Unveiling ecological/evolutionary insights in HIV viral load dynamics: Allowing random slopes to observe correlational changes to CpG-contents and other molecular and clinical predictors

Abstract

In the context of infectious diseases, the dynamic interplay between ever-changing host populations and viral biology demands a more flexible modeling approach than common fixed correlations. Embracing random-effects regression models allows for a nuanced understanding of the intricate ecological and evolutionary dynamics underlying complex phenomena, offering valuable insights into disease progression and transmission patterns. In this article, we employed a random-effects regression to model an observed decreasing median plasma viral load (pVL) among individuals with HIV in Mexico City during 2019-2021. We identified how these functional slope changes (i.e. random slopes by year) improved predictions of the observed pVL median changes between 2019 and 2021, leading us to hypothesize underlying ecological and evolutionary factors. Our analysis involved a dataset of pVL values from 7325 ART-naïve individuals living with HIV, accompanied by their associated clinical and viral molecular predictors. A conventional fixed-effects linear model revealed significant correlations between pVL and predictors that evolved over time. However, this fixed-effects model could not fully explain the reduction in median pVL; thus, prompting us to adopt random-effects models. After applying a random effects regression model-with random slopes and intercepts by year-, we observed potential "functional changes" within the local HIV viral population, highlighting the importance of ecological and evolutionary considerations in HIV dynamics: A notably stronger negative correlation emerged between HIV pVL and the CpG content in the pol gene, suggesting a changing immune landscape influenced by CpG-induced innate immune responses that could impact viral load dynamics. Our study underscores the significance of random effects models in capturing dynamic correlations and the crucial role of molecular characteristics like CpG content. By enriching our understanding of changing host-virus interactions and HIV progression, our findings contribute to the broader relevance of such models in infectious disease research. They shed light on the changing interplay between host and pathogen, driving us closer to more effective strategies for managing infectious diseases. SIGNIFICANCE OF THE STUDY: This study highlights a decreasing trend in median plasma viral loads among ART-naïve individuals living with HIV in Mexico City between 2019 and 2021. It uncovers various predictors significantly correlated with pVL, shedding light on the complex interplay between host-virus interactions and disease progression. By employing a random-slopes model, the researchers move beyond traditional fixed-effects models to better capture dynamic correlations and evolutionary changes in HIV dynamics. The discovery of a stronger negative correlation between pVL and CpG content in HIV-pol sequences suggests potential changes in the immune landscape and innate immune responses, opening avenues for further research into adaptive changes and responses to environmental shifts in the context of HIV infection. The study's emphasis on molecular characteristics as predictors of pVL adds valuable insights to epidemiological and evolutionary studies of viruses, providing new avenues for understanding and managing HIV infection at the population level.

Keywords: Biological markers; CpG islands; HIV viral load; Prognostic factors; Regression analysis.

Figure A1.

Scatterplot of individual scores for sequences at PC1 and PC2, the extremes of the horizontal and vertical axes have been labeled with the most-contributing mutations—and their loadings— that should render larger positive or negative scores.

Figure A2.

PC 10 and 104 scores distribution in our sample, showing the most important positive and negative HAPs (HLA-associated mutations) and their corresponding loadings contributing to these PCs. Viral mutations in HIV-pol sequences may co-occur in combination (mutations with the same-signed loadings in a PC) or exclude each other (mutations with opposite-signed loadings in a PC) in our population of sequences. Numbers indicate the loading of each given mutation within PC_10 or PC_104. Blue bars (positive numbers) will add to the corresponding PC score, whenever a sequence possesses such mutation; red bars will subtract whenever an observation possesses that mutation. Mutations labeled “M” correspond to mutations associated to HLA alleles of the Mexican population, and labeled “H” correspond to HLA alleles in Caucasian populations (Deeks et al., 2001).

Figure A3.

PC 10 and 133 scores distribution in our sample, showing the most important positive and negative HAPs (HLA-associated mutations) and their corresponding loadings contributing to these PCs. Viral mutations in HIV-pol sequences may co-occur in combination (mutations with the same-signed loadings in a PC) or exclude each other (mutations with opposite-signed loadings in a PC) in our population of sequences. Numbers indicate the loading of each given mutation within PC_10 or PC_133. Blue bars (positive numbers) will add to the corresponding PC score, whenever a sequence possesses such mutation; red bars will subtract whenever an observation possesses that mutation. Mutations labeled “M” correspond to mutations associated to HLA alleles of the Mexican population, and labeled “H” correspond to HLA alleles in Caucasian populations (Deeks et al., 2001).

Fig. 1.

T Boxplot of HIV pVL yearly time-changes (2011 – 2021) for 7325 observations with complete data on all variables included in our models. The long horizontal red line indicates the overall pVL median. Horizontal lines within boxes indicate each year’s median (Q2), boxes comprise the Q1 and Q3 quartiles (IQR, interquartile range), while the extreme lines (whiskers) show Q3 + 1.5*IQR to Q1.

Fig. 2.

Z-standardized coefficients and their significance from a fixed-effects linear regression of pVL versus relevant predictors, including Year and all clinical and molecular variables. Colored bars represent the comparative signed quantities of each coefficient. Red bars indicate negative coefficients and blue bars indicate positive coefficients. Adjusted R-squared: 0.1303.

Fig. 3.

Independent linear models of molecular and clinical variables as functions of pVL and Year.

Fig. 4.

Boxplot shows the modelled yearly distributions of calculated pVLs, from a fixed-effects linear regression that included all predictors except the year of sampling.

Fig. 5.

A) Random coefficients by year of pVL versus clinical and molecular predictors. Colored bars in each column represent comparable yearly variation within each column. Slopes are non-standardized and not comparable to each other. B) Boxplot distributions of predicted (calculated) pVLs versus Year. Adjusted R-squared: 0.1422.

Fig. 6.

Comparison of intercepts in a random-intercepts-only versus a random intercepts and random slopes model.