• The largest HIV resistance database available. EuResist's first key achievement was the creation of the EuResist Integrated Data Base (EIDB), which is now among the largest databases of HIV genotypes and clinical response available. It integrates data from the project partners (ARCA from Italy, AREVIR from Germany and Karolinska from Sweden), and has recently expanded with data coming from Luxembourg, Belgium (Leuven) , Spain (Badalona) and Portugal (Lisbon). The data collected include demographics, viral loads, CD4 counts, treatment history and HIV genotypes for almost 50.000 patients.

​

• EuResist Resistance prediction system: better than the state of the art. 

 

A standard datum for training the engines has been defined based on the notion of Treatment Change Episode (TCE). In its essential form, a TCE comprises the baseline viral load and genotype at the start of a new treatment, the treatment itself and a follow-up viral load obtained while still on that treatment. The time intervals in the EuResist TCE are compliant with the Forum for HIV Research definition focusing on short-term 8-week response. Data mining techniques have discovered several additional features and derived features that can have an impact on response to treatment and improve the accuracy of the prediction engine.

​

Three engines have been developed and found to perform similarly but not identically in validation tests. Techniques such as Support Vector Machines, Fuzzy logic, Case Based Reasoning, and Random Forests have been investigated, while finally all of the three engines have ended up with using Logistic Regression models for the classification of the therapies as successes or failures. However, the three engines use different approaches to derive such models from the EuResist database plus derived features providing extra information not directly contained in the standard datum.

​

Several sophisticated information fusion methods have been studied and compared to deliver one single combined prediction starting from the three individual engines. Finally a simple mean combiner has been chosen, which shows a better learning curve than the individual prediction systems or more complex information fusion methods. Based on validation studies done with historical cases, the final EuResist Prediction engine currently predicts the correct outcome in 76% of cases. Interestingly, a recent analysis has also indicated that the accuracy of the prediction of the 24-week outcome equals that of the 8-week outcome even without retraining the system with 24-week follow-up data.

​

The EuResist engine has been compared with the state-of-the-art tools available for assisting the selection of antiretroviral therapy in clinical practice, such as ANRS (France), Rega (Belgium) and HIVdb (Stanford, USA).

Different methods produce different types of predictions and this complicates the comparison. In order to allow for reasonably fair comparisons among the systems, careful statistical analysis has been performed. Comparison results show that EuResist prediction is more accurate than state-of-the-art systems.

​

Background. While combination antiretroviral therapy has made HIV infection a treatable condition, eradication of infection is not yet achievable and antiretroviral therapy needs to be administered as a prolonged, possibly lifelong treatment . Long-term toxicity, difficulty in adhering to complex regimens, possible pharmacokinetics problems, and intrinsically limited potency are all factors favoring the selection of drug-resistant viral strains . Development of drug resistance is nowadays a major cause for treatment failure.

​

The "old" solution. The basic method for defining the drug resistance profile of a given virus population implies culturing the virus in the presence of each single drug and measuring the extent of virus replication in a well controlled virus-cell system. Unfortunately, considerable efforts are required to develop and standardize a test of this kind (‘phenotypic' assay) to be performed on a routine basis. Indeed, reliable drug resistance phenotyping is currently provided only by specialized companies at very high cost . Moreover, evolution of resistance in vivo can be different from what observed in vitro and inferring clinical resistance from in vitro data is not straightforward, particularly for some drugs.

​

Recent advances in molecular diagnostics have made available a more practical and cost-effective approach to drug resistance testing, based on direct sequencing of the relevant regions of the virus genome (genotype). Mutations detected in the virus genotype are then used to infer the drug susceptibility profile based on known correlations between genotype and phenotype.

​

A number of algorithms (e.g. stanford, retrogram, trugene, anrs, Rega) have been developed in recent years based on sets of rules which assign each drug to one of different resistance categories depending on the presence of defined mutations or combinations of mutations. These methods, still immersed in a phenotype-based theoretical framework, provide predictions which can be inaccurate for clinical purposes, particularly considering that both phenotypic and genotypic assays provide an estimate of susceptibility to single drugs while three- or four-drug combinations are commonly used in vivo.