Evolution as Optimization

Evolution is the explanation for why every biological system we've examined in this book exists. The cell cycle has checkpoints because cells that didn't have them accumulated mutations and died. Proteins fold because ones that didn't fold were degraded. The immune system recognizes pathogens because organisms without immune systems didn't survive long enough to reproduce.

Every feature of molecular biology is the output of 4 billion years of optimization — not by a designer, but by the relentless differential reproduction of entities with heritable variation. Understanding this process mathematically is the domain of population genetics; understanding its products in sequence space is phylogenetics. Both are foundational to modern bioinformatics.

The Logic of Natural Selection

Three conditions are necessary and sufficient for natural selection to occur:

1. Heritable variation: individuals differ, and those differences are transmitted to offspring
2. Differential fitness: some variants reproduce more than others in a given environment
3. Sufficient population size: variants must persist long enough to spread

Given these conditions, selection will increase the frequency of fitter variants and decrease less fit ones. It's not a force — it's a logical consequence of differential reproduction.

Genetic Drift: Random Walks in Frequency Space

Natural selection is not the only evolutionary force. Genetic drift — random fluctuations in allele frequency due to finite population size — can fix or eliminate alleles regardless of fitness.

The key relationship: the strength of genetic drift relative to selection is determined by effective population size (Ne) and selection coefficient (s):

• When |s| >> 1/Ne: selection dominates (allele fate determined by fitness)
• When |s| << 1/Ne: drift dominates (allele fate is essentially random)

Most amino acid changes in proteins are nearly neutral — they have very small fitness effects that are dominated by drift, especially in populations with small Ne (humans, large mammals). This is the basis of neutral theory (Kimura, 1968): most observed sequence variation is selectively neutral, maintained or eliminated by drift.

Practical implication: when comparing homologous genes between species, synonymous substitution rate (dS) reflects drift (neutral), while nonsynonymous substitution rate (dN) reflects selection + drift. The ratio dN/dS (ω):

• ω < 1: purifying selection (most amino acid changes are deleterious; removed by selection)
• ω ≈ 1: neutral evolution
• ω > 1: positive selection (amino acid changes are beneficial; actively accumulated)

dN/dS analysis is used to identify proteins under positive selection (rapidly evolving under adaptive pressure) and to test whether specific codons in an alignment are under adaptive evolution.

Coalescent Theory: Looking Backward in Time

Instead of asking "where will this allele go?" we can ask "when did all copies of this allele share a common ancestor?" Coalescent theory models the genealogy of gene copies backward in time.

Key result: in a population of Ne diploid individuals, two randomly chosen gene copies share a common ancestor, on average, 2Ne generations ago. The deeper implication:

• Humans have Ne ≈ 10,000–20,000 (bottleneck from ancient population history)
• Most pairs of human gene copies coalesce ~200,000–400,000 years ago
• This defines the time depth of human sequence variation

Coalescent-based methods underlie:

• Divergence time estimation in phylogenetics
• Demographic inference from population genomic data (detecting ancient bottlenecks, expansions, admixture)
• Genealogical reasoning in forensic genetics

Molecular Clocks: Sequences as Time Records

Neutral mutations accumulate at a roughly constant rate per generation per site — the molecular clock. This allows sequences to function as molecular clocks: the more different two sequences are, the longer ago they diverged.

Molecular clock applications:

• Dating phylogenetic divergences: when did humans and chimpanzees last share a common ancestor? (5–7 Ma, calibrated by the molecular clock from genomic divergence)
• Dating viral outbreaks: Bayesian phylogenetic methods (BEAST, TreeTime) use sampling dates and substitution rates to date the origin of viral outbreaks (SARS-CoV-2 origin estimated at November–December 2019)
• Dating cancer mutations: with tumor mutation rates, it's possible to estimate when the first driver mutation occurred (some cancers begin 10–20 years before diagnosis)

The molecular clock is not perfectly constant — it varies with mutation rate (higher in RNA viruses), generation time, and selection pressure. Relaxed clock models account for rate variation across lineages.

Phylogenetics: Reading Evolutionary History from Sequences

Phylogenetics reconstructs the evolutionary relationships between sequences (and the organisms or genes that carry them). The output is a phylogenetic tree — a branching diagram showing relationships and divergence times.

Distance-Based Methods

Compute pairwise sequence distances (percent divergence, Jukes-Cantor corrected), then cluster using algorithms:

• UPGMA: assumes a molecular clock (all lineages evolve at the same rate)
• Neighbor-joining (NJ): does not assume a clock; the standard fast method for large datasets

NJ is used in preliminary analyses and large-scale phylogenetics where parsimony and likelihood methods are too slow.

Maximum Parsimony

Selects the tree that minimizes the total number of evolutionary events (mutations) required to explain the observed sequences. Computationally hard (NP-hard for large trees). Used for closely related sequences.

Maximum Likelihood (ML)

Selects the tree and model parameters that maximize the probability of observing the sequence data under an explicit substitution model. The gold standard for phylogenetic accuracy. Tools: RAxML, IQ-TREE (both implement fast ML heuristics; IQ-TREE is now preferred for most analyses).

Substitution models describe the rates at which nucleotides or amino acids change:

• JC69 (Jukes-Cantor): simplest; all substitutions equally probable
• GTR+G+I: General Time Reversible with Gamma-distributed rates and invariant sites; most flexible and commonly used

ModelTest or IQ-TREE's built-in model selection identifies the best-fit model for a given dataset.

Bayesian Methods

BEAST, MrBayes: incorporate prior distributions on parameters and sample from the posterior using MCMC. Can estimate divergence times, population sizes, and migration rates simultaneously. Computationally intensive but powerful — the standard for dated phylogenies and demographic inference.

Positive Selection in the Human Genome

Not all human evolution is neutral drift. Regions of the genome under recent positive selection show characteristic signatures:

• Selective sweeps: when a beneficial allele rapidly rises to fixation, it carries surrounding variants with it (hitchhiking). The result: a region of reduced diversity and extended haplotype homozygosity around the selected variant. Detected by extended haplotype homozygosity (EHH) statistics and iHS (integrated haplotype score).

• Population-specific sweeps: variants at high frequency in one population but rare in others suggest recent local adaptation. Classic examples: LCT (lactase persistence in dairy-farming populations), HbS (sickle cell allele in malaria-endemic regions), EPAS1 (altitude adaptation in Tibetans).

• Balancing selection: some variants are maintained at intermediate frequencies by selection that favors heterozygotes or alternates over time. HLA genes show extreme balancing selection — diversity is maintained because a diverse MHC repertoire protects against a diverse pathogen landscape.

Population Genomics: Mapping Human History

Modern population genomics uses genome-wide SNP data from thousands of individuals to infer:

• Population structure: principal component analysis (PCA) and ADMIXTURE reveal clusters corresponding to ancestral populations and admixture proportions
• Migration patterns: F-statistics and D-statistics test for gene flow between populations
• Bottlenecks and expansions: effective population size trajectories inferred from the distribution of pairwise coalescence times (PSMC, SMC++ methods)
• Archaic admixture: Neanderthal and Denisovan sequences introgressed into modern humans at levels of 1–4% in non-African populations — detectable from ancient introgressed haplotypes

The 1000 Genomes Project, gnomAD, and the UK Biobank provide population genomic reference panels used routinely for variant frequency lookup, ancestry estimation in clinical genetics, and GWAS interpretation.

Evolutionary Thinking in Bioinformatics Practice

Evolutionary concepts pervade bioinformatics:

• Ortholog vs. paralog: genes related by speciation (orthologs) vs. duplication (paralogs). Ortholog identification is essential for comparative genomics and gene function prediction.
• Synteny: conserved gene order between chromosomes of different species, reflecting ancestral genome organization
• Conservation scores: evolutionarily conserved positions in multiple sequence alignments are functionally important — the basis of GERP, PhastCons, and PhyloP scores used in variant pathogenicity prediction
• Ancestral sequence reconstruction: inferring the sequence of an ancestral protein to study the evolution of function

Phylogenetic conservation is one of the strongest lines of evidence for variant pathogenicity: an amino acid perfectly conserved in 100 vertebrate species is very likely to be functionally important, and a variant disrupting it is likely to be damaging.