M3

M3 (the modified mixture model) is a two-stage genotype-calling algorithm for high-density SNP microarray data, developed to increase call accuracy across the full minor-allele-frequency spectrum—especially for rare variants—on platforms such as Illumina BeadArray/BeadStation and Affymetrix arrays. Genotype calling converts observed probe (allele) intensity signals into discrete genotypes (e.g., AA/Aa/aa), a preprocessing step that directly affects downstream quality control, Hardy–Weinberg equilibrium (HWE) diagnostics, and statistical power in phenotype–genotype association studies.

The method is motivated by complementary failure modes of two major strategy classes. Population-based callers genotype one SNP at a time using cohort-wide clustering models (implicitly assuming three genotype clusters are well-populated and well-separated); performance degrades for low-MAF loci because one or more clusters may be sparsely represented or shifted, increasing missingness and miscalls. SNP-based (per-individual) callers instead infer genotypes for all SNPs within one individual under assumptions of shared probe-response structure across the genome; this can improve rare-variant calls without large sample sizes, but may elevate HWE failures if genome-wide homogeneity assumptions are violated.

M integrates these ideas in a two-stage procedure that couples a population-based component (e.g., GenCall-style clustering) with an individual/SNP-based component (e.g., GenoSNP-like modeling). By blending cohort-level cluster information with within-sample intensity structure, M3 aims to stabilize genotype assignment when clusters are underpowered (rare variants) while reducing systematic distortions that can inflate HWE deviations.

In evaluation on International HapMap Project quality-control samples used in a large case–control setting, M3 improved genotype calling and increased association-study power, with gains typically larger for rare variants than for common variants (model-dependent).