|Title / Titel||Large-scale statistical analysis of treatment change episodes in the SHCS database|
|Abstract (PDF, 14 KB)|
|Summary / Zusammenfassung||Despite an increasing arsenal and improved potency of antiretroviral drugs, the optimal use of combination therapy remains challenging. Complicating factors include the development of drug resistance, severe side effects, long-term toxicity and adherence to therapy. Because genotypic drug resistance testing is done on a routine basis today and because mutational patterns are unique for each individual patient, treatment choices are, in principle, highly personalized. In practice, however, it can still difficult to identify an optimal drug combination for each individual patient. Optimal selection of regimens is complicated by complex mutational patterns involving dozens of mutations and hundreds of feasible drug combinations. Treatment outcome can be analyzed in large clinical databases that collect and integrate viral sequence data, drug usage as well as virological and immunological follow-up data. Statistical and computational analysis of these high-dimensional data is challenging because of unobserved confounding factors, noisy observations, and sparse data sets.
The aim of this study is to analyze all treatment change episodes (TCEs) available in the SHCS database using state-of-the-art statistical learning methods. We will develop a computational model for the prediction of clinical response from viral genotype and drug combination. Important features (i.e., drugs, mutations, and combinations thereof) discriminating therapeutic success from failure will be identified.
We will extract all eligible TCEs from the SHCS database and develop appropriate filters for quality control. The resulting dataset is expected to be large and of exceptional quality due to the careful design and maintenance of the SHCS database. We will apply a range of modern statistical learning techniques to obtain computational models of therapeutic outcome based on viral genotype and regimen choice. We have developed several such methods in the past and we are currently undertaking a comprehensive comparison of existing methods on a publically available TCE dataset from the Stanford HIV Drug Resistance Database. The applied methods include linear and non-linear sparse regression and classification methods (LASSO, support vector machines), ensemble methods (random forests), and non-parametric methods (k-nearest neighbour). We will also explore the integration of phenotypic drug resistance predictions and of the genetic barrier to drug resistance.
|Keywords / Suchbegriffe||HIV, antiretroviral treatment, treatment change episode, drug resistance|
|Project leadership and contacts /
Projektleitung und Kontakte
|Funding source(s) /
|SNF (Personen- und Projektförderung)
Swiss HIV Cohort Study
|Duration of Project / Projektdauer||Apr 2010 to Mar 2013|