HIV Drug Resistance: Current Status/Future Direction Brendan Larder - PowerPoint PPT Presentation

HIV Drug Resistance: Current Status/Future Direction Brendan Larder PhD Chair of the RDI Scientific Core Group

Resistance testing today • Resistance testing has become routine in HIV management – especially for ‘difficult’ cases • Genotyping & phenotyping technology has considerably improved & is widely commercially available – Genotyping is much more commonly used: more rapid & cheaper than phenotyping – Commercial, FDA approved sequencing kits are available

However…. • Interpretation is the key issue – It is likely that interpretation problems limits the use of resistance testing

Phenotyping • Not technically feasible to assess multiple drug combinations – Any potential drug interactions not seen • Cut-offs are an issue - how should they be derived & used? – ‘Technical’ cut-offs - limit of assay variability – ‘Biological’ cut-offs - natural variation in virus from untreated patients – ‘Clinical’ cut-offs: categorical, based on small numbers & specific to drug context

Interpreting resistance mutations • Expert view & panels • ‘Rules-based’ algorithms from consensus • Rules driven software • Phenotype-genotype database matching • Predicting viral phenotype using artificial intelligence

Generation of simple rules • 215Y/F = AZT resistance • 184V/I = 3TC resistance • 103N = EFV resistance • 30N = NFV resistance • 50V = APV resistance

Rules have become more complex… • ABC resistance = 215Y/F + 1 or more of: 65R, 69D, 74V, 70R, 115F, 210W, 219E/Q, 184V/I + 1 or more of: 41L, 67N • 3TC resistance = 44D + 118I + 4 or more of: 41L, 67N, 69D, 70R, 210W, 215F/Y, 219E/Q • EFV resistance = 181C + any of: 100I, 101E, 179D, 230L • NFV resistance = 82A/F/T + 2 or more of: 10I/R/V, 20M/R, 36I, 54L/M/V, 71V/T • SQV resistance = 84V + 2 or more of: 10I/R/V, 20M/R, 36I, 54L/M/V,71V/T

How well do algorithms predict phenotype?

Error rates for protease inhibitors 30 25 % 20 15 10 5 0 IDV SQV RTV NFV APV Geno2pheno RetroGram Stanford Korn et al Scottsdale, 2001

Error rates for RT inhibitors 40 35 % 30 25 20 15 10 5 0 ZDV DDC DDI D4T 3TC ABC NVP DLV EFV Geno2pheno RetroGram Stanford Korn et al Scottsdale, 2001

Predicting phenotype: systematic approaches • ‘Virtual Phenotype’ – Requires a substantial genotype-phenotype database – Based on mutation pattern recognition & phenotype retrieval – Relies on analysis of pre-defined mutational clusters, e.g., 41 + 67 + 210 + 215 in RT • Neural networks – Large data-sets are needed for training & testing

Correlation of actual with virtual PT n = 500 per group (random selection) Virtual Phenotype (log 10 ) Observed Phenotype (log 10 ) (Correlation coefficients: 0.86 - 0.89)

Neural networks • Computer learning technique, where the network learns to connect complex data & identify patterns by being ‘fed’ many examples • Networks are ‘trained’ with large datasets • Can be used to relate resistance mutations to phenotype or clinical response • Neural networks are particularly useful to analyse resistance because of the many combinations of mutations that are possible

Predicting Lopinavir phenotype using neural networks 2.5 R 2 = 0.88 2 (n=1322) 1.5 1 0.5 0 -1 -0.5 0 0.5 1 1.5 2 2.5 -0.5 -1 Actual fold (log) 28-Mutation model

Predicting D4T phenotype using a 26- mutation neural network model 30 25 y = 0.6705x + 2.0149 R 2 = 0.6766 20 15 10 5 0 0 5 10 15 20 25 30 Actual fold increase

How well do rules predict clinical response?

Response to ABC at Week 4 by Stanford Rules 100 80 <400 copies 60 > 0.5 log 40 No Response 20 0 Sensitive (N=18) Contributes/Low Resistant (N=46) (N=102) Lanier et al Scottsdale, 2001

Standarised comparison of different rules (odds ratio of virological failure) (analysis = drop outs as failures) Hammer et al Scottsdale, 2001

The next logical step ……. • Develop large database(s) to correlate mutation patterns with clinical responses – Not just a correlation of genotype with phenotypic drug resistance – Addresses response to combinations of drugs

HIV Resistance-Response Database Initiative (RDI) Aim: “To improve the clinical management of HIV infection by developing & making freely accessible a large clinical database & bioinformatic techniques that define with increased precision & reliability the relationships between drug resistance & virologic response to treatment”.

RDI status • Independent not-for-profit organization • Scientific Core Group established • Small executive group running analyses • Range of large cohorts committed to support with data • Range of experts in the field pledged support – e.g. Julio Montaner, Rob Murphy, Joep Lange, Brian Gazzard, Bonaventura Clotet, Jose Gatell

RDI approach • Collecting genotype, treatment & clinical outcome data from large numbers of patients • Variety of data analysis methodologies to relate resistance to clinical response • Wide access to enable the database to be queried via the internet

Status of data analysis • Data identified from about 7,000 patients – More when additional genotyping performed • Database now: 1,000 patients from BC Centre, Italian cohort, US Military, NIAID • Power calculations performed to estimate number of required data points • Neural network models constructed & tested

Initial neural network model Mutations Neural On Network therapy BL VL VL Training Therapies

Initial neural network model Mutations Neural Predicted Network BL VL VL change Training Therapies

Initial neural network model • Input variables – 20 PI & 29 RT codon positions (based on prevalence) - 12 drugs - Therapy duration • Output variable – Viral load change at on-therapy time points (up to 6 months) Wang et al Seville, 2002

Example result of NN model Training set (n=639) 2 = 0.85 4 P re dic te d vira l loa d R 2 c ha nge 0 -4 -3 -2 -1 0 1 2 3 -2 -4 Actual viral load change Wang et al Seville, 2002

Example result of NN model Validation set (n=63) 3 R 2 = 0.55 Predicted viral load 2 1 change 0 -4 -3 -2 -1 -1 0 1 2 3 -2 -3 -4 Actual viral load change Wang et al Seville, 2002

Predicting VL trajectory NN Prediction: 75%±1.8% correct (86/115) Viral Load Time Wang et al Seville, 2002

Predicting VL failure (Dichotomous Model, <, >400 copies) NN Prediction: 82%±1.6% correct Viral Load <400 Time Wang et al Seville, 2002

Example of predicting response • Baseline genotype for a ‘virtual patient’ - RT: 41, 67, 118, 210, 215 - PI: 10, 46, 82, 90 • Alternative therapy regimens a. D4T, ddI, Kaletra b. D4T, ddI, indinavir c. AZT, 3TC, Kaletra d. AZT, 3TC, indinavir Wang et al Seville, 2002

Example of predicting response 3 2 Viral Load (log 10 ) D4T/ddI/IDV 1 0 AZT/3TC/IDV D4T/ddI/Kal -1 AZT/3TC/Kal -2 Baseline week 8 week 16 Wang et al Seville, 2002

Summary • Resistance rules & algorithms have increased in complexity – BUT these are not systematically designed to predict clinical response • Methodologies such as neural networks can enhance the accuracy of predictions • RDI is establishing relationships between genotype & virological response to combination therapy via analysis of a large clinical dataset – Approaches of this nature are likely to improve the accuracy of predicting outcome from genotype