Volume 13 Supplement 4

Abstracts of the Tenth International Congress on Drug Therapy in HIV Infection

Open Access

O116. The EuResist expert model for customised HAART optimisation: 2010 update and extension to newest compounds

  • A Pironti1,
  • A Sönnerborg2,
  • M Zazzi3,
  • R Kaiser4,
  • D Struck5,
  • B Clotet6,
  • ÀM Vandamme7,
  • F Incardona8,
  • T Lengauer1,
  • M Rosen-Zvi9 and
  • M Prosperi10
Journal of the International AIDS Society201013(Suppl 4):O6

DOI: 10.1186/1758-2652-13-S4-O6

Published: 8 November 2010


The design of an optimal highly active antiretroviral therapy (HAART) customised on patient's background and viral genotyping, is still a challenge. EuResist has been the first data-driven system to be implemented as a free web-service for customised HAART optimisation, and was proven to be superior to all existing genotypic interpretation systems (GIS) since it takes into account not only genotype but multiple other variables.

Data and methods

The EuResist database stores and updates periodically demographic, clinical, and genomic information of HIV+ patients from several countries in Western Europe. The EuResist system is trained on treatment change episodes (TCE) drawn from the EuResist data base, composed of a new drug regimen with a baseline HIV-1 RNA load and a CD4+ count, a baseline viral pol genotype, demographic and previous treatment information. Each TCE is associated to an HIV-1 RNA measurement after 8 weeks, which is used for the definition of virological success (below 500 copies/ml or >2 Log10 reduction from baseline HIV-1 RNA). The system is a combination of three independent machine learning models (based on logistic regression, random forests, and Bayesian networks). The 2010 update has been trained on >5,000 TCE, composed of 20 FDA/EMEA approved nucleoside/tide, non-nucleoside, and protease inhibitors, including the recently approved compounds (RAC) tipranavir, etravirine and darunavir.


The EuResist combined system performance in predicting the correct virological success of a TCE after 8-weeks (validation set, n=561) exhibited an area under the receiver operating characteristic (AUROC) of 0.8 (whereas Stanford HIVdb GIS assessed to 0.73, p=0.002). The inclusion of therapy history, clinical, and demographic covariates was shown to increase significantly prediction performance. In the subset of regimens containing RAC (n=151), the EuResist AUROC was 0.7 and Stanford HIVdb AUROC was 0.63, not allowing to assess a significant difference owing to the small sample size. See Figure 1.

Figure 1


Based on patient’s information and virus genotype, the EuResist web-service ranks the most effective HAART regimens by the probability to achieve an undetectable HIV-1 RNA load after 8 weeks. Thus, it might be useful in clinical practice. The 2010 update includes also RAC and achieved fair performance, which is expected to increase with expanding training data.

Authors’ Affiliations

Max-Planck-Institut für Informatik, Computational Biology and Applied Algorithmics
Karolinska Institutet, Clinical Virology/Infectious Diseases
Università degli Studi di Siena, Molecular Biology Department
Universitätsklinikum Köln, Institut für Virologie
Centre de Recherche Public de la Santé, Laboratory of Retrovirology
Katholieke Universiteit Leuven, Clinical and Epidemiological Virology
Informa S.r.l., Research and Design
BM Haifa Research Labs, Machine Learning and Data Mining
Catholic University of the Sacred Heart, Clinic of Infectious Diseases


© Pironti et al; licensee BioMed Central Ltd. 2010

This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.