Mattia CF Prosperi, PhD

ahnven@yahoo.it Clinic of Infectious Diseases

Catholic University of Sacred HeartLargo F. Vito, 1 – 00168 - Rome, Italy

HIV-1 biology, treatment and resistanceEpidemiology

Surveillance of resistance trendsPhylogenetics, HIV-1 evolutionClustering of HIV-1 mutations

Intra-host analysisHIV-1 replication, natural genetic drift and selective drug pressureDifferential equation modelling

Optimising treatments with machine learningPrediction of HIV-1 co-receptor usagePrediction of in vivo HIV-1 virologic response to treatments

Genotype-based modelsTreatment history-based models

PerspectivesModelling time to viral rebound, and resistance emergenceModelling epidemic with complex networks

Usually focus on specific problemsIncidence of new infectionsTemporal trends of treatment efficacyDeterminants of virologic or immunologic failure…

Standard modelsUnivariable analysis (chi-square, t-test, rank-sum)Linear, logistic regressionSurvival analysis (Kaplan-Meier, Cox proportional hazard)Often limited when considering predictive ability of the models

Complex Network modelsHIV-1 is peculiar!

not only sexually transmitted, long asymptomatic stage, high rate of evolution, integration into host genome, no natural eradication, rapid development of drug resistance…

SIR-like models, adjusted for MANY other factors like the drug resistance emergence (Smith, Blower et al., Science, 2010)

Several epidemiological studies carried out at our instituteTemporal trends of drug resistance in Europe, considering different inhibition classes

Resistance to Drug Classes per Calendar Year

00.10.20.30.40.50.60.70.8

1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008

NRTI NNRTI PI_major

We assessed the temporal trends of resistance by fitting a linear model, adjusting for potential confounders such as age, gender, mode of HIV transmission, introduction of new drugs…

Several scales of analysisSubtype evolutionTransmission events (and drug resistance transmission)

Not always the phylogenetic reconstruction is able to trace the epidemiological evidence, due to sparse sampling (intra-host evolution does not proceed at a constant rate)

126@

J 0

64@37 0

0.01

Note: phylogenetic analysis constructs a hierachy of sequences that represents the evolution from an hypothetical ancestor. Several techniques are available, from distance-based clustering to maximum likelihood, to bayesian clustering

Probability of DR transmissionintrinsic efficiencyviral loadfrequency and modality of exposition

≈12,000 group M subtype B HIV-1 polymerase sequences collected from Italian ARCA DBMaximum-likelihood parallel phylogeny, computationally intensiveThere is no methodology for automatic cluster identificationNew technique for partitioning a phylogenetic tree

Depth first visit with constraints on node reliability and intra/inter-cluster patristic distance distributionsValidated on a set of known transmission events

11,541 sequences9,855 patients

Tree rooted on outgroup subtype J (ancient differentiation)3D-hyperbolic geometry view

Fractal dimension ≈1.6Sustained level of differentiation

A more recent calendar year of sequencingPatients from central-Italy vs those from northern- or southern-ItalyHeterosexuals and homosexuals vs injecting drug usersYounger patientsPatients with more recent infections (<4 years vs those infected >=14 years)Presence of resistance mutations in the protease gene

Similar to phylogenetic analysis, but performed on transposed sequence alignmentsUseful to find associations among mutations under particular drug pressures

Basis for structural analysis

(Prosperi et al, ARHR 2009)

Description of patient’sviral dynamicsimmune response

Suitable for control theory (if equations could be treated analytically)Difficulties in dealing with prediction of therapy outcomes (see the constant η values, indeed they should change!)Difficulties modelling resistance outbreak (stochasticity, multi-strain models)

System of differential equations

New approach to account both for treatment New approach to account both for treatment administration and viral evolutionadministration and viral evolution

Stochastic model for viral natural evolution in absence Stochastic model for viral natural evolution in absence of treatmentof treatmentCalculation of instant resistance at different time Calculation of instant resistance at different time points using in-vitro known drug susceptibilitypoints using in-vitro known drug susceptibilityUsage of time-varying resistance [Usage of time-varying resistance [η= η(t)] in the ] in the differential equations and approximation with numerical differential equations and approximation with numerical solutionssolutions

Calculation of number of virions in the next replicationCalculation of number of virions in the next replicationSelection of resistant strains with roulette wheel Selection of resistant strains with roulette wheel procedure (from genetic algorithm theory)procedure (from genetic algorithm theory)

Different combination treatments evaluated, along with therapy sequencing policiesAlthough theoretically complex and sound, the model was not suitable for clinical practice

(Prosperi et al, Bioinformatics 2008)

We do not attempt to define an explicit modelExtensive use of machine learning

Linear and non-linear modelsFeature selectionRobust validation

In-vitro: prediction of HIV-1 co-receptor usageIn-vivo: prediction of virological response to combination antiretroviral therapy (cART)

With viral genotypic informationWithout viral genotypic information

Designed for low/middle income countries

HIV-1 can use two different co-receptors (CCR5/CXCR4)Entry inhibitors block only the CCR5 co-receptorThe model helps to decide if a patient can be given an entry inhibitor or not, given his viral sequenceAnalysis using whole envelope region and other patient’s characteristicsLogistic regression is a suitable model

Not inferior to complex non-linear modelsPerformance (with robust validation) up to

93% accuracy0.77 sensitivity0.93 AUC (Prosperi et al, ARHR 2009)

Predicting the actual viral load changes following treatment switches is a challenging task

Individual variability of immune response to infections add noise to the systemLarge number of possible therapeutic combinations leads to complex viral evolutionary pathwaysOther treatment-related factors such as pharmacokinetics and patient adherence to therapy play a crucial role in the control of virus replication and the development of resistance

We focused on fixed patient’s follow up times (n-weeks of therapy)

EuResist is a no-profit foundation (formerly EU-funded project), a consortium were hospitals, biology labs and universities cooperate

Karolinska institute, University of Siena, University of Cologne, Max Planck Institute, IBM…

It is the largest data base in the world comprising clinical, demographic and genomic data of HIV+ patients from national cohorts of Western Europe (at now Belgium, Italy, Germany, Sweden, Spain, Luxembourg)

≈34’000 patients≈500’000 CD4 and ≈400’000 HIV-RNA measurements≈100’000 antiretroviral therapies≈31’000 HIV sequences (polymerase)

Open to any kind of collaboration and data exchange

Treatment Change Episodes (TCE) with a new cARTBaseline HIV RNA load, CD4+ T cell countsBaseline HIV polymerase genotype and subtypePatient’s demographics (age, gender, ethnicity, mode of HIV transmission…)Previous drug usages (>1 year usage) for each drug class and each single drug8-weeks and 24-weeks HIV RNA responseSuccess defined as the achievement of <500 cp/ml (or >2 Log decrease from baseline at 8-weeks)

The EuResist repository was queried and generated more than 3,000 TCE that were used for training and validating a prediction engine

Statistical learning modelsLogistic Regression with higher-order interactions (LR)

AIC stepwise selectionRandom Forests (RF)

Feature importance evaluation with Strobl’s methodBayesian networks (BN)Three independent models were merged improving performance

Extra sample error estimationMultiple ten-fold cross validation (MCV)

Adjusted t-test on MCV performance distributions for model comparison

External independent test set evaluationComparison against human expertsComparison against rule-based algorithms (Stanford, Rega, ANRS)

Statistical models

Web-service

Customised cART sequencing

Patient’s

Age, gender

HIV RNA

CD4

Experienced drugs

HIV genotype

The EuResist model outperforms the whole set of state-of-the-art techniques (i.e. rule-bases) and is as good as the world’s best human expertsThe average accuracy on validation is 76%, and AUC is 0.77

Best ExpertEuResistMean ExpertWorst Expert

0 10 20 30 40 50 60 70 80 90 100

100

90

80

70

60

50

40

30

20

10

0

100-Specificity

Sen

sitiv

ity

(Prosperi et al, Antivir Ther 2009)

In high-income countries, guidelines recommend genotypic resistance testing (GRT) both before starting antiretroviral therapy (ART) and at ART failureAppropriate funding and/or facilities to perform GRTs may be not available in low-middle income countries (LMIC), leaving physicians to switch therapy based solely on the clinical/immunological conditions (sometimes even without virological monitoring)Treatment history (TH) is one of the most crucial factors to play a role in the response to a new treatment.

Other important factors are virologic and immunologic monitoring

GRT-based vs TH-based models were compared to see if there were sensible loss in performancePerformance of the model were tested in extra-EU-like scenarios

Tests on a larger set of TCE without the mandatory GRT baseline attribute were carried out

No statistically significant differences found by comparing GRT and TH models

We want to design and test a model that predicts viral load rebound over time using

Patient’s viral genotypic informationPatient’s clinical and demographic background

Suitable models: Cox regression, random survival forestsNeed to define an appropriate goodness of fitPreliminary inquire on the EuResist DB gave a considerable number of training instances

Also, we might be interested in a model that predicts drug resistance emergence

Design and test an epidemic model for HIV-1 using complex networksStart from Science paper and from other models presented in literature

New insightsCapability to handle dynamics at a regional, national and international levelEffective description of

Infection incidence over different risk group strataHomogeneous vs heterogeneous mixing?

Drug resistance trendsPrevision of trends with the introduction of new inhibition classesPrevision of HIV-1 evolution with respect to drug resistance prevalence in the treatment-naive populationAccount for transmitted drug resistance from treatment-naive and treatment-experienced patients

How much shall we go into details as concerns the intra/inter-host genetic HIV-1 evolution?