Dose selection in solid tumor trials:�Making use of all available tumor data.

PhUSE Boston – June 20, 2013

Andrew Stein

andrew.stein@novartis.com

​

​

Dose Response

​

​

Response Criteria for

Solid Tumors (RECIST)

​

Overview

• Oncology Phase III trials have a >40% failure rate¹
• When choosing the dose for Phase III, drug development teams do not take full advantage of the moderately rich tumor data that is available.
• We are developing methodologies for dose selection that take into account all available data, which could potentially lead to reduced failure rates, more efficient trials, and better characterization of the benefit-risk ratio.

​

​

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

2

1. DiMasi JA and Grabowski, J Clin. Oncol., 25, 209 (2007)

Background: Solid tumors are imaged using 3d Xray images, i.e. computed tomography (CT)

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

3

Background: RECIST¹ lesion types

• Target Lesions
• 1-2 “representative” lesions per organ
• Lesions must be measurable
• Longest diameter recorded and summed (SLD²)

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

4

• Nontarget Lesions
• When >1-2 lesions exist per organ
• “Unmeasurable” lesions
• Categorical response recorded

• New Lesions
• Detectable only after the start of therapy
• Binary response (yes/no) recorded

1. RECIST = Response Evaluation Criteria In Solid Tumors

2. Sum of the Longest Diameters

Background: Patient response is summarized by single value measurements; some information is lost

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

5

1. RECIST = Response Evaluation Criteria In Solid Tumors SD = Stable Disease

2. SLD = Sum of Longest Diameters of target lesions PR = Partial Response

PD = Progressive Disease

10 mg

5 mg

0 mg

6 cm

4 cm

2 cm

• Summary statistics based on single value measurements are then used to compare trial arms

Background: How can we use all available tumor data to better select the dose of new drugs?

• Standard efficacy endpoints do not incorporate all available data.
• Median Overall Survival (OS)
• Median Time to Progression (TTP)
• Best Overall Response
• Missing data includes
• Phase I trials can have many dose interruptions and dose escalations, but individual dosing history is ignored.
• Heterogeneous lesion response is ultimately summarized by only 1-2 binary variables (Progression: yes/no, Response yes/no).

​

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

6

We are developing new methods for modeling tumor data for identifying the most efficacious dose.

1. Using longitudinal tumor data, we characterize a difference between two doses, which was not possible using standard statistical approaches.
2. Taking into account multivariate lesion assessment is essential for predicting overall survival.
3. We have developed multivariate, longitudinal models to capture the full dose-efficacy dataset of all lesion types.

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

7

Background Example: clinical trial comparing everolimus to placebo in renal cell cancer

• RAD001C2240:�Phase III pivotal trial comparing everolimus 10mg to placebo in 2^nd/3^rd line renal cell cancer patients.
• Progression free survival was the �primary end point.
• Results demonstrated superiority of everolimus 10mg over placebo.

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

8

Placebo (N=139);

Median PFS = 1.9 mo

% of pts. who have not progressed

Everolimus (N=277)

Median PFS = 4.9 mo

Time after randomization (months)

Comparison of trial arms:

Log rank p-value < 0.001

Background: The RAD001C2240 trial with no 5mg arm contains information about the 5mg response.

• Many patients (N=98) were dose- reduced from 10mg to 5mg due to adverse events
• Is 10mg superior to 5mg? Should patients who are dose-reduced to 5mg be re-challenged with 10mg?
• Standard statistical tools cannot answer these questions because there was no prospective 5mg arm in the trial.
• Model-based tools can address these questions by relating the daily dosing history to tumor response over time.

​

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

9

Observed solid tumor data from an individual patient: In this patient, target lesions appear to shrink under 10mg, but grow under 5mg.

10mg

5mg

Dose reduction due to an adverse event

Individual Patient

#1 We first focus on using the longitudinal, continuous, target lesion data to compare doses.

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

10

1. RECIST = Response Evaluation Criteria In Solid Tumors SD = Stable Disease

2. SLD = Sum of Longest Diameters of target lesions PR = Partial Response

PD = Progressive Disease

10 mg

5 mg

0 mg

6 cm

4 cm

2 cm

Methods: model based approach for describing individual patient data.

• Modeling approaches describe the daily change in tumor size for each individual patient.

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

11

• Using each individual’s dosing history and longitudinal tumor size data, we fit all patients from the trial using a nonlinear mixed effect model.

Results: The model well describes representative patient data. 10mg was more effective than 5mg.

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

12

1. SLD = Sum of Longest Diameters of target lesions

Placebo patients that cross over to Everolimus

Patients on 10mg

Patients reduced from 10mg to 5mg

10mg

5mg

0mg

10mg

5mg

0mg

10mg

5mg

0mg

Conclusions: Modeling approaches can describe response to 5mg, even without a 5mg arm

• Analysis supports 10 mg as �starting dose.
• Analysis demonstrates that 5 mg�is superior to placebo when dose reductions are necessary.
• Analysis suggests that patients who are dose-reduced to 5 mg may benefit from a re-challenge at 10 mg.
• Results were published in BMC Cancer¹.

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

13

Stein et al., BMC Cancer, 12:311, 2012

Error bars denote intersubject-variability (±35%)

#2: We next focus on predicting survival using the first post-baseline multivariate assessment

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

14

1. RECIST = Response Evaluation Criteria In Solid Tumors SD = Stable Disease

2. SLD = Sum of Longest Diameters of target lesions PR = Partial Response

PD = Progressive Disease

10 mg

5 mg

0 mg

6 cm

4 cm

2 cm

Methods: Cox regression is used to establish the effect of each tumor measure on survival

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

15

Survival

[Days]

=

Change in Target Lesion Size

(Δy = %)

Progression of Nontarget Lesion

(Non = 0/1)

Appearance of New Lesion

(New = 0/1)

• {a₁, a₂, a₃} are Cox coefficients linking each patient measurement to survival.

Survival Hazard

~

a₁*Δy

+

+

a₂*Non

a₃*New

~

• Only patients on the everolimus arm are analyzed.
• Response at first assessment (month 2) is correlated with overall survival using Cox regression.

Results: Change in the sum of the longest diameters (SLD) was not predictive of survival.

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

16

SLD Reduction

SLD Growth

SLD = Sum of Longest Diameters of target lesions

Target Lesion

Sum of Longest Diameters

(SLD)

Results: Nontarget lesion progression at month 2 was predictive of survival

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

17

SLD Growth

SLD Reduction

Target Lesion SLD

Nontarget Lesion

Progression

No Progression

Progression

p = 0.0001

SLD = Sum of Longest Diameters of target lesions

Results: New lesion appearance at month 2 was predictive of survival

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

18

SLD Growth

SLD Reduction

Target Lesion SLD

New Lesion

Appearance

p = 0.0001

No New Lesions

New Lesion

SLD = Sum of Longest Diameters of target lesions

Is there another way we can look at target lesion data that would make it predictive?

• Proposal: look at the “worst lesion” instead of the sum to account for known heterogeneity in tumor response¹

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

19

SLD = Sum of longest diameters of target lesions

Lesion #1

Lesion #3

Resistant

1. Gerlinger et al., NEJM, 366, 883 (2012).

​

Lesion #2

Results: Looking at growth/shrinkage of the “least responsive” lesion was predictive of survival

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

20

SLD Reduction

Target Lesion SLD

Behavior of

“worst” lesion

All lesions

shrink

At least one

lesion grows

p = 0.03

SLD Growth

SLD = Sum of Longest Diameters of target lesions

Conclusions

• This work¹, adds to a growing body of literature^2-4, demonstrating the importance of accounting for all lesion types (target, nontarget, new).
• The heterogeneity of the target lesion response may also be useful in predicting survival.
• Dose-response modeling methods should take into account the additional available data.
• There is a need to further investigate use of image processing algorithms for capturing the dynamics of all lesions types

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

21

1. Stein et al., European Urology, in press, 2013. 2. Suzuki et al., Annals of Oncology, 23.948, 2012.

3. Mietlowski et al., J Clin. Oncol., abstr 2543, 2012 4. Litiere et al., J. Clin. Oncol., 30, abstr 10602, 2012

#3: Finally, we focus on modeling the full, multivariate tumor time-course.

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

22

1. RECIST = Response Evaluation Criteria In Solid Tumors SD = Stable Disease

2. SLD = Sum of Longest Diameters of target lesions PR = Partial Response

PD = Progressive Disease

10 mg

5 mg

0 mg

6 cm

4 cm

2 cm

Methods: Can we use the same model we had for the target lesions to also describe the new lesions?

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

23

Time (days)

New Lesions

Tumor Dynamics

Lesion diameter (cm)

dose

no

Continuous Data

Binary Data

Probability

of Detection

Probability of

Detection

Prob. of Appearance

Methods: The new lesion detection function can be estimated from literature data.

• Imaging resolution has improved since 1999, so it is assumed that lesions that were 0.5 cm in diameter were detected with 50% probability.
• Model also requires of an assumption of the initial lesion size (we choose 0.1cm)

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

24

1. Lorentz et al, Chest, 115; 720-724 (1999)

Methods: model based approach for describing individual patient data.

• We model all three lesions types for each individual patient using the target lesion model.

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

25

• Target lesions: fit model to sum of longest diameters (as before)
• New lesions: use the “detection function” to transform the diameter (y) to a probability of detection between 0 and 1.
• Nontarget lesions: construct a similar “progression function” to transform the diameter (y) to a probability of progression.

​

Results: The model well describes representative patient data.

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

26

Target

Lesion

Change

(%)

Nontarget

Lesion

Progression

(yes/no)

New

Lesion

Appearance

(yes/no)

Patient 1

Results: Model can be qualified using the primary endpoint (Progression Free Survival)

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

27

Stein et al., J Clin Oncol, Abs#4602, 2011 (ASCO Poster)

Conclusions: 10 mg everolimus was superior to placebo, but 5- and 10-mg could not be differentiated

• For nontarget and new lesions, 5mg and 10mg could not be differentiated.
• There may be a true lack of difference between the two doses in the nontarget and new lesions.
• Poor resolution of categorical lesion data may make the dose-response relationship more difficult to detect.
• Improved methodologies for characterizing nontarget and new lesions are warranted.

​

​

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

28

Nontarget

Lesion

Probability of

Progression

New

Lesion

Probability of

Appearance

Next Steps

• Add PK to model (has been done in other indications)
• Improve resolution of the tumor data
• Model each individual lesion instead of the sum
• Integrate information from multiple readers
• Apply automated algorithms for assessing the size of all available lesions (not just target lesions).
• Characterize therapeutic window using safety data.

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

29

Recap: We have developed new methods for modeling tumor data for identifying the most efficacious dose.

1. Using longitudinal target lesion data, we characterize the difference in efficacy between two doses, which was not possible using standard statistical tools.
2. Taking into account multivariate lesion assessment is essential for predicting overall survival.
3. We have developed multivariate, longitudinal models to capture the full dose-efficacy dataset of all lesion types.
4. Next Steps: incorporate PK and explore the use of richer tumor data to capture additional efficacy information for dose selection.

​

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

30

Acknowledgements

• M&S Oncology Cluster
• Gabriel Helmlinger
• Celine Sarr
• Jerry Nedelman
• Bill Sallas
• Wenping Wang
• Ovidiu Chiparus
• Alison Carter
• Varun Goel
• Shu Yang
• Marina Savelieva
• Olesya Melnichenko
• Kyle Lemoi (Intern)
• Daniel Lusk (Intern)

• BDM (Statistics)
• Bill Mietlowski
• Norbert Hollaender
• Clinical Team
• Dennis Kim
• Sandra Chica
• MPI Workshop (NJIT)
• Linda Cummings
• Richard Moore
• Tom Witelski + team

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

31

BACKUPS

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

32

Everolimus showed similar effect on target lesions among patients with and without dose reductions.

• Because there was not a randomized 5mg arm, the patients who are dose-reduced to 5mg represent a biased sample of patients who experience an adverse vent.
• We compare the 10mg drug effect for patients with and without dose-reductions to see if tumor shrinkage is different across the different subgroups.
• The right hand plot demonstrates that the effect of 10mg on tumor size is similar for both subgroups of patients, irrespective of whether the patient had an adverse effect.

| PhUSE Boston | A. Stein | June 20, 2013 | Modeling solid tumor data

33

Drug Effect in patients without (left) and with (right) dose interruptions is similar for both groups.