Arevir Meeting, Bonn, 22-23 April 2010 M. Zazzi on behalf of the - PowerPoint PPT Presentation

Arevir Meeting, Bonn, 22-23 April 2010 M. Zazzi on behalf of the EuResist Network (www.euresist.org)

EuResist status Funded by the EU JAN-06 to JUN-08, then set as a  European Network (legal entity) GOAL: to develop and make freely available an on-line  expert system for prediction of response to antiretroviral treatment Data collected from ~40,000 patients (Italy,  Germany, Sweden, Luxembourg, Belgium, Spain) Data modeling by IBM Israel, Max Planck Institute  for Informatics, Informa & Rome TRE University

EuResist – type of data collected Integration of clinical and laboratory data from multiple sources THERAPY AIDS EVENTS PatientID PatientID Treatment regimen GENOTYPE Event Date of start PatientID Date Date of stop Date Reason for change/stop Sequence Method PATIENTS Patient ID CD4 Gender HIV RNA PatientID Year of birth PatientID Date Country of origin STATUS Date CD4/cmm Risk group PatientID Copies/ml CD4% HCV status Date <LLD (undetectable) HBV status Followed Method Lost Died

From genotype to response CD4 Treatment switch HIV RNA Baseline Follow-up viral GENOTYPE, viral load, CD4, ... load, CD4, ... Model training

EuResist – Treatment Change Episode (TCE) definition CD4 Viral load Treatment switch Genotype Viral load time 0 to 12 weeks Short-term model: 4-12 weeks Pre-therapy HIV RNA Patient demographics (age, gender, race, route of infection) Past AIDS diagnosis Past treatments Past genotypes

EuResist – labeling therapies Baseline data HIV genotype at 0 to 12 weeks before treatment VL at 0 to 12 weeks before treatment Additional variables when available Treatment switch VL at 4 to 12 weeks (8-week outcome) SUCCESS FAILURE Undetectable or >2 log Detectable and not >2 log decrease VL decrease VL

The data funnel…

EuResist – HIV clades distribution

EuResist – engines Three prediction engines developed  independently Generative-Discriminative (by IBM)  Evolutionary (by Max-Planck Institute)  Mixed effects (by Rome TRE & Informa)  Then, combined into a unique engine  and made freely available on the web

Eu Resist Generative/Discriminative engine Model response to treatment in the absence of genotype with a Bayesian network For any defined regimen, compute a probability of success (Generative step) Use the probability as an additional feature for logistic regression together with genotype and other covariates (Discriminative step)

Eu Resist Evolutionary engine Model HIV evolution under therapy from longitudinal and cross-sectional sequence data For any defined genotype, neighbor mutants can be computed in silico and the contribution of the expected mutants to resistance can be calculated Functions weight for probability and expected time for mutants to occur Probability to remain susceptible to a drug (below a defined phenotypic threshold) Altmann et al, AVT 2007 GENETIC BARRIER

Eu Resist Mixed-Effects engine Focuses on interactions among variables drug x drug • drug x drug x drug • drug x mutation • drug x previous drug class exposure • drug x previous drug exposure • mutation x mutation •

The EuResist combined engine 3143 therapies, Short-term outcome (8 weeks) Rosen-Zvi, Bioinformatics 2008

Eu Resist Combined engine The combination  (mean) of the engines performs equal to or better than the individual engines The combined  engine learns faster, i. e. it is more accurate when trained on limited data sets Altmann et al, PLoS ONE 2008

Virtual response on the web INPUT QUESTION Remote What treatment(s) will be successful for users my patient? RESPONSE Ordered list of the best treatments for that patient Web server OUTPUT

Connections used during project life and then for system updates Connections used by the final users Combined Web interface predictive Individual system engines End users Merged EuResist DB Merged EuResist DB Merged EuResist DB D I L S D I L S D I L S Feeding DBs from different countries

INPUT

INPUT OUTPUT

The Computerworld Honors Program Honoring Those Who Use Information Technology to Benefit Society EuResist • Laureate Award plus • Special Recognition

EuResist vs. Expert interpretation (EVE study) Form the invitation letter: The requested response include a categorical (C) answer and a quantitative (Q) estimate: C) Given this HIV genotype and patient information, will the indicated therapy be successful (i. e. will it make HIV RNA decrease by at least 2 logs or to undetectable levels in 8 weeks) ? Q) Given this HIV genotype and patient information, what probability of success would you estimate for the indicated therapy?

Distribution of wrong and correct calls 25 HAART cases randomly selected form the EuResist db: • Obsolete therapies excluded • Wild type genotype excluded • All clinical and virological information available 12 experts enrolled, response obtained from 10: • On ‐ line anonymous rating • Only European (E) vs. non ‐ European (N) setting traceable • Use of any interpretation system allowed (and declared)

Why do we err (so much)? Limitations in the definition of success can  “create” prediction errors (e. g. success at a later time point) Expected adherence cannot be accounted for  based on presently available training data set Genotype shortcomings  Impact of past mutations  Short-term drug activity on partially resistant strains  (e. g. 215 revertants, Y181C with EFV) Subtype differences?  Unweighted factors (host genetics) 

Prediction based on treatment history vs. genotype Based on HIV Based on genotype treatment history Preliminary analysis  Further work warranted in the setting of missing genotype information 

EuResist – Data sources Data regularly refreshed from original  sources & later contributors Belgium, Italy, Germany, Luxembourg, Spain  Data from new countries  Other European countries (Greece, Portugal)  Non-European countries  Data format conversion and upload provided by  IBM as EuResist contractor

EuResist – Accessibility Official website www.euresist.org  Interface in Russian now available  Incorporated into InfCare, the Swedish HIV  data management and remote consulting system Used in all Scandinavia and Baltic states, being  adopted in several low-middle income countries EuResist on Facebook (coming soon)  EuResist entry in Wikipedia (coming soon) 

EuResist within InfCARE

EuResist engine update From ~3,000 to >5,000 TCEs  Small increase in accuracy  AUC from 0.76 to 0.78  Upper bound of accuracy already reached? ○ Outperforms HIVdb with the training set (AUC 0.69)  List of regimens considered in output  expanded from 100 to >400 Tipranavir, Darunavir, Etravirine now computed  Lower accuracy, AUC 0.70 (HIVdb 0.63) ○ Data from clinical trials coming! ○

EuResist A treasure for collaborative research Extraordinary collection of clinical and  virological data Ideal for resistance related investigations but  suitable also for other research topics Data available for the scientific community  Many requests already satisfied  Networks (Virolab, CHAIN, CORONET) as well as ○ single research units Cooperation preferred over the simple task of  providing data

EuResist Partners Expectations & opportunities Partnership regulations set (see web site)  Benefits for the prediction engine  Increased size of training data set  Expanded representation of different scenarios  (e. g. drug prescriptions, HIV clades, different epidemiology & demographics) Testing the system in different geographic areas  Benefit for clinical research  Role as data providers  Role as investigators 

EuResist Partners & Scientific Board The EuResist Network offers partnership to new  parties of proved or potential scientific value All partners become members of the EuResist  Network Scientific Board The EuResist Network Scientific Board evaluates  the scientific adequacy of research proposals, made by third parties, that require the access to the data stored in the EuResist Integrated Data Base (EIDB) and takes decisions on approval or rejection of such research proposals

EuResist Partners Partner’s activities and data are protected by the  data and authorship policies Data provided by a partner always remain the  property of the partner Data providers are acknowledged in all  publications Authorship is built based on the proportion of  cases contributed for a specific study. Credits are saved and used subsequently