History-aware models for predicting outcomes of HIV combination - PowerPoint PPT Presentation

History-aware models for predicting outcomes of HIV combination therapies Jasmina Bogojeska AREVIR-GenaFor-Meeting 2011

Problem setting Prediction of the outcome of combination therapies given to HIV patients Develop methods that can deal with: Different trends in treating patients Uneven therapy representation Current therapy Level of therapy experience

Data occurrence of Viral genotype 0 1 0 0 1 … resistance mutations drugs used in Current treatment 1 0 0 1 0 … current treatment drugs used in all Treatment history 1 1 0 0 1 … previous treatments 1 or -1 Label (success or failure) 6336 labeled samples with different 638 combination therapies

Treatment history-aware model Problems in the available HIV data: Only dominant strain sequenced- no information on latent virus population Uneven sample representation regarding level of therapy experience Idea: use treatment history information Extract knowledge on latent virus population Balance the data regarding level of therapy experience How? Quantify similarity of treatment histories by using drug resistance mutations

Similarity of treatment history Treatment record ordered by therapy starting time => therapy sequence    ( ) { | ( ) ( ) and ( ) ( )} r t z start z start t patient z patient t Desired properties Accounts for length of therapy history Order matters! Prediction targets current therapy Adapt sequence alignment methods to align therapy sequences!

Therapy sequence alignment Therapy sequence similarity Quantify pairwise therapy similarity Use drug resistance mutations

Treatment history-aware model Train a separate model for each therapy sequence by using knowledge from similar therapy sequences Sample weighted regularized logistic regression 1     T arg max ( ( ), ( )) ( ( , , , ), ) S r z r t l f x z h w y w w i i i i t i t t | | D w  t ( , , , ) x z h y D i i i i seq similarity loss function

Time-oriented evaluation scenario Special evaluation setup to address evolving trends in treatments over time 80% training 20% test

Results AUC performance stratified for the level of therapy experience

AUC performance stratified for the therapy abundance Results

Results

Conclusions Information extracted from treatment history enhances the performance for samples originating from both therapy-experienced patients and rare therapies Balance the uneven therapy-history representation in clinical data Time-aware evaluation setup to encounter changing trends in HIV treatment over time

Acknowledgements Thomas Lengauer HIV group @ MPII Saarbrucken Joachim Büch Rolf Kaiser & Group

Thank You!

Results