Lecture 25 — DNA, genetics, and gene regulation

Lecture 25 — DNA, genetics, and gene regulation#

⚠️ AI-synthesized; not fully reviewed by course staff. Treat as a study aid; released slides, notebooks, and lecture recordings are authoritative.

Tue Apr 21, 2026 · Part 5 — Molecules & biological AI

What this lecture is about#

Genetic data are the highest-dimensional, most-structured, and most population-dependent data the course will see. This lecture introduces the genome as an ML modality — what’s measured, what the data look like, and why naïve “just throw a neural net at it” fails almost immediately because of population structure and multiple-testing pathologies. The lecture is dense, but the payoff is large: by the end you can read a GWAS paper, you can interpret a Manhattan plot, and you understand why polygenic risk scores transfer poorly across ancestries.

We work top-down from biology to ML.

  • Central dogma and DNA basics. DNA → RNA → protein. ~3 billion base pairs in the human genome, ~99.5-99.9% identical across humans (the higher figure counts only SNVs; the lower includes indels and structural variants), ~4-5 million variants per individual relative to the reference (mostly SNVs plus indels). Variants come in different flavors: SNPs (single-nucleotide), indels, structural variants, copy-number variants. Most ML on genetics works with SNPs.

  • Measurement technologies. SNP arrays (cheap, ~1M variants, biased toward known variants) vs. sequencing (more expensive, full coverage of detected variants — ~4-5M per individual for WGS). Population-scale catalogs that aggregate variants across many people (gnomAD, 1000 Genomes) reach 30-100M+ unique variants total — that’s a catalog size, not a per-individual count. Modern biobanks combine arrays and sequencing.

  • Population structure. Allele frequencies vary by ancestry. The first few principal components of a SNP matrix correspond, almost literally, to geographic origin. This is not a bug — it’s a feature of the data — and ignoring it is one of the most common errors in genetic ML.

  • Linkage disequilibrium (LD). Nearby variants are correlated because they’re inherited together. LD structure varies by ancestry. The “effective number of independent variants” is much smaller than the raw count.

  • GWAS — Genome-Wide Association Studies. The standard analysis: for each of ~10⁶ variants, test for association with phenotype. Effect sizes are tiny, multiple-testing burden is enormous, p-value threshold is conventionally 5 × 10⁻⁸ (Bonferroni for ~10⁶ independent tests).

  • Manhattan plots and QQ plots. The two diagnostic plots every GWAS paper produces. Manhattan = -log₁₀(p) along the genome; QQ = observed vs. expected p-value distribution. A QQ plot that lifts off the diagonal globally suggests population stratification (uncorrected confounding) rather than real signal.

  • Population stratification correction. Include the first 10-20 principal components of the genotype matrix as covariates in the GWAS regression. This is the canonical correction.

  • Polygenic risk scores (PRS). A weighted sum of risk alleles across the genome: PRS_i = Σⱼ βⱼ Xᵢⱼ where βⱼ comes from GWAS effect sizes. PRS predicts disease risk modestly.

  • PRS transferability. PRS computed in one ancestry transfers poorly to others — losing 30-90% of predictive power. Because GWAS samples are >80% European-ancestry, PRS clinical utility is systematically worse for non-European patients. This is a fairness story (link to L24).

  • Gene expression and regulation. DNA → RNA is regulated by transcription factors, enhancers, chromatin state. Expression is context-dependent: tissue, cell type, time, environment.

  • eQTLs. Expression quantitative trait loci — genetic variants associated with expression levels. Provide a mechanistic bridge between GWAS hits and actual biology.

The paradox flagged: “first or middle principal components?” — a teaser for which PCs to keep when correcting for population structure.

Why it matters#

Three reasons this lecture is more than “the biology you should remember from undergrad”:

Population structure is the most important confounder in genetic ML. Two cases that look like signal are actually structure: (a) a variant whose frequency tracks ancestry will appear “associated” with any phenotype that also tracks ancestry; (b) a non-causal variant in LD with a true causal one will appear associated and will not generalize when LD structure differs (across populations).

PRS transferability is a fairness problem. The clinical use of PRS — “this patient is in the top 5% of genetic risk for X” — has different reliability for different ancestries. Deploying PRS clinically without acknowledging this risks widening, not closing, health disparities.

Genotype is not phenotype. Most genetic effects on disease run through expression, regulation, context. A variant that “causes” a disease in some sense usually does so via complex pathways. Models that ignore this layer (just-mapping-genotype-to-disease neural networks) leave a lot on the table.

Things you should walk away believing#

  • Genetic data are high-D, structured, and population-dependent.

  • Population structure looks like signal until you correct for it. PCA-correction is canonical.

  • LD makes “number of variants” misleading. The effective dimensionality is much smaller.

  • GWAS p-values must clear ~5 × 10⁻⁸ for genome-wide significance. Anything between ~5 × 10⁻⁸ and ~10⁻⁵ is “suggestive” at best (smaller p-values are more significant).

  • A QQ plot that doesn’t lie on the diagonal is the first thing a reviewer will check.

  • PRS transferability is poor across ancestries, and this matters clinically.

  • Gene regulation, expression, and eQTLs are the mechanistic layer between genotype and phenotype.

  • Public databases (dbSNP, GTEx, ENCODE, UK Biobank) are the substrate for almost all current genetic ML.

How this connects to the rest of the course#

  • L26 (proteins, molecules) builds on the molecular biology established here.

  • L27 (modern biological AI) covers DNA foundation models (Enformer, Evo) — the deep-learning extensions of GWAS and gene-regulation modeling.

  • L24 (fairness) — PRS transferability is a textbook causal-fairness story.

  • L23 (population health) — UK Biobank and other biobanks are the substrate for GWAS at scale.

  • L2 (linear algebra) — population PCA is literally an eigendecomposition of the centered genotype covariance matrix.

  • L4 (information theory / multiple testing) — GWAS multiple-testing correction is the largest such correction the course will see.

Study guide#

Key terms, self-check questions, and additional resources for active recall.

Key Terms#

Term

Definition

Central dogma

DNA → RNA → protein. The information flow that ML on biological sequences ultimately models.

SNP

Single-Nucleotide Polymorphism — variation at a single base. The variant type most ML on genetics works with.

Indel

Small insertions or deletions in DNA.

Structural variant / CNV

Larger rearrangements; copy-number variants are gains/losses of large chromosomal segments.

SNP array

Microarray genotyping technology. ~1M variants, biased toward known common variants. Cheap.

Whole-genome / Whole-exome sequencing

Read all (WGS) or coding-region (WES) DNA. Far more variants, more expensive.

Population structure

Allele frequencies vary by ancestry. Top PCs of the genotype matrix correspond, almost literally, to geographic origin.

Linkage disequilibrium (LD)

Nearby variants are inherited together; correlation between adjacent SNPs. Effective number of independent variants is much smaller than raw count.

GWAS (Genome-Wide Association Study)

For each of ~10⁶ variants, test for association with phenotype. Effect sizes are tiny; multiple-testing burden is enormous.

Genome-wide significance

Conventional p < 5 × 10⁻⁸ (Bonferroni for ~10⁶ independent tests).

Manhattan plot

-log₁₀(p) plotted along the genome. Spikes mark associated loci.

QQ plot

Observed vs. expected p-value distribution. Diagonal = no signal; lift-off → signal or uncorrected confounding.

Population stratification

Confounding when both allele frequency and phenotype vary by ancestry. The first thing a GWAS reviewer checks.

Polygenic risk score (PRS)

Σⱼ βⱼ Xᵢⱼ — a weighted sum of risk alleles using GWAS effect sizes. Predicts disease risk modestly.

PRS transferability

PRS computed in one ancestry transfers poorly to others — losing 30-90% of predictive power. A fairness story.

Gene expression

Amount of mRNA produced from a gene. Tissue-, cell-type-, and condition-dependent.

eQTL

Expression Quantitative Trait Locus — a genetic variant associated with expression levels. Mechanistic bridge between GWAS hits and biology.

The Most Important Confounder#

Population structure looks like signal until you correct for it

A variant whose frequency tracks ancestry will appear “associated” with any phenotype that also tracks ancestry. Always include the first ~10 PCs of the genotype matrix as covariates in a GWAS regression. A QQ plot that doesn’t lie on the diagonal globally is the canonical sign you forgot.

PRS Transferability Is a Fairness Story#

Why this matters clinically

GWAS samples are >80% European-ancestry. PRS computed from those GWASs predicts disease risk worse for non-European patients — and the gap is large. Deploying PRS clinically without acknowledging this risks widening, not closing, health disparities.

Genotype Is Not Phenotype#

Don’t skip the regulatory layer

Most genetic effects on disease run through expression, regulation, chromatin context, cell type, environment. Models that map genotype directly to disease ignore the mechanism — and leave a lot on the table. eQTLs are the mechanistic bridge. GTEx and ENCODE are the substrate.

Self-Check Questions#

  1. A QQ plot shows almost all p-values lifting off the diagonal globally. What three explanations should you consider, in order?

  2. LD makes “1M variants” misleading. What is the effective number of independent variants in a typical European-ancestry GWAS, roughly?

  3. A PRS for coronary artery disease was developed in UK Biobank (~95% European ancestry). It will be deployed in a Singaporean clinic. Walk through the failure modes you’d expect.

  4. Gene expression varies by tissue and cell type. What does this mean for an “expression-prediction model” that ignores tissue context?

  5. dbSNP, GTEx, ENCODE, UK Biobank: match each to “regulatory-element catalog,” “common-variant catalog,” “tissue-specific eQTL atlas,” “phenotype-rich biobank.”

Additional Resources#