Positive Selection

Summary

Positive selection is the evolutionary process by which mutations that confer a fitness advantage increase in frequency within a population. In cancer, positive selection acts on driver-mutations — somatic alterations that increase proliferation, survival, migration, or immune evasion — causing the mutant lineage to expand relative to competitors. The average selective advantage of a cancer driver is small (s ≈ 0.4%; Bozic et al., 2010), which means selection operates slowly and subtly: most new driver mutations are lost to genetic-drift before they can establish, the observable signature is a shift in the variant frequency spectrum rather than an abrupt sweep, and distinguishing weak selection from neutral-evolution requires careful calibration of the neutral null (Graham & Sottoriva, 2017; Bozic et al., 2016). Positive selection is the directional force in clonal-evolution — it is what makes cancer progression adaptive rather than merely random.

Definition

In evolutionary biology, positive selection (also called directional selection or Darwinian selection) is the process by which an allele that increases fitness rises in frequency. The allele confers a reproductive or survival advantage, so carriers leave more descendants than non-carriers, and the allele’s frequency increases over generations. This is distinct from negative-selection (purifying selection, which removes deleterious variants) and from genetic-drift (random frequency changes with no fitness effect).

In cancer, the “allele” is a somatic mutation, the “population” is the community of cells within a tissue or tumor, and “fitness” is the net replication rate — the balance of cell division and cell death. A mutation that tips this balance in favor of the mutant lineage is positively selected. Nowell (1976) described this without using the term “positive selection”:

“Within such a mutant subpopulation, an additional mutant may arise with an additional selective advantage with respect to the original tumor cells as well as normal cells, and this mutant becomes the precursor of a new predominant subpopulation.” (p. 24)

The explicit positive/negative selection terminology was imported from population genetics into cancer biology with the driver/passenger framework (Greaves & Maley, 2012).

Mechanism

The Fitness Advantage

A positively selected mutation must increase the net growth rate of the mutant lineage. This can occur through:

Mechanism	Effect on fitness	Examples
Increased proliferation	More cell divisions per unit time	KRAS G12D, EGFR L858R
Reduced death	Fewer cells lost to apoptosis	BCL2 overexpression, TP53 loss
Immune evasion	Escape from T-cell/NK-cell killing	HLA LOH, B2M loss
Metabolic reprogramming	Growth in resource-limited environments	IDH1/2 mutations
Replicative immortality	Bypass of telomere shortening	TERT promoter mutations
Angiogenesis	Improved oxygen and nutrient supply	VEGF amplification
Invasion and metastasis	Colonization of new tissue territories	EMT program activation

All of these increase the expected number of surviving daughter cells per parent cell — which is the operational definition of fitness in the Bozic-Nowak branching process framework (see branching-process-model).

The Selective Advantage Is Small

Bozic et al. (2010) estimated the average selective advantage of a driver mutation at s ≈ 0.4% (0.004 ± 0.0004). This was derived by fitting a discrete-time branching process model to glioblastoma and pancreatic cancer sequencing data, and validated against independent clinical data from familial adenomatous polyposis (FAP). The estimate is robust to the driver mutation rate: varying u from 10⁻⁶ to 10⁻⁴ shifts s only between 0.32% and 0.65%.

A 0.4% advantage means a driver-bearing cell produces ~1.004 surviving descendants for every 1.000 produced by a non-driver-bearing competitor. This is extraordinarily weak by organismal standards — a cheetah with a 0.4% speed advantage would catch one extra gazelle every 250 hunts. But in a population of millions of dividing cells, compounded over hundreds of generations, a 0.4% advantage produces detectable clonal expansion.

Three implications of small s:

1. Most new drivers are lost to drift. The probability that a new driver-bearing cell escapes stochastic extinction to establish a lasting lineage is P(survival) ≈ 2s (for small s). For s = 0.004, only ~0.8% of new driver mutations survive — 99.2% are erased by genetic-drift within a few generations (Bozic et al., 2010).

2. Selection is hard to detect over short intervals. A 0.4% advantage takes many cell generations to produce a measurable frequency shift. In vitro validation experiments lasting days to weeks — spanning only tens of cell divisions — cannot reliably detect such weak effects (Turajlic et al., 2019).

3. Progression takes decades. The cumulative time to acquire k drivers grows logarithmically with k (Σ 1/i ≈ log k), but the waiting time for the first successful driver is already ~8.3 years for typical parameters. Tumor progression is slow not because drivers are rare but because most are lost before they can establish (Bozic et al., 2010).

Selection Inference from Sequencing Data

Positive selection leaves a characteristic signature in the variant-allele-fraction distribution: an excess of mutations at high frequency compared to the neutral expectation. The standard method is:

1. Establish the neutral null. Under neutral-evolution in a growing population with no cell death (δ = d/b = 0), the number of mutations as a function of frequency f follows a 1/f² distribution (Graham & Sottoriva, 2017).

2. Calibrate for δ. When the death-birth ratio δ > 0, the neutral spectrum shifts upward by a factor proportional to 1/(1 − δ). δ must be estimated independently — from the low-frequency tail of the VAF distribution or from orthogonal growth-rate data — before selection can be inferred (Bozic et al., 2016).

3. Test for deviation. Mutations exceeding the δ-calibrated neutral expectation are candidates for positive selection. The test is conservative: only strong or sustained selection produces statistically significant deviation.

This two-step procedure (estimate δ, then test) resolves the apparent tension between the 1/f test (Graham & Sottoriva, 2017) and the δ-generalized spectrum (Bozic et al., 2016): they are nested models, not competing ones. See neutral-evolution §The Frequency Spectrum for the full mathematical relationship.

flowchart TD
    M["Somatic mutation arises<br>in a single cell"] --> Q{"Fitness effect?"}

    Q -->|"s > 0<br>(advantageous)"| PS["POSITIVE SELECTION"]
    Q -->|"s ≈ 0<br>(neutral)"| ND["NEUTRAL DRIFT"]
    Q -->|"s < 0<br>(deleterious)"| NS["NEGATIVE SELECTION"]

    PS --> D{"Survive stochastic drift?<br>P(survival) ≈ 2s"}

    D -->|"YES (~0.8% for s=0.4%)"| Est["Lineage establishes"]
    D -->|"NO (~99.2%)"| Ext["Extinction within<br>a few generations"]

    Est --> Cond{"τ_k > sweep_time?"}

    Cond -->|"YES<br>(small N, early tumor)"| Sweep["[[clonal-sweep]]<br>Clean fixation,<br>diversity collapse"]
    Cond -->|"NO<br>(large N, late tumor)"| Interfere["[[clonal-interference]]<br>Competing clones,<br>no single sweep"]

    ND --> Freq["Frequency determined by<br>timing of occurrence + drift"]
    Freq --> VAF["Observed as 1/f² tail<br>in VAF distribution"]

    NS --> Elim["Clone eliminated<br>by immune surveillance<br>or intrinsic disadvantage"]

    Sweep --> Result["Reduced diversity<br>Next round of mutation accumulation"]
    Interfere --> Result

Figure: The fate of somatic mutations as a function of fitness effect. Positive selection (s > 0) can lead to clonal expansion, but only if the mutant lineage survives stochastic drift (~0.8% chance for s = 0.4%). Even after establishment, the outcome depends on the τ_k vs. sweep_time competition: in small populations (early tumors), clean sweeps occur; in large populations (established tumors), clonal interference prevents any single clone from dominating. Neutral mutations drift passively, producing the 1/f² VAF tail. Deleterious mutations are eliminated. Synthesized from Bozic et al. (2010), Greaves & Maley (2012), and Turajlic et al. (2019).

Relationship to Driver Mutations

Positive selection is the process; a driver-mutation is the entity the process acts on. Not all mutations in cancer genes are under positive selection — some occur as passengers in genes that could be drivers but lack the functional consequence or clonal evidence — and not all positively selected mutations are in known cancer genes (PCAWG, 2020). The operational definition of a driver is precisely “a mutation showing evidence of positive selection” — through recurrence (more frequent than the background mutation rate predicts), clustering (hotspot mutations in functional domains), or VAF enrichment (excess high-frequency mutations in the clone).

Gerstung et al. (2020) showed that clonal drivers (those under positive selection early in tumor evolution) are drawn from a restricted repertoire — TP53, PIK3CA, KRAS, PTEN, APC — while subclonal drivers (selected later) involve a nearly fourfold broader set of genes. This suggests that the “low-hanging fruit” of cancer selection is limited: only a handful of genes can initiate tumorigenesis through positive selection, but once the tumor is established, many more genes can provide incremental selective advantage.

Positive Selection vs. Neutral Evolution

Positive selection and neutral-evolution are not mutually exclusive modes — they are interleaved. Neutral dynamics dominate between selection events: mutations accumulate, drift shifts frequencies, and the population explores the fitness landscape without directional pressure. When a new driver arises and survives stochastic extinction, positive selection operates — the mutant lineage expands, diversity contracts, and the neutral clock resets. After the sweep (or partial sweep), neutral dynamics resume.

This interleaving is captured in Turajlic et al.’s (2019) observation that neutral evolution is both a mode in its own right and the “inter-sweep default” that operates within the other three modes (linear, branching, punctuated). Positive selection is episodic; neutral evolution is continuous.

The clinical correlate: in multiple myeloma, detection of neutral dynamics correlated with better survival and was associated with a single strong clonal (truncal) oncogenic driver — the tumor had “found its peak” and stopped exploring (Johnson et al., 2017, cited in Turajlic et al., 2019). Tumors with ongoing positive selection — repeated sweeps by new drivers — represent a more aggressive, adapting population.

Limitations

1. Detectability threshold. With s ≈ 0.4%, positive selection is near the detection floor of moderate-depth sequencing (~100×). Many weakly selected drivers go undetected, and the boundary between weak selection and neutral drift is blurry (Turajlic et al., 2019).

2. Passenger confound. The 1/f² tail of neutral passengers is abundant at low frequencies. Distinguishing a weakly selected subclone at 2–5% VAF from a neutral passenger at the same frequency requires either massive cohort sizes (for recurrence-based inference) or orthogonal functional validation.

3. Time-limited observable window. Bulk sequencing at 100× depth can only detect mutations present in >1% of cells, which — at typical tumor sizes — means mutations that arose early in the tumor’s history. Late-arising drivers under positive selection are invisible because their lineage hasn’t had time to reach the detection threshold (Graham & Sottoriva, 2017).

4. Copy-number confounding. CNA alters VAF independently of clone abundance. An SNV under positive selection may have its VAF inflated (if on an amplified segment) or deflated (if on a deleted segment), producing false-positive or false-negative selection signals. CNA-aware CCF correction is essential for reliable selection inference (see cancer-cell-fraction).

5. Gompertzian correction. The neutral null is derived under exponential growth. Under gompertzian-growth, fewer cell divisions occur late in growth, depleting the low-frequency tail — a Gompertzian-growing tumor without selection could deviate from 1/f² and be misclassified as showing positive selection. The magnitude of this effect is unquantified (see neutral-evolution §Growth Model Caveat).