DIGITAL GENETICS DRIVING PRECISION NEUROSCIENCE WITH AI AND BIOSENSORS

Transcript

• 00:00Our next speaker is a
• 00:02true yearly. It really epitomizes
• 00:05the kind of rising star,
• 00:07that the Adam Center is
• 00:09looking
• 00:10forward to,
• 00:12recruit,
• 00:12in in the future. So
• 00:14if you are interested,
• 00:16shoot me an email.
• 00:18Jason Liu,
• 00:20has an a bachelor's degree
• 00:22in applied mathematics from Yale
• 00:24and then received his PhD
• 00:26in computational biology
• 00:28and biomedical informatics,
• 00:30again from Yale, and is
• 00:32now a postdoctoral associate
• 00:34in Marc Gerstein's lab.
• 00:37And
• 00:38looking forward to your talk.
• 00:40Jason is
• 00:41connecting genomics,
• 00:44to,
• 00:44biosensors
• 00:45and really, opening a new
• 00:48new area of genetics that
• 00:50I think of as digital
• 00:52genetics or biosensor genetics, and
• 00:54that sort of allows
• 00:56to link what,
• 00:58this digital twin model,
• 01:01back to biosensors in live
• 01:03patients, which has enormous applications.
• 01:07Great. Thank you so much,
• 01:08doctor, for the introduction.
• 01:11And,
• 01:12I'm very honored to be
• 01:12here today to speak to
• 01:13you all, about some of
• 01:15our exciting work using AI
• 01:17and biosensors
• 01:18to drive precision medicine.
• 01:20So I'll just kind of
• 01:21give a quick introduction in
• 01:23terms of the terminology that
• 01:24I'll be using today.
• 01:26On the left here, you'll
• 01:26see,
• 01:27we'll refer to the clinical
• 01:29diagnosis as macrophenotype,
• 01:31this brown m here.
• 01:33And then, of course, we
• 01:34have the genotype here in
• 01:36the gray g. And a
• 01:37lot of the research here
• 01:39at the Adam Center and,
• 01:40many other people are interested
• 01:42in is linking together
• 01:43the macrophenotype
• 01:45to the genotype to understand
• 01:47the genetic architecture of disease.
• 01:49And, of course,
• 01:50this has been done quite
• 01:52successfully
• 01:53through many different case control
• 01:55GWAS.
• 01:56So just here as a
• 01:57couple of examples,
• 01:58Parkinson's GWAS,
• 02:00over fifty thousand cases,
• 02:02recently published in Nature Genetics
• 02:03and and also
• 02:05neuropsychiatric,
• 02:06conditions here at ADHD has
• 02:08been widely studied by the
• 02:09PGC.
• 02:10So today, most of my
• 02:12talk is gonna focus on
• 02:14ADHD,
• 02:15not directly Parkinson's disease, but
• 02:17I,
• 02:18hope to kind of show
• 02:19that a lot of this
• 02:20work is,
• 02:22adaptable and applicable
• 02:23to the research going on
• 02:25with Parkinson's disease.
• 02:26So
• 02:27despite all this, GWAS that
• 02:29has been occurring, you know,
• 02:31there there still is this,
• 02:33question of missing heritability, which
• 02:35is that if we look
• 02:36at the heritability given by
• 02:38a twin study, it's very
• 02:39high, especially for things like
• 02:41Parkinson's, Alzheimer's, ADHD.
• 02:43But
• 02:44what we actually capture using
• 02:45these case and control GWAS
• 02:47is only a fraction of
• 02:48that. And that difference, the
• 02:50missing heritability, is kind of
• 02:51what we're after.
• 02:52And so the question is,
• 02:54are there ways that we
• 02:55can improve this, capture more
• 02:57of that missing heritability?
• 02:59And so one of the
• 03:00ways, that we wanna do
• 03:02so is through this concept
• 03:03of precision phenotyping
• 03:05or intermediate phenotypes.
• 03:07So in the previous speakers,
• 03:08we've kind of heard already
• 03:09a little bit how genomics
• 03:11is used using, for example,
• 03:12single cell,
• 03:14RNA. You can do EQTLs,
• 03:16to link to the genetics.
• 03:18But today, what I'm gonna
• 03:19mainly speak about is how
• 03:21we can use digital technology
• 03:23and in particular, wearable and
• 03:25smartwatches.
• 03:26There's been kind of an
• 03:27emergence of this,
• 03:29the popularity of smartwatches.
• 03:30Many, many people have it.
• 03:31It's more accessible
• 03:33in terms of cost.
• 03:34And so the question is,
• 03:36if we use the information
• 03:38captured by the smartwatch,
• 03:40can we first, number one,
• 03:42link it to the macro
• 03:43phenotype or the disease, gain
• 03:45some sort of clinical insight?
• 03:46And then second of all,
• 03:48if we can do that
• 03:49successfully,
• 03:49can we then relink it
• 03:51back to the genotype
• 03:52to hopefully gain more statistical
• 03:54power in terms of our
• 03:56ability for genetic discovery?
• 04:01So for the rest of
• 04:02the talk, I'm gonna, kind
• 04:03of give an overview
• 04:04into one of our recent,
• 04:06projects.
• 04:07This is, digital phenotyping with
• 04:09wearables,
• 04:10for a cohort,
• 04:11known as the ABCD,
• 04:13adolescent brain cognitive development study.
• 04:16And, it was recently published
• 04:17in Cell. And,
• 04:19so just the kind of
• 04:20a high level overview of
• 04:22the data, it's a few
• 04:23thousand individuals with psychiatric diagnosis.
• 04:26They also have digital data.
• 04:27In this case, it's Fitbit
• 04:29smartwatches.
• 04:30And then finally, we have,
• 04:32genetic information for all of
• 04:33these individuals.
• 04:37And so oftentimes, the question
• 04:39that people ask is, well,
• 04:41what does that data look
• 04:42like? What does Fitbit data
• 04:43look like? So just on
• 04:45the left here, I'm
• 04:46showing, three individuals,
• 04:48their signal tracks across a
• 04:50variety of different modalities.
• 04:52So for these three individuals,
• 04:53we can see their heart
• 04:54rate, their calories, activity, steps,
• 04:56so on and so forth.
• 04:57These are things that,
• 04:59probably if you have a
• 05:00smartwatch, then you can also,
• 05:02look at as well.
• 05:03But, you know, this data
• 05:05is very noisy, and it
• 05:06has,
• 05:07problems with it in terms
• 05:09of downstream analysis. So how
• 05:10how can we prepare or
• 05:11process this data in a
• 05:13way that makes it suitable
• 05:14for downstream analysis?
• 05:16Well, one of the, very
• 05:18straightforward ways to do so
• 05:19is to summarize the data.
• 05:21So if I have somebody's
• 05:22heart rate, it changes across
• 05:24the day, I can do
• 05:25something very simple, which is
• 05:26to say, just take take
• 05:27the mean heart rate during
• 05:28the day or the mean
• 05:29heart rate at night. And
• 05:31we can look at a
• 05:32variety of different statistical measures
• 05:34and then summarize them into
• 05:35what we're calling the static
• 05:37features.
• 05:38We call them static features
• 05:39because,
• 05:41by nature of being a
• 05:42time series and then summarized,
• 05:43we lose some of that
• 05:45temporal resolution.
• 05:46But the good thing about
• 05:48the static features is they're
• 05:49very easy to understand. They're
• 05:51easy to work with. We
• 05:52can make this nice matrix
• 05:53of individuals by features,
• 05:56and then we can attach
• 05:57on a set of covariates
• 05:58that we would typically use
• 06:00for these types of studies.
• 06:01And so the static features
• 06:03are are one of the
• 06:04ways that we're gonna process
• 06:05the data.
• 06:06But as I just mentioned,
• 06:09the static features, we lose
• 06:10some of the temporal
• 06:12dynamics, meaning,
• 06:14how somebody is changing behaviorally
• 06:16or physiologically
• 06:17on a seconds or minute
• 06:18level. And so in order
• 06:20to preserve some of that,
• 06:22we still want to kind
• 06:23of,
• 06:24move to some sort of
• 06:25feature set that is temporally
• 06:27resolved. And, we're calling that
• 06:29temporally resolved feature set the
• 06:31dynamic features.
• 06:33And there's a variety of
• 06:34steps that we,
• 06:35perform on the raw data
• 06:37to achieve that,
• 06:38dynamic features, and I'll just
• 06:40briefly go over some of
• 06:41those steps now.
• 06:42So first of all, one
• 06:44of the challenges of this,
• 06:46wearable data is everyone's data
• 06:47looks very different. It's collected
• 06:49at different times. And so
• 06:50in the first step, what
• 06:52we're really interested is in
• 06:54aligning individuals
• 06:55across different days, seasonalities.
• 06:59And after we've aligned it,
• 07:00then we want to take,
• 07:02optimal window selection or a
• 07:04slice of that data.
• 07:05Some individuals, they may have
• 07:07three weeks of data. Others
• 07:09might just have a few
• 07:10days of data. And it's
• 07:11so it's important that we're
• 07:13kind of making a fair
• 07:14comparison, and and we have
• 07:16this empirical optimal window selection.
• 07:18And that results, on the
• 07:19right hand side, you can
• 07:20see, with about sixty seven
• 07:22percent of the data remaining,
• 07:25and for kind of over
• 07:26two thousand individuals. You can
• 07:28see if we if we
• 07:29go to a higher threshold
• 07:30of inclusion,
• 07:31the sample size on the
• 07:32y axis drops quite significantly.
• 07:33So this was, kind of
• 07:35an empirically derived,
• 07:37optimal window selection.
• 07:40And then after we've done
• 07:41the optimal window selection,
• 07:43of course, there's still some
• 07:44small amounts of missing data,
• 07:46and we do a very
• 07:47simple linear imputation here. But
• 07:49in addition to the linear
• 07:50imputation,
• 07:51we're also tracking which variables
• 07:54at what time points were
• 07:55actually imputed and which ones
• 07:57are actually observed.
• 07:59And the point of that
• 08:00is to let downstream
• 08:02AI and machine learning models
• 08:04actually decide how reliable these
• 08:06imputed values are.
• 08:08And this kind of process
• 08:09in aggregate gives us what's
• 08:11known as the dynamic features.
• 08:13So we have static features
• 08:14and dynamic features. This is
• 08:15how we've processed smartwatch data
• 08:17to build kind of a
• 08:18digital phenotype.
• 08:20And the next question is,
• 08:21how how do we use
• 08:21it in in modeling?
• 08:24So for the static features,
• 08:25it's it's quite straightforward. You
• 08:27have a a matrix of
• 08:28individuals by features,
• 08:30and we can use traditional
• 08:32machine learning,
• 08:33models here. In this case,
• 08:35we've, tested XGBoost and RandomForce,
• 08:37which are quite popular,
• 08:39machine
• 08:40machine learning models.
• 08:42And the idea is, can
• 08:43we use those static features
• 08:45to then,
• 08:46predict individuals with a particular,
• 08:49disease or if they're a
• 08:50healthy control.
• 08:52And one of the byproducts
• 08:53of doing machine learning like
• 08:54this is,
• 08:55you can see I've I've
• 08:56kind of marked here the,
• 08:58green d score, digital phenotyping
• 09:01score, or an AI generated
• 09:03risk score, it is to
• 09:04move away from binary
• 09:06definitions
• 09:07of disease.
• 09:08We're very used to using
• 09:10zeros and ones to identify
• 09:12people who have or do
• 09:13not have a disease, But
• 09:14this is, kind of one
• 09:15of the byproducts of using
• 09:16machine learning models is we
• 09:18can move towards kind of
• 09:19a continuum or a spectrum,
• 09:21and that can be more
• 09:22inclusive and, more precise in
• 09:24terms of defining individuals.
• 09:28Now in terms of the
• 09:29dynamic features, those are
• 09:31a time series,
• 09:32and they're a little bit
• 09:34different. We we can't use
• 09:35a traditional machine learning model
• 09:36to deal with them. And
• 09:38so instead, what we've adopted
• 09:40here is a convolutional neural
• 09:41net. So this is a
• 09:42a deep learning model.
• 09:44And on the left side
• 09:45here, you'll see those dynamic
• 09:46features. We we treat them
• 09:47actually very similar to an
• 09:49image.
• 09:50You have many different channels,
• 09:52across different time points.
• 09:55And, very similar to, like,
• 09:57a image classification task,
• 09:58we would want to use
• 10:00this stack of digital phenotypes
• 10:02to, again, predict the macro
• 10:04phenotype.
• 10:05Again, we're able to generate
• 10:06this
• 10:07continuous based risk score to
• 10:09hopefully more precisely characterize an
• 10:12individual.
• 10:13And and one technical detail
• 10:15that I'll just highlight here
• 10:16is the use of a
• 10:18variable size convolutional filter here
• 10:20in the bottom. And and
• 10:22the idea of the variable
• 10:23size convolutional filter is behavioral
• 10:26and physiological changes
• 10:28may occur on a minute
• 10:30or a very small time
• 10:31scale, and we're definitely interested
• 10:33in those changes.
• 10:34But we're also interested in
• 10:35more global changes that perhaps
• 10:37occur on the day or
• 10:38the weekly level.
• 10:40And so by using this
• 10:41variable size convolutional filter, we're
• 10:43actually able to capture more
• 10:45of those behavioral and physiological
• 10:47changes at different time scale.
• 10:50So what what are the
• 10:52main results or kind of,
• 10:54findings of using these models
• 10:55in this data?
• 10:57So, again, we're here look
• 10:58not looking exactly at Parkinson's.
• 11:00We're looking at ADHD and
• 11:01anxiety disorder, but, again, many,
• 11:04applications and extensions
• 11:06to Parkinson's.
• 11:07So in in the case
• 11:08of ADHD, the top row
• 11:09in blue,
• 11:11and and anxiety disorder, in
• 11:13the first column here, we
• 11:14have three different box plots.
• 11:15And and these are showing
• 11:16the accuracy of identifying individuals
• 11:19with or without the disease.
• 11:20And we can see that
• 11:22and we can see that
• 11:24the baseline model at the
• 11:25first model is without any
• 11:27wearable information.
• 11:28The second model is using
• 11:30the static features, and then
• 11:31the third model is using
• 11:32actually those deep learning based,
• 11:35features, the time series. And
• 11:37we can see the a
• 11:38kind of a increase in
• 11:39performance and accuracy of the
• 11:40model as we incorporate more
• 11:42of those temporally resolved features.
• 11:45But in addition to just
• 11:46improving the accuracy,
• 11:48we're also able to identify
• 11:49what are the the actual
• 11:51physiological features that drive that
• 11:53prediction.
• 11:54So
• 11:55for example, in ADHD,
• 11:57we find that heart rate
• 11:58is really kind of the
• 11:59the key driver,
• 12:00but in anxiety disorder, sleep
• 12:02quality is.
• 12:03And not only that, we're
• 12:05able to temporarily resolve that
• 12:07importance.
• 12:08Meaning, in in this curve
• 12:10here, we're showing where was
• 12:12heart rate important. Is heart
• 12:13rate important all the time
• 12:14or just some of the
• 12:15time? And and we're able
• 12:16to temporarily resolve the heart
• 12:17rate importance to kind of
• 12:18the early to, late afternoon.
• 12:21And then, of course, for
• 12:22sleep, it's during the night.
• 12:23But in particular, you can
• 12:25see a peak there around
• 12:27five AM. And and, again,
• 12:28these are adolescents with anxiety
• 12:30disorder, and so we're kind
• 12:31of showing that perhaps there's
• 12:32sleep disturbance
• 12:33as they're waking up going
• 12:35to school, and that's really
• 12:36driving
• 12:37potential,
• 12:38phenotypic traits.
• 12:41Okay. So that was kind
• 12:42of the,
• 12:43using this wearable data to
• 12:44make predictions,
• 12:46about phenotype.
• 12:47But then, of course, we're
• 12:48very interested in genetic discovery.
• 12:50So the question then becomes,
• 12:52if we move away from
• 12:53binary traits and go to
• 12:55these continuous based digital phenotypes,
• 12:57are we actually able to
• 12:58gain statistical power in terms
• 13:00of genetic discovery?
• 13:03So in order to evaluate
• 13:05that, the first thing we
• 13:06wanna do is establish kind
• 13:07of a baseline comparison, meaning
• 13:10let's use just the individuals
• 13:12we have and perform a
• 13:13traditional GWAS, meaning a case
• 13:15control study.
• 13:16And this would be using
• 13:17a zero or one oops.
• 13:19Sorry. A zero or one,
• 13:21for the disease and then
• 13:23the genotype here.
• 13:24And when we do such,
• 13:26analysis, this is using twelve
• 13:28hundred individuals.
• 13:29Perhaps unsurprisingly,
• 13:31we we don't find any
• 13:32genetic loci at genome wide
• 13:34significance above this blue line.
• 13:36And twelve hundred individuals is
• 13:38is really not that many
• 13:39individuals. Most GWAS studies might
• 13:41have hundreds or millions of
• 13:43individuals.
• 13:44So,
• 13:45again, this is not using
• 13:47any of the wearable data.
• 13:49Now instead,
• 13:51let's imagine we have this
• 13:52wearable data as kind of
• 13:53a multidimensional
• 13:54array, a digital phenotype,
• 13:57and this vector d.
• 13:59And instead, now let's model
• 14:00that, against the genotype.
• 14:03And we also include this,
• 14:05genotype by macrophenotype
• 14:07interaction term to ensure that
• 14:09any changes with the digital
• 14:10phenotype are actually tied to
• 14:12the disease itself.
• 14:14And what we find here
• 14:15in the same exact set
• 14:16of individuals that you previously
• 14:18saw, the twelve hundred,
• 14:19individuals, now we're able to
• 14:21see much more enrichment in
• 14:22statistical power. We identified two
• 14:24significant genetic loci.
• 14:26And just to highlight that
• 14:28a little bit, this this
• 14:29particular loci
• 14:30is,
• 14:31related to sedentary time, and
• 14:33you can see that there's
• 14:34a clear difference between individuals
• 14:37that are healthy controls and
• 14:38those individuals that have ADHD.
• 14:40But, additionally,
• 14:42in the group that,
• 14:43individuals of ADHD, you can
• 14:45see that sedentary time drop
• 14:47significantly
• 14:48as we go through, the
• 14:49different genotypes.
• 14:51And so, really, the the
• 14:52result here is kind of
• 14:53establishing this relationship
• 14:55between the disease itself, the
• 14:57genetics,
• 14:58but also these digital phenotypes.
• 15:03And I'll present kind of
• 15:05one other way that we've
• 15:06tackled this problem,
• 15:08which is,
• 15:09if we recall from before,
• 15:11using those machine learning and
• 15:13deep learning models, we're able
• 15:14to generate these digital phenotype
• 15:16or AI generated risk scores.
• 15:18And and those scores really
• 15:20are aggregating all of the
• 15:21information from the smartwatch.
• 15:22Instead, now let's use those
• 15:24as a,
• 15:26target for the genetic discovery.
• 15:28So here, the d, the
• 15:29digital phenotyping score as a
• 15:31function of the genetics.
• 15:33And
• 15:34using this method, again, we
• 15:35we increase the statistical power
• 15:37even more. So now ten
• 15:38genetic loci.
• 15:40And many of these, the
• 15:41pink and the blue represent
• 15:43psychiatric associated genes that are
• 15:44known.
• 15:45Some are ADHD associated,
• 15:48but also we identify,
• 15:50an array of novel targets
• 15:52as well.
• 15:53And so just kind of
• 15:54in in in summary here,
• 15:56really, the idea for the
• 15:58digital phenotypes
• 15:59is, number one, we can
• 16:01use it to drive clinical
• 16:03discovery,
• 16:04clinical characterization of individuals and
• 16:06their subtypes,
• 16:07but also to really drive
• 16:09forward the,
• 16:11the genetics discovery.
• 16:13And I would just wanna
• 16:15emphasize that this,
• 16:17this
• 16:18result here of course, there's
• 16:19there's many questions about causality,
• 16:22and that's kind of ongoing
• 16:24work that we we wanna
• 16:25address and, of course, that
• 16:26this is just the starting
• 16:27point of using digital phenotypes.
• 16:29We there's many much more
• 16:31translational work that needs to
• 16:32be done, and validation as
• 16:34well.
• 16:36So,
• 16:37to kind of end and
• 16:38wrap up here, the the
• 16:40future work,
• 16:41of course, is expanding to
• 16:43other modalities and diseases.
• 16:45The hope here is that
• 16:47this framework, this triangle here,
• 16:49can be used for things
• 16:50like Parkinson's disease and for
• 16:52other kind of, related movement
• 16:53disorders and other neurodegeneration.
• 16:56And the idea is to
• 16:57move away from directly associating
• 16:59the genotype to the macro
• 17:01phenotype, but instead
• 17:03using this intermediate phenotype to
• 17:05to characterize
• 17:06both molecularly, physiologically, and behaviorally
• 17:09individuals
• 17:10better, provide those clinical sorry.
• 17:12Provide those clinical insights,
• 17:14and, of course, most importantly,
• 17:16is to retrace
• 17:17back to the genetic information
• 17:19and, link to potential molecular
• 17:22mechanisms.
• 17:23And and there's a lot
• 17:24of ongoing work right now.
• 17:26Of course, we heard a
• 17:27little bit about the spatial
• 17:28data.
• 17:30I'm very excited about the
• 17:31digital health sphere using wearables,
• 17:34smartwatches, but also video data,
• 17:36video capture data,
• 17:37as well as brain imaging
• 17:38data.
• 17:39And so,
• 17:41yes, happy to, collaborate and
• 17:43and,
• 17:44work on this with others
• 17:45more. So I'll I'll end
• 17:46here just with the acknowledgment
• 17:48slide. In particular, I just
• 17:50wanna highlight Mark Gerstein,
• 17:52whose lab I'm in, and
• 17:53also Walter Roberts from Yale
• 17:55Psychiatry who
• 17:57we,
• 17:58they they're the co co
• 18:00senior authors on the study.
• 18:01And then also I just
• 18:02wanna highlight Beatrice who co
• 18:04led this study with me
• 18:05and, Yun Yang, a very
• 18:06talented graduate student who's been
• 18:08working with me, on this
• 18:09project as well, as well
• 18:10as some of the collaborators,
• 18:12from Barcelona and California. So
• 18:14with that, thank you very
• 18:15much.
• 18:16Yep.