*2.5. ART Experience*

Among the 1497 patients receiving atazanavir, 907 (60.6%) were previously ART-naïve and 590 (39.4%) were ART-experienced (Table 2). The proportion of sequences containing one or more PI-associated DRMs was 21.7% in previously ART-experienced patients and 15.0% in previously ART-naïve patients (*p* = 0.001). Among those with one or more PI-associated DRMs, the number of DRMs was not significantly different in previously ART-experienced patients (median 2 DRMs; IQR: 1–4 DRMs) compared with previously ART-naïve patients (median 3 DRMs; IQR: 1–4 DRMs; *p* = 0.3).

Among the 907 previously ART-naïve patients, atazanavir was administered with 2 nucleoside RT inhibitors (NRTIs) in 840 (92.6%) patients. Among the remaining 67 patients, the co-administered antiretroviral drugs (ARVs) were not provided for 44 (4.9%), while 23 (2.5%) received a variety of other ARVs.

Among the 590 previously ART-experienced patients, atazanavir was administered with 2 NRTIs in 345 (58.5%) patients. Among the remaining 245 patients, the co-administered ARVs were not provided for 163 (27.6%), while 82 (13.9%) received a variety of other ARVs. Only four patients received atazanavir plus one additional ARV.

The year of ART-initiation was available for 1127 (75.3%) of all patients. The patients could be pooled into four time periods containing approximately equal numbers spanning the years between 1993 and 2018 (Table 2). The proportion of patients with one or more PIassociated DRMs decreased over time (binomial coefficient = −0.26; 95% CI: −0.45 to −0.07; *p* = 0.007), but the number of DRMs in patients with one or more DRMs did not change.

Using just those patients for whom the year of ART initiation was available, a multivariate logistic regression analysis was performed to assess the association between four factors and the development of a PI-associated DRM. The four factors included the year of ART initiation, subtype (B vs. non-subtype B), the use boosted vs. unboosted atazanavir, and previous ART (naïve vs. experienced). The analysis found that a later year of ART initiation (OR: 0.62; 95%CI: 0.49–0.79; *p* = 0.0001) and the administration of boosted atazanavir (OR: 0.57; 95%CI: 0.35–0.93; *p* = 0.02) were associated with a decreased risk of developing a PI-associated DRM.

#### *2.6. Bayesian Network Analysis of Correlated Mutations*

We used the 1437 (96%) sequences containing 0 to 4 PI-associated DRMs (i.e., sequences with ≥5 PI-associated DRMs were excluded) to calculate Jaccard similarity coefficients and their standard Z scores for all pairs of DRMs and NP-TSMs. Eleven pairs of mutations comprising six major DRMs (M46I, I50L, I54V, V82A, N88S and L90M), three accessory DRMs (K20T, L33F and G73S), and the NP-TSM L89T participated in one or more significant pairwise correlations (*p* < 0.01). We then performed a Bayesian network analysis to determine the conditional dependency between the mutations in each of the pairwise correlations (Figure 2).

**Figure 2.** Bayesian network analysis of positively correlated mutation pairs with a hill-climbing search. The Bayesian network analysis yielded 11 mutation pairs, including 6 major DRMs (red), 3 accessory DRMs (yellow), and an additional nonpolymorphic treatment-selected mutation (light blue) with a significant Jaccard correlation coefficient (*p* < 0.01). The thickness of the arrows indicates the strength of the probabilistic relationship of the two mutations. The direction of the probabilistic causation is shown with an arrowhead. For the direction between V82A and I54V for which the probabilistic causation is not greater than the probabilistic causation of the opposite direction by 0.1, the arrowhead is not shown.

#### *2.7. Estimated Cross Resistance to LPV/r and DRV/r*

Among the 264 sequences with 1 or more PI-associated DRMs, there were 182 distinct DRM patterns, including 124 patterns (164 sequences; 62.1% of 264) interpreted by HIVDB as causing high-level atazanavir resistance, 19 patterns (20 sequences; 7.6% of 264) as causing intermediate atazanavir resistance, and 29 patterns (51 sequences; 19.3% of 264) as causing low- or potential low-level atazanavir resistance. The remaining 10 DRM patterns (n = 29 sequences patterns; 11.0% of 264) consisting primarily of singe accessory DRMs (e.g., K20T, Q58E) were not interpreted as causing reduced atazanavir susceptibility.

A total of 56 distinct DRM patterns (58 sequences; 22.0% of 264) were interpreted as causing high-level lopinavir resistance, 40 patterns (43 sequences; 16.3% of 264) as causing intermediate lopinavir resistance, and 44 patterns (62 sequences; 23.5% of 264) as causing low- or potential low-level lopinavir resistance. A total of 3 distinct DRM patterns (3 sequences; 1.1% of 264) were interpreted as causing high-level darunavir resistance, 14 patterns (14 sequences; 5.3% of 264) as causing intermediate darunavir resistance, and 32 patterns (34 sequences; 12.9% of 264) as causing low- or potential low-level darunavir resistance.

#### *2.8. Virological Failure with Resistance*

Five of the thirty studies included participants from three clinical trials and from two clinical cohorts for which genotypic resistance testing was routinely available (Table 1). Together, these five studies included 1037 (69.3%) of all 1497 patients from whom sequences were available. Of these 1037 patients, 63.0% and 37.0% received boosted and unboosted atazanavir, respectively. In these studies, the proportion of sequences containing one or more PI-associated DRMs ranged from 2.9% to 14.5% and the overall proportion of sequences containing one or more PI-associated DRMs in patients receiving boosted and unboosted atazanavir were 7.2% and 13.5%, respectively.

#### *2.9. Studies Not Included in the Analysis*

We identified 32 additional studies reporting sequences from 1089 previously PInaïve patients receiving boosted or unboosted atazanavir-containing regimens (Table S3). Approximately 10% of the sequences in these studies were reported to contain one or more PI-associated DRMs. However, as the sequences were not available and as different mutations were reported in different studies, we did not include the data from these studies in our analysis.
