The search for N_eff (and other HIV evolutionary parameters) has gone on for almost two decades, following every turn and hitting each pothole on the eventful road of HIV modeling [9]. The rapidity of resistance to monotherapy (in 1–2 weeks) was explained by the deterministic selection of alleles that preexist therapy in minute quantities [1]. The large numbers of virus-producing cells (∼10⁸) in the lymphoid tissue of experimentally infected macaques seemed to confirm this simple Darwinian selection model [10]. However, the Darwinian view has faced challenges. Tajima's “neutrality test” applied to HIV sequences in untreated patients assumed that selection was neutral and predicted much smaller “effective” populations, of N_eff∼10³ [11]. Since Tajima's approach was designed to detect isolated selective sweeps at one or a few mutant sites—while HIV exhibits hundreds of diverse sites in vivo—two groups re-tested the result. A linkage disequilibrium (LD) test [12] and analysis of the variation in the time to drug resistance [13] arrived at the same value, N_eff = (5–10)×10⁵, for an average patient (with the mutation rate ∼10⁻⁵ per base). Such populations are sufficiently large for deterministic selection to dominate, yet not large enough to neglect stochastic effects altogether. The LD test [12] is affected by recombination, and HIV's recombination rate had not been well measured at that time. The recent measurement of 5×10⁻⁶ crossovers per base per HIV replication cycle in an average untreated individual [14]–[16] updates N_eff to (1–2)×10⁵, not far from the original value. A recent study of the pattern of diversity accumulation in early and late HIV infection confirms the range of N_eff [17]. However, all these estimates of N_eff are lower bounds.

Pennings et al. [8] continue this quest for an effective population size of HIV using a new method based on a theoretical calculation of the probability of multiple introductions of a beneficial allele at a site before it is fixed in a population [18]. The prediction does not depend on whether mutations are new or result from standing variation prior to therapy. The authors use sequence data obtained from 30 patients who failed suboptimal antiretroviral regimens, including efavirenz [19]—a non-nucleoside reverse transcriptase (RT) inhibitor (NNRTI)—and who exhibited a rise of drug-resistant alleles in RT. The sequence data reveal fixation of two alleles, both corresponding to an amino-acid replacement K103N. Pennings et al.'s analysis focuses on the genetic composition at RT codon 103 and the adjacent 500 nucleotides. Based on the changes in the genetic diversity in this region, 30 fixations are classified into “hard” selective sweeps with a single parental sequence, or “soft” sweeps with multiple parental sequences. Observing that both types of sweep occurred at similar frequencies (also confirmed by observations in other resistance codons), the authors predict N_eff = 1.5×10⁵, in agreement with the LD test.

Pennings et al. also discuss why “selectively neutral” methods based on synonymous diversity underestimate the population size. It is well known that a selection sweep lowers the diversity at linked sites (hence the term “sweep”) and any method assuming selective neutrality translates lower diversity to smaller N_eff. The interesting part is the dynamic component of this effect. Pennings et al. demonstrate that rapid sweeps are followed by long periods when the diversity recovers at the linked sites (for synonymous sites, these periods are very long). From another angle, we can add that selection shortens the time to the common ancestor, which decreases the sequence divergence. The ancestral-tree argument is rather general and also applies to a large number of linked sites evolving under selection [20]–[23].

The previous estimates [12], [13], [17] were lower bounds on N_eff. In contrast, the Pennings et al. study puts a number on N_eff. However, this number (N_eff = 1.5×10⁵) raises a question: why is N_eff so far below the census population size of 10⁸ or more? Pennings et al. offer an elegant explanation of this relatively small N_eff in the spirit of the “traveling wave” approach [24]–[27]. They note that resistant alleles at different sites emerge against different fitness backgrounds. To be fixed, alleles conferring a small benefit must emerge in the most-fit genomes [28], [29]; hence, the effective N_eff for these alleles is small. Alleles with a larger beneficial effect can explore a larger fraction of population (larger N_eff). Conceptually, this idea is quite correct; quantitatively, in the context of drug resistance, some problems arise. For example, the fitness benefit from a resistance mutation (under drug) is almost 100%, while the difference between the fittest and the average genome (in untreated patients) is a modest ∼10% [14]. Indeed, the average selection coefficient is quite small, ∼0.5% [14], [15].

There may be several other reasons for N_eff<10⁸, as follows.

For these reasons, the value N_eff = 1.5×10⁵ obtained in the study of Pennings et al. should probably still be regarded as a lower bound. At the same time, the study solidifies our understanding of HIV evolution as a Darwinian process and leads to important questions regarding the structure of HIV population, which are still waiting for new insights.