How to handle LD?

[davmlaw]

Presumably, markers in LD are only 1 informative marker in terms of genotype? But - they can presumably be used individually (and maybe in combination?) to work out errors

[davmlaw]

DR says there is a paper that mentions how you can treat it as 1 but reduce sequencing errors

Chimerism Assay Using Single Nucleotide Polymorphisms Adjacent and in Linkage-Disequilibrium Enables Sensitive Disease Relapse Monitoring after Hematopoietic Stem-Cell Transplantation
doi: 10.1093/clinchem/hvab010

A technical application of quantitative next generation sequencing for chimerism evaluation
DOI: 10.3892/mmr.2016.5593

[davmlaw]

Decision for the tool paper (2026-06-15): scoped in Discussion, not blocking

Added an LD subsection to paper/discussion.md ("Marker independence and linkage disequilibrium") plus a reference (Kim2021ld, Clin Chem 2021, 10.1093/clinchem/hvab010) in references.bib. Summary of the reasoning so we don't re-derive it:

Two distinct LD angles (both were buried in this issue):

1. LD as a nuisance. The MLE likelihood is a product over markers => assumes independence. Correlated markers carry less information than the same count of independent ones, so the point estimate stays unbiased but the profile-likelihood CIs are too narrow.
2. LD as a feature (the David Ross paper = Kim et al. 2021). Phased SNPs in tight LD on the same read used as a concordance filter (linked alleles must agree before a read counts) suppress sequencing error and push the detectable host fraction lower.

Why it does not block the tool paper:

• Panel design defuses (1) for the intended use. Sample-ID / ancestry SNP panels are deliberately genome-spread, in approximate linkage equilibrium, one independent identity bit each. Both panels we use (76-SNP rhAmpSeq SID, 1062-SNP MIP) are of that design. LD only bites on clustered panels (multiple SNPs/amplicon, dense tiling), which is a lab's panel choice.
• Not a silent failure. Residual correlation inflates per-marker variance like overdispersion does, so it is partly absorbed by the shared overdispersion (rho) term and surfaces in the goodness-of-fit REVIEW flag (same mechanism that flagged the SRP434573 mixtures). A panel dense enough for LD to matter tends to be caught as an overdispersed fit, not reported as a confidently-wrong interval.

Future work (background, after the paper lands):

• Haplotype-aware error suppression (Kim et al. 2021 approach). Needs multiple phased markers per amplicon, read-backed phasing, haplotype counting. It is a sensitivity upgrade for the residual-host presence test, not a correctness fix. Pairs naturally with the per-site empirical error model (issue context) and the relapse-detection direction.
• Optional cheap insurance if a reviewer pushes: measure pairwise spacing / r^2 across the 76-SNP and 1062-SNP panels and report "X of Y pairs with r^2 > 0.1" (expected ~0) as a one-line supp check, turning the independence assumption into a measurement. Not done yet; only worth it on reviewer request.

Second ref from this issue (mmr 2016.5593, 'technical application of quantitative NGS for chimerism') not cited; Kim 2021 is the relevant LD-method one.

[davmlaw]

Added a bit of tips to paper:

A laboratory whose panel does carry clustered markers can restrict allomix to an LD-pruned subset with a simple BED file or marker list, keeping one marker per linked cluster, so the independence assumption holds without changing the panel.