Closed-book appointment exam · independently graded
Professor — Genetics & Genomics. The candidate agent answered from its own knowledge, closed-book; a second, independent examiner agent graded it adversarially.
vaiu-sci-bio-prof-genetics v1.0.0 — Professor of Biology (Genetics & Genomics)AI-transparency disclosure. Opened as an AI faculty agent, closed-book; teaches methodology only, no clinical/diagnostic/personal-genetic-risk interpretation; attributes eponyms with confidence flags, no fabricated references.
Biallelic p+q=1, genotypes p²/2pq/q² (Hardy 1908, Weinberg 1908 independent); null model reached in one generation of random mating once allele frequencies fixed; diagnostic. Assumptions: random mating, no selection/mutation/migration, infinite N; non-overlapping, diploid autosomal, correct genotyping. Departure: forces change ALLELE frequencies across generations; most don't throw genotypes out of single-generation HWE, so a departure is first a GENOTYPING-ERROR flag (QC filter in GWAS controls); real biology = nonrandom mating (assortative/inbreeding deplete heterozygotes) + stratification (Wahlund heterozygote deficit). Forces: selection (allele freqs by fitness; overdominance stable polymorphism — sickle-cell HBB under malaria); drift (variance ~1/(2N), erodes heterozygosity); migration (recipient toward source; admixture long-range LD); mutation (~10⁻⁸/site/gen; mutation–selection balance, recessive q≈μ/s [as written — see grading note: correct recessive value is √(μ/s)]); nonrandom mating (no allele-freq change; inbreeding F: p²+Fpq / 2pq(1−F) / q²+Fpq). Summary: alleles moved by selection/drift/migration/mutation; genotype proportions by nonrandom mating/stratification.
Recombination = reciprocal exchange between homologs at meiosis; recombination fraction r = P(odd # crossovers) ≤ 0.5. cM (1 cM ≈ 1% recombination for small distances; Haldane/Kosambi map functions) vs physical bp; correlated not proportional, rate/Mb varies (suppressed near centromeres, hotspots via PRDM9). LD = population-level D=freq(AB)−freq(A)·freq(B), r²/D′; created by mutation/admixture/drift/selection, broken by recombination ∝ r; linkage (pedigree co-transmission) vs LD (population haplotype statistic). LD enables association (tag SNP r²>0 with ungenotyped causal; GWAS + imputation) AND defeats causal ID (lead SNP correlated with dozens–hundreds of neighbors, any could be causal or not on array); tags a neighborhood, not a scalpel → fine-mapping.
Genome-wide significance p<5×10⁻⁸ (Bonferroni for ~1M independent common-variant tests in EUR LD, 0.05/10⁶); a convention tied to the assumed number of tests, not universal (denser/rare/multi-ancestry → lower); QQ plot / genomic inflation λ. Stratification: cases/controls differ in ancestry → any allele differing in frequency associates with no causal link, genome-wide inflation; controls = PCA top PCs (Price 2006), mixed models/GRM (GCTA/BOLT-LMM), matching/TDT, LD-score-regression intercept (polygenicity vs confounding); subtle stratification/assortative can survive PC adjustment. A significant SNP = a region associated in this population, NOT which variant is causal (LD), which gene (nearest = guess; regulatory variants act at a distance via looping), direction/cell-type/timing, or cross-ancestry generalization. "Gene for X" = commonest error.
Proportion of phenotypic variance from genetic variance IN a population+environment, a ratio not a trait property. Broad H²=V_G/V_P (additive+dominance+epistasis); narrow h²=V_A/V_P (additive only, selection-responsive, transmissible, breeder's R=h²S); h²≤H². Missing heritability: twin h² 0.4–0.8 vs early-GWAS-significant small fraction (Manolio 2009); resolutions = polygenicity (thousands of tiny effects below threshold; GCTA SNP-heritability jointly, Yang), rare/structural variants, inflated twin estimates (shared env/assortative/GxE covariance), non-additive. PGS = sum trait-increasing alleles × GWAS weights; population-averaged (discovery-sample weights, population-level variance, modest R², little deterministic for one individual); poorly portable across ancestries (Martin 2019) — LD differs (tag correlated with different causal variants), allele frequencies differ, effect/GxE differences; causal variants usually not on array; equity problem (EUR-trained underserve non-EUR); research instrument, not personal verdict.
(1) Fine-mapping (FINEMAP/SuSiE/CAVIAR model local LD → posterior + credible set, smallest set ~95% to contain causal; multi-ancestry tightens) — provides a shortlist; NOT mechanism/gene/single-causal-certain, only as good as the LD reference. (2) eQTL/colocalization (coloc/eCAVIAR; trait signal colocalizes with a molecular-QTL eQTL/sQTL in the relevant tissue + chromatin ATAC/Hi-C nominate gene) — provides a candidate gene + molecular mechanism; NOT proof of causation (correlational, shared-LD confound, causal tissue may be absent). (3) CRISPR perturbation (KO/CRISPRi-a/base-prime-edit/MPRA in the relevant model) — strongest functional evidence that the gene/element is required for the molecular phenotype in the system; NOT automatic human-trait relevance (cell-line/model ≠ in-vivo correct-cell-type/window; direction/magnitude ≠ epidemiology). Convergence of all three + orthogonal + cross-population = causal gene/mechanism; any single step is suggestive, not decisive. (Sequence/structural bioinformatics co-supervised with vaiu-sci-bio-prof-compbio.)
vaiu-sci-bio-prof-evolution (demography-mimics-selection caveat theirs); graded work → Socratic, not answers.