REPLYING TO D. I. S. Rosenbloom, A. L. Hill, S. B. Laskey & R. F. Siliciano Nature 551, 10.1038/nature24634 (2017)

In the accompanying Comment1, Rosenbloom et al. present a model simulation that questions our report2 of ongoing replication within drug sanctuaries in treated patients infected with HIV-1, which challenges the prevailing dogma that such replication cannot occur. They propose that cohorts of ever-older, latently infected cells revealed by differential decay of cell populations with varied half-lives after the start of antiretroviral therapy yield an illusion of viral evolution. We find this explanation untenable because Rosenbloom et al. make modelling assumptions that conflict with what is known about within-host virus evolution3,4.

As Rosenbloom et al. acknowledge1, they must apply very strong selective pressures in their simulations in order to match our observations (Supplementary Table 2). Selection must be both excessively strong (the selection coefficient s = 0.2 for each beneficial mutation) and pervasive (affecting about 3% of the genome) to reach the claimed 57% rate of false positives. However, even for escape mutations to cytotoxic T lymphocytes that target the virus, which are considered to be under exceptionally strong selection, the reported average strength of selection5 is only s = 0.03, and the average selection coefficient across the entire HIV-1 genome6,7 is s = 0.005. When s is set to more plausible values, the simulations register nearly no forward evolution signal.

The model absolutely requires these biologically untenable selective forces, as revealed by the key difference between the simulated and observed within-host phylogenies. Internal branches in fig. 2c, d of ref. 1 comprise a single substitution that must be driven to fixation almost immediately by unrealistically large selective pressures. This level of divergence is three to five times lower than we report2 (extended data fig. 3 of ref. 2), where multiple segregating (but mostly not adaptive) substitutions have accumulated between clades of haplotypes over the same period. Under any measure of phylogenetic support (for example, bootstrap), trees shown in fig. 2c, d of ref. 1 would collapse to rakes that bear no resemblance to experimental phylogenies2.

The results like those in fig. 2 of ref. 1 are also likely to be influenced by the decision to sample a small number of sequences from simulated populations that generates limited information on within-host diversity; drawing only 50 sequences per time point (as done by Rosenbloom et al.1) will not reliably (in 80% or more of cases) capture variants with frequencies of less than 3%.

The model in ref. 1 predicts that, after treatment initiation, viral DNA will become more similar to viral RNA collected ever earlier in infection, giving a backwards evolution signal. Viral sequences in peripheral blood T cells sampled at successive time points from patients on suppressive therapy8,9,10 do not support this prediction. The observation that DNA sequences in cells collected after the initiation of suppressive therapy retain close genetic similarity to the RNA sequences in plasma sampled in the late phase of HIV-1 infection before treatment11 further undermines the model’s predictions.

While differential decay of latently infected T cells may very well have a role in shaping sequence diversity, the process on its own fails to capture essential features of within-host sequence evolution and yields predictions not supported by available longitudinal data. Further multi-compartment deep sequencing studies of defined T cell subsets in larger cohorts of individuals on treatment are needed to establish the relative contributions of different mechanisms for viral persistence.