Immune Signatures of Vaccination and Infection Responses

Transcript

• 00:00So, kicking us off in
• 00:02session two, it's,
• 00:03Steve, Kleinstein,
• 00:05to who's the Anthony Brady
• 00:06professor of pathology here at
• 00:08the Yale School of Medicine.
• 00:10He also co directs the
• 00:11grad program in computational biology
• 00:11and
• 00:13biomedical informatics,
• 00:15with secondary appointments in immunobiology
• 00:18and the new department of
• 00:19biomedical informatics and data science.
• 00:22And so Steve, came to
• 00:23immunology from computer science, and
• 00:25he's been a leader working
• 00:27on
• 00:29various aspects of the immune
• 00:30system for the last, two
• 00:31decades. So welcome, Steve.
• 00:36Thank you.
• 00:39It's, re really nice to
• 00:41see all all the energy
• 00:42and all the all the,
• 00:43collaborations that are starting up,
• 00:45here at Yale in this
• 00:46area of, computational systems, immunology,
• 00:49AI and engineering.
• 00:50So I am going to,
• 00:52hit on a lot of
• 00:53the themes that so a
• 00:54a bunch of the previous
• 00:55speakers already, already touched upon.
• 00:58And broadly speaking, we are,
• 00:59we are interested in how
• 01:01to read an individual's immunological
• 01:03state. So there's lots of
• 01:04new and emerging experimental methods
• 01:06that allow one to profile
• 01:07somebody's immune state,
• 01:09at at very high throughput
• 01:10ranging from things like transcriptomics,
• 01:13single cell transcriptomics, metabolomics,
• 01:16proteomics,
• 01:17and also profiling of the
• 01:18adaptive immune receptor repertoire, b
• 01:20and t cell receptor repertoires.
• 01:23My lab both, develops novel
• 01:25computational methods to, analyze,
• 01:28and, deal with a lot
• 01:29of these new, and emerging
• 01:30data types,
• 01:31as well as working with
• 01:32clinical and experimental groups to
• 01:34apply those,
• 01:35those techno those techniques,
• 01:37to real to real data
• 01:38to understand immunological
• 01:40state. And the kinds of
• 01:41questions we're broadly interested in
• 01:42are some of the kind
• 01:43some of the questions that
• 01:44John already already touched on.
• 01:46So how do we you
• 01:47know, can we take these
• 01:48immunological measurements,
• 01:49and say something about the
• 01:51exposure history,
• 01:52of the individual? Right? The
• 01:53immune system encodes to some
• 01:55degree our history of prior
• 01:56exposure. So can we say
• 01:57something about whether an individual
• 01:58has been exposed to, say,
• 01:59SARS COV two? Does the
• 02:01individual carry protective memory from
• 02:02this, from this infection?
• 02:05Talking about current immunological state,
• 02:06can we profile somebody's immune
• 02:08system, and make an inference
• 02:10about whether they have are
• 02:11are currently undergoing
• 02:13an acute infection? With what
• 02:14virus are they currently infected
• 02:15with? Can we say something
• 02:17about the future? Will the
• 02:18outcome of that, that infection
• 02:19be sort of a mild
• 02:20infection that resolves quickly, or
• 02:22will it be more more
• 02:23severe?
• 02:24And we can we also
• 02:25wanna be able to ask
• 02:26questions about, the the the
• 02:28outcome of different,
• 02:29interventions or things like vaccinations.
• 02:31We give someone a vaccine,
• 02:33how does their immunological state
• 02:34can we predict what the
• 02:35outcome of that vaccine will
• 02:37be? Will they generate good
• 02:38protective, immunological
• 02:39memory? So I'm gonna talk
• 02:41at, somewhat high level about
• 02:43a few of the different
• 02:44projects going on in my
• 02:45lab to touch on on
• 02:46these different aspects.
• 02:47So first, let's talk about
• 02:48identifying immunological exposures from somebody's
• 02:51host, host response profile.
• 02:53So the idea here is
• 02:55that we're going from a,
• 02:56a population of people. We
• 02:57take some kind of blood
• 02:58or tissue sample, and we
• 03:00profile that sample. So here,
• 03:01for example, we're looking at
• 03:03transcriptional profiles that we take,
• 03:05so we measure different gene
• 03:07expression levels, genes on the
• 03:08x axis, samples from those
• 03:09individuals on the y axis.
• 03:11So we get these different
• 03:11gene expression profiles, and we
• 03:13wanna make some inferences. So,
• 03:14for example, can we look
• 03:15at those gene expression profiles
• 03:17and discern a pattern that
• 03:18might tell us, hey. This
• 03:19group of individuals
• 03:21is currently undergoing an acute
• 03:22response to West Nile virus
• 03:23infection.
• 03:24So there are a lot
• 03:25of publications, a lot of
• 03:26studies that try to develop
• 03:28these types of signatures. Right?
• 03:29You recruiting a cohort of
• 03:30people with an infection,
• 03:32some kind of control populations,
• 03:34profile them, do some machine
• 03:35learning, some differential expression analysis,
• 03:37come up with some set
• 03:38of genes or metabolites or
• 03:39proteins that are characteristic of
• 03:41that, of that infection.
• 03:43So we set out a
• 03:44few years ago to evaluate
• 03:45some of these signatures that
• 03:46have been proposed in literature.
• 03:48CBB graduate student, Jan Chawla,
• 03:51wanted to ask a couple
• 03:52of questions.
• 03:53Particularly, he wanted to look
• 03:54at the the signatures that
• 03:56were in the literature and
• 03:57assessed how, robust they were.
• 03:59So to what extent you
• 04:00could take a signature from
• 04:01one study and, apply that
• 04:03signature to an independent,
• 04:04cohort and have it be
• 04:05predictive and also cross reactivity.
• 04:07So to what extent could
• 04:08you take a signature of,
• 04:09say, for some particular virus
• 04:11infection,
• 04:12like a like an influenza
• 04:13infection?
• 04:14And to what extent was
• 04:15that signature specific to that
• 04:17infection or would it what
• 04:18we call cross react? Would
• 04:20that signature also come up,
• 04:21in a lot of other
• 04:22viral infections or bacteria or
• 04:24or other types of infections
• 04:25like a bacterial?
• 04:26So we put together a
• 04:27cohort of about, at the
• 04:29time of the publication, it
• 04:29was a hundred and fifty
• 04:30datasets and a a a
• 04:32bit over seventeen thousand,
• 04:34transcriptional profiles
• 04:35encompassing a whole variety of
• 04:37viral, viral infections, bacterial infections,
• 04:40as well as some noninfections,
• 04:42conditions that we thought might,
• 04:43might be similar, might look
• 04:45similar transcriptionally to a,
• 04:47to an infection response.
• 04:48And he evaluated those signatures,
• 04:50for their robustness and cross
• 04:52reactivity
• 04:53and brought you know, high
• 04:54level what we found from,
• 04:55from the study is one,
• 04:56a lot of the signatures
• 04:57that were,
• 04:59proposed in literature, in fact,
• 05:00were quite robust. Generally speaking,
• 05:02if you developed a signature
• 05:04that was supposed to be
• 05:04predictive of some virus response,
• 05:07it was reproducible in in,
• 05:09other studies of the of
• 05:10that virus response. But there
• 05:11was also a large amount
• 05:12of cross reactivity, and maybe
• 05:14this is something,
• 05:15that's not too surprising.
• 05:17A lot of these a
• 05:18lot of these studies, you
• 05:19will get, you know, for
• 05:20example, like an interferon signature
• 05:22that will come up in
• 05:23almost every virus,
• 05:24virus response that you study.
• 05:26And you might expect if
• 05:27you don't account for that,
• 05:28that signature is also gonna
• 05:29be predictive of multiple other
• 05:30viruses.
• 05:31We also found those signatures
• 05:33cross react with a lot
• 05:34of noninfectious conditions. So here
• 05:35you see, for example, we
• 05:36we evaluated a bunch of
• 05:37different, signatures for virus infections
• 05:39along the x axis and
• 05:41evaluated their ability to predict,
• 05:43aging. So just comparing young
• 05:44versus older older individuals. And
• 05:46you can see a lot
• 05:47of these signatures,
• 05:48as measured by the area
• 05:50under the curve for predicting
• 05:51old versus young are highly
• 05:52predictive of age, and only
• 05:54a subset of signatures is
• 05:55not. So for so we
• 05:56are getting a lot of
• 05:57cross reactivity, and you have
• 05:58to account for this if
• 05:59you wanna use these signatures,
• 06:01for, predictive responses.
• 06:03And in general, we are
• 06:04interest often interested in what
• 06:05makes a response unique. Right?
• 06:07In general, we don't necessarily
• 06:08wanna pick out, oh, we're
• 06:09getting an interferon response,
• 06:11in in in coming up
• 06:12in response to a viral
• 06:13infection. We wanna know what's
• 06:14specific about a SARS COV
• 06:16two infection that makes it
• 06:17different from a typical antiviral
• 06:19response, or what's different about
• 06:20this,
• 06:21vaccine response, this influenza vaccine
• 06:23response,
• 06:24when it doesn't generate protective
• 06:25memory that makes it different
• 06:27from other types of, responses?
• 06:28So we've also worked on
• 06:29methods in collaboration with, Elena
• 06:31Daslavsky's group at Mount Sinai
• 06:33to as methods to generate
• 06:35robust signatures that are not
• 06:36cross reactive. And this takes
• 06:38advantage of the massive public
• 06:40data that's available,
• 06:41profiling different types of responses.
• 06:43So if we have our
• 06:44response, our our our dataset
• 06:46of interest, so for example,
• 06:47disease versus a control, we
• 06:49can also integrate all sorts
• 06:51of control samples from other
• 06:52diseases that are in the
• 06:53public domain to come up
• 06:55with our signature that's specific.
• 06:56So for example, if we're
• 06:57profiling the influenza,
• 06:59an influenza infection response, we
• 07:00might pull from the literature
• 07:02lots of other
• 07:12specificities, so gene expression,
• 07:14protein expression signatures are are
• 07:16nice, but the immune system,
• 07:17we can take advantage already
• 07:19has a a sort of
• 07:20component built in that's highly
• 07:22specific,
• 07:22for the for the response
• 07:24being generated, and that is
• 07:25the the the adaptive immune
• 07:26response, the b cells and
• 07:28the t cells, that compose
• 07:29the adaptive arms of the
• 07:30immune response. I'm gonna focus
• 07:32on b cells. My lab
• 07:33focuses works a lot on
• 07:34b cell receptor, responses.
• 07:36So b cells have antibody
• 07:38receptors on their surface that
• 07:39give them specificity,
• 07:41to recognize different pathogens.
• 07:43The naive repertoire coming out
• 07:44of the bone marrow, so
• 07:45our our immune system constantly
• 07:47produces naive b cells with,
• 07:49pretty much a unique receptor
• 07:50on every cell that's generated
• 07:52from the bone marrow, and
• 07:53that gives those cells their,
• 07:54a unique ability to recognize
• 07:56different pathogens. So some cells
• 07:58coming out of the bone
• 07:59marrow might, have a have
• 08:00a,
• 08:01innate specificity for the influenza
• 08:03virus. Others might have an
• 08:04innate specificity for,
• 08:06for SARS CoV two. Once
• 08:08you get once you encounter
• 08:10a a pathogen, you're confronted
• 08:11with pathogenic challenge, whether it
• 08:13be a natural infection, or
• 08:14a vaccine.
• 08:15Some of those naive b
• 08:16cells are stimulated, activated into
• 08:18the response, and they undergo
• 08:20a process of, rapid tonal
• 08:22expansion and somatic hypermutation.
• 08:25So these cells introduce point
• 08:26mutations into the DNA that
• 08:28codes for their receptor. This
• 08:29is a really scary process
• 08:30when I first learned about
• 08:31it,
• 08:32but it's actually quite,
• 08:34quite good because it allows
• 08:35our bodies to learn about
• 08:36the pathogenic environment.
• 08:38And through a process of
• 08:39mutation and selection, we can
• 08:41generate higher and higher affinity
• 08:42receptors for the pathogen that
• 08:44we're confronted with. So we
• 08:45can learn about the pathogenic
• 08:46environment through our, throughout our
• 08:48lifetime.
• 08:49And this occurs in specialized
• 08:51structures called, germinal centers in
• 08:53our secondary lymphoid organs.
• 08:55And here you see actually
• 08:56a a germinal center. The
• 08:58b cells are ingrained from
• 08:59a live anesthetized mouse, undergoing
• 09:01an immune response. This is
• 09:02work we did with Anne
• 09:03Haberman,
• 09:04with two photon microscopy over,
• 09:05over a decade ago. So
• 09:07this process results in affinity
• 09:09maturation of the b cell
• 09:11repertoire, and you can imagine
• 09:12that we can learn a
• 09:13lot about the history of
• 09:14immunological
• 09:15exposures by studying these, the,
• 09:17the set of b cell
• 09:18receptors carried by an individual.
• 09:20Since two thousand nine, we've
• 09:21had the ability to sequence
• 09:23the b cell receptors at
• 09:24a high throughput,
• 09:26both with,
• 09:27both with bulk sequencing technologies
• 09:29where we can take a
• 09:30tissue sample,
• 09:31sequence the b cell receptor
• 09:32using PCR amplification and high
• 09:34throughput sequencing. That gives us
• 09:35anywhere from hundreds of thousands
• 09:37to millions of b cell
• 09:39receptors in a sample,
• 09:40and more recently through single
• 09:42cell technologies where we can
• 09:43get individual from a single
• 09:45cell, the b cell receptor
• 09:46along with the pair transcriptome
• 09:48for tens of thousands of
• 09:49cells, at a time.
• 09:51And, of course, the challenge
• 09:52is now once we're once
• 09:53we have these receptors, how
• 09:55do we
• 09:56how do we interpret them?
• 09:57So the typical data we
• 09:58get kinda looks like this.
• 10:00Right? We get the b
• 10:01cell receptor sequence. Each row
• 10:02is a b cell receptor
• 10:03sequence,
• 10:05associated with the cell. And
• 10:06now we wanna we wanna
• 10:07know things like, well, which
• 10:08one of these b cells
• 10:09which one of these b
• 10:09cell receptors,
• 10:11is is specific or binds
• 10:13to the influenza hemagglutinin,
• 10:14molecule, or which one of
• 10:15these are specific for the
• 10:16SARS CoV two spike protein.
• 10:18And I wanna thank Deepa
• 10:20for for introducing these,
• 10:22representation learning approaches because that's
• 10:24one of the things we
• 10:25do with b cell receptors
• 10:26is we take advantage of
• 10:28these, language based language based
• 10:30models
• 10:30to, develop,
• 10:32embedding approaches where we can
• 10:34learn we can take those
• 10:35b cell receptors,
• 10:36map them into a lower
• 10:37dimensional space where we can
• 10:39then,
• 10:40develop classification models and do
• 10:41analysis and learn, learn all
• 10:43sorts of interesting things about
• 10:44the b cell receptor.
• 10:46So one of the things
• 10:47we wanted to learn about
• 10:48and one of the, sort
• 10:49of important key challenges for
• 10:50the community is learning specificity.
• 10:52So we would like to
• 10:53be able to take a
• 10:54b cell receptor sequence and
• 10:55predict that this b cell
• 10:57receptor is going to bind,
• 10:58for example, to SARS CoV
• 10:59two spike protein. A graduate
• 11:01student, Mimi Wang, from the
• 11:03CBB program,
• 11:04did a did a study
• 11:05a couple of years ago,
• 11:07looking actually, just just was
• 11:09published this, this past year
• 11:10early early this year, actually.
• 11:12Looking at multiple embedding methods.
• 11:14Some of these are general
• 11:15protein embedding methods that were
• 11:17generated not for antibody receptors,
• 11:18but to embed general protein
• 11:19sequences,
• 11:21and others, such as this
• 11:22AnteBERTy were developed specifically to
• 11:24embed antibody
• 11:25receptors.
• 11:26And she evaluated the ability
• 11:28of these, of these different
• 11:29embedding methods to predict SARS
• 11:31CoV two spike binding.
• 11:33There's a background control where
• 11:34we scrambled up the sequences
• 11:35shown in gray here, that
• 11:37has the expected
• 11:39f ones. And you can
• 11:40see on the real data
• 11:42shown in the colored, in
• 11:43the colored bar too, we
• 11:44actually get a pretty good
• 11:45ability,
• 11:46a very good ability to
• 11:47predict SARS CoV two spike
• 11:48binding,
• 11:49from these, from the, from
• 11:50these embedded b cell receptors.
• 11:52And in fact, some of
• 11:53the, you know, the the
• 11:54generic method, the general methods
• 11:56actually work pretty well. We
• 11:57do get a little bit
• 11:58of a performance boost, from
• 11:59the models that were developed
• 12:01specifically on antibody receptors.
• 12:03We have a newer study
• 12:04that's on bioRxiv now where
• 12:05we've shown that if you
• 12:06take these, general methods and
• 12:08you actually do a little
• 12:08bit of fine tuning on
• 12:09them specific to the problem,
• 12:11we can actually boost, that
• 12:12power to predict specificity,
• 12:14even further.
• 12:16So that's the power to
• 12:17predict specificity on a single
• 12:18sequence. What about whole repertoires?
• 12:20So when we take a
• 12:21blood sample or a tissue
• 12:22sample from a person, we
• 12:23don't get a single sequence.
• 12:25We actually get tens of
• 12:26thousands to to potentially millions
• 12:28of sequences.
• 12:29It looks something like this.
• 12:30This is actually a blood
• 12:31repertoire from somebody undergoing, an
• 12:32actual vaccine response seven days
• 12:34post the response,
• 12:36where each of the points
• 12:36in these trees a leaf
• 12:37and each of those,
• 12:39and each of those,
• 12:40sort of circle diagrams represent
• 12:42a clonal expansion starting from
• 12:44a starting from a naive
• 12:45b cell. So you can
• 12:46see there is a bunch
• 12:47of clonal expansions in the
• 12:48blood.
• 12:49So that's pretty indicative of
• 12:51of, an actual ongoing infection.
• 12:53Generally speaking, in a in
• 12:54a healthy human, at rest,
• 12:56you don't have large clonal
• 12:57expansions,
• 12:58in your blood. But now
• 12:59can we tell what infection
• 13:01or what vaccination this person
• 13:02is actually responding to? So
• 13:04one simple thing we can
• 13:05do is take all of
• 13:06the b cell receptor sequences
• 13:08in these clones, calculate the
• 13:09probability of it being a
• 13:10SARS CoV-two specific clone, and
• 13:12just average over the repertoire.
• 13:14And you can see that
• 13:15actually gives us a pretty
• 13:16good power to predict, individuals
• 13:17who are twenty eight days
• 13:18post SARS COV two vaccine,
• 13:20shown here on the right,
• 13:21versus a control population at
• 13:23rest pre vac pre vaccination.
• 13:26We do a lot of
• 13:27work like this with b
• 13:28cell receptors. As I mentioned,
• 13:29we we do all sorts
• 13:30of methods for preprocessing, population
• 13:32structure inference, and repertoire analysis.
• 13:34And BILab makes available a
• 13:36a computational framework called called
• 13:38incantation,
• 13:39to deal with all sorts
• 13:40of analysis of these kinds
• 13:41of data,
• 13:42which is pretty widely used
• 13:44by the by the community.
• 13:45And I'll just take a
• 13:46moment to plug. We are
• 13:47having our twenty twenty five
• 13:48users group meeting on January
• 13:50thirtieth. If you are interested
• 13:51in joining us for that,
• 13:52please do please do sign
• 13:53up. It's free. If you've
• 13:55ever used our tools and
• 13:56are interested in submitting an
• 13:57abstract, the abstract deadline is
• 13:59just about a week from
• 13:59now.
• 14:01So now let me move
• 14:02to an, beyond just sort
• 14:04of looking at the current
• 14:05state of an individual,
• 14:07into the problem of can
• 14:08we predict how an individual
• 14:09respond to a to a
• 14:10certain immunological exposure? John already
• 14:13introduced this idea of of
• 14:14variable you know, human variability
• 14:16to the same immunological exposure.
• 14:18So for example, in a
• 14:19vaccine response, if we take
• 14:21a bunch of young adults,
• 14:23shown along the x axis
• 14:24here, we give them the
• 14:25seasonal flu vaccine and we
• 14:26measure their antibody response on
• 14:28the y axis. You can
• 14:29see there's a whole,
• 14:31wide variety of different, responses.
• 14:33Some people barely generate any,
• 14:36or actually generate no boost
• 14:37in their antibody titers, while
• 14:39other individuals generate, you know,
• 14:41orders of magnitude increase in
• 14:42their, in their titers. So
• 14:43we wanna know, can we
• 14:44predict the response to this
• 14:46vax to to influenza vaccination?
• 14:48And more broadly, can we
• 14:50predict the response to other
• 14:51vaccinations? Is there a universal
• 14:52signature? Is this are the
• 14:53same biological mechanisms, the same
• 14:56signatures that are associated with
• 14:57good flu vaccine responses? Do
• 14:59those also hold for things
• 15:00like yellow fever, for smallpox,
• 15:02for SARS CoV two?
• 15:04And to do that, we
• 15:05leveraged the power of a
• 15:06consortium that we're involved in.
• 15:07John already mentioned, HIPSI, the
• 15:09human human immunologic human immunology
• 15:11project consortium.
• 15:13This is some a consortium
• 15:14that's been funded since two
• 15:16thousand and ten. We've had
• 15:17one of the centers here
• 15:18at Yale.
• 15:18It's currently co directed by
• 15:20Ruth Montgomery and and David
• 15:21Hafler, and we've been part
• 15:22of it since the beginning
• 15:24in twenty ten.
• 15:26HiPSI does a lot of
• 15:27high throughput profiling of of,
• 15:29humans in diverse immunological states.
• 15:32All of the data is
• 15:33deposited into a repository called
• 15:35import.
• 15:36There's over a hundred different
• 15:37studies,
• 15:38over, almost eight thousand different
• 15:40individuals and lots and lots
• 15:41of measurements.
• 15:43As part of this consortium,
• 15:44we put together a data
• 15:45resource of all the studies
• 15:47that were done, profiling human
• 15:49vaccine responses.
• 15:50In the end, we profiles
• 15:52of thirteen different vaccines where
• 15:53we could relate transcriptional profiles,
• 15:56to antibody tighter, tighter responses.
• 15:58From the analysis of those
• 16:00data, we did two different
• 16:01things. One is we looked
• 16:02for baseline immune states.
• 16:04So could you take somebody's
• 16:06immunological state prior to vaccination
• 16:08and predict how they were
• 16:09gonna respond to the to
• 16:10the to the vaccine?
• 16:12So we could, in fact,
• 16:14generate identify those kinds of
• 16:15profiles as well as temporal
• 16:17profiles. So could you look
• 16:18a day or a week
• 16:19post vaccination
• 16:21and predict the longer term
• 16:22antibody response to vaccine? It
• 16:24turns out you can do
• 16:24that, you can do that
• 16:26as well.
• 16:27And just to show you
• 16:28one,
• 16:29one aspect
• 16:30of that and the just
• 16:31to emphasize the importance of
• 16:33kinetics and and and sort
• 16:34of timing of the response.
• 16:35So here you see a
• 16:37a bunch of those different
• 16:38vaccines that we profiled. Each
• 16:39one is shown by a
• 16:40different line, days post vaccination
• 16:42on the x axis, and
• 16:43the y axis shows the
• 16:45ability of a plasma cell
• 16:46signature to predict the antibody
• 16:48titers of the response at
• 16:49each at different days post
• 16:50vaccination.
• 16:52And you can see for
• 16:52the flu vaccine, if you
• 16:54look seven days post vaccination,
• 16:56we know that there's a
• 16:57burst of plasma blasts that
• 16:58are generally observed in the
• 16:59blood, post vaccination in the
• 17:01flu response. And in fact,
• 17:02those that plasma blast is
• 17:04just highly predictive of the
• 17:05ultimate antibody response, to flu
• 17:08vaccine.
• 17:08A lot of other vaccines
• 17:10share those characteristics,
• 17:11but notably not the yellow
• 17:12fever vaccine. So in yellow
• 17:14fever, if you look at
• 17:15seven days and you look
• 17:15for that signature, you don't
• 17:17see it. But if you
• 17:18look a little bit later,
• 17:19say say,
• 17:20ten days or or a
• 17:21couple weeks after the vaccine
• 17:23response, then in fact, you
• 17:24actually do see that signature
• 17:26and coming out as predictive.
• 17:27So in fact, it's the
• 17:28same biology in both those
• 17:29cases that drives a successful
• 17:31response, but it's very important
• 17:33sort of when during the
• 17:34course of the response,
• 17:35that we, that we look.
• 17:37If you're interested in those
• 17:39kinds of of predictions,
• 17:40I'll make one other plug
• 17:41is that we are part
• 17:42of a center,
• 17:44center for modeling immunity, along
• 17:46with Bjorn Peters at the
• 17:47La Jolla Institute,
• 17:49and we run an annual
• 17:50challenge contest for predicting the,
• 17:52vaccine response. So the CMI
• 17:54PB challenge.
• 17:55And so the idea is
• 17:57that, we provide you data,
• 17:59multiomics,
• 18:00profiles,
• 18:01at baseline, so prior to
• 18:03getting a vaccine,
• 18:04and it's your job to
• 18:05try to come up with
• 18:06a model to predict the
• 18:07vaccine response. We provide a
• 18:08couple of cohorts of data
• 18:09where we have the,
• 18:11profiles and also the endpoints,
• 18:12and there's some data that's
• 18:14held out, and there are
• 18:15there are prizes for the
• 18:16teams that, that do the
• 18:18best. So if you're interested
• 18:19in that kind of work,
• 18:20please do, check it out.
• 18:21The submission deadline is, for
• 18:23on November twenty second. You've
• 18:24got just about a month
• 18:25to develop, to develop your
• 18:27models.
• 18:29Okay. And so now I
• 18:30wanna touch, for a minute
• 18:31on multiomics.
• 18:33So previously, I was just
• 18:35talking, for for HiPSI. We
• 18:36looked a lot of transcriptional
• 18:37signatures, and those are, those
• 18:39are very powerful. But it
• 18:40was mentioned a couple times
• 18:41already. Increasingly,
• 18:43a lot of our experimental
• 18:44data encompasses multiple different types
• 18:46of measurements, transcriptional profiles, met
• 18:49metabolomic profiles, proteomic profiles,
• 18:52all together. And one one
• 18:53of the challenges, how do
• 18:55we integrate those data together
• 18:57in order to come up
• 18:58with, predictive
• 18:59responses and understand the the
• 19:01immunological
• 19:02mechanisms that are driving those
• 19:03differential,
• 19:04differential outcomes.
• 19:06And here I wanna highlight
• 19:07a a method that, we've
• 19:08been developed in collaboration, with
• 19:10Liang who spoke spoke earlier,
• 19:12and a graduate student, Jeremy
• 19:14Geege, another CBB graduate student.
• 19:16And the idea here is
• 19:17a lot of times where
• 19:18we have these multi omics
• 19:19data, the common way to
• 19:20analyze those data is to
• 19:22take the data and do
• 19:23some dimensionality reduction. Right? So
• 19:25you have those, you know,
• 19:27potentially hundreds of thousands of
• 19:28measurements of genes and proteins
• 19:30and and and metabolites
• 19:31to reduce the dimensionality of
• 19:33those into some smaller, some
• 19:35smaller space and then do
• 19:36machine learning. And one of
• 19:37the things we found is
• 19:39that doing doing it that
• 19:40way actually could be suboptimal
• 19:41because you would actually you
• 19:42could actually miss variation
• 19:44that was important for the
• 19:45response,
• 19:46of interest. And so we
• 19:47developed a method called, SPEAR,
• 19:49which is a sparse supervised
• 19:50Bayesian factor model,
• 19:52where in generating the lower
• 19:53dimensional representation,
• 19:55we took into account both
• 19:56the variation in the multi
• 19:57omic data as well as
• 19:59the response of interest. And
• 20:00the the method learns to
• 20:02adaptively balance how much to
• 20:04try and explain the variability
• 20:05in the data and how
• 20:06much to try and explain
• 20:07the response of interest in
• 20:09coming up with those lower
• 20:10dimensional,
• 20:11lower dimensional factors.
• 20:13And finally, what I'll end
• 20:14with is the the importance
• 20:15of genetics.
• 20:16So so all of these,
• 20:18responses and all of these
• 20:19signatures are often modulated by
• 20:21the genetic background,
• 20:22of of the individual. And
• 20:24in particularly, it's an emerging
• 20:26area to understand to what
• 20:28extent the genetics of the
• 20:29b cell receptor,
• 20:31repertoire
• 20:32and the components of the
• 20:33b cell receptor
• 20:34influence the ability of people
• 20:36to respond to infection,
• 20:37and vaccination.
• 20:39And I'll show one example
• 20:40here,
• 20:41from, showing the influence of
• 20:43a b cell receptor variable
• 20:45gene, a gene called, IGH
• 20:46v one one sixty nine,
• 20:48on the flu vaccine response.
• 20:50So here you see,
• 20:52there's a position, in this,
• 20:54in this gene, which can
• 20:55either be an f or
• 20:56an l. And depending on
• 20:57the genotype of the individual,
• 21:00that individual will generate different
• 21:01titers of broadly neutralizing antibody
• 21:04in response to, in response
• 21:06to the flu vaccine. You
• 21:06see the micro
• 21:08titers, on the y on
• 21:09the y axis here.
• 21:11And so it's really an
• 21:12emerging area to try and
• 21:13understand how the genetics of
• 21:15the antibody repertoire
• 21:16influence the response to vaccination,
• 21:18and infection.
• 21:19And there's not been a
• 21:20lot of associations,
• 21:22that have been discovered so
• 21:23far. In part, I think
• 21:24this is because in all
• 21:25of the genome wide association
• 21:27studies that are done, the
• 21:28immunoglobulin locus is generally not
• 21:30included in any of those
• 21:32studies. So it's not there's
• 21:33very low representation
• 21:34on the, SNP arrays that
• 21:36are that are used.
• 21:37And in general, one can't
• 21:38get this from the short
• 21:39read sequencing
• 21:42data because the region is
• 21:43highly, highly repetitive.
• 21:46And so with the postdoc
• 21:47in the lab,
• 21:48Dylan, who's one of the,
• 21:50Gale BI data science fellows.
• 21:52He's actually developed a,
• 21:54an approach now where we
• 21:55can go in from short
• 21:57read DNA sequencing data and
• 21:59genotype the immunoglobulin
• 22:00locus. And the way he's
• 22:02able to do this is
• 22:03by leveraging,
• 22:04a pangeneome graph. So this
• 22:05is a graphical representation of
• 22:07a genome where instead of
• 22:08having a single linear reference
• 22:10as is commonly used today,
• 22:11we can actually represent and
• 22:13encode the diversity,
• 22:14of the population into a
• 22:15single graphical representation of the
• 22:17genome.
• 22:18And by and by leveraging
• 22:20that structure, we can take
• 22:21those short read, DNA sequencing
• 22:23data,
• 22:24map it to this pan
• 22:25genome graph, and discover variability
• 22:27in the, in the immunoglobulin
• 22:28locus. And what we're gearing
• 22:30up to do now is
• 22:31to actually use this method
• 22:32on biobank data because we
• 22:33now have public data sources
• 22:36of massive numbers of individuals
• 22:37with, short read DNA sequencing
• 22:39data, and a lot of
• 22:40interesting clinical and other, and
• 22:42other phenotypes.
• 22:43And so we're we're planning
• 22:45to use this approach to
• 22:46mine those biobanks and find
• 22:47interesting,
• 22:48connections between the immunoglobulin locus,
• 22:51and and those outcomes.
• 22:53And so I think I
• 22:53will, stop there because I'm,
• 22:56out of time. And,
• 22:57thank you and take any
• 22:58questions.