Towards a New Therapy-Success Definition for the EuResist Prediction - - PowerPoint PPT Presentation

Towards a New Therapy-Success Definition for the EuResist Prediction Engine

Alejandro Pironti

Computational Biology and Applied Algorithmics Max-Planck-Institut für Informatik

Motivation

• Therapy-success rates in resource-rich settings have steadily

increased over time

• Reason: Improved potency, drug-resistance profile, and tolerability
• f therapies
• Prediction of initial therapy response has become less challenging
• Dichotomization of therapies into successes and failures based on a

single (or a few) viral-load measurements problematic

2

Figure: Kaplan-Meier probabilities of therapy continuation stratified by therapy start year. EuResist therapies with baseline genotypes.

Challenges

3

Turbulent first therapy years

Figure: Example therapy viral-load trajectory from the EuResist Database

Looking at first therapy year only can lead to mislabeling

• f therapies. Predict whether the viral load will become

suppressed at all?

Challenges

4

Lack of adherence

Figure: Example therapy viral-load trajectory from the EuResist Database

Using only one (or a couple) of viral-load measurements not sufficiently discriminative of good and bad therapies. Predict time to viral-load rebound?

Challenges

5

Transient viral-load rebound

Figure: Example therapy viral-load trajectory from the EuResist Database

Time to first viral-load rebound is inadequate in some cases. Predict area under viral-load curve?

Challenges

6

Differential viral-load monitoring intervals

Figure: Example viral load trajectories for a Therapy from the EuResist Database

Due to differential viral-load monitoring intervals, area under viral-load trajectory is not comparable among therapies.

Proposed Solution

• Use quantitative measure of

therapy success

7

Figure: Example therapy viral-load trajectory

Proposed Solution

• Use quantitative measure of

therapy success

• Organize therapy viral-load

trajectory into semesters

8

Figure: Example therapy viral-load trajectory

Proposed Solution

• Use quantitative measure of

therapy success

• Organize therapy viral-load

trajectory into semesters

• Use mean viral load for each

semester

9

Figure: Example therapy viral-load trajectory

Proposed Solution

• Use quantitative measure of

therapy success

• Organize therapy viral-load

trajectory into semesters

• Use mean viral load for each

semester

• Count number of semesters

with mean viral load under some threshold (aviremic semesters)

• Working threshold:

125 copies per milliliter of blood serum

• To the right: 9 aviremic

semesters

10

Figure: Example therapy viral-load trajectory

Right Censoring of the Number of Aviremic Semesters

• Intent-to-treat (ITT) criterion

labels a number of aviremic semesters as censored if: – Therapy is still ongoing – There are semesters without viral-load measurements

11

• On-treatment (OT) criterion

labels a number of aviremic semesters as censored if: – Therapy is still ongoing – There are semesters without viral-load measurements – Therapy was interrupted while viral load was suppressed

Methods for performing regression with right-censored data are required

Proof-of-Concept Analysis

• Dataset:

– 11,394 therapies from EuResist database – 2,211 therapy-failure genotypes

• Therapy-failure genotypes are

included as therapies with uncensored zero aviremic semesters

• Protease and reverse-

transcriptase genotypes

• Integrase-inhibitor-use history

as a surrogate for integrase genotype

12

Figure: Histogram of the numbers of aviremic semesters for dataset

ITT criterion: 4,529 censored (33%) OT criterion: 8,658 censored (64%)

Proof-of-Concept Analysis

• Use following features for

predicting the log number of aviremic semesters: – Drug compounds – Protease and reverse- transcriptase genotypes – Integrase-inhibitor-use history – Drug-exposure scores for protease and reverse- transcriptase inhibitors

• Separate analyses for ITT and

OT criteria

13

• Regression method:

– Linear support vector machines for right- censored data

• Performance assessment:

– Patient-wise disjoint 10-fold cross validation

Proof-of-Concept Analysis: Results

Performance measure: Harrell’s concordance index

• Inspired on receiver-operating-

characteristic curves

• Compares pairs of pairs of

measured and predicted numbers of aviremic semesters

• Two pairs are either unusable,

concordant, or discordant m1 < m2 ⇒ p1 < p2 ? (mi, pi): pair of measured and predicted number of aviremic semesters

14

OT criterion:

• 833,619 (91,411) usable pairs
• 645,102 (64,996) concordant

pairs

• Concordance probability: 0.77

(0.01) ITT criterion:

• 1,084,045 (32951) usable pairs
• 782,716 (31,806) concordant

pairs

• Concordance probability: 0.72

(0.01)

Numbers averaged (standard deviation) across cross-validation folds

Outlook

• What do you think of the number of aviremic semesters as a

measure for therapy success?

• Perform detailed analysis including test set
• Include further variables as predictors:

– Gender – HIV transmission mode – Baseline viral load – Baseline CD4 – Integrase baseline genotype

15

Alejandro Pironti

Acknowledgements

Max-Planck-Institut für Informatik

Thomas Lengauer Nico Pfeifer Joachim Büch Prabhav Kalaghatgi Joachim Büch

University of Düsseldorf

Björn Jensen

University of Cologne

Rolf Kaiser Mark Oette Saleta Sierra Aragon Elena Knops Maria Neumann-Fraune Eugen Schülter Eva Heger Claudia Müller Nadine Lübcke

Medizinisches Labor Berg

Hauke Walter Martin Obermeier

Institut für Immunologie und Genetik Kaiserslautern

Martin Däumer Alexander Thielen Berhard Thiele

EuResist

Francesca Incardona Maurizzio Zazzi Mattia Prosperi

Robert-Koch-Institut

Claudia Kücherer