YSPH Biostatistics Seminar: "Innovations in Immune-Oncology Early-Phase Trial Designs: Theory, Practice and Next Steps"

November 08, 2021

Information

Codruta Chiuzan, PhD, Associate Professor, Center for Personalized Health, Institute of Health System Science, The Feinstein Institutes for Medical Research, Northwell Health

November 2, 2021

ID7137

To CiteDCA Citation Guide

00:00<v Wei>Okay, hello everyone.</v>
00:02Today, we are very fortunate to have Dr. Codruta Chiuzan
00:11as our speaker.
00:14Dr. Chiuzan is Associated Professor,
00:17Institute of Health System Science
00:20at Northwell Health New York.
00:23So before that, she was an Assistant Professor
00:26in the Department of Biostatistics
00:29at Mailman School of Public Health Columbia University.
00:33In her research area focus on earning phase
00:37clinical trial designs and an average aging real-world
00:40evidence to prove all the cons and increased diversity
00:44of population in clinical trials.
00:47Now, she is in receipt of Junior Faculty Research Award
00:52and the Columbia Public Health Innovation Award
00:55from the Mailman School of Public Health.
00:58So, Dr. Chiuzan, has a very strong record or mentoring,
01:03both master's students, PhD students and clinical fatal.
01:08She is an active committee member
01:10of JSM, Diversity Mentoring Program
01:13and she had held leadership positions
01:15as the President of the American Statistical Association
01:19and New York City Metropolitan Area
01:22and she as the Chair over the Student Scholars Committee
01:26at the Society of Clinical Trials.
01:29So, welcome Dr. Chiuzan and time is yours.
01:35<v ->Thank you so much Wei, for the invitation,</v>
01:39it's a pleasure, I'm going to share my...
01:47Hello.
01:50Hello, I hear some echo in the background.
01:56<v Student>Is anybody (indistinct).</v>
02:10<v ->The host, can you disable the screen sharing</v>
02:13so I can share the slides?
02:25Oh, perfect.
02:40Okay, can everybody see the screen, the full screen?
02:57Okay.
03:01All right, so it's a pleasure to be with you
03:07even virtually and I'm glad to see so many people in person
03:13with academic semester ongoing.
03:20Today I'm going to talk about one area of my research
03:25and that is early phase designs for immunotherapies
03:33or cancer immunotherapies and I will take you
03:39through a journey, through a story by giving some examples
03:43and explanation of what are these cancer immunotherapies,
03:47what are the promises, what are the challenges
03:50and how do they actually reflect in the early phase designs?
03:57Then I will talk about current models
04:01and a model that we developed on has been implemented
04:06and has been implemented in an R package and the Shiny app
04:12and I will conclude with a practical demonstration.
04:18Please, if anybody has any questions at any point,
04:24please feel free to raise a hand or just ask,
04:27I like us to have an interactive session
04:32and to have a continuous dialogue if you're able please.
04:39So what is immunotherapy?
04:42The New York times called it a long awaited reality
04:45because immunotherapy has been developed
04:50since the early 1900s, actually by a New York surgeon
04:56that saw that in cancer patients that develop flu,
05:03had a better anti-cancer response.
05:07So immunotherapy works on a different paradigm
05:11compared to cytotoxic agents and by cytotoxic agents,
05:15I mean, chemotherapy or radiations.
05:19So immunotherapy boosts or leverages the body's own immune
05:22system to fight cancer, to recognize it, to attack it
05:26and ultimately to kill the cancer cells.
05:29The three Rs of cancer immunotherapies
05:32are reverse tolerance, rejuvenate the immune system
05:39and restore the internal environment homeostasis.
05:46So, you'll probably hear more and more updates
05:51and FDA approvals for cancer immunotherapies.
05:55Between 2017 and 2020, over 65% increase has been seen
06:01in the number of immunotherapies and these immunotherapies,
06:05most of them have been approved
06:06for immune checkpoint inhibitors, Ipilimumab, Nivolumab
06:12and Pembrolizumab that works by incubator thing,
06:17the relationship, the association between the PD-1
06:20and the PD-L1 receptors, but the largest growth
06:27has been seen for cell therapies.
06:31And what are these cell therapies?
06:34The most frequent and the most study one is called T-cells,
06:40so far, which will be approvals.
06:42So that these old therapies use the T-cells in the body
06:48to fight cancer and then you do that by first,
06:54taking blood from the patient, isolating the T-cells
06:58in the lab in the Petri dish and genetically modify
07:02these T-cells to display a specific receptor
07:07that then is introduced and after that when the cells
07:14are being introduced into the body, this T-cell receptor
07:17will bind to specific antigen present on the cancer cells
07:22and trigger an anti-tumor reaction.
07:26So these are the T-cells, these are the cell therapies
07:33that are being studied and they're very promising
07:36in terms of prolonging overall survival
07:39and also in lowering the toxicity,
07:43killing cancer without killing the patient.
07:48Last year update from the Cancer Research Institute
07:52has shown that, as I said, that the most promising field
07:57in cancer therapies are,
07:59has been seen in this T-cell therapies,
08:02most of them for a solid cancer, nonsmall,
08:07renal cancer, colorectal cancer,
08:10but it's moving into the non-solid cancer as well.
08:17So, Hype versus Hope, I entitled this slide,
08:22immunotherapy, is it the holy grail
08:27is the answer to all cancer therapies, yes or no,
08:32because of course it comes with some challenges.
08:36Some of the challenges are immunotherapy sometimes
08:41can trigger delayed responses,
08:44meaning that the treatment has to continue
08:46even if initial response has not been seen,
08:50there's been cases of hyper-progression
08:53where the cancer tumor seen this rapid growth
08:57in the early stages and that can be the problematic
09:04on diminished overall survival.
09:07And most importantly, we're not really sure exactly
09:11what to measure and how to incorporate end points
09:17into all phases of drug development from early phase,
09:22phase one and two, where we're looking at identifying
09:24the optimal dose to later phases.
09:29So, because it's still under development,
09:33there is a lack of biomarkers to predict the responders
09:37versus not the responders and the difficult to correlate
09:41these immunological biomarkers with outcomes,
09:44clinical outcomes, overall survival response progression,
09:48free survival compared to cytotoxic agents, chemotherapy,
09:52immunotherapies have different toxicity profiles,
09:56meaning they can have lower toxicity
09:59and also different grades and different profiles,
10:06is called by, also called as immune-related adverse event.
10:13So if you're familiar with drug development phases,
10:16as you know that usually it will start with early phase.
10:20Phase one, identifying the maximum tolerated dose
10:24to be carried forward then for establishing efficacy
10:28and in later phases.
10:32In the old paradigm, so the objective was to find the MTD
10:37and the MTD was mainly based on toxicity as binary,
10:42yes or no DLT, the patient has after receiving treatment,
10:48we quantify the number of those limiting toxicities,
10:51unacceptable toxicity within a certain interval.
10:55However, the new immunotherapies have,
10:59as I mentioned before,
11:00they have different toxicity profiles,
11:02so this old paradigm of finding the MTD,
11:07does longer stint, there are a lot of trials
11:10where the dose escalation moved quickly
11:14to the maximum dose level and MTD.
11:17There were no DLTs, the MTD was not identified
11:21and most importantly, toxicity and efficacy
11:26might not necessarily be those dependents.
11:28So you might be able to find a safe dose,
11:31but that might not necessarily be the most promising one
11:36in terms of efficacy.
11:37In many cases, we actually see this plateau trend
11:41where after a certain level,
11:43the efficacy levels out plateaus
11:47and we don't see any effect.
11:52Another challenge, so in this context of different toxicity,
11:58is different levels of toxicity incorporation of efficacy
12:00into the dose finding process, we need to reconsider
12:06the definition and think of more in terms of identifying
12:11the optimal biological dose versus the MTD,
12:15a dose that is acceptable in terms of toxicity,
12:18but also this place, a good efficacy profile.
12:25So in terms of methodology, again in early-phase,
12:33research has been dominated in the past decades
12:38by algorithmic designs by algorithmic,
12:41I mean the three plus three,
12:42which is definitely not preferred by,
12:49and actually strongly disapproved by statisticians
12:55and even until 2014, when we did the last review
13:02of early-phase methodology, we saw that over 90%
13:05of this trials have implemented a rule-based design.
13:10Rule-based working only on toxicity with absolutely
13:14no statistical background.
13:20From 2012, we saw more than 60% of the trials
13:25that the tested targeted or immunotherapies
13:29and only 7.6 actually used a model-based design.
13:34So what do I mean by model-based design?
13:36Well, we criticize in three plus three,
13:39but are there alternatives actually several.
13:45One alternative that addresses the matter
13:50of late onset toxicities that is usually seen
13:54in immunotherapies, meaning you see DLTs on toxicity
13:58outside of DLT window, that is usually 28 days
14:04so longer toxicities, for that we have the time to event
14:09continual reassessment method that was proposed
14:12by Jenga Chapelle in 2000.
14:16The problem with multiple toxicities
14:19across different varying grades and moving away
14:22from the binary DLT has been tackled by Ezzalfani and others
14:27by using, by incorporating these types,
14:30different toxicity types and different grades
14:33into the total toxicity score,
14:35which is a quasi continuous measure.
14:40As I mentioned for immunotherapies,
14:43it makes more sense to incorporate both toxicity
14:46and efficacy and for that we have models that look at,
14:53that incorporate both toxicity and efficacy
14:56and these are the F-stocks designs method
15:00or the bivariate continual reassessment method.
15:05And more recently, other measures have been looked
15:10at in immunotherapies and these are the PK,
15:14the pharmacokinetics or the pharmacodynamics
15:17and these have been incorporated by, for example,
15:20Ursino, in a patient design proposed in 2017.
15:25So this is to present the status quo
15:29of what's being proposed out there, what we are suggesting
15:33is also a design that is specific
15:38or for immunotherapy trials and this was published in 2018
15:44and since then we have added a different measure
15:47of toxicity, we have implemented it into an R package
15:51that is on a available on cram, iAdapt
15:55and also can be tried using charmia.
15:59So this design for immunotherapies uses both toxicity
16:04and efficacy to identify the optimal dose.
16:09Optimal dose meaning unacceptable dose
16:11with promising efficacy profile.
16:16The design is unique in the sense it can incorporate
16:19both binary or quasi continuous toxicity scores,
16:23and it's looking at the continuous efficacy outcomes.
16:26Most of the designs that I mentioned before are using
16:29either binary or ordinal efficacy.
16:32In this one we're looking at continuous outcomes,
16:35such as T-cell persistence at followup compared to baseline
16:39why the cell persistence, as I mentioned before,
16:43well, about this engineered T-cells
16:46when they are being put into the body,
16:48they maintain the steer soul store,
16:51then the genetic information and the trigger
16:58and tumor response and it's been shown,
17:02there's some studies shown that the number of T-cells
17:06that are still present, still survive in the blood
17:10at one or two months after being reinfused tends to predict
17:17response on the overall survival on the long-term.
17:22The design has, does not impose any monotonicity assumption
17:26in terms of those efficacy relationship and does not account
17:30for dependence between toxicity and efficacy.
17:34So now let's take a look at the two,
17:37the difference between incorporating toxicity only
17:40and looking at efficacy also.
17:43So the cartoon on the left shows the dose toxicity
17:47relationship or five dose level.
17:50So in this graph, let's say we have five dose levels
17:53and we have a threshold of unacceptable toxicity set at 40%.
17:58So based on this graph, we have about four dose levels
18:04that are below the threshold, one dose level that is above.
18:11So if we have 40% toxicity threshold, dose number four
18:18would be identified as the MTD, the maximum tolerated dose.
18:22However, if we are to look also at efficacy
18:27and in this case, the dose efficacy has this umbrella,
18:30this non-monitoring trend.
18:32We will see that by looking at the MTD,
18:35we would totally miss the optimal dose
18:39because dose number four has actually a lower efficacy
18:46as compared to dose number three.
18:49So this is to pretty much justify the need to incorporate
18:53both toxicity and efficacy into the dose finding process.
18:59So the design has two stages.
19:01In stage one, we're establishing the safety profile
19:05at each dose, after we establish the safety profile,
19:09the acceptable doses are carried to stage number two,
19:13where we using efficacy driven randomization
19:15to allocate patients to acceptable doses
19:19that emphasis towards more promising efficacious ones.
19:25So, now let's take a look at stage one,
19:27establishing the safety profile.
19:29We have a number of pre-specified dose levels
19:33and we start by defining the set of hypothesis,
19:37where hypothesis one represents the unacceptable DLT rate
19:42and hypothesis two, represent unacceptable DLT rate.
19:46DLT, meaning the Dose Limiting Toxicity.
19:53So the quantification of this evidence
19:59in favor of hypothesis one or hypothesis two,
20:04is done by the likelihood ratio V, the evidential paradigm.
20:10So to give you a little bit of a background,
20:15in statistics there pretty much three school of thoughts,
20:18we have the frequencies approach based on Pearson,
20:22you'll have the patient school of thought,
20:24and then you have the evidential paradigm.
20:27The evidential paradigm and the frequent tests
20:30are somehow similar, but the difference between the two
20:35is the evidential paradigm based on the law of likelihood
20:39the couples, the strength of evidence from uncertainty.
20:43So the strength of the evidence is quantified
20:45by the likelihood ratio and our certainty
20:48is quantified by the probability of misleading evidence
20:54and the probability of observing weak or strong evidence
20:59in favor of the other two.
21:03Yeah, in comparison the frequent is the approach,
21:05that's not the couple, the strength of evidence
21:09and from uncertainty.
21:12So evidential paradigm used to establish acceptability
21:20in stage number one.
21:21And how do we do that?
21:22Let's say we have a certain number of levels, each show,
21:27and we treat cohorts of size M patients
21:32to each of these dose levels based on toxicity information,
21:38we calculate the likelihood ratio and evaluate evidence
21:44as one of the three, either we have strong evidence
21:48in favor of hypothesis two,
21:51declaring that the dose is acceptable.
21:54Either we have strong evidence in favor of H1,
21:57it's unacceptable or we conclude the weak evidence
22:02that doesn't support either of the hypothesis,
22:05the likelihood ratio is compared to a threshold, okay?
22:10So how do we, let's take a look at an example
22:13to see how we set the hypothesis
22:15and how we started this threshold, okay?
22:18So let's say that we have hypothesis one 40%,
22:23this is unacceptable, the DLT rate toxic
22:27and hypothesis two 15%, that is an acceptable DLT rate
22:34and we want, we evaluate each dose
22:37based on these two hypothesis.
22:40And for that one, in this case, we use a k threshold
22:46equal to two, so there's been a lot of literature
22:51written on this evidential paradigm and the k thresholds
22:56can vary, we can take values from two, four, eight
23:02all the way to 32, depending on the sample size,
23:05the bigger the sample size, the bigger the k thresholds.
23:11Because in phase one, we tend to deal with limited
23:15sample sizes, 30 maybe all the way to 50 number of patients,
23:22of k threshold or two or four seems to be sufficient
23:28to be able to quantify the strength of evidence.
23:31So I teach dose levels based on cohorts of three patients,
23:36we compare the likelihood ratio and if the likelihood ratio
23:40is greater than one over k in this case two,
23:44we decide that the dose is acceptable and safe
23:46and it will be carried forward to station number two.
23:50Otherwise, the dose is considered unacceptably toxic
23:55and it's being discarded and will not be considered
23:59for further evaluation.
24:01So in this case, let's say that we have two or more
24:06the maximum doses in stage one, we continue to stage two
24:11to employ an adaptive randomization.
24:16So in stage two, we use a linear model
24:19to calculate the randomization probabilities
24:21based on efficacy and in this case we use indicator variable
24:26for each dose level and Y represents the continuous
24:32immunological response and as I mentioned before
24:36in our application, that response is a T-cell persistence
24:41at one month after infusion.
24:44So based on the estimated Ys,
24:48we calculate the randomization probability for each dose
24:53and allocate patients sequentially based on this path.
24:58So how does this look, let's say again,
25:01that we have for dose levels and three patients treated
25:05at each dose level, we measure the T-cell persistence
25:10for all patients within the cohort
25:12and we feed the linear model
25:16to generate the estimated efficiencies.
25:21Based on the estimated efficiencies,
25:23we calculate the randomization probabilities.
25:27So for each dose, in this case,
25:30the randomization probability for dose one is 5%,
25:34the highest is for dose number 49.
25:37So the next patient that will be allocated,
25:43that will be randomized, will probably be randomized
25:48to dose number four,
25:49because this one has the highest randomization probability,
25:54and the process continues in stage two
25:56and feel you have reached the maximum sample size
26:00that you've specified for the trial.
26:05So how did we evaluate the model behavior?
26:09Well, we looked at two different sample sizes,
26:1225 and 50 patients in total for the trial.
26:17The number of levels varied from three to five,
26:20we don't recommend using this design
26:25for less than three of dose levels is just not enough
26:28and the design that you don't gain anything by using it
26:33if you have less than three.
26:36We use the 5,000 simulations for each scenario
26:41and in terms of establishing the operating characteristics,
26:46we quantified two things.
26:48One, we looked at present those allocation
26:52and we looked at estimation of efficacy outcomes,
26:55because that is the main goal actually, after implementing.
26:59you're looking at both toxicity and efficacy
27:03in determining the optimal dose
27:05with the goal of allocating more patients,
27:08skewing the allocation to a dose that is acceptably safe
27:13and has promising efficacy,
27:17in this case has a higher percentage of these helper system.
27:24So this is our combination of efficacy and toxicity scenario
27:29in panel A you see that we have five dose level
27:35on the x-axis and on the y we have the T-cell persistence
27:39as a functional skills
27:41and the scenarios vary from completely flat to monotonic
27:47increasing to non-monotonic and also to plateau,
27:53which is scenario number three.
27:55As I mentioned, this is a pretty frequent scenario
27:59at a certain dose level, the efficacy, that's not,
28:05we don't see any big increases in the efficacy.
28:08In terms of toxicity, again, different flat trends
28:15on the steeper dose toxicities scenario.
28:22In terms of beta stimulation, toxicity was stimulated
28:26from a Bernoulli distribution, persistence was simulated
28:31from a beta-binomial and for variance,
28:35the variance was assumed to be a constant,
28:40but we then read the values of small and large,
28:44and to give you an idea of what means
28:45a large variance at 1%, that is equivalent to about 20%
28:50deviation from the mean T-cell persistence
28:55toxicity and efficacy are modeled independently.
29:01So some results, the paper on had the has several scenarios,
29:08but just to illustrate, this is for a total sample size
29:14of 25 patients both in stage one and in stage two
29:18and the design was compared to MTPI,
29:21which is the Modified Toxicity Probability Interval.
29:25And we wanted to compare and this is one of the operating
29:29correctors, 6% patient allocation.
29:34How does the design allocate patients based on toxicity
29:38and efficacy?
29:39So in this case, that toxicity we see is toxicity three,
29:46we have five, dose levels, the three dose levels
29:51have toxicities lower than 15%.
29:56Dose number four has a toxicity between 15 and 40%,
30:02and dose number five pretty much
30:05is at the maximum DLT threshold,
30:08'cause remember that we had two hypothesis,
30:1115% acceptable toxicity and 40%, so dose one, two, three
30:15are considered acceptable.
30:17Dose four is within the interval,
30:1915, 40 and 40% as shown in red is considered a toxic dose.
30:28So this is the same toxicity
30:31for different efficacy scenarios increasing,
30:35this is non-monotonic this umbrella plateau,
30:41this umbrella trend,
30:42this efficacy three is the plateau trend
30:46and the efficacy four is constant efficacy across all doses.
30:54So what we noticed that the design does a good job
30:59allocating most of the patients
31:01to a dose that is considered safe
31:04and also has the optimal efficacy.
31:10So this would be dose number three in panel A,
31:15again, dose number three in panel B and similar for C
31:22and so ultimately we allocate most of the patients
31:29to this optimal level.
31:35Now let me, I would like to show you an illustration
31:40of how can we use this design,
31:44its implementation in the R package and in the Shiny app,
31:50because the package has two purposes.
31:53One, is to run simulation, to observe to quantify
31:57the operating correctly sticks for different dose,
32:01toxicity dose efficacy scenarios and another benefit
32:07is that you can implement it to actually allocate,
32:11to run the trial, to allocate the next patient
32:17to the optimal dose.
32:18So simulation and implementation and the scenario
32:23that I will discuss next is inspired from a real study
32:29that we worked on when I was at the Cancer Center
32:35at Columbia and this was a phase-one trial
32:38evaluating modified autologous T-cells
32:41this genetically modified T-cells in patients
32:46where the recurrent solid tumors.
32:48So if you think that all these designs
32:51are usually theoretical or just great statistical proposals,
32:55actually this one had an implementation
33:03and had a real setting.
33:05So the initial design that was proposed
33:12was of course, an role-based design derivation,
33:16this was a two-by-two with up to five dose levels,
33:19and I wanna bring your attention to the dose levels.
33:23So in immunotherapy, especially in this T-cell therapies,
33:28the dose levels are not quantities of dose age
33:34of the medication are not milligrams,
33:36but they are actually the number of T-cells
33:41that are being infused back into the body.
33:44So in this case the dose levels vary from 50 to 10
33:53to the six to 500 and 10 the six viable T-cells,
33:58so millions of cells.
34:05And we wanted to explore what is the optimal dose
34:13for this regimen.
34:16So, this is a snapshot of the Shiny app,
34:26we redesigned the trial to incorporate both toxicity
34:29and continuous efficacy and in this case continue,
34:33this T-cell persistence was a reasonable biomarker,
34:38we looked at five dose levels, we had in simulations,
34:43you have to specify what are the toxicity to toxicity rates
34:47and what are the T-cell persistence levels.
34:52The true toxicity rates varied from five to 40%,
34:55the T-cell persistence varied from 15 to 40%.
35:00We didn't see that we're going to see more than 40%
35:03persistence followup and we looked at the total sample size
35:08of 30 patients for the trial, this was feasible
35:14and practical.
35:16So the setup it's very simple,
35:19you just put the number of dose levels,
35:21you specify the toxicities, you specify the mean efficacy,
35:25the variance for the efficacy, we chose 1%,
35:29but this can take different values,
35:35then of course, this is a dialogue
35:37that you have with your clinical investigator,
35:42hopefully with data supported from previous studies
35:45and what is considered to be to very these parameters.
35:50Then for stage one, you have to specify the two hypothesis,
35:54acceptable and unacceptable DLT.
35:58For this one, we set it at 15 and 40%,
36:02the likelihood ratio was set at two
36:05because of the sample size of 30 with k equals to two
36:12is reasonable and three cohorts, three patients per cohort,
36:17meaning each of the five dose levels we have three patients
36:22allocated each.
36:24Total sample size of 30 on the stopping rule,
36:28stopping rule meaning if none of the doses
36:32are considered acceptable in stage one,
36:35we're going to allocate up to nine patients
36:40at the first dose to further establish toxicity,
36:43and this can be changed to six or other number.
36:52So these are, how did the scenario looks?
36:56This is the graphs actually generated by the app
37:01toxicity and efficacy as a function of dose level,
37:05and now I'd like to ask you, ask for your participation,
37:12looking at these two scenarios,
37:14what do you think would be an optimal dose level,
37:19the recommended to be studied in phase-two or later?
37:35Okay, maybe we need some hint,
37:39so we want an acceptable dose.
37:44<v ->[] Three.</v>
37:46<v ->Dose number three, dose number three,</v>
37:49because dose number three is way outside
37:55of the acceptability range and also dose number three
38:00tends to have a good efficacy after dose number three
38:06we don't see any improvement in terms of efficacy.
38:11Dose number four is between the 15 and 40%
38:14then dose number five was probably toxic.
38:18So, dose number three is the optimal dose
38:21and what we would like to see,
38:26is most patients being allocated at this level.
38:29So in simulations, this is simulations for stage one,
38:37where based on observed the DLTs,
38:40we calculate the likelihood ratio and mark the doses
38:49as being acceptable or unacceptable.
38:53So in this case, based on the simulations, we see dose one,
38:57two, three and four are considered acceptably safe
39:03and they will be carried forward to stage number two,
39:07to be considered for that different organization,
39:10dose number five will be discarded
39:13and will not be used in stage two.
39:16And why is that?
39:17Because the likelihood ratio is less than 0.05.
39:29Now these are the simulations for a stage number two,
39:33so in the first part, we had five dose levels,
39:37three patients each, so that's a total of 15 patients
39:40starting with patient number 16, we moved to stage two
39:45and we do this adaptive randomization
39:47until we reach the maximum sample size of 30.
39:51So the app actually gives you simulations and allocations
39:58for all the patients from 16 to 30 dose assignments
40:02and efficacy outcome and this gives you a graph
40:07of the estimated efficacy and the medians
40:13and inter quartile ranges.
40:17So we repeated this a hundred times, you can repeat it more
40:211,000, 5,000 in terms of allocation based on the setting,
40:29based on the parameters, the hypothesis, the k threshold,
40:34the toxicity and efficacy scenarios scenario,
40:38we see that dose three tends to be favored
40:43in terms of allocation, where the highest media allocation,
40:4826.7 and going all the way to 33.3 for the 75th percentile.
40:57In terms of efficacy estimation, dose number three
41:04or if you remember when we specify the true mean
41:07efficacy was 40, the median estimated efficacy
41:12in this case is 39.75,
41:15the 75 percentile goes all the way to 45,
41:19but of course that will be improved with,
41:23as the sample size increases.
41:28So in conclusion, what does iAdapt proposals?
41:37It's an option, it's a viable option
41:40for incorporating toxicity and efficacy outcomes,
41:44especially for immunotherapy trials.
41:49The novelty is in the designing allows
41:56to model toxicity, both of binary
42:01and also as quasi-continuous measures
42:04and this was actually updated this year in the package
42:08to use the several types of toxicities and several grades,
42:17that continuous efficacy outcome
42:22is very relevant for immunotherapies and I really showed
42:26the example from the trial with T-cell persistence,
42:29it is a relevant biomarker, but you can use for example,
42:35absolute counts, you can use a full changes,
42:39so design is flexible in incorporating
42:41other continuous outcomes.
42:43As far as I know, this is the only design at this point
42:48that uses continuous efficacy outcomes.
42:53So in terms of operating characteristics,
42:58the design as well and allocating,
43:00skewing the allocation to optimal doses,
43:04estimation is marginally improved depends of course,
43:08on the level of various and the sample size
43:11and if anybody wants to try, you can use the R package,
43:14you can use the Shiny app to simulate
43:18to look at the behavior of different scenarios
43:22to put that in trial and of course,
43:26to use it to run the trial.
43:30I'd like to thank two former students, Alyssa and Laura,
43:34that helped in uploading the R package
43:40and Laura has created the Shiny app.
43:43And I do have some references in case you're interested,
43:47but I can also share the slides later on.
43:53So I think that is it and I wanted to allow some time
43:57for questions and comments and feedback from you.
44:02Thank you so much.
44:06<v ->Thank you very much professor.</v>
44:07(applauds)
44:13Do you ave any questions in the room here?
44:28Does anyone from Zoom have any questions?
44:35<v ->Cody, thank you for the presentation,</v>
44:37I think it's very useful to talk her way,
44:40as well as my future like your possible designs
44:46of the trials related to immunotherapy, I have one question.
44:53So you mentioned toxicity and...
45:01Does anywhere in the design actually dependent
45:08on the independence of toxicity and efficacy profile?
45:17<v ->It's a great point to talking</v>
45:18if we actually modeled jointly toxicity and efficacy.
45:23<v ->So I'm actually looking at the slide</v>
45:26that you have the toxicity and like also efficacy profile,
45:33I think like say if we have an ordinal categorical,
45:37lets say for example, like say we pick a number three
45:42because it's intolerable toxicity
45:47and maximize the efficacy, right?
45:50And anything above that will be too much
45:52and anything below that will be like,
45:55say not like effective enough.
46:00So if they are actually,
46:04how about certain joint distribution,
46:09is there any thing we can do in order to like,
46:13'cause that actually affect the simulation much?
46:18<v ->So the current model does not account</v>
46:22for the joint distribution,
46:24it models toxicity and efficacy separately.
46:27But as a next step we can look under to try to model
46:35that dependency between toxicity and efficacy,
46:38and especially for this novel agents,
46:41we've seen that most of the times toxicity as is related
46:48to efficacy, a stronger ethical, a stronger response
46:54does come with some higher levels of toxicity,
47:00but the current model does not look,
47:02they twist them independent.
47:06<v ->So do we, so is there, do you consider any,</v>
47:11like penalty for example,
47:12like say when there isn't such a good, like a compromise
47:16between like minimizing toxicity
47:20while maximizing efficacy, right?
47:23So in the demonstration we have a compromise,
47:26which is dose level number three,
47:30if there's, for example, if there is like a conflict
47:34between dose two and we can, like we don't really have,
47:39like the obvious optimal, like say optimized solution.
47:46Do we constantly there, like, for example, penalties
47:49or do we always pay for like toxicity,
47:53like say minimizing toxicity over like,
47:57say maximizing efficacy?
48:01<v ->And then think it's always, I think so for example,</v>
48:07in that situation, so you could actually take both
48:10dose three and dose number four,
48:14you could consider both to be considered for future trials.
48:20So, it's not the definite that the dose selected,
48:30that you're always gonna reach a minimum of toxicity
48:33and maximum efficacy, but you can look at different options
48:39with as long as toxicity is acceptable,
48:44you can consider maybe in phase-two
48:47to look at randomized trial, look at dose level combo one
48:52and dose levels combo-two based on efficacy
48:55and we've actually seen this in a lot of trials,
49:00the immune check point inhibitors review that I talked,
49:05that it's now in progress, we looked at phase-one and two
49:09and the rate of success and the design that are being used
49:12and the doses that are being carried forward from phase-one
49:16and phase-two, and surprisingly only 30%,
49:21in 30% of phase-two trials the MTD was used from phase-one,
49:30the rest either they use a lower dose
49:33or they use a higher dose, but not the MTD.
49:38So absolutely we can have like a range,
49:42because if you think about it,
49:43we have a limited sample size, right?
49:46We need more information for efficacy,
49:48so to complete the clear, the winner based on efficacy
49:53might not be sufficient at this level.
49:57<v ->Okay, great, thank you.</v>
49:58So the answer is before phase-three
50:01and as long as it's below the MTD,
50:04the efficacy is important, is more important to prove,
50:08like, say to move on to next stage.
50:10Thank you.
50:11<v Codruta>Yes.</v>
50:29Just connection.
50:30(chuckles)
50:34<v Moderator>Sorry, we're still having a little weird audio</v>
50:37issues obviously,
50:45but does anybody in the room have any other questions
50:48for the professor?
51:02Or even we end the Zoom.
51:11<v Wei>Hello, we have a question there.</v>
51:16<v Moderator>Hold on.</v>
51:21<v Student>Hi professor, I know we probably mentioned</v>
51:22this already, but I probably didn't typed that,
51:26can you repeat, maybe repeat what it was,
51:28what do you consider would be like an advantage
51:32of having a continuous efficacy
51:35compares non-continuous efficacy in your model?
51:40<v ->Yes, lots of information, so a lot of the lines</v>
51:48are looking at the efficacy as a binary or ordinal,
51:53there is actually one,
51:55I don't know if you've heard of the Boyne,
51:57that's also was published for immunotherapies
52:00and that's using you take the efficacy levels
52:06and you either dichotomized to represent
52:09what is a successful or promising efficacy versus not,
52:13and you pretty much modeled the probability of a response,
52:18right one versus zero or at an ordinal level.
52:22Number one, I think we were losing some information
52:25when we do this categorization,
52:28number two might be difficult to actually establish
52:32this cutoffs and what represents a success
52:35or how do we partition this efficacy range
52:38for this novel agents.
52:40So by looking at the continuous values,
52:43we make the most out that information and we let it on,
52:48we modeled it as such,
52:54plus in the last couple of years,
52:58this T-cell persistence has been shown
53:01to be a promising biomarker.
53:03So it's right on par with our proposal.
53:18I know this might be a tough topic to digest for students
53:24with early finding it's not such a...
53:29It's a (chuckles) framework on its own.
53:33So maybe not that everybody's familiar
53:38with the whole terminology on the landscape.
53:48<v Student>During that stimulation, you specifically...</v>
53:52So the toxicity was stimulated
53:55from a continuity distributions,
53:58is there any specific reason
53:59why you choose these distribution versus there's,
54:03and if we similarly from a different distribution,
54:08well, how about different conclusion like,
54:11well, there would be any dependence
54:13between efficacy and toxicity.
54:17<v ->Yes, so that's, and so in this case,</v>
54:21the results that I showed you were for toxicity,
54:25for binary toxicity, yes or no.
54:27So in a cohort of three patients for each patient,
54:31you observed either a zero or a one response,
54:36given the binary structure, it makes sense to use
54:39this Bernoulli right distribution
54:42and that sums up to binomial zero or one.
54:46In terms of dependency is with what Dr. Cheng
54:50was mentioning, we did not specify any correlation
54:55between toxicity and efficacy
54:57and did not look at the joint distribution between the two,
55:00we modeled them separate and probably
55:09that would be a good point moving forward.
55:12What's difficult is how do we, what would be interesting
55:17is looking at different levels of correlation
55:21and see how in this joint distribution,
55:25how the results with change, if we would capture that.
55:38<v Wei>Okay, so any more questions?</v>
55:50Okay, so thank you Dr. Chuizan,
55:53for your wonderful presentation.
55:58<v ->Thank you.</v>
55:59<v ->Thank you, and if you have any questions,</v>
56:02please email me anytime.
56:03(chuckles)
56:04And I'm sorry that you (indistinct).
56:10Okay, I'll see you shortly, bye.
56:13<v Wei>Thank you.</v>