Predicting and Understanding HIV-1 Resistance to Broadly Neutralizing Antibodies Anna Feldmann Max Planck Institute for Informatics
Motivation ● HIV-1 drug target space is limited ● Drug resistance emergence under HAART ● Consistent change of treatment for chronic patients Need for new drug targets De Clercq et al. , Nature Reviews Drug Discovery 2007 1 Anna Feldmann
Motivation De Clercq et al. , Nature Reviews Drug Discovery 2007 2 Anna Feldmann
Motivation Burton et al. , Nature Reviews Immunology 2002 (new) target: envelope spike treatment option: broadly neutralizing antibodies ( bNAbs ) De Clercq et al. , Nature Reviews Drug Discovery 2007 2 Anna Feldmann
Broadly Neutralizing Antibodies PG9 (V1/V2-loop) (Walker et al. , 2009) PGT128 (V3-loop) (Walker et al., 2011; Mouquet et al. , 2012) Glycans 45-46G54W (CD4bs) (Diskin et al. , 2011) 4E10 (MPER) (Cardoso et al. , 2005) Burton et al. , Science 2012 3 Anna Feldmann
Broadly Neutralizing Antibodies PG9 (V1/V2-loop) (Walker et al. , 2009) Developed in 10-30% of infected patients PGT128 (V3-loop) BUT: too little too late (Walker et al., 2011; Mouquet et al. , 2012) Glycans 45-46G54W (CD4bs) (Diskin et al. , 2011) HIV-1 vaccine HIV-1 research: treatment: 4E10 (MPER) (Cardoso et al. , 2005) Induce bNAb bNAb Burton et al. , Science 2012 development immunotherapy 3 Anna Feldmann
bNAbs for HIV-1 Treatment 4 Anna Feldmann
bNAbs for HIV-1 Treatment 4 Anna Feldmann
bNAbs for HIV-1 Treatment 4 Anna Feldmann
bNAbs for HIV-1 Treatment 1. Resistance to bNAbs Challenges ( Will it work? ) 2. Optimal Personalized Treatment ( Which will work best? ) Goal Given HIV-1 variants of the patient and a bNAb Predict susceptibility or resistance 5 Anna Feldmann
Data Neutralization assay data covering 4 major epitopes: V1/V2-loop: gp41-gp120: 35O22 PG9, PG16 V3-loop: CD4bs: PGT128, VRC01, PGT121, VRC-PG04, 10-1074, NIH45-46, 10-996 3BNC117 Doria-Rose et al. , J.Virol. 2009; Mouquet et al. , PNAS 2012; Huang et al. , Nature 2014 6 Anna Feldmann
Data Neutralization assay data covering 4 major epitopes: V1/V2-loop: gp41-gp120: 35O22 PG9, PG16 V3-loop: CD4bs: PGT128, VRC01, PGT121, VRC-PG04, 10-1074, NIH45-46, 10-996 3BNC117 115-230 envelope (Env) sequences with corresponding IC50 values per bNAb Doria-Rose et al. , J.Virol. 2009; Mouquet et al. , PNAS 2012; Huang et al. , Nature 2014 6 Anna Feldmann
Building Prediction Model 1. Learning 6 Anna Feldmann
Building Prediction Model 1. Learning 6 Anna Feldmann
Building Prediction Model 1. Learning Binarize label into ● resistant ( – ), if IC50 ≥ 50μg/mL ● susceptible ( + ), if IC50 < 50μg/mL 6 Anna Feldmann
Building Prediction Model 1. Learning Binarize label into ● resistant ( – ), if IC50 ≥ 50μg/mL ● susceptible ( + ), if IC50 < 50μg/mL 6 Anna Feldmann
Building Prediction Model 1. Learning 2. Predicting 6 Anna Feldmann
Prediction Performance ● Model performance was tested in a 10 - times nested cross- validation ● Overall high prediction performance ( up to 0.84 AUC ) ● Classifiers for the same epitope achieve similar performances 7 Anna Feldmann
Prediction Performance Questions: Can we interpret the models? Can we interpret the classification result? 7 Anna Feldmann
Understanding the Classifier Learnt discriminant positions of the classifiers susceptible resistant 8 Anna Feldmann
Understanding the Result AA susceptible aa resistant Residues of the test sequence that contributed the most (strongest 5%) to the classification result of the PG9 classifier. 9 Anna Feldmann
bNAbs for HIV-1 Treatment 1. Resistance to bNAbs Challenges ( Will it work? ) 2. Optimal Personalized Treatment ( Which will work best? ) Goal Given HIV-1 variants of the patient and a bNAb Predict susceptibility or resistance 10 Anna Feldmann
bNAbs for HIV-1 Treatment 1. Resistance to bNAbs Challenges ( Will it work? ) 2. Optimal Personalized Treatment ( Which will work best? ) Goal Given HIV-1 variants of the patient and a bNAb Predict the corresponding IC50 value 10 Anna Feldmann
Building Regression Models Setup: ● Same input data ● Instead of binarization, log transformation used ● Instead of classification, the corresponding IC50 value is predicted using support vector regression 11 Anna Feldmann
Building Regression Models Setup: ● Same input data ● Instead of binarization, log transformation used ● Instead of classification, the corresponding IC50 value is predicted using support vector regression Result: Positive correlations of 0.3 – 0.5 for all bNAbs apart from 35O22 11 Anna Feldmann
Continuous Drift Towards Resistance Studied population: ● 40 Caucasian men having sex with men, subtype B ● similar distribution of viral loads and CD4-T cell counts ● b12, VRC01 , VRC03, NIH45-46G54W , PG9 , PG16 , PGT121 , PGT128 , PGT145 ● Over 20 years (1987–1991/ 1996–2000/ 2006–2010) ● French ANRS PRIMO and SEROCO cohorts 12 Anna Feldmann
Continuous Drift Towards Resistance Studied population: ● 40 Caucasian men having sex with men, subtype B ● similar distribution of viral loads and CD4-T cell counts ● b12, VRC01 , VRC03, NIH45-46G54W , PG9 , PG16 , PGT121 , Questions: PGT128 , PGT145 Only for subtype B? ● Over 20 years (1987–1991/ 1996–2000/ Does it hold for global viral population? 2006–2010) ● French ANRS PRIMO What about other time periods? and SEROCO cohorts 12 Anna Feldmann
Time Analysis over LANL Env Seqs Setup: ● ~36.000 Env Seqs from LANL, different subtypes 1981 1987 1992 1996 2000 2006 2010 2013 ● Time covered: 1981-2013 Paper vs our time ART cocktail partitioning LPV/r before ART (NRTIs) Maraviroc/ Raltegravir ● Predicted IC50 value HAART cocktail with PIs using support vector ART monotherapy regression models Doria-Rose et al. , J.Virol. 2009; Mouquet et al. , PNAS 2012; Huang et al. , Nature 2014 13 Anna Feldmann
Time Analysis over LANL Env Seqs ● Continuous trend towards resistance NIH45-46 PG9 for all antibodies but PG9 and PG16 (Bonferroni correction threshold 0.05/22=~0.002, umbrella test) ) ● Considering non-B subtype (vs B): 0 5 C I similar trend, but PGT121, PGT128 ( g o l not significant anymore (Bonferroni correction threshold 0.05/22=~0.002, umbrella test) ● Over the whole available time period Time periods 14 Anna Feldmann
Conclusion ● Well performing classification models for HIV-1 resistance to bNAbs ● Reliable classifiers identifying potential binding site residues ● Visualization of data relationships and motif logos improve biological understanding of the classification result ● Regression models provide more fine-grained predictions ● Useful as recommendation device for bNAb combination therapy ● Extendable to new HIV-1 bNAbs or HCV bNAbs 15 Anna Feldmann
Thanks to ... Max Planck Institute for Informatics, Saarbrücken ● Thomas Lengauer ● Nico Pfeifer ● Alejandro Pironti ● Nora Speicher and Rolf Kaiser 16 Anna Feldmann
Thanks to ... Max Planck Institute for Informatics, Saarbrücken ● Thomas Lengauer ● Nico Pfeifer ● Alejandro Pironti you for listening! … you for listening. ● Nora Speicher Questions? and Rolf Kaiser 16 Anna Feldmann
Recommend
More recommend
