Genetic Drift

Summary

Genetic drift is the random fluctuation of allele frequencies due to stochastic sampling in finite populations — each generation, only a subset of individuals reproduce, and chance determines which alleles survive. In cancer, drift is the dominant force determining the fate of individual mutations: the probability that a new driver-mutation escapes stochastic extinction is only ~0.8% (P(survival) ≈ 2s, for s ≈ 0.4%), meaning 99.2% of potentially advantageous mutations are erased by drift within a few cell generations (Bozic et al., 2010). Drift is the mechanism underlying neutral-evolution — it operates continuously, on every mutation, in every cell division — while positive-selection and negative-selection operate episodically on the small fraction of mutations with non-zero fitness effects. The death-birth ratio δ = d/b amplifies drift: as δ → 1, stochastic extinction events become more frequent, and neutral mutations can reach fixation (clonal status) through drift alone (Bozic et al., 2016).

Definition

In population genetics, genetic drift is the change in allele frequency due to random sampling of gametes at each generation. In an ideal Wright-Fisher population of size N, the variance in allele frequency change per generation is p(1 − p)/N — smaller populations experience larger frequency fluctuations. Drift is directionless: an allele is equally likely to increase or decrease in frequency in any given generation, and in the long run, every allele either fixes (frequency = 1) or goes extinct (frequency = 0).

In cancer, the “population” is the community of tumor cells, the “generation” is a cell division cycle, and the “allele” is a somatic mutation. Because tumor cell populations are finite — ranging from ~10³ cells (microscopic lesion) to ~10¹¹ cells (clinically detected tumor) — drift operates at every stage of tumor evolution. The smaller the population, the stronger the drift.

Mechanism

Stochastic Extinction of New Mutations

When a new mutation arises in a single cell, its immediate fate is determined almost entirely by drift, regardless of its fitness effect. The mutant cell must divide to produce at least one surviving daughter cell; if it dies before dividing, or if both daughters die before dividing further, the mutation is lost. This is a branching process problem: from a single starting cell, what is the probability that the lineage survives indefinitely?

For a mutation with selective advantage s, the survival probability in a birth-death process is (Bozic et al., 2010):

$P(\text{survival})\approx \frac{2s}{1+s}\approx 2s(\text{for small }s)$

For the average driver (s ≈ 0.004), P(survival) ≈ 0.008 — less than 1%. For a truly neutral mutation (s = 0), P(survival) = 0 in a population with any cell death (δ > 0): every neutral lineage eventually goes extinct, though some may persist for many generations and reach substantial size before doing so. In a pure-birth process (δ = 0), neutral lineages never go extinct — the population only grows — but their relative frequency declines as the population expands (1/N of cells carry the mutation).

This stochastic bottleneck at the single-cell stage is the primary reason tumor progression is slow: the vast majority of mutations — including most driver mutations — are lost to drift before they can establish. The progression rate is determined not by how often drivers arise but by how rarely they survive (Bozic et al., 2010).

Drift in Expanding Populations

The tumor population is not constant-sized — it expands, often approximately exponentially in early stages. In an expanding population, the effective strength of drift changes with population size:

• Small N (early tumor, N ~ 10³–10⁵): Drift is strong. A mutation arising in a single cell represents a non-trivial fraction of the population. Frequency fluctuations are large relative to the population mean. This is when drift most strongly shapes the mutation landscape — mutations that survive this stage become pervasive (high VAF) simply because they arose early, not because they were selected.

• Large N (late tumor, N ~ 10⁸–10¹¹): Drift is weaker per-mutation but operates on vastly more mutations. The number of new mutations per cell division is proportional to N, so many more mutations are “trying their luck” against drift. Most still go extinct, but the absolute number of surviving mutations is higher. This produces the 1/f² VAF distribution: many more low-frequency mutations (recent, large-N background) than high-frequency mutations (ancient, small-N background).

The Death-Birth Ratio Amplifies Drift

When cell death is common (δ = d/b close to 1), drift effects are dramatically amplified. The expected number of subclonal passenger mutations above frequency α is:

$m_s=\frac{u(1-\alpha )}{(1-\delta )\alpha }$

and the expected number of clonal passenger mutations (fixed by drift alone) is:

$m_c=\frac{\delta u}{1-\delta }$

For δ = 0.99 and u ≈ 0.015 (exome-wide passenger rate per cell division), there are ~150 subclonal passengers at >1% frequency and ~1.5 clonal passengers — mutations at 100% frequency that fixed through drift during clonal expansion, without ever conferring a fitness advantage (Bozic et al., 2016). For δ = 0.999 (premalignant lesions), these numbers rise tenfold.

This has a critical implication: clonal does not equal selected. A mutation present in all tumor cells may be a drift-fixed passenger, not a truncal driver. Equating clonality with functional significance overestimates the number of driver events and misdates the timing of branching events. See neutral-evolution for the full frequency spectrum derivation and passenger-mutation for the clonal-vs-truncal distinction.

flowchart TD
    M["New somatic mutation<br>arises in 1 cell"] --> D{"Stochastic drift:<br>does the lineage survive?"}

    D -->|"Survives (rare)"| Est["Lineage established<br>Frequency grows (or shrinks)<br>by random walk + selection"]
    D -->|"Goes extinct (common)"| Lost["Lineage lost<br>99.2% of new drivers<br>100% of new passengers<br>(when δ > 0)"]

    Est --> Fate{"Long-term fate?"}

    Fate -->|"s > 0 + survives"| FixSel["Fixation by selection<br>[[clonal-sweep]]"]
    Fate -->|"s ≈ 0, δ close to 1"| FixDrift["Fixation by drift<br>Clonal passenger<br>(not truncal)"]
    Fate -->|"s ≈ 0, δ ≪ 1"| LowFreq["Remains subclonal<br>Part of 1/f² tail"]
    Fate -->|"s < 0"| Elim["Eliminated by<br>[[negative-selection]]"]

    Lost --> Invis["Invisible to sequencing<br>No trace in VAF distribution"]

    FixDrift -->|"Recorded as"| ClonalPass["Clonal passenger mutation<br>m_c = δu/(1 − δ)<br>~1.5 at δ = 0.99"]
    LowFreq -->|"Recorded as"| Subclonal["Subclonal passenger<br>m_s = u(1−α)/((1−δ)α)<br>~150 at δ = 0.99, α = 0.01"]

Figure: The central role of genetic drift in determining mutation fate. Every new mutation — whether driver or passenger — faces a stochastic bottleneck at the single-cell stage. Most lineages go extinct. Survivors face a long-term trajectory shaped by the interplay of drift strength (population size), death-birth ratio (δ), and fitness effect (s). Clonal status is not evidence of selection: at δ close to 1, drift alone can fix neutral mutations. Synthesized from Bozic et al. (2010, 2016) and Greaves & Maley (2012).

Drift vs. Selection

The three evolutionary forces — drift, positive-selection, and negative-selection — are distinguished by their relationship to fitness:

Force	Fitness effect	Direction	Timescale	Signature in VAF data
Genetic drift	s ≈ 0 (neutral)	Random	Continuous	1/f² distribution (δ = 0); elevated by factor 1/(1 − δ) when δ > 0
Positive selection	s > 0 (advantageous)	Directional (increase)	Episodic	Excess high-frequency mutations above δ-calibrated neutral null
Negative selection	s < 0 (deleterious)	Directional (decrease)	Episodic	Deficit of mutations in constrained regions vs. neutral expectation

In practice, drift and weak positive selection are on a continuum. A mutation with s = 0.1% is almost indistinguishable from a neutral mutation in all but the largest cohorts. The boundary is statistical, not ontological: a “driver” is a mutation for which the evidence of positive selection exceeds a significance threshold, not a mutation that is fundamentally different in kind from a passenger.

Clinical Significance

Explains Heterogeneity in Progression Rates

Even among patients with identical tumor types and driver mutations, progression rates vary enormously. Bozic et al. (2010) demonstrated this through simulation: with identical parameters (u = 10⁻⁵, s = 0.01, T = 4 days), one simulated patient had only acquired a second driver after 20 years with <10⁵ cells, while another had three drivers and >10¹¹ cells by 25 years. The difference was purely stochastic — which driver lineages survived drift and which didn’t. This means a patient’s progression rate is not fully predictable from their mutational profile; it depends on the unobservable history of drift events.

Intermediate Clones Are Invisible

Each clonal passenger fixation event (m_c ≈ 1.5 at δ = 0.99) records a bottleneck where competing intermediate-clones went extinct. The surviving lineage’s mutations become clonal; the losers’ mutations are erased. The intermediate clones that bridged successive sweeps — carrying the adaptive mutation but not yet fixed — are invisible to standard sequencing because their lineages were outcompeted. This is why intermediate-clones are a theoretical necessity but an empirical ghost: drift eliminates them, and only the winner’s genome survives to be sampled.

Population Bottlenecks Reset Drift

Therapies, metastases, and physical barriers impose population-bottlenecks that amplify drift. A bottleneck reduces N — sometimes to a handful of surviving cells — which increases the variance of allele frequency change (∝ 1/N). Mutations that were rare in the pre-bottleneck population can become common in the post-bottleneck population purely by chance (founder effect). A resistance mutation present in 0.1% of cells pre-therapy can become the dominant clone post-therapy not because it was selected during treatment but because it happened to be among the few cells that survived the initial kill (see therapy-resistance and population-bottleneck).

Limitations

1. Drift is invisible. We observe the outcome — which mutations are present at which frequencies — but not the process. The extinction of 99.2% of driver lineages leaves no detectable trace. We infer drift’s operation from its statistical consequences (the VAF distribution, the heterogeneity of progression rates), not from direct observation.

2. δ must be estimated. The strength of drift — and therefore the expected neutral frequency spectrum — depends on δ, which varies across tumor types and evolutionary stages. Applying a δ = 0 model (1/f²) when the true δ ≈ 0.99 overestimates the evidence for selection and underestimates the role of drift (Bozic et al., 2016).

3. Drift-selection boundary is statistical. For weak selective advantages (s < 0.5%), distinguishing drift from selection requires cohort sizes that exceed most published studies. Many mutations currently classified as “drivers” on recurrence grounds may in fact be passengers that happen to hit the same gene multiple times by chance in large cohorts.

4. Population structure complicates drift. Real tumors are not well-mixed populations. Spatial structure (territorial segregation of clones), microenvironmental heterogeneity, and differential blood supply create subpopulations with different effective sizes, making drift strength spatially variable — a mutation may drift to fixation in one region while going extinct in another (Turajlic et al., 2019).