Human Immunodeficiency Virus 1 (HIV-1) enters cells in a complex process involving interactions of viral envelope protein gp120 with the cellular receptor CD4 and a co-receptor, typically one of the chemokine receptors CCR5 or CXCR4 [1]. According to their co-receptor usage or "tropism", viruses are classified as "R5" (interacting with CCR5) or "X4" (interacting with CXCR4). Additionally, there are dual-tropic "R5X4" strains that use both co-receptors for cell entry. Tropism is mainly determined by the sequence of the variable loop 3 (V3) of gp120. In initial infection, R5 viruses dominate the viral quasi-species [2]. As the disease progresses, about 50% of the patients develop X4 virus [3]. CCR5 blocking drugs, such as Maraviroc or Vicriviroc [4, 5] are ineffective against X4 virus, and thus it is advisable to test tropism prior to treatment with these drugs. The current state-of-the-art is testing by phenotypic assays such as Trofile^® (Monogram Biosciences, CA) [6] or enhanced sensitivity Trofile^® assay (ESTA) [7]. However, their restriction to specialized laboratories, high cost and long turn-around are limiting availability. Moreover, phenotypic assays have been reported to fail in delivering any result in more than 15% of the cases [8]. An alternative for routine diagnostics is genotypic testing: the genomic sequence of V3 from a patient is interpreted using computational models that relate V3 sequence and tropism. These models are typically derived by machine learning methods from a training set of V3 sequences and corresponding phenotypic test results [9–14]. Genotypic predictions can be made available via the Internet, and they are fast and cheap. Failure rates have been estimated to be around 7.5% [8]. In clinical settings with tropism predictions based on single sequences from bulk sequencing, genotypic methods tend to perform less well [15], which is mostly attributed to low detection rates of X4 minorities by bulk sequencing [16]. Genotypic testing based on so-called "next generation sequencing" or "deep sequencing" methods may not suffer from this limitation [17] as they provide detailed data for the whole viral quasi-species. In fact, Vandenbroucke et al.[18] have demonstrated that a combination of deep sequencing of V3 with a genotype interpretation algorithm [11, 19] can be used for determination of tropism even in cases where phenotypic testing fails. In their study, the error rate of prediction methods was a limiting factor.

Recently, we have devised a two-level machine learning approach (T-CUP) [14] for the prediction of HIV-1 co-receptor usage from V3 sequences. At the first level, two independent predictions are made, based on the electrostatic potential and hydropathy values [20] of the V3 loop, respectively. The predictions are then combined, and a final decision is reached. The method is accurate, provides predictions for all subtypes and is robust with respect to insertions and deletions. In this study, we applied T-CUP to deep sequencing data from Tsibris et al.[21], compared the predictions with results of phenotypic assays, and analyzed features of the viral quasi-species that could indicate tropism switches. Tsibris et al. had generated deep sequencing data for four patients at three time points during treatment with Vicriviroc, and concurrently had measured phenotypic tropism with the standard Trofile^® assay. These assays had shown R5 virus at treatment start (Week 0 in Figure 1 and 2) for all patients. The common feature of all four patients was failure of Vicriviroc therapy.

We next exploited the property of T-CUP to provide in the first level two independent tropism predictions based on physical properties (electrostatics and hydropathy) of V3. The corresponding probabilities span a plane ("probabilities plane") in which every V3 sequence is represented by a point and the quasi-species by a cloud of such points. Figure 2 shows this plane for all twelve datasets from Ref. [21] with the points colored according to frequency of the respective sequence in the deep sequencing data.

The dynamics of the quasi-species in the probabilities plane has several remarkable features. First, all sequences in week 0 cluster in the lower left corner of the plane as is expected for a quasi-species that is R5 tropic. Second, the movement of the clouds indicates the dynamics of tropism. For Subjects 07 and 18 the clouds move towards the upper right, i.e. to more X4 tropism. For Subject 19 this movement is also seen for the first two time points but then reverts again to the lower left, i.e. to more R5 tropism. Subject 47 shows no marked movement to the upper right but remains localized in the lower left, in agreement with a quasi-species that remains R5 tropic. Third, for the patients 07, 18, and 19 where a co-receptor switch had been observed, there is only one clearly dominating X4 strain in the probabilities plane, and this strain is already present at therapy start with considerable frequency (bright spots with green arrows). This "X4 seed strain" is specific for each of the patients - the seed strains for different patients are clearly located in different regions of the probabilities plane. Additionally, the X4 seed strain is accompanied from the beginning by a growing halo of local minor variants. Note that Subject 47 who remains R5 tropic throughout all time points does not have such a cluster.

We interpret the seed strain with its halo as a variant that has established itself in the quasi-species even at week 0 so that it can generate a considerable number of copies and also generates variant offspring. This interpretation is supported by the high homology of the sequences in the cluster (Figure 3). Using the statistical properties of this cluster in the probability plane (strong seed strain, halo of neighbor strains), it may become possible to predict a future co-receptor switch and therapy failure many weeks earlier than the switch of tropism becomes manifest in a phenotypic test. For such an early detection of a later switch of tropism, the resolution of deep sequencing and the accuracy of the prediction method is critical. For all three subjects where a switch occurs, the X4 seed strain initially accounts for 0.5% to 1% of the quasi-species (Table 2). The following progression towards X4 tropism is almost exclusively due to the expansion of these seed strains. For the three patients with a tropism switch a linear correlation of development of the total X4 fraction over time with the development of the fraction of the seed strain over time yields R² = 0.98 (p = 7.7 · 10^-10).

Although the high cost of deep sequencing will probably prevent its use in routine diagnostics in the near future, the combination of this powerful method with accurate predictions could be applied when phenotype testing fails and to study evolution of viral quasi-species under selective pressure, and thus contribute to the development of sustainably effective treatments.