WEBVTT
00:00:01.199 --> 00:00:14.640
Welcome to the Mind the Methods Podcast, brought to you by the International Society for CNS Clinical Trials and Methodology, your gateway to the latest insights, innovations, and conversations at the forefront of central nervous system clinical research.
00:00:14.880 --> 00:00:22.640
This series dives into the challenges, progress, and people shaping the future of CNS therapeutics development and trial methodology.
00:00:22.879 --> 00:00:31.839
In this first season, we'll explore innovations in CNS trial design, from digital endpoints and adaptive protocols to patient-centric models and generative AI.
00:00:32.079 --> 00:00:41.200
Here we bring together researchers, clinicians, regulators, and industry leaders to discuss the challenges and breakthroughs shaping the future of CNS therapeutics development.
00:00:41.359 --> 00:00:50.159
From advancing trial methodology to improving outcomes for patients, our conversations dive deep into the science and strategy behind CNS research.
00:00:50.399 --> 00:00:51.439
Thanks for joining us.
00:00:51.520 --> 00:00:52.399
Let's get started.
00:00:52.719 --> 00:00:59.200
Mind the Methods brought to you by the International Society for CNS Trials and Methodology.
00:01:01.359 --> 00:01:04.640
Welcome to today's ISCTM Mind the Methods podcast.
00:01:05.200 --> 00:01:10.719
I'm joined today by Sean Ratcliffe SmithHurst, who is president-elect of the ISCTM.
00:01:11.040 --> 00:01:12.879
Sean, welcome back to the podcast.
00:01:13.519 --> 00:01:13.840
Thank you.
00:01:14.400 --> 00:01:17.040
Today we'll discuss two related questions.
00:01:17.280 --> 00:01:24.239
First, how can biomarkers accelerate drug development in neuroscience without creating false confidence?
00:01:24.480 --> 00:01:32.480
And second, what practical methods can sponsors use to create regulatory flexibility while preserving scientific rigor?
00:01:32.719 --> 00:01:35.519
So, Sean, I'd like to start with some terminology.
00:01:35.760 --> 00:01:42.079
We use the word biomarker so broadly sometimes, and it can obscure more than it clarifies.
00:01:42.319 --> 00:01:47.920
How should we be thinking about defining the term a biomarker as drug developers?
00:01:48.319 --> 00:01:49.920
Really important question, Kemi.
00:01:50.400 --> 00:01:54.799
And this is where I'll reference the FDA NIH best framework.
00:01:55.040 --> 00:01:58.400
That framework is for biomarkers, endpoints, and other tools.
00:01:58.480 --> 00:02:10.639
And this framework helps to define a biomarker as a measurable characteristic that indicates a normal biological process, a disease process, or a response to an exposure or intervention.
00:02:10.879 --> 00:02:16.159
It can be molecular, imaging-based, histologic, physiologic, or digital.
00:02:16.400 --> 00:02:23.199
Importantly, a biomarker is not itself a direct assessment of how a patient feels, functions, or survives.
00:02:23.360 --> 00:02:28.719
That latter category really belongs to clinical outcome assessments and clinical endpoints.
00:02:29.039 --> 00:02:30.000
Thank you so much.
00:02:30.159 --> 00:02:36.639
You know, that category matters quite a bit because not all biomarkers are being asked to do the same job, correct?
00:02:36.960 --> 00:02:37.680
Absolutely.
00:02:37.919 --> 00:02:49.360
The best framework includes susceptibility or risk, diagnostic, monitoring, prognostic, predictive, pharmacodynamic or response and safety biomarkers.
00:02:49.599 --> 00:03:05.919
And what's been fantastic is that advances in neuroimaging, electrophysiology, fluid biomarkers, neurochemical biomarkers, and physiologic measures have really opened up new possibilities, particularly in CNS clinical trial development.
00:03:06.159 --> 00:03:21.840
Great examples with functional MRI patterns of activation in certain circuits, such as in the prefrontal cortex, the amygdala, and default mold network, can predict response to antidepressants, psychotherapy, or neuromodulation.
00:03:22.000 --> 00:03:34.879
And EEG signatures such as alpha, asymmetry, theta activity, and event-related potentials can be linked to response in depression, ADHD, anxiety, and other psychological conditions.
00:03:35.120 --> 00:03:46.960
A single bar marker might fit more than one category, but every regulatory discussion should be anchored to one proposed use in one defined population and stage of development.
00:03:47.120 --> 00:03:48.319
That's really critical.
00:03:48.639 --> 00:03:49.759
Well, thank you so much.
00:03:50.000 --> 00:04:01.120
I think about the context of use principle, your point about proposing the use in a particular population that's predefined and a particular stage of development.
00:04:01.439 --> 00:04:02.000
Absolutely.
00:04:02.159 --> 00:04:14.479
I think that context of use is really critical because it describes what the biomarker is, the population and setting in which it will be measured, what decision it will inform, and how the results should be interpreted.
00:04:14.719 --> 00:04:20.000
And something that we'll refer to later on in the podcast is neurofilament light.
00:04:20.160 --> 00:04:25.680
Saying neurofilament light is a biomarker is not a sufficient development plane per se.
00:04:25.920 --> 00:04:39.279
Plasma neurofilament light measured by a specific assay at a specific time will support pharmacodynamic interpretation in adults with a defined neurodegenerative disease is much closer to a usable context of use.
00:04:39.680 --> 00:04:43.839
So, Sean, where do surrogate endpoints fit into this hierarchy?
00:04:44.160 --> 00:04:51.360
A surrogate endpoint is particularly special because it really has a much higher stakes use of a biomarker.
00:04:51.519 --> 00:05:07.759
In the context of surrogate endpoints, this is substituting for a direct clinical outcome in an approval decision because it is either known to predict clinical benefit or is considered reasonably likely to predict clinical benefit under accelerated approval.
00:05:07.920 --> 00:05:15.759
Many biomarkers are excellent development tools without actually meeting this standard of being able to suffice as a surrogate endpoint.
00:05:16.160 --> 00:05:17.839
Yeah, that's a great point.
00:05:18.319 --> 00:05:25.120
One of the discipline checks really for development teams is to stop asking if a biomarker is validated.
00:05:25.439 --> 00:05:38.079
And rather than asking that question, you may want to really think about validated for what decision and in which population and which assay and what are the consequences if we are wrong.
00:05:38.480 --> 00:05:39.600
Absolutely, 100%.
00:05:39.920 --> 00:05:41.600
That's exactly the right question.
00:05:41.920 --> 00:05:52.800
Analytical validity, biological plausibility, clinical association, responsiveness to treatment, and surrogacy are related, but they're not all interchangeable.
00:05:53.600 --> 00:05:55.759
So, Sean, let's let's get more granular.
00:05:55.920 --> 00:06:01.360
I like this example and discuss the many uses of neurofilament light chain.
00:06:01.680 --> 00:06:10.639
Why do you think NFL has become so prominent across ALS, MS, and even other neurodegenerative diseases?
00:06:10.959 --> 00:06:19.199
Yeah, this is really an exciting use case of both bar marker being able to meet that evidentiary standard of a surrogate endpoint.
00:06:19.360 --> 00:06:24.639
So neurofilament light is a structural protein in large myelinated axons.
00:06:24.720 --> 00:06:31.360
When axons are injured, neurofilament light is released into cerebrospinal fluid and then into blood.
00:06:31.519 --> 00:06:41.120
The practical appeal is that blood-based assays can provide a relatively scalable and dynamic measure of neuronal injury in this way.
00:06:41.360 --> 00:06:48.160
The limitation is that neurofilament light is not disease fit per se because it's a marker of damage to these long neurons.
00:06:48.399 --> 00:06:55.360
An elevated value will tell us that injury is occurring, but it doesn't necessarily say why it is occurring or where it is occurring.
00:06:55.680 --> 00:06:57.120
Yeah, that's a great point.
00:06:57.360 --> 00:07:04.800
Potentially many use cases, but definitely not a specific unless you really predefine what the context of use is.
00:07:05.040 --> 00:07:23.360
I think neurofilament is potentially powerful either as a monitoring or pharmacodynamic biomarker, but that its interpretation requires the exact clinical context, reference data, standardized sampling across either blood or CSF and an assay that has an adequate amount of precision.
00:07:23.680 --> 00:07:24.079
Correct.
00:07:24.319 --> 00:07:33.120
Age, comorbidities, acute neurologic events, sample handling, assay platform, and the timing of measurement can all influence interpretation.
00:07:33.360 --> 00:07:47.759
So it is really key for development programs to prospectively define what magnitude and timing of change would be considered biologically meaningful rather than simply measuring neurofilament light and deciding later what the result means.
00:07:48.000 --> 00:07:53.279
It really is that predefinition and putting all of those components in context.
00:07:53.600 --> 00:07:55.040
Exactly, exactly.
00:07:55.279 --> 00:08:05.040
And the most important regulatory example is with TOFersin, which is currently marketed as Calcity Pasad1 Associated ALS.
00:08:05.360 --> 00:08:09.759
Sean, can you walk us through why that particular Toeferson case is relevant here?
00:08:10.079 --> 00:08:10.720
Absolutely.
00:08:10.959 --> 00:08:18.879
And for those of you who have attended one of our ISCTM meetings in the last year, we did have some discussion of this across two different panels.
00:08:19.040 --> 00:08:20.800
So this might be familiar to some.
00:08:20.959 --> 00:08:33.279
So forgive me going all the way back to the beginning, but Tofersin is an antisense oligonucleotide or an ASO that's designed to reduce SOD1 messenger RNA and SOD1 protein.
00:08:33.440 --> 00:08:36.799
SOD1 here being superoxide dismutase one.
00:08:37.200 --> 00:08:47.120
In a randomized trial conducted by Biogen called Valor, 108 participants were assigned two to one to Tofersin or placebo.
00:08:47.279 --> 00:08:56.799
At 28 weeks, the primary clinical endpoint, the change in ALS functional rating scale in the faster progressing subgroup was not statistically significant.
00:08:56.960 --> 00:09:06.799
However, tofersin reduced both CSF SOD1 and produced a substantial reduction in plasma neurofilament light compared to placebo.
00:09:07.120 --> 00:09:20.879
That's such an important nuance to point out that even though this population seems to have been enriched for a faster progressing subgroup of patients, they still didn't meet their primary endpoint in terms of the clinical endpoint.
00:09:21.039 --> 00:09:25.679
However, the mechanistic as well as the biomarker moved in the expected direction.
00:09:26.000 --> 00:09:26.960
Yeah, absolutely.
00:09:27.039 --> 00:09:37.279
And those are key pieces, really, when we think about the biomarker in the context of both disease biology as well as a meaningful change that goes with sort of target engagement as well.
00:09:37.440 --> 00:09:53.279
And based on this, FDA granted accelerated approval in April 2023, based on this reduction in the plasma neurofilament light, concluding that the biomarker effect was reasonably likely to predict clinical benefit in this specific genetic form of AMS.
00:09:53.519 --> 00:10:03.519
So, Sean, what made the NFL argument more persuasive for the FDA than a generic association between just having a lower NFL and better outcomes?
00:10:03.840 --> 00:10:04.320
Absolutely.
00:10:04.399 --> 00:10:07.679
And this is where the totality of evidence really comes in.
00:10:07.840 --> 00:10:33.759
There was a genetically defined disease, a therapy directed at the causal pathway, evidence of target engagement through lower CSFs, odd one, a downstream injury marker that moved strongly and consistently, natural history evidence linking neurofilament light to disease activity and prognosis, and ultimately longer-term observations that were directionally more favorable with earlier treatment.
00:10:34.000 --> 00:10:40.799
The regulatory conclusion was not that every neurofilament light reduction equals clinical benefit.
00:10:40.960 --> 00:10:52.480
It was that this particular chain of evidence in this population and with a confirmatory plan to confirm the clinical benefit downstream, that was what supported accelerated approval in this case.
00:10:52.799 --> 00:10:54.480
Yeah, thanks for elaborating on that.
00:10:54.559 --> 00:10:58.960
But the totality of evidence seems to have been favorable in this particular instance.
00:10:59.440 --> 00:11:00.000
Absolutely.
00:11:00.159 --> 00:11:02.159
I think that those are key points.
00:11:02.320 --> 00:11:15.200
It's a precedent for really a context-specific reasoning and very clearly laid out, but it's not a universal surrogacy for neurofilament like for all broad ALS, not until it's been studied in that way.
00:11:15.440 --> 00:11:25.440
A sponsor developing a treatment for sporadic ALS, MS, or other conditions can't simply cite this example and assume that it'll have the same regulatory status.
00:11:25.600 --> 00:11:32.480
It's really important for those sponsors to work through the totality of evidence in the same way it was laid out for Calcity.
00:11:32.799 --> 00:11:40.320
I think something that is encouraging, however, it does open the door for other areas where neurofilament light might also play a role.
00:11:40.480 --> 00:11:47.279
And a great example here is the growing body of evidence in SMA, spinal muscular atrophy.
00:11:47.440 --> 00:12:01.120
You know, however, there are some considerations here because neurofilament light is most informative in infants and young children where levels dramatically elevate before treatment and fall rapidly with effective treatment.
00:12:01.360 --> 00:12:12.639
The magnitude of decline really correlates with clinical improvement on the motor milestones and motor endpoints, as well as the chop-in 10 scores, which are sort of around development milestone acquisition.
00:12:12.879 --> 00:12:24.639
However, you know, if we look across older age groups, older children and adults with SMA, neurofilament light is less often less useful in that regard because it's often lower or near normal.
00:12:24.720 --> 00:12:28.720
So you haven't got that ability to sort of really discern the treatment effect.
00:12:28.879 --> 00:12:38.480
And so this is another great example, even within a disease or an illness, where one biomarker has a very age-specific context that needs to be taken into account.
00:12:38.799 --> 00:12:39.679
Yeah, I agree.
00:12:39.840 --> 00:12:41.360
That's another excellent example.
00:12:41.519 --> 00:12:42.639
Thank you so much for that.
00:12:42.879 --> 00:12:48.000
So let's contrast NFL with amyloid lowering and Alzheimer's disease.
00:12:48.159 --> 00:12:54.159
Amyloid can be used diagnostically for enrichment or as a pharmacodynamic measure.
00:12:54.240 --> 00:13:01.440
And in some settings, as a surrogate point, what should listeners take from that history of amyloid and NAD?
00:13:01.840 --> 00:13:02.320
Absolutely.
00:13:02.399 --> 00:13:07.440
And this is another one that's really exciting, and we had a good chance to explore in one of our recent panels.
00:13:07.600 --> 00:13:13.120
Again, I think one of the key lessons is that one biomarker can serve several functions in the same program.
00:13:13.360 --> 00:13:18.000
Amyloid PET or CSF biomarkers can confirm the underlying pathology.
00:13:18.159 --> 00:13:22.240
So they can be diagnostic as well as enrich the study population.
00:13:22.480 --> 00:13:29.279
Serial amyloid pet over time can demonstrate target engagement and it can quantify plaque removal.
00:13:29.440 --> 00:13:33.840
So this is where you have multiple ways in which this biomarker can function.
00:13:34.080 --> 00:13:49.279
And under the accelerated approval paradigm, reduction in amyloid plaque here was accepted as a surrogate endpoint that's reasonably likely to predict clinical benefit because it also shows the removal of the amyloid plaques part of the disease hallmark.
00:13:49.519 --> 00:13:50.320
Yeah, exactly.
00:13:50.480 --> 00:13:57.759
And Lakinabab is particularly instructive here because this drug moved from accelerated to traditional approval, correct?
00:13:58.080 --> 00:13:58.639
Absolutely.
00:13:58.879 --> 00:13:59.360
Absolutely.
00:13:59.519 --> 00:14:04.000
This is another great example, very similar to the Toferson example we shared earlier.
00:14:04.159 --> 00:14:09.200
Lakembi initially received accelerated approval based on amyloid plaque reduction.
00:14:09.279 --> 00:14:25.039
And then based on data from the Clarity AD trial that showed benefit on the clinical measures of cognition as well as function, that this approval then became a traditional approval with that confirmatory evidence from the Clarity AD trial.
00:14:25.200 --> 00:14:41.600
And I think what this tells us is having this sequence of an accelerated approval and then further confirming it in a longer and bigger study is a great way to provide patient access sooner and earlier and shows the regulatory flexibility to do that.
00:14:41.840 --> 00:14:51.919
I think the confirmatory evidence is critical to point out here because often we're so excited to think about what the surrogate biomarker is going to be and getting that accelerated approval.
00:14:52.000 --> 00:14:59.519
But along with that comes commitments to make sure that that confirmatory evidence is really gathered and supports the accelerated approval.
00:14:59.679 --> 00:15:05.600
So thinking about DanaMab, this offers a slightly different design lesson than Lacanomab, correct?
00:15:05.840 --> 00:15:06.720
Absolutely, yeah.
00:15:06.879 --> 00:15:20.480
And this is sort of where that kind of growth of evidence in the amyloid lowering pathway, I think, really helped the Donanomab program, who were able to receive traditional approval, which is in the US marketed as consumer.
00:15:20.879 --> 00:15:29.440
This was because those studies showed both amyloid lowering as well as statistically significant slowing cognitive and functional decline.
00:15:29.600 --> 00:15:36.799
So those studies were able to demonstrate both the effect on the biomarker, but it also showed the effect on clinical outcomes.
00:15:37.039 --> 00:15:40.159
Amyloid PET was still very central in these trials.
00:15:40.240 --> 00:15:51.519
It confirmed pathology, measured pharmacodynamic effect, and it was used in the trial to also stop active treatment when participants had reached a certain pre-specified low amyloid level.
00:15:51.679 --> 00:16:05.759
The approval case was really grounded in both those components as well as the clinical outcomes and the amyloid aspect has been leveraged in the label to inform who to treat, whether the drug engaged its target, and when treatment could be discontinued.
00:16:05.919 --> 00:16:10.480
So slightly different development path and outcome to Lacanomat.
00:16:10.799 --> 00:16:14.320
So let's pivot a little bit here to digital biomarkers.
00:16:14.559 --> 00:16:24.480
So this includes wearables, sensors from smartphones, voice, ectigraphy, and other continuously collected measures.
00:16:24.799 --> 00:16:28.960
How should a sponsor begin to address this particular type of firework?
00:16:29.440 --> 00:16:30.240
Oh, absolutely.
00:16:30.320 --> 00:16:35.519
And this I think is really central to really begin with the clinical concept, not the device.
00:16:35.679 --> 00:16:39.279
Many of these come in a form of a device, and that's not the starting point.
00:16:39.360 --> 00:16:51.279
The real starting point is what aspect of disease or function is being measured, why the digital measure is better or complementary to an established assessment, and what decision will it inform?
00:16:51.600 --> 00:16:55.039
A digital health technology is a data acquisition system.
00:16:55.279 --> 00:17:00.960
The digital endpoint or biomarker is the interpretable measure that's derived from those data.
00:17:01.120 --> 00:17:04.799
So those two concepts are related, but they're also not the same thing.
00:17:05.039 --> 00:17:12.799
Correct me if I'm wrong, FDA has a guidance on remote data acquisition that emphasizes fit-for-purpose selection and validation, correct?
00:17:13.039 --> 00:17:13.759
Correct, yeah.
00:17:13.920 --> 00:17:40.960
And this is a really helpful guidance because it really outlines how sponsors should address verification of the sensor or the software function, the analytical validation of the derived measures, clinical validation where appropriate, usability in the intended population, data transmission and storage, as well as missing data, algorithm version control as well, which can be really important, privacy as well as participant burden.
00:17:41.119 --> 00:17:43.440
So all of that is covered in that guidance.
00:17:43.599 --> 00:17:50.240
A consumer wearable can be appropriate, but only if its performance is adequate for the proposed context of use.
00:17:50.400 --> 00:17:59.680
And you mentioned ectigraphy as one of those examples, risk-worn activity device has a number of different great uses across a number of different conditions.
00:17:59.839 --> 00:18:09.519
But again, it has to be a very specific context of use because if you're measuring ambulation and foot strides, that has a in non-ambulatory patients, that could be very different.
00:18:09.839 --> 00:18:11.519
Yeah, that's a very good point.
00:18:12.000 --> 00:18:19.200
I'm thinking about our audience, uh, Sean, a lot of uh people who work with ICTM are involved in psychiatry clinical trials.
00:18:19.359 --> 00:18:33.279
And this really is an attractive use case for digital biomarkers because episodic clinic visits typically miss changes in sleep, mobility, social interaction, other things routinely occurring in daily life.
00:18:33.599 --> 00:18:34.079
Absolutely.
00:18:34.160 --> 00:18:39.039
And I think that's the promise of passive sensing with that improved ecological validity.
00:18:39.200 --> 00:18:54.480
There are some great examples, including the cost check program, which combines self-report and smartphone-derived measures on activity, mobility, communication, and device use in people with schizophrenia spectrum disorders.
00:18:54.640 --> 00:18:59.519
And it demonstrated that there are individualized changes around relapse events.
00:18:59.680 --> 00:19:05.599
It also illustrated some challenges that these patterns can be highly person-specific.
00:19:05.839 --> 00:19:12.400
Adherence and data completeness can vary, and small observational data sets can produce unstable algorithms.
00:19:12.559 --> 00:19:18.079
So I think this is helpful as it gives us some key lessons that we can take forward in other programs.
00:19:18.400 --> 00:19:36.000
Yeah, so the lesson here is that digital biomarkers can be valuable for exploratory signal detection and even endpoint development, but sponsors should resist converting an algorithm into a pivotal endpoint before its measurement properties and its interpretability are fully established.
00:19:36.319 --> 00:19:42.160
So let's switch gears and talk a little bit about prognostic adjustment and statistical methods.
00:19:42.319 --> 00:19:50.000
We recently discussed prognostic adjustment and other digital twin methods at our last ISCTM meeting.
00:19:50.240 --> 00:19:53.119
What is the cleanest way to explain the distinction?
00:19:53.440 --> 00:20:05.759
A prognostic model uses baseline information to predict a participant's likely outcome under standard of care or placebo, because it's largely been built based on longitudinal data set.
00:20:06.079 --> 00:20:13.680
In a randomized trial, that prediction can be included as a pre-specified baseline prognostic covariate.
00:20:13.839 --> 00:20:29.279
If the prediction is really strongly associated with the observed outcome, this prognostic adjustment of the analysis can produce a more precise treatment effect estimate and a narrower confidence interval while preserving randomization.
00:20:29.519 --> 00:20:36.400
So it can really, it's an individual patient, a digital twin built from a prognostic model.
00:20:36.640 --> 00:20:39.440
So hopefully that can help explain to a degree.
00:20:39.759 --> 00:20:40.720
Yeah, absolutely.
00:20:40.880 --> 00:20:48.559
And this is aligned with FD's final guidance, encouraging pre-specified adjustment for prognostic baseline covariates.
00:20:48.880 --> 00:20:49.519
Absolutely.
00:20:49.680 --> 00:21:17.200
And and it and just sort of as I introduced there, it has an important caveat that 2023 guidance doesn't specifically endorse machine learning or digital twin methods, but it really covers that broader statistical principle that well-chosen prognostic covariates can improve trial efficiency, whether that is through then generating a digital twin score that fits and this will fit if the model is locked before I'm blinding.
00:21:17.359 --> 00:21:24.480
It's externally trained, adequately validated and calibrated, and supported by sensitivity analysis as well.
00:21:24.720 --> 00:21:25.039
Yeah.
00:21:25.200 --> 00:21:31.359
So that's quite different from saying that a digital twin is going to replace a concurrent placebo participant.
00:21:31.839 --> 00:21:32.640
Exactly.
00:21:32.799 --> 00:21:39.279
And I think you highlight here sort of the words digital twin often get used to mean multiple different things.
00:21:39.359 --> 00:21:44.400
And so it really is important to clarify what is being conveyed and meant by a digital twin.
00:21:44.480 --> 00:21:53.039
And what I've been describing here is this prognostic score as an individual patient's digital twin to improve the precision of estimate in the trial.
00:21:53.279 --> 00:22:11.759
Once you're leveraging a digital twin to replace a control arm, or if you're using historical Data set to replace a control arm, the design becomes vulnerable to differences in patient selection, standards of care, endpoint assessment, calendar time, missing data, and unmeasured confounding.
00:22:11.920 --> 00:22:19.200
And so this is where actually following FDA's draft guidance on externally controlled trials is really key.
00:22:19.440 --> 00:22:34.079
It makes it very clear that the comparability and data quality need to be really central with that external control or historic data set that are being leveraged and that they can carry greater risk of bias than concurrent randomized controls.
00:22:34.240 --> 00:22:48.720
However, you know, this is a really important guidance because it also paves the way for being able to use large historic data sets as you know historical control arms, very important in rare disease and rare CNS disease in particular.
00:22:49.039 --> 00:22:49.440
Yeah.
00:22:49.680 --> 00:22:54.640
And I'm thinking about one particular example in a prevalent disease in Alzheimer's.
00:22:54.960 --> 00:23:11.279
There was a in 2025 a data set with an AI digital twin model that harmonized historical data from over 6,000 patients that were treated and evaluated in a 450-plus phase two trial.
00:23:11.519 --> 00:23:22.400
And the predicted projectories for placebo actually correlated modestly with the observed outcomes and they supported a potential use as a prognostic covariate.
00:23:22.559 --> 00:23:30.079
So not as an individual counterfactual truth and not as a regulatory approval precedent, but as a prognostic covariate.
00:23:30.319 --> 00:23:30.799
Absolutely.
00:23:30.880 --> 00:23:33.039
And I think it's a great interpretation.
00:23:33.279 --> 00:23:45.920
I think digital twins prognostic school models in this regard are most defensible today as either precision-enhancing covariates, scenario planning tools, possibly components of a Bayesian or adaptive design.
00:23:46.079 --> 00:24:05.200
And I think again, are very much something that the agencies, both EMA and FDA, are willing to talk about and encourage for use, particularly in phase two trials, and definitely need to see discussions on how they will be used in pivotal studies if that's the direction that would be taken.
00:24:05.440 --> 00:24:18.559
And I think it's, you know, again important that they're not necessarily a synthetic patient that eliminates the need for actual controls, unless there truly is the strength of evidence and regulatory agreement that's very strong in that regard.
00:24:18.720 --> 00:24:24.000
So always consult with the agencies when you're planning to sort of leverage this in your trials.
00:24:24.319 --> 00:24:26.319
I wholeheartedly agree with that.
00:24:26.799 --> 00:24:31.119
And while we could talk for another hour, Sean, I'm conscious of the time.
00:24:31.200 --> 00:24:34.799
So maybe let's just close this discussion with some practical approaches.
00:24:35.039 --> 00:24:45.119
So if a sponsor wants flexibility from regulators, what should the team development team do before the pivotal file in terms of planning?
00:24:45.440 --> 00:24:48.480
I think, you know, consult early, I think is really key.
00:24:48.640 --> 00:24:56.640
I think we've talked about a couple of guidances here today with prognostic modeling, the use of external control arms.
00:24:56.799 --> 00:25:04.640
And we've also talked about surrogate endpoints as a way to drive earlier access to treatment through the accelerated regulatory pathways.
00:25:04.880 --> 00:25:13.359
So really important to consult early, define the decision in the context of use if you're going for a biomarker as a surrogate endpoint.
00:25:13.599 --> 00:25:22.880
Really, that context of use should be well articulated before finalizing the protocol, be linked to the assay plan and the statistical analysis plan.
00:25:23.039 --> 00:25:27.200
The biomarker plan should really state what the measure will be measuring.
00:25:27.440 --> 00:25:41.599
Will it be measured in all participants, select participants, really confirm the mechanism, choose dose, support the monitoring, any adjustment to the analysis, or whether it will ultimately contribute to an approval endpoint.
00:25:41.839 --> 00:25:54.880
Really getting all of that clarity up front with an early consultation because ultimately, you know, ambiguity at the beginning really moves downstream more as multiplicity and interpretability problems at the end.
00:25:55.279 --> 00:25:55.839
Yeah.
00:25:56.319 --> 00:26:06.880
And I would I would add also, you know, don't forget to connect the biomarker to the mechanism and to the ultimate clinical meaning of that biomarker in the context of use.
00:26:07.440 --> 00:26:08.799
Absolutely, absolutely.
00:26:08.880 --> 00:26:28.400
And whenever we've spoken about these on our panels, it's been very, very clear to really you need to use that triangulated evidence package, non-clinical data, target engagement, dose response, longitudinal natural history, treatment associated biomarker changes, and relationships to clinically meaningful outcomes.
00:26:28.640 --> 00:26:32.559
You can't rely solely on a single correlation in this regard.
00:26:32.720 --> 00:26:44.640
And, you know, one of the things that machine learning has given us an advantage of here in biomarker discovery and prediction is that machine learning approaches can really deconvolute and uncover complex patterns.
00:26:44.799 --> 00:26:50.559
You know, you can look at high-dimensional data sets, so you get many more correlations than a single correlation.
00:26:50.640 --> 00:27:09.039
And this is where, you know, model aspects such as the feature selection or dimensionality reduction with principal components analysis and/or predictive modeling using these model validation are really important aspects to also include in this triangulated evidence package.
00:27:09.279 --> 00:27:13.039
The confirmatory question really should not be postponed until after approval.
00:27:13.119 --> 00:27:19.440
It really needs to be at thought a priori and that plan underway as you go through discussions for accelerated approval.
00:27:20.000 --> 00:27:43.359
And I just thought about one more thing, just making sure that the data is generalizable, because biomarkers can unfortunately sometimes unintentionally narrow participation if you have, you know, reference range differences, algorithms, or thresholds that are developed in unrepresented populations, these can have knock-on effects in the development program and ultimately the intended users of the drug.
00:27:43.680 --> 00:27:44.880
Absolutely, absolutely.
00:27:44.960 --> 00:28:02.400
And this is, you know, it's a scientific issue, not only an equity issue as well, a bar marker that performs differently across age, sex, ancestry, language, geography, disability, device ownership or comorbidity can really undermine the internal validity and that generalizability.
00:28:02.720 --> 00:28:09.920
So really any subgroup performance and calibration should be part of that development planning right from the get-go and right from the beginning.
00:28:10.240 --> 00:28:13.039
Well, Sean, it's been fantastic speaking with you.
00:28:13.200 --> 00:28:21.519
This is an exciting time for CNS clinical trials, where advances in validated biomarkers are providing objective measures of treatment response.
00:28:21.599 --> 00:28:29.440
And they're really paving the way for what we call precision medicine and improved trial efficiency and ultimately supporting regulatory decision making.
00:28:29.680 --> 00:28:31.200
Thank you again for joining me today.
00:28:31.519 --> 00:28:32.240
Thank you, Kemi.
00:28:34.160 --> 00:28:40.319
Mind the methods brought to you by the International Society for CNS Trials and Methodology.
00:00:01.199 --> 00:00:14.640
Welcome to the Mind the Methods Podcast, brought to you by the International Society for CNS Clinical Trials and Methodology, your gateway to the latest insights, innovations, and conversations at the forefront of central nervous system clinical research.
00:00:14.880 --> 00:00:22.640
This series dives into the challenges, progress, and people shaping the future of CNS therapeutics development and trial methodology.
00:00:22.879 --> 00:00:31.839
In this first season, we'll explore innovations in CNS trial design, from digital endpoints and adaptive protocols to patient-centric models and generative AI.
00:00:32.079 --> 00:00:41.200
Here we bring together researchers, clinicians, regulators, and industry leaders to discuss the challenges and breakthroughs shaping the future of CNS therapeutics development.
00:00:41.359 --> 00:00:50.159
From advancing trial methodology to improving outcomes for patients, our conversations dive deep into the science and strategy behind CNS research.
00:00:50.399 --> 00:00:51.439
Thanks for joining us.
00:00:51.520 --> 00:00:52.399
Let's get started.
00:00:52.719 --> 00:00:59.200
Mind the Methods brought to you by the International Society for CNS Trials and Methodology.
00:01:01.359 --> 00:01:04.640
Welcome to today's ISCTM Mind the Methods podcast.
00:01:05.200 --> 00:01:10.719
I'm joined today by Sean Ratcliffe SmithHurst, who is president-elect of the ISCTM.
00:01:11.040 --> 00:01:12.879
Sean, welcome back to the podcast.
00:01:13.519 --> 00:01:13.840
Thank you.
00:01:14.400 --> 00:01:17.040
Today we'll discuss two related questions.
00:01:17.280 --> 00:01:24.239
First, how can biomarkers accelerate drug development in neuroscience without creating false confidence?
00:01:24.480 --> 00:01:32.480
And second, what practical methods can sponsors use to create regulatory flexibility while preserving scientific rigor?
00:01:32.719 --> 00:01:35.519
So, Sean, I'd like to start with some terminology.
00:01:35.760 --> 00:01:42.079
We use the word biomarker so broadly sometimes, and it can obscure more than it clarifies.
00:01:42.319 --> 00:01:47.920
How should we be thinking about defining the term a biomarker as drug developers?
00:01:48.319 --> 00:01:49.920
Really important question, Kemi.
00:01:50.400 --> 00:01:54.799
And this is where I'll reference the FDA NIH best framework.
00:01:55.040 --> 00:01:58.400
That framework is for biomarkers, endpoints, and other tools.
00:01:58.480 --> 00:02:10.639
And this framework helps to define a biomarker as a measurable characteristic that indicates a normal biological process, a disease process, or a response to an exposure or intervention.
00:02:10.879 --> 00:02:16.159
It can be molecular, imaging-based, histologic, physiologic, or digital.
00:02:16.400 --> 00:02:23.199
Importantly, a biomarker is not itself a direct assessment of how a patient feels, functions, or survives.
00:02:23.360 --> 00:02:28.719
That latter category really belongs to clinical outcome assessments and clinical endpoints.
00:02:29.039 --> 00:02:30.000
Thank you so much.
00:02:30.159 --> 00:02:36.639
You know, that category matters quite a bit because not all biomarkers are being asked to do the same job, correct?
00:02:36.960 --> 00:02:37.680
Absolutely.
00:02:37.919 --> 00:02:49.360
The best framework includes susceptibility or risk, diagnostic, monitoring, prognostic, predictive, pharmacodynamic or response and safety biomarkers.
00:02:49.599 --> 00:03:05.919
And what's been fantastic is that advances in neuroimaging, electrophysiology, fluid biomarkers, neurochemical biomarkers, and physiologic measures have really opened up new possibilities, particularly in CNS clinical trial development.
00:03:06.159 --> 00:03:21.840
Great examples with functional MRI patterns of activation in certain circuits, such as in the prefrontal cortex, the amygdala, and default mold network, can predict response to antidepressants, psychotherapy, or neuromodulation.
00:03:22.000 --> 00:03:34.879
And EEG signatures such as alpha, asymmetry, theta activity, and event-related potentials can be linked to response in depression, ADHD, anxiety, and other psychological conditions.
00:03:35.120 --> 00:03:46.960
A single bar marker might fit more than one category, but every regulatory discussion should be anchored to one proposed use in one defined population and stage of development.
00:03:47.120 --> 00:03:48.319
That's really critical.
00:03:48.639 --> 00:03:49.759
Well, thank you so much.
00:03:50.000 --> 00:04:01.120
I think about the context of use principle, your point about proposing the use in a particular population that's predefined and a particular stage of development.
00:04:01.439 --> 00:04:02.000
Absolutely.
00:04:02.159 --> 00:04:14.479
I think that context of use is really critical because it describes what the biomarker is, the population and setting in which it will be measured, what decision it will inform, and how the results should be interpreted.
00:04:14.719 --> 00:04:20.000
And something that we'll refer to later on in the podcast is neurofilament light.
00:04:20.160 --> 00:04:25.680
Saying neurofilament light is a biomarker is not a sufficient development plane per se.
00:04:25.920 --> 00:04:39.279
Plasma neurofilament light measured by a specific assay at a specific time will support pharmacodynamic interpretation in adults with a defined neurodegenerative disease is much closer to a usable context of use.
00:04:39.680 --> 00:04:43.839
So, Sean, where do surrogate endpoints fit into this hierarchy?
00:04:44.160 --> 00:04:51.360
A surrogate endpoint is particularly special because it really has a much higher stakes use of a biomarker.
00:04:51.519 --> 00:05:07.759
In the context of surrogate endpoints, this is substituting for a direct clinical outcome in an approval decision because it is either known to predict clinical benefit or is considered reasonably likely to predict clinical benefit under accelerated approval.
00:05:07.920 --> 00:05:15.759
Many biomarkers are excellent development tools without actually meeting this standard of being able to suffice as a surrogate endpoint.
00:05:16.160 --> 00:05:17.839
Yeah, that's a great point.
00:05:18.319 --> 00:05:25.120
One of the discipline checks really for development teams is to stop asking if a biomarker is validated.
00:05:25.439 --> 00:05:38.079
And rather than asking that question, you may want to really think about validated for what decision and in which population and which assay and what are the consequences if we are wrong.
00:05:38.480 --> 00:05:39.600
Absolutely, 100%.
00:05:39.920 --> 00:05:41.600
That's exactly the right question.
00:05:41.920 --> 00:05:52.800
Analytical validity, biological plausibility, clinical association, responsiveness to treatment, and surrogacy are related, but they're not all interchangeable.
00:05:53.600 --> 00:05:55.759
So, Sean, let's let's get more granular.
00:05:55.920 --> 00:06:01.360
I like this example and discuss the many uses of neurofilament light chain.
00:06:01.680 --> 00:06:10.639
Why do you think NFL has become so prominent across ALS, MS, and even other neurodegenerative diseases?
00:06:10.959 --> 00:06:19.199
Yeah, this is really an exciting use case of both bar marker being able to meet that evidentiary standard of a surrogate endpoint.
00:06:19.360 --> 00:06:24.639
So neurofilament light is a structural protein in large myelinated axons.
00:06:24.720 --> 00:06:31.360
When axons are injured, neurofilament light is released into cerebrospinal fluid and then into blood.
00:06:31.519 --> 00:06:41.120
The practical appeal is that blood-based assays can provide a relatively scalable and dynamic measure of neuronal injury in this way.
00:06:41.360 --> 00:06:48.160
The limitation is that neurofilament light is not disease fit per se because it's a marker of damage to these long neurons.
00:06:48.399 --> 00:06:55.360
An elevated value will tell us that injury is occurring, but it doesn't necessarily say why it is occurring or where it is occurring.
00:06:55.680 --> 00:06:57.120
Yeah, that's a great point.
00:06:57.360 --> 00:07:04.800
Potentially many use cases, but definitely not a specific unless you really predefine what the context of use is.
00:07:05.040 --> 00:07:23.360
I think neurofilament is potentially powerful either as a monitoring or pharmacodynamic biomarker, but that its interpretation requires the exact clinical context, reference data, standardized sampling across either blood or CSF and an assay that has an adequate amount of precision.
00:07:23.680 --> 00:07:24.079
Correct.
00:07:24.319 --> 00:07:33.120
Age, comorbidities, acute neurologic events, sample handling, assay platform, and the timing of measurement can all influence interpretation.
00:07:33.360 --> 00:07:47.759
So it is really key for development programs to prospectively define what magnitude and timing of change would be considered biologically meaningful rather than simply measuring neurofilament light and deciding later what the result means.
00:07:48.000 --> 00:07:53.279
It really is that predefinition and putting all of those components in context.
00:07:53.600 --> 00:07:55.040
Exactly, exactly.
00:07:55.279 --> 00:08:05.040
And the most important regulatory example is with TOFersin, which is currently marketed as Calcity Pasad1 Associated ALS.
00:08:05.360 --> 00:08:09.759
Sean, can you walk us through why that particular Toeferson case is relevant here?
00:08:10.079 --> 00:08:10.720
Absolutely.
00:08:10.959 --> 00:08:18.879
And for those of you who have attended one of our ISCTM meetings in the last year, we did have some discussion of this across two different panels.
00:08:19.040 --> 00:08:20.800
So this might be familiar to some.
00:08:20.959 --> 00:08:33.279
So forgive me going all the way back to the beginning, but Tofersin is an antisense oligonucleotide or an ASO that's designed to reduce SOD1 messenger RNA and SOD1 protein.
00:08:33.440 --> 00:08:36.799
SOD1 here being superoxide dismutase one.
00:08:37.200 --> 00:08:47.120
In a randomized trial conducted by Biogen called Valor, 108 participants were assigned two to one to Tofersin or placebo.
00:08:47.279 --> 00:08:56.799
At 28 weeks, the primary clinical endpoint, the change in ALS functional rating scale in the faster progressing subgroup was not statistically significant.
00:08:56.960 --> 00:09:06.799
However, tofersin reduced both CSF SOD1 and produced a substantial reduction in plasma neurofilament light compared to placebo.
00:09:07.120 --> 00:09:20.879
That's such an important nuance to point out that even though this population seems to have been enriched for a faster progressing subgroup of patients, they still didn't meet their primary endpoint in terms of the clinical endpoint.
00:09:21.039 --> 00:09:25.679
However, the mechanistic as well as the biomarker moved in the expected direction.
00:09:26.000 --> 00:09:26.960
Yeah, absolutely.
00:09:27.039 --> 00:09:37.279
And those are key pieces, really, when we think about the biomarker in the context of both disease biology as well as a meaningful change that goes with sort of target engagement as well.
00:09:37.440 --> 00:09:53.279
And based on this, FDA granted accelerated approval in April 2023, based on this reduction in the plasma neurofilament light, concluding that the biomarker effect was reasonably likely to predict clinical benefit in this specific genetic form of AMS.
00:09:53.519 --> 00:10:03.519
So, Sean, what made the NFL argument more persuasive for the FDA than a generic association between just having a lower NFL and better outcomes?
00:10:03.840 --> 00:10:04.320
Absolutely.
00:10:04.399 --> 00:10:07.679
And this is where the totality of evidence really comes in.
00:10:07.840 --> 00:10:33.759
There was a genetically defined disease, a therapy directed at the causal pathway, evidence of target engagement through lower CSFs, odd one, a downstream injury marker that moved strongly and consistently, natural history evidence linking neurofilament light to disease activity and prognosis, and ultimately longer-term observations that were directionally more favorable with earlier treatment.
00:10:34.000 --> 00:10:40.799
The regulatory conclusion was not that every neurofilament light reduction equals clinical benefit.
00:10:40.960 --> 00:10:52.480
It was that this particular chain of evidence in this population and with a confirmatory plan to confirm the clinical benefit downstream, that was what supported accelerated approval in this case.
00:10:52.799 --> 00:10:54.480
Yeah, thanks for elaborating on that.
00:10:54.559 --> 00:10:58.960
But the totality of evidence seems to have been favorable in this particular instance.
00:10:59.440 --> 00:11:00.000
Absolutely.
00:11:00.159 --> 00:11:02.159
I think that those are key points.
00:11:02.320 --> 00:11:15.200
It's a precedent for really a context-specific reasoning and very clearly laid out, but it's not a universal surrogacy for neurofilament like for all broad ALS, not until it's been studied in that way.
00:11:15.440 --> 00:11:25.440
A sponsor developing a treatment for sporadic ALS, MS, or other conditions can't simply cite this example and assume that it'll have the same regulatory status.
00:11:25.600 --> 00:11:32.480
It's really important for those sponsors to work through the totality of evidence in the same way it was laid out for Calcity.
00:11:32.799 --> 00:11:40.320
I think something that is encouraging, however, it does open the door for other areas where neurofilament light might also play a role.
00:11:40.480 --> 00:11:47.279
And a great example here is the growing body of evidence in SMA, spinal muscular atrophy.
00:11:47.440 --> 00:12:01.120
You know, however, there are some considerations here because neurofilament light is most informative in infants and young children where levels dramatically elevate before treatment and fall rapidly with effective treatment.
00:12:01.360 --> 00:12:12.639
The magnitude of decline really correlates with clinical improvement on the motor milestones and motor endpoints, as well as the chop-in 10 scores, which are sort of around development milestone acquisition.
00:12:12.879 --> 00:12:24.639
However, you know, if we look across older age groups, older children and adults with SMA, neurofilament light is less often less useful in that regard because it's often lower or near normal.
00:12:24.720 --> 00:12:28.720
So you haven't got that ability to sort of really discern the treatment effect.
00:12:28.879 --> 00:12:38.480
And so this is another great example, even within a disease or an illness, where one biomarker has a very age-specific context that needs to be taken into account.
00:12:38.799 --> 00:12:39.679
Yeah, I agree.
00:12:39.840 --> 00:12:41.360
That's another excellent example.
00:12:41.519 --> 00:12:42.639
Thank you so much for that.
00:12:42.879 --> 00:12:48.000
So let's contrast NFL with amyloid lowering and Alzheimer's disease.
00:12:48.159 --> 00:12:54.159
Amyloid can be used diagnostically for enrichment or as a pharmacodynamic measure.
00:12:54.240 --> 00:13:01.440
And in some settings, as a surrogate point, what should listeners take from that history of amyloid and NAD?
00:13:01.840 --> 00:13:02.320
Absolutely.
00:13:02.399 --> 00:13:07.440
And this is another one that's really exciting, and we had a good chance to explore in one of our recent panels.
00:13:07.600 --> 00:13:13.120
Again, I think one of the key lessons is that one biomarker can serve several functions in the same program.
00:13:13.360 --> 00:13:18.000
Amyloid PET or CSF biomarkers can confirm the underlying pathology.
00:13:18.159 --> 00:13:22.240
So they can be diagnostic as well as enrich the study population.
00:13:22.480 --> 00:13:29.279
Serial amyloid pet over time can demonstrate target engagement and it can quantify plaque removal.
00:13:29.440 --> 00:13:33.840
So this is where you have multiple ways in which this biomarker can function.
00:13:34.080 --> 00:13:49.279
And under the accelerated approval paradigm, reduction in amyloid plaque here was accepted as a surrogate endpoint that's reasonably likely to predict clinical benefit because it also shows the removal of the amyloid plaques part of the disease hallmark.
00:13:49.519 --> 00:13:50.320
Yeah, exactly.
00:13:50.480 --> 00:13:57.759
And Lakinabab is particularly instructive here because this drug moved from accelerated to traditional approval, correct?
00:13:58.080 --> 00:13:58.639
Absolutely.
00:13:58.879 --> 00:13:59.360
Absolutely.
00:13:59.519 --> 00:14:04.000
This is another great example, very similar to the Toferson example we shared earlier.
00:14:04.159 --> 00:14:09.200
Lakembi initially received accelerated approval based on amyloid plaque reduction.
00:14:09.279 --> 00:14:25.039
And then based on data from the Clarity AD trial that showed benefit on the clinical measures of cognition as well as function, that this approval then became a traditional approval with that confirmatory evidence from the Clarity AD trial.
00:14:25.200 --> 00:14:41.600
And I think what this tells us is having this sequence of an accelerated approval and then further confirming it in a longer and bigger study is a great way to provide patient access sooner and earlier and shows the regulatory flexibility to do that.
00:14:41.840 --> 00:14:51.919
I think the confirmatory evidence is critical to point out here because often we're so excited to think about what the surrogate biomarker is going to be and getting that accelerated approval.
00:14:52.000 --> 00:14:59.519
But along with that comes commitments to make sure that that confirmatory evidence is really gathered and supports the accelerated approval.
00:14:59.679 --> 00:15:05.600
So thinking about DanaMab, this offers a slightly different design lesson than Lacanomab, correct?
00:15:05.840 --> 00:15:06.720
Absolutely, yeah.
00:15:06.879 --> 00:15:20.480
And this is sort of where that kind of growth of evidence in the amyloid lowering pathway, I think, really helped the Donanomab program, who were able to receive traditional approval, which is in the US marketed as consumer.
00:15:20.879 --> 00:15:29.440
This was because those studies showed both amyloid lowering as well as statistically significant slowing cognitive and functional decline.
00:15:29.600 --> 00:15:36.799
So those studies were able to demonstrate both the effect on the biomarker, but it also showed the effect on clinical outcomes.
00:15:37.039 --> 00:15:40.159
Amyloid PET was still very central in these trials.
00:15:40.240 --> 00:15:51.519
It confirmed pathology, measured pharmacodynamic effect, and it was used in the trial to also stop active treatment when participants had reached a certain pre-specified low amyloid level.
00:15:51.679 --> 00:16:05.759
The approval case was really grounded in both those components as well as the clinical outcomes and the amyloid aspect has been leveraged in the label to inform who to treat, whether the drug engaged its target, and when treatment could be discontinued.
00:16:05.919 --> 00:16:10.480
So slightly different development path and outcome to Lacanomat.
00:16:10.799 --> 00:16:14.320
So let's pivot a little bit here to digital biomarkers.
00:16:14.559 --> 00:16:24.480
So this includes wearables, sensors from smartphones, voice, ectigraphy, and other continuously collected measures.
00:16:24.799 --> 00:16:28.960
How should a sponsor begin to address this particular type of firework?
00:16:29.440 --> 00:16:30.240
Oh, absolutely.
00:16:30.320 --> 00:16:35.519
And this I think is really central to really begin with the clinical concept, not the device.
00:16:35.679 --> 00:16:39.279
Many of these come in a form of a device, and that's not the starting point.
00:16:39.360 --> 00:16:51.279
The real starting point is what aspect of disease or function is being measured, why the digital measure is better or complementary to an established assessment, and what decision will it inform?
00:16:51.600 --> 00:16:55.039
A digital health technology is a data acquisition system.
00:16:55.279 --> 00:17:00.960
The digital endpoint or biomarker is the interpretable measure that's derived from those data.
00:17:01.120 --> 00:17:04.799
So those two concepts are related, but they're also not the same thing.
00:17:05.039 --> 00:17:12.799
Correct me if I'm wrong, FDA has a guidance on remote data acquisition that emphasizes fit-for-purpose selection and validation, correct?
00:17:13.039 --> 00:17:13.759
Correct, yeah.
00:17:13.920 --> 00:17:40.960
And this is a really helpful guidance because it really outlines how sponsors should address verification of the sensor or the software function, the analytical validation of the derived measures, clinical validation where appropriate, usability in the intended population, data transmission and storage, as well as missing data, algorithm version control as well, which can be really important, privacy as well as participant burden.
00:17:41.119 --> 00:17:43.440
So all of that is covered in that guidance.
00:17:43.599 --> 00:17:50.240
A consumer wearable can be appropriate, but only if its performance is adequate for the proposed context of use.
00:17:50.400 --> 00:17:59.680
And you mentioned ectigraphy as one of those examples, risk-worn activity device has a number of different great uses across a number of different conditions.
00:17:59.839 --> 00:18:09.519
But again, it has to be a very specific context of use because if you're measuring ambulation and foot strides, that has a in non-ambulatory patients, that could be very different.
00:18:09.839 --> 00:18:11.519
Yeah, that's a very good point.
00:18:12.000 --> 00:18:19.200
I'm thinking about our audience, uh, Sean, a lot of uh people who work with ICTM are involved in psychiatry clinical trials.
00:18:19.359 --> 00:18:33.279
And this really is an attractive use case for digital biomarkers because episodic clinic visits typically miss changes in sleep, mobility, social interaction, other things routinely occurring in daily life.
00:18:33.599 --> 00:18:34.079
Absolutely.
00:18:34.160 --> 00:18:39.039
And I think that's the promise of passive sensing with that improved ecological validity.
00:18:39.200 --> 00:18:54.480
There are some great examples, including the cost check program, which combines self-report and smartphone-derived measures on activity, mobility, communication, and device use in people with schizophrenia spectrum disorders.
00:18:54.640 --> 00:18:59.519
And it demonstrated that there are individualized changes around relapse events.
00:18:59.680 --> 00:19:05.599
It also illustrated some challenges that these patterns can be highly person-specific.
00:19:05.839 --> 00:19:12.400
Adherence and data completeness can vary, and small observational data sets can produce unstable algorithms.
00:19:12.559 --> 00:19:18.079
So I think this is helpful as it gives us some key lessons that we can take forward in other programs.
00:19:18.400 --> 00:19:36.000
Yeah, so the lesson here is that digital biomarkers can be valuable for exploratory signal detection and even endpoint development, but sponsors should resist converting an algorithm into a pivotal endpoint before its measurement properties and its interpretability are fully established.
00:19:36.319 --> 00:19:42.160
So let's switch gears and talk a little bit about prognostic adjustment and statistical methods.
00:19:42.319 --> 00:19:50.000
We recently discussed prognostic adjustment and other digital twin methods at our last ISCTM meeting.
00:19:50.240 --> 00:19:53.119
What is the cleanest way to explain the distinction?
00:19:53.440 --> 00:20:05.759
A prognostic model uses baseline information to predict a participant's likely outcome under standard of care or placebo, because it's largely been built based on longitudinal data set.
00:20:06.079 --> 00:20:13.680
In a randomized trial, that prediction can be included as a pre-specified baseline prognostic covariate.
00:20:13.839 --> 00:20:29.279
If the prediction is really strongly associated with the observed outcome, this prognostic adjustment of the analysis can produce a more precise treatment effect estimate and a narrower confidence interval while preserving randomization.
00:20:29.519 --> 00:20:36.400
So it can really, it's an individual patient, a digital twin built from a prognostic model.
00:20:36.640 --> 00:20:39.440
So hopefully that can help explain to a degree.
00:20:39.759 --> 00:20:40.720
Yeah, absolutely.
00:20:40.880 --> 00:20:48.559
And this is aligned with FD's final guidance, encouraging pre-specified adjustment for prognostic baseline covariates.
00:20:48.880 --> 00:20:49.519
Absolutely.
00:20:49.680 --> 00:21:17.200
And and it and just sort of as I introduced there, it has an important caveat that 2023 guidance doesn't specifically endorse machine learning or digital twin methods, but it really covers that broader statistical principle that well-chosen prognostic covariates can improve trial efficiency, whether that is through then generating a digital twin score that fits and this will fit if the model is locked before I'm blinding.
00:21:17.359 --> 00:21:24.480
It's externally trained, adequately validated and calibrated, and supported by sensitivity analysis as well.
00:21:24.720 --> 00:21:25.039
Yeah.
00:21:25.200 --> 00:21:31.359
So that's quite different from saying that a digital twin is going to replace a concurrent placebo participant.
00:21:31.839 --> 00:21:32.640
Exactly.
00:21:32.799 --> 00:21:39.279
And I think you highlight here sort of the words digital twin often get used to mean multiple different things.
00:21:39.359 --> 00:21:44.400
And so it really is important to clarify what is being conveyed and meant by a digital twin.
00:21:44.480 --> 00:21:53.039
And what I've been describing here is this prognostic score as an individual patient's digital twin to improve the precision of estimate in the trial.
00:21:53.279 --> 00:22:11.759
Once you're leveraging a digital twin to replace a control arm, or if you're using historical Data set to replace a control arm, the design becomes vulnerable to differences in patient selection, standards of care, endpoint assessment, calendar time, missing data, and unmeasured confounding.
00:22:11.920 --> 00:22:19.200
And so this is where actually following FDA's draft guidance on externally controlled trials is really key.
00:22:19.440 --> 00:22:34.079
It makes it very clear that the comparability and data quality need to be really central with that external control or historic data set that are being leveraged and that they can carry greater risk of bias than concurrent randomized controls.
00:22:34.240 --> 00:22:48.720
However, you know, this is a really important guidance because it also paves the way for being able to use large historic data sets as you know historical control arms, very important in rare disease and rare CNS disease in particular.
00:22:49.039 --> 00:22:49.440
Yeah.
00:22:49.680 --> 00:22:54.640
And I'm thinking about one particular example in a prevalent disease in Alzheimer's.
00:22:54.960 --> 00:23:11.279
There was a in 2025 a data set with an AI digital twin model that harmonized historical data from over 6,000 patients that were treated and evaluated in a 450-plus phase two trial.
00:23:11.519 --> 00:23:22.400
And the predicted projectories for placebo actually correlated modestly with the observed outcomes and they supported a potential use as a prognostic covariate.
00:23:22.559 --> 00:23:30.079
So not as an individual counterfactual truth and not as a regulatory approval precedent, but as a prognostic covariate.
00:23:30.319 --> 00:23:30.799
Absolutely.
00:23:30.880 --> 00:23:33.039
And I think it's a great interpretation.
00:23:33.279 --> 00:23:45.920
I think digital twins prognostic school models in this regard are most defensible today as either precision-enhancing covariates, scenario planning tools, possibly components of a Bayesian or adaptive design.
00:23:46.079 --> 00:24:05.200
And I think again, are very much something that the agencies, both EMA and FDA, are willing to talk about and encourage for use, particularly in phase two trials, and definitely need to see discussions on how they will be used in pivotal studies if that's the direction that would be taken.
00:24:05.440 --> 00:24:18.559
And I think it's, you know, again important that they're not necessarily a synthetic patient that eliminates the need for actual controls, unless there truly is the strength of evidence and regulatory agreement that's very strong in that regard.
00:24:18.720 --> 00:24:24.000
So always consult with the agencies when you're planning to sort of leverage this in your trials.
00:24:24.319 --> 00:24:26.319
I wholeheartedly agree with that.
00:24:26.799 --> 00:24:31.119
And while we could talk for another hour, Sean, I'm conscious of the time.
00:24:31.200 --> 00:24:34.799
So maybe let's just close this discussion with some practical approaches.
00:24:35.039 --> 00:24:45.119
So if a sponsor wants flexibility from regulators, what should the team development team do before the pivotal file in terms of planning?
00:24:45.440 --> 00:24:48.480
I think, you know, consult early, I think is really key.
00:24:48.640 --> 00:24:56.640
I think we've talked about a couple of guidances here today with prognostic modeling, the use of external control arms.
00:24:56.799 --> 00:25:04.640
And we've also talked about surrogate endpoints as a way to drive earlier access to treatment through the accelerated regulatory pathways.
00:25:04.880 --> 00:25:13.359
So really important to consult early, define the decision in the context of use if you're going for a biomarker as a surrogate endpoint.
00:25:13.599 --> 00:25:22.880
Really, that context of use should be well articulated before finalizing the protocol, be linked to the assay plan and the statistical analysis plan.
00:25:23.039 --> 00:25:27.200
The biomarker plan should really state what the measure will be measuring.
00:25:27.440 --> 00:25:41.599
Will it be measured in all participants, select participants, really confirm the mechanism, choose dose, support the monitoring, any adjustment to the analysis, or whether it will ultimately contribute to an approval endpoint.
00:25:41.839 --> 00:25:54.880
Really getting all of that clarity up front with an early consultation because ultimately, you know, ambiguity at the beginning really moves downstream more as multiplicity and interpretability problems at the end.
00:25:55.279 --> 00:25:55.839
Yeah.
00:25:56.319 --> 00:26:06.880
And I would I would add also, you know, don't forget to connect the biomarker to the mechanism and to the ultimate clinical meaning of that biomarker in the context of use.
00:26:07.440 --> 00:26:08.799
Absolutely, absolutely.
00:26:08.880 --> 00:26:28.400
And whenever we've spoken about these on our panels, it's been very, very clear to really you need to use that triangulated evidence package, non-clinical data, target engagement, dose response, longitudinal natural history, treatment associated biomarker changes, and relationships to clinically meaningful outcomes.
00:26:28.640 --> 00:26:32.559
You can't rely solely on a single correlation in this regard.
00:26:32.720 --> 00:26:44.640
And, you know, one of the things that machine learning has given us an advantage of here in biomarker discovery and prediction is that machine learning approaches can really deconvolute and uncover complex patterns.
00:26:44.799 --> 00:26:50.559
You know, you can look at high-dimensional data sets, so you get many more correlations than a single correlation.
00:26:50.640 --> 00:27:09.039
And this is where, you know, model aspects such as the feature selection or dimensionality reduction with principal components analysis and/or predictive modeling using these model validation are really important aspects to also include in this triangulated evidence package.
00:27:09.279 --> 00:27:13.039
The confirmatory question really should not be postponed until after approval.
00:27:13.119 --> 00:27:19.440
It really needs to be at thought a priori and that plan underway as you go through discussions for accelerated approval.
00:27:20.000 --> 00:27:43.359
And I just thought about one more thing, just making sure that the data is generalizable, because biomarkers can unfortunately sometimes unintentionally narrow participation if you have, you know, reference range differences, algorithms, or thresholds that are developed in unrepresented populations, these can have knock-on effects in the development program and ultimately the intended users of the drug.
00:27:43.680 --> 00:27:44.880
Absolutely, absolutely.
00:27:44.960 --> 00:28:02.400
And this is, you know, it's a scientific issue, not only an equity issue as well, a bar marker that performs differently across age, sex, ancestry, language, geography, disability, device ownership or comorbidity can really undermine the internal validity and that generalizability.
00:28:02.720 --> 00:28:09.920
So really any subgroup performance and calibration should be part of that development planning right from the get-go and right from the beginning.
00:28:10.240 --> 00:28:13.039
Well, Sean, it's been fantastic speaking with you.
00:28:13.200 --> 00:28:21.519
This is an exciting time for CNS clinical trials, where advances in validated biomarkers are providing objective measures of treatment response.
00:28:21.599 --> 00:28:29.440
And they're really paving the way for what we call precision medicine and improved trial efficiency and ultimately supporting regulatory decision making.
00:28:29.680 --> 00:28:31.200
Thank you again for joining me today.
00:28:31.519 --> 00:28:32.240
Thank you, Kemi.
00:28:34.160 --> 00:28:40.319
Mind the methods brought to you by the International Society for CNS Trials and Methodology.
