaliana panel used by GrENE-Net is available on request.

---

Input variants (the panel VCF)

Only relevant if you are building a panel. Both matrices are built from one VCF, and it must satisfy all of the following.

• Multi-sample: one column per founder haplotype. The sample order becomes the haplotype order and is shared by both matrices.
• Haploid: exactly one allele per genotype (0, 1, .), not 0/1. A panel column is a haplotype; see Haplotypes and windows. The builders stop with an error on diploid input.
• Biallelic: one ALT per record. The builders stop with an error otherwise.
• Sequence-resolved: real REF/ALT sequences. Symbolic alleles like <DEL> cannot be turned into k-mers.
• Non-overlapping: one variant per locus per haplotype. For graph VCFs this means top-level bubbles only (what vcfbub produces).
• bgzip-compressed and indexed: .vcf.gz plus .tbi or .csi.
• Segregating only, recommended: drop variants carried by nobody or by everybody; they cost memory and add nothing.

Building a panel shows how to get from a raw VCF to one that satisfies these.

---

Input reference

• A FASTA matching your VCF's coordinates and contig names, indexed (.fai).
• One reference for every step. IUPAC ambiguity codes are fine: kMate converts any non-ACGTN base to N internally, the same thing a pangenome graph build does.