October 26, 2023 • 44 mins
Welcome to the Olink® Proteomics in Proximity podcast!
Below are some useful resources from this episode:
Published study of primary focus
Koprulu M, Carrasco-Zanini J, Wheeler E, Lockhart S, Kerrison ND, Wareham NJ, Pietzner M, Langenberg C. Proteogenomic links to human metabolic diseases. Nat Metab. 2023 Mar;5(3):516-528. doi: 10.1038/s42255-023-00753-7. Epub 2023 Feb 23. Erratum in: Nat Metab. 2023 Mar 19;: PMID: 36823471; PMCID: PMC7614946. https://pubmed.ncbi.nlm.nih.gov/36823471/
Laboratory, first author, and corresponding author of the study
· Public Health University Research Institute (PHURI), a multidisciplinary research center to drive personalized healthcare: https://www.qmul.ac.uk/phuri/about/
· Mine Koprulu (first author), PhD student, University of Cambridge: https://www.linkedin.com/in/mine-koprulu-497659b9/
· Dr. Claudia Langenberg (corresponding author); Director of PHURI, Queen Mary, University of London; Professor of Computational Medicine, Berlin Institute of Health at Charité: https://www.qmul.ac.uk/phuri/our-people/professor-claudia-langenberg/
Olink tools and software
· Olink® Explore 3072, the platform that measured proteins in this study with a next-generation sequencing (NGS) readout: https://olink.com/products-services/explore/
UK Biobank Pharma Proteomics Project (UKB-PPP), one of the world’s largest scientific studies of blood protein biomarkers conducted to date, https://www.ukbiobank.ac.uk/learn-more-about-uk-biobank/news/uk-biobank-launches-one-of-the-largest-scientific-studies
Genotype-Tissue Expression (GTEx) project, a biobank and open-access database to study tissue-specific gene expression and regulation: https://www.gtexportal.org/home/
European Prospective Investigation into Cancer (EPIC)-Norfolk study, a prospective cohort of middle-aged individuals from Eastern England: https://www.epic-norfolk.org.uk/
Genome Aggregation Database (gnomAD), the largest publicly available collection of population variation from harmonized exome and genome sequencing data: https://gnomad.broadinstitute.org/
Would you like to subscribe to the podcast on your favorite player or app? You can do so here:
Apple Podcasts: https://apple.co/3T0YbSm
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:05):
Welcometo the Proteomics and Proximity Podcast
where your co-hosts, Cindy Lawleyand Sarantis Chlamydas, from Olink Proteomics
talk about the intersection of proteomicswith genomics for drug target
discovery, the application of proteomicsto reveal disease biomarkers,
and current trends in using proteomicsto unlock biological mechanisms.
(00:26):
Here we have your hosts, Cindy and Sarantis.
Hello, everyone.
Thanks so much for joiningProteomics in Proximity.
I'm one of your co-hosts, Cindy Lawley,and I have with me my other co-host.
Hello, everybody.
I'm Sarantis Chlamydas, happy to be here.
Very good.
So so today we are joined by Mine Kopruluand Claudia Langenberg.
(00:51):
We are talking today abouta wonderful paper that came out in Nature
Metabolism in March [2023] called "Proteo-genomic links to human metabolic diseases."
Very exciting paper. Byby Claudia Langenberg and Mine
Koprulu standards,a modest sample size but amazing findings,
so I think it's really enabling.We'll dig into that,
(01:13):
but first, let's introduce our guests.
Sarantis,do you want to do the honors?
Absolutely, we'll start with Mine.Mine is a PhD
student on Gates Cambridge scholarshipand supervised by Dr.
Langenberg.
Before her PhD, she received a Bachelorof Studies in Human Genetics at UCL [University of College London]
and then subsequently she completed
(01:35):
a Masters of Philosophy in GenomicMedicine at University of Cambridge
working in genomics data and
proteomics data and dealing with population studies
and the UK Biobank data with us.
And we are really looking forward to hearingabout your project and the paper about proteogenomics.
Yeah,
and digging into how you gotwhere you are.
(01:57):
Anything you want to add to that, Mine,that you want people to know going in?
No, I think that's a very clearintroduction and it's lovely
to be here and I'm looking forwardto sharing a bit more about
both our journeyand our recent publication.
Awesome, fantastic.
So then I have the privilege and and the
(02:18):
opportunity to introduceClaudia Langenberg.
Now, Claudia is probably someonewho needs no introduction, but
the I think the impact
that she's had in the publicationspace is huge.
Last I saw, she had I think this is in
2021 over300 peer reviewed publications at her
(02:40):
very young age and and I think
over 40,000 citations at that point.
And that's already several years old.
So this is athis is a world-renowned
scholar that is working with Mine,
And of course, many of her publicationsare in
journal serieslike Nature, JAMA, Lancet,
(03:02):
very, very
prestigious and very impactful,working with massive datasets.
So I think of her as a large scalepopulation epidemiologist,
such a big word,but it integrates many of the omics
and I think she keeps an eyeon the technology advances to be able to
build systematic approachesto understanding human biology.
(03:25):
So Claudia, I think now you're in you'reat Queen Mary, University of London.
I think when I met you originally,you were at the Berlin Institute of
of Health and Charité.
So I'd love for you to just tell usa little bit
about your position nowand your affiliations, if you don't mind.
Well, thank you so muchfor the extremely kind introduction.
(03:48):
And I think the most complimentary thingwas about the young age,
which is not only a complimentbut a lie,
so thisI think, puts everything in perspective.
And yes, also I should say the workthat we've done that
Mine is a studentat the University of Cambridge
where I've had a dual affiliationand a long career.
And prior to me arrivingat the Berlin Institute of Health,
so originally be Germanyand training as a doctor in Germany.
(04:12):
I came to the UK and I was in Londonoriginally where I did my Ph.D.
in epidemiology and a Master's of hygiene.
And before that I went to Cambridge
to focus more specifically on genomicsand then later other omics.
And so yeah, it's a huge privilegeto be here
and it's an even bigger privilegeto work with talented people like Mine
and others in our team.
So it's a real team effortand a lot of fun there to do that.
(04:36):
So it does seem like you have fun.
And I'll tell you what, wheneverI talk to Mine and when I've seen her
present,she's always got a big smile on her face.
So I would love to dig into your journeyto end up in Claudia's lab,
Mine.
Do you mind giving us a little backgroundand how you got interested in science?
(04:57):
I mean, feel free to startwherever you want.
Yeah, sure.
I've always been very interestedin science, and especially biology.
But especially I always wanted to dosomething with my career, which improves
life, supports and contributesto the society in a meaningful way.
And I've said givenI was quite interested in biology,
human geneticsespecially quite interested me
(05:21):
because a lot of the sciencesare quite established in the sense
that, you know, thehuman genome is very recent compared
to many other scientific fieldsand there's so much unknown.
And I think that is especiallywhat interested me
as well as its potential to, you know,improve the current healthcare system
and the clinical translational potentialof the human genetics findings.
(05:45):
So that was my inspirationto go into that.
And unlike some other people,I think I've been in that quite
narrow specific path ever since then.
So I did my bachelor's at UCL
as Sarantis mentioned in human genetics -it was just strictly.
And after that I did my master'sin University of Cambridge
(06:05):
in genomic medicine, which is where I gotintroduced to bioinformatics.
So I did my Masters projectat Sanger Institute working on UK Biobank
as well as whole genome sequencing datafor each population isolates
at least
in that lab at the time.And after that I moved back to back home,
(06:26):
which is Turkey for me to work on raredisease genetics.
So I spent two years back in Turkeyat a
universityresearching rare disease genetics.
So that was more studyingwhole exome sequencing data
from consanguineous marriages
or consanguineous parents havingdisabled children trying to better
(06:50):
pinpoint the exact genetic changesthat caused such severe, rare disorders.
And in the meantime, while I was backhome, I was thinking about my PhD
and what I might want to pursueand multi-omics integration.
So better understanding the different "omic" layersand how that can actually improve
our understanding of genomic studiesand their clinical translation
(07:14):
was what interested me.
So I got in touch with Claudia
and we had a brief interviewand that was the beginning
of such a beautiful collaboration,at least on my part.
And then starting alittle bit about your paper -
What were you able to understand about the data,about this EPIC-Norfolk study?
(07:36):
The EPIC-Norfolk cohort. What are the characteristicsof this cohort and what makes it so special?
Sure, the EPIC-Norfolk studyis a population cohort
that was first establishedin beginning of 1990s,
so it has approximately 30,000
participants that were followed up.
Phenotypic research in terms of different
(07:59):
behavioral and sort of healthcharacteristics for several years,
as well as linkage to their health records.
And we had proteomics samples
measured from 3000 ofEPIC-Norfolk participants.
But Claudia, if you have anything moreto add on the EPIC reference study ...
I think I can say one of the opportunitiesthis study has offered
(08:21):
and I think the beautyof such perspectives cohorts such as EPIC-Norfolk,
and UK Biobankand so on, things that were set up
in the past is that we benefit from the samplesthat were stored at baseline
in liquid nitrogen.
And those samples are so valid
because anything you can do in the futureand now you measure 3000 [proteins].
And you measure 5000, you measure tenand more thousand proteins.
So anything we usenow we won't be able to use in the future.
(08:44):
So we're always very cautious
of that and hencecautious with our sample use.
So it's really important that proteomictechnologies now use such a little sample.
So you could have used more samplein the past to measure three proteins.
So to be able to [measure] thousands
in such a small amount of bloodis absolutely amazing.
(09:05):
And to have the opportunity, you know,
PIs when our PIs of thosecohorts had the foresight of setting this up.
Nick Wareham in Cambridge and before thatKay-Tee Khaw, others who let us study.
And other cohorts that we hadthe privilege of using the samples for.
It's really,really amazing to be able to do that.
That's one thing.
But the other thingthat's important is as time
elapses, these codesbecome even more valuable because,
(09:27):
you know, sadly people have events, theythey develop diseases, they die.
So if you then look backwardsbecause you have samples stored,
provides the opportunity to have suchan efficient design of doing it,
you don't have to measureeveryone. But you can have a kind of a
very nifty design in choosingpeople who have developed a disease, i.e.
(09:47):
new onset disease, butthe sample was stored before they had it.
So you avoid thiskind of reverse causation by which disease
impacts your proteome and henceyou can't really just associate.
This is what we exactly didin this context.
Mine's study was one of the projectsthat we did on context of that design.
But you can also look at
different diseasesbased on the people who developed them
(10:09):
and then studied. Use that baseline sampleand then you use a big control cohort
of people who serve as the controlsfor many different diseases.
It's called a nested case cohort study.
And it's a really beautiful design,but that can only be used
if you had the foresight of setting upa large prospective cohort like that.
So we were very gratefulto all the participants.
EPIC-Norfolk and the PIs, of course,who's enabled the use of that,
(10:31):
and also coming to the costof some of these kind of very,
you know,
informative molecular technologysuch as yours is.
Of course they're not cheap
and hence that meansif you can use a design that minimizes
the number of samplesyou have to use for given purpose,
of coursethat's incredibly useful for us. So
and I think I just
(10:52):
wanted to comment on this,this liquid nitrogen aspect, this ability,
I think this foresight -I work with a lot of groups
that do a lot of different thingswith large population health studies.
There are very few cohortsthat had that much
attentionto reducing pre-analytical variation
and tracking itas is documented in this cohort.
(11:16):
So that's exciting to see.
I think it's maybe less importantfor proteomics
than metabolomics as an example,but you work in all of that.
And so yeah,I just wanted to underscore that point.
Yeah, sorry, Sarantis, please go ahead.
Okay.
I will like Mineto summarize the take-home messages
of your paper.
and what's so excitingabout this paper.
(11:40):
Yeah.
First, and just alsoclarify a bit [about] the paper,
I was the first author,
but of course it was a team workand everyone contributed a lot
making this work possible,including the EPIC-Norfolk participants
of course.
Just to summarize the paper,
I think I could just briefly saywe looked at the sort of systematically
(12:01):
linking genetic variants, blood proteinlevels as well as disease risk data,
to be able to pinpoint causal genesand proteins that underlie diseases.
So just to give a bit of background sinceenablement
of the genotyping technology,little studies have been conducted
(12:22):
associating genetic variantswith different disease risks or
disease susceptibility.
And today 200,000 genetic variant disease
associations have been establishedand are publicly available.
Looking at such numbers and figures,
we would assume we knowthe biological basis for all diseases.
(12:44):
However, we all know that's not at allthe case, and I think that's
where proteomics and differentsorts of omic layers come into play:
in helping us better understandthe disease mechanisms of
actually what's happeningunderlying the diseases in our body.
So in this studywe had samples from 3000 individuals
(13:06):
measuring three approximately 3000 blood protein levels.
So we first looked at the geneticregulation of the cis regions.
So the cis regions are the protein-encoding regions
and sort of flanking regions aroundthe gene for the protein target itself,
because of our moderate samplesize, as we've already mentioned.
(13:28):
So we looked at the cis genetic regulationor different blood
protein levels,some of which had never been targeted
before, given the recentOlink platform [Olink Explore 3072]
and then we have used that knowledge,that sort of protean genomic knowledge,
to better understand
causal genes of proteins thatunderlie diseases in a systematic manner.
(13:49):
So we have firstlooked at shared genetic regulation
for different disease outcomes and bloodprotein level regulation,
pQTLs as we call it - protein Quantitative trait loci.
And we identified 224 targets
that regulates 500,
(14:11):
approximately 500 different traits.
And we also refined the causal genesor proteins
for 40% of the previously establishedgenetic risk loci,
which was which sort of highlightedthat even moderately sized
proteogenomic studies can contributeto our covering of novel biology
(14:36):
for consanguineous risk loci that werepublished in the literature.
And finally, we looked at the convergence
of the pQTLsstudies and the regulation
wherethe rare variant gene burden analysis.
So, on one hand, comparing the lossof function of the genes with that sort
(14:56):
of genetic regulation of the proteins.
And so I wanted to click back on this: the disease associations (15:00):
undefined
that were enabled by genetics. I think the
the ability to make those associations
in small sample sizes was likelow hanging fruit in the early days
of that sort of GWAS era,which I think of maybe early 2000s, right?
(15:21):
2005, 2006
that's when those things startedreally ramping up discoveries
and then they became harder and harderto make those associations
as those really strong,
strong associationswere discovered and documented
and then larger and larger populationswere needed to make those associations.
(15:45):
I think what you're saying
is that these other layers are helping you
to to do morewith those more modest populations.
And Claudia already made the pointabout the costs of layering
on metabolomics or proteomicsor these additional omics.
Can you say something about what
(16:05):
what that enables and what you thinkwill happen in the future there?
Yeah, of course.
So basically, as I mentionedand as you've mentioned already, thousands
of genetic variants,a disease associations are being made
and those were very valuableand contributed to our understanding
of diseases to some extent.
However, the majority of those associationsfell into noncoding regions of the genome,
(16:28):
meaning it was quite difficultto actually interpret which causal gene
or protein were acting incausing the disease.
Hence, it was quite difficultto understand the pathways
and the mechanisms.
And, well, ifwe don't have a functional target,
it is difficult to actually buildmore effective and safer
therapies or repurpose existing ones.
(16:50):
So in that sense, having the initial layerof the proteomics states actually
helps us to pinpoint a functional entity
that plays a role in the disease.
And so
I said there's
certain statistical methodsthat help us better understand
the shared genetic regulationof both the disease risk
(17:13):
and the blood protein levelsor abundance of certain proteins.
And we see strong statistical evidencefor a shared genetic
regulation done within a large regionof many candidate genes.
We can actually refinethat to a single candidate gene or protein
for particular diseasesin a systematic way
which can then be usedfor as I mentioned, intervention
(17:37):
or more targetedtherapy, safer therapies.
So in that sense, even moderately sized,more regular pQTL studies
can really help us better understandthe disease and pathways that are involved
and also build more effective therapies.
I think we're on a path.
I think this is a path for discoverythat is, I think just going to ramp up.
(17:59):
I think very similarly to what we've seen,those discoveries enabled by genetics.
I'm really excited about that.
I always think about
these different diseasesthat might share pathways in common,
almost like a Venn diagramand understanding the complexities of
of why certain proteins or protein
(18:22):
pathways are related to like,I don't know, caspase,
I think is like showing upin breast cancer and asthma. It's
an example. I thinkthose diseases are so disparate to me.
I wonder if youthink that proteins or pathways
that are showing up as causaland I think you make a good point about
(18:42):
why causality is so importantfor therapies.
Do you expect those same proteinsto be causal in other diseases?
Like what doyou think about that?
You're much more versedin seeing the data.
So I think that is sort of the beautyof our study design
that we sort of approachwith different layers
(19:02):
of biological datastarting from the genome
all the way to the phenomein a hypothesis-free manner.
So doing that and looking at the datasystematically without any sort of prior
hypothesisallows us to see what the data tells us.
So as you mentioned, certainly in lookingat what we have discovered in our papers,
(19:23):
that genetically anchored proteogenomics studies so the proteomics
can help us discovermolecular hubs.
All associations of proteinswith diseases that we wouldn't
have predicted otherwise from justthe general literature or prior knowledge.
And that is actually quite interestingto follow up
because, yeah, the data uncoverssomething that was unexpected.
(19:47):
But likewise we can also seequite specific protein disease pairs,
which can also allow usto have quite specific therapies
for potentially not interveneddiseases before.
But just one more thing to addis that we
currently do not have not a very full completepicture
(20:08):
of the proteomics, so we're only ableto do that for the proteins
we are able to target.
So for future directionsyou asked about,
having a more complete ideaabout the full landscape of proteomics
would be ideal to better understandthose molecular hubs
that we have just talked about.
But so maybeI can add one more thing to this,
(20:31):
because as Mine's workother proteogenomics work
than we previously did in the pasthas shown is exactly the beauty is that, as
you scale up the number of proteinsand as you scale up
the number of diseases,you can look at this kind of proteo-
genomic approach that doesn't requirethat each of these layers is measured
in the same sample size.
You can utilize the power thatyou have across any biggest genomic study
(20:54):
for diseases that you have to geta reaction, a proteogenomic study.
And that is the beauty of looking exactlylike you say, Cindy,
at the overlap of a specific gene to proteinacross any disease that you can look at
because the community is sharingdata openly or the summary statistics
for each of these diseases is sharedopenly for most diseases. Cancer is
(21:16):
lagging behind sadly,a little bit in terms of the openness.
But for many diseases that is there,it enables us to do exactly
that (21:23):
to draw a whole map from gene to protein to disease
and is that link coincidental
or is possibly causaland a rare genetic signal?
And that's such a good wayof prioritizing what you then use for
experimental work downstream
because of coursethis is in the computational approach
(21:44):
and not an outcome of proof,but it's such a good and data driven way
of prioritizing
links betweendiseases, links for where we have already,
you know, specific drug targets, new ones,potentially adverse effects.
So it's a really versatile wayof looking at all of that.
And then just maybe to add one thingagain, as you said, is
(22:06):
you can kind of ramp this upas you increase the number of proteins
or as you increase the density
of your genomic arrayor the coverage of rare variants
and by sequencing,
and so on. Of coursean important part of really making
this more usefulis increasing the phenotypic spectrum.
(22:27):
And that really is only possibleif we move away from diseases. We all
study opportunistically to diseasesthat you can't really easily measure
or nobody's interested in, but they're stillreally important for patients and a lot
less headway has been made in termsof understanding that genetics
releasing the summary statisticsfor those studies, and that's really where
(22:48):
huge studies come in that have electronichealth record data from GPs [general practitioners],
from hospitals, from death certificates,and bring all of this together.
And that's why, for example, UK Biobankbut also Fingen
and many other endeavors around the worldcan really help us
to not just increase the molecular sidebut the phenotypic side.
(23:09):
And that is so important to alsolink diseases which have not really been
so much in the center of attentionand but need to be.
Can you give an example of one of those?
Are we thinking rare disease here?
So I think it can be
across the frequency spectrum,it can be across specialties.
I think these diseaseexamples that I would choose,
(23:30):
others that possibly are not as easily
diagnosed, are not severe enoughto always require
hospitalization,because I think most people around
the world have tried to gettheir hospitalization data
I see decoded is a relatively easy,but the data, for example, in the UK
that comes from England, from primary carerecords, is harder to map,
(23:51):
is kind of bit more diversity in termsof systems, the data structures and so on.
So diseases that are predominantly managedand diagnosed in
primary care are morewhere we need to move towards.
Oh, that's so interesting.
Yeah. I have a bit of a
technical question here.
How do you make the selection
(24:11):
of false positivesand how do you define
these in your analysiscompared to
a lot of technologies, for example, that you applied before?
You're
asking about false positives? Yes.
So in our analysis, we try to be careful
given we're workingwith data and diseases, as you mentioned.
(24:33):
So in terms of our analysis,we always use quite a rigorous
threshold to reportwhat is statistically significant.
So in terms of our statistical threshold,we always called for genome wide
significance times the number of proteinsthat we are including in the study
to sort of minimize the error, includingthousands of targets in this case.
(24:59):
So we usually go for quitea rigorous statistical threshold
as well as goodQC before we feed in the data,
of course.
So that's great.
So some of these pQTLs
may fall in non-coding regions.
(25:20):
Do you have any idea about the type of these regions?Maybe have a promoters there,
I don't know the answers.Do you have an idea and can you map these regions?
So we mapped
all the variants in terms of whatthey're predicted to be in the genome.
So we look at the proportion of proteinaltering variants, those variants
(25:40):
that fall into the proteincoding region itself and alter
the shape or the structure ofthe protein.
And the percentagethat fall into non-coding regions.
However, there's quite limited knowledgeabout the functional characterization
of the non-coding variants themselvesacquired from particular groups studying
(26:01):
particular genes and more sort of cellularor functional models.
So in terms of our work so farwe have only computationally annotated
the predicted consequence of the variants.
However, of course further follow upof analytes and what exactly those
non-coding variants do in termsof cis regulation of the protein itself.
(26:23):
That's great.
So I think a lot of studiesand then from the literature,
as you as you mentioned, for differentproteins from different
mice knockouts,
you can have an ideaabout how they are regulated.
Is there any plan to follow up, like moremice knockouts
in the future to followsome of these targets or collaborations
(26:45):
in this respect?
If you can share this?
I mean, we of course always welcome
any sort of collaborationson the functional characterization.
The idea of doing these studiesis not to put in a long Excel sheet,
as we have done so far
in terms of our supplementary table,but rather genuinely understand
the biological mechanismsthat underlie diseases
(27:06):
and how that can contributeto clinical translation
of some of the findingsand the sort of longer runs.
So, with that, an open callto anyone who finds any of
that targetsthat we highlight interesting.
We are of course
very keen to functionally characterizeand better understand the mechanisms.
I thinkin the paper you mentioned something about -
(27:28):
I was just looking for the reference,something about cellular models
and the history ofof demonstrating functional associations
or understandinga function around cellular models
and how this approach
can be a complementary wayto add an understanding of function.
Do you mind talking a little bitabout that,
(27:50):
like where this fitsinto our traditional approaches?
Like where it might fit inin understanding
what we're learning from single cellor spatial
work that some of those technologiesare really advancing right now?
I'm curious your thoughts on those.
Of course.
So what we do in terms of our teamis more bioinformatics.
(28:14):
So what we work with is what the data and
sort of predictionsand better understanding the computational
and statistical conclusionswe can draw from the data.
However, as we all knowand as you've highlighted,
they are complementary waysin better understanding these.
And basically we are very interestedto see whether
(28:35):
what we observe computationallyand statistically actually
translates into biology,starting from cellular models,
building up to animal modelsand sort of human biology.
So what we're observing almost
quite isolated
looking at singularsort of targets and singular
genetic variants. It will be quite interestingin whether our conclusions hold or
(29:00):
sort of variations fromour conclusions we see in the different models.
So I definitely agreethat they're complementary
and there's ways of integrating thoseknowledge to build a more comprehensive
or sort of holistic understandingof the biological mechanisms.
Oh, I love it.
I love the backgroundyou have both in an understanding
(29:22):
of epidemiologyand your mad bioinformatics skills.
Where does machine learningfit into all of this?
Is it something you've -
yeah, I'll just stop with the question.
No, of course, machinelearning and artificial intelligence
is certainly fieldsthat are very quickly developing
and are grabbing a lot of attentionat the moment. So what
(29:46):
we use are considered basic machinelearning models themselves.
But in terms of employing machinelearning or artificial intelligence to
sort of prettier genomics for theseor more biological studies at the moment
is rather challengingbecause as we all spoken about, this large
high triplet biological datais recently being generated
(30:09):
and they are so recent,but we're spending a lot of time
better understanding what they meanand what they actually are.
And frankly, the biological datathe way they are actually violates
a lot of assumptions that machinelearning and artificial intelligence
models make.
So I think in the future,they can certainly be very useful,
but we certainly need to be cautiousand better understand
(30:33):
thembefore feeding them into the models,
I personally say.And maybe I can add something to this.
So I'm currently sitting inwhat's called the DERI.
So when I came to QMU, or QueenMary University of London last September,
I had this opportunity to set up thisnew institute for precision healthcare.
And it's kind of quite uniquebecause it's a cross functional institute.
(30:55):
So it really draws in a very broadrange of expertise and so [there are] big plans
and developments for a new life sciencesbuilding, and that's all quite away.
So we have to wait for this.
So we need an interim space and we're verylucky and it's actually not coincidence
WhereI'm sitting right now is called the DERI.
It's the only other cross functionalinstitute that QMU has and DERI stands
(31:16):
for the Digital Environment ResearchInstitute, which is focused on AI.
And so it's not just in healthcare, it'sAI across a broad range of applications.
You know, from games - we're sitting onthe second floor with the games people,
which my my kids think I have the best jobin the world.
Actually what they don't know
when they have conversationsabout board games is
(31:37):
I have not heard of a single one of them,so I can't mention really anything else.
But anyway, having said that, this is theenvironment in which it can flourish.
So we focus on healthand on healthcare data
and DERI hasas part of its remit as well.
So it's really importantto bring this together.
And, you know,going away from,
(31:58):
you know, this pitch onhow it's important, the concrete thing
that we're already doing isthe molecular data, which is so complex,
you need efficientand unbiased ways of data reduction.
That, at the moment, is most of the usefor machine learning in our
arena of work.
(32:18):
So where we kind of try to, let's say,
predict diseases,how are you going to prioritize?
Nobody wants to go into the clinicand measure 3000 proteins
to predict a disease.
Do you even have any benefit of
what's the incremental valueof measuring proteins
once you have a core set of tenmost important ones?
(32:39):
So those are all questionsthat machine learning can help us to
to address more systematically,and that's incredibly useful.
So is a steep learning curve for us.
This one, it's not our bread and butter.
We're not the methodological expertsof developing it,
but being in an institutelike this one, for example, means
that we're only very a short way removed
(33:00):
from the people who do develop thesemethods and that we at the right time
can employ them, can test other useful,could they be biased.
And so that's really, reallyyeah, it's a great opportunity, I think.
Well,I think those people that are developing
those methods aren't going to havethe biological know-how to know
when they've overtrained those methodsand are going off on a spurious tangent
(33:21):
so the context I think is essential.
I think Mine makes that point.
Yeah, exactly.
It's this exactly.It's a team effort, right?
And you need the people, havethe samples, and the clinical knowledge.
You need the people who havesome other kind of bioinformatics,
computational and you need the,you know, method developers.
So it's beautiful.
Nobody can do it all.
(33:41):
So, you know, I think it's a great.
It's CQ, right?
It's collaboration quotient, right?
You have your IQ, your EQand CQ. From my experiences,
once you develop your collaboration quotient,
which you're teaching your studentsand your postdocs and all of that
if they don't already have it,
Claudia, is once you have it,you can't go back?
(34:03):
I don't think
once you experiencethe joy and the ability to work together
cross-functionally, it's reallyremarkable what can be accomplished.
Well, coming back to the question that
Sarantis was asking,and I think it's a very good one
and Mine has said so.
The problem that we have encounteredis, you know, is Mine
has come upwith these beautiful candidates.
And it's almost there's two thingsthat are important to kind of what
(34:27):
she said
so well, which is that now using humansas the model organism
is a great way of prioritization.That's number one.
But the second one is kind of, you know,if you come from a Bayesian framework,
you really want to find out what'syour greatest chance of success?
How do you increase your priorof this having success?
And given that the experimental procedureis expensive, it's lengthy
(34:48):
and it can go wrong in so many ways,
It is really a great opportunity.
to prioritizeon the basis of human proteomics
as Mine has said.The problem is: how do you get people
who have the experimental set upto come to our results and take it?
(35:10):
Because that's not as easy as we thought.
We think, "Oh, here!"We're so excited to see it.
We go to the world expert in thisand bring them our example,
it's hard.
It's hard to motivate peopleto step out of what they're already doing
and focus on your finding.
It's hard if you have somebodywho's focused their whole life
on a given pet protein to say,how about this one?
(35:32):
So we need help,I think, to try and learn
how we engage the relevantclinicians and molecular biologists
and other people to also use ourresults rather and at least,
you know,complement what they're already doing
in order to be able to follow some of itup, because we do depend
on the functional valueand the expertise of those people.
(35:55):
So how do we reach out?How can you help us
to reach out to the relevant peopleto really make it worthwhile?
What are the titles of those people?
I mean, what are their jobs?
You know, like,I know there's more than one, but
I think of translational,
certainly translational scientistor implementation scientist
or maybe those are pieces of it.
(36:17):
But you know, who are the onesthat we need to
listen to this?
So I think it depends on the stageof translation
that you're at. For really earlyexperimental work,
it's a very different set of peoplethan it would be for the people
if you have a different kind of work
where you really are ready to maybe,possibly already consider an initial trial,
(36:38):
then you would need a relevant clinicianor a head of that clinical department
to enable that kind of set up within that,let's say, hospital.
So I think it really dependson what level you're talking.
The initial level I think iswhat's the experimental follow up
and validation, because what we do isyeah, the statistics and probability an an error,
but it's still an associationwhere we're under no doubt we do not prove causality.
(37:01):
And so that kind of functionalfollow-up is crucial.
And those are the people at the first stage.
Think actually you don't knowthe cause of some of these diseases,
you don't knowthe cause of the effect, right?
You don't know if this is what you seein the protein level. Is it a cuase
or the effect of the disease?It could be both. The last comment
to me I really like when
you see the co-localization with eQTLsfrom different tissues
(37:24):
like does what we see in the plasma proteome
reflect what happens in the tissues?
And do you have any comment
on how you see these correlateand what is the distribution
of the singular featureto the plasma proteome?
You have some examples,some ideas on that?
And a shout out to gnomAD and GTEX
(37:46):
I will say fantastic resources. Yes.
No, we are very gratefulfor all the publicly available resources,
ranging from the eQTL studiesall the way to the GWAS summary statistics,
which have made our work possible as well.In terms of the eQTL overlap,
that is rather challengingin our field in the sense that now
(38:07):
we see, for example, in our study
that only approximately 4% of the pQTLsthat we see
are in close linkage disequilibrium
that we deal with an eQTL
and certain examples,we see beautiful overlap, right?
Like we find an examplewith a disease for that
(38:28):
act in a particular tissueand we see co-localization,
quite strong co-localizationwith a tissue of interest
where the story becomes quite beautifuland everything makes sense.
However, there is also the flip of the coin
where we see in some examplesquite limited overlap
between the pQTL and eQTL data.
And although our current paperhasn't really focused on that,
(38:51):
I think that is something that weas a community need to better understand
where there is overlapand where there's a lack of
and what might be the underlyingreasons that would be.
I think this is a great
but to mention the really interestinglink to type two diabetes
that you found in here that that we wantsomeone to follow up on, as Claudia mentioned.
(39:16):
Do you want to just summarizethat really quickly, that pathway?
Of course,one of the examples that we have noticed
or we have highlighted in our paperis the gastrin-releasing
peptide, or GRP for shortand its link with type two diabetes.
So what we saw in our paper is that
(39:36):
beautifully different layersof biological data,
so evidence from my studies,evidence from human data as we have been talking
about and different sort of similar modelsall overlapped with the same conclusion.
So what we've identifiedwas that the higher levels of GRP in human
(39:56):
plasma were co-localizing
with lower risk of type two diabetes.
And when we integratedin the different sort of
intermediate layers,which are different body sets and
distribution traits,so where we actually accumulate
fat in our bodyas well as the overall fat,
(40:16):
what we observedwas that higher levels of GRP
was leading to less fataccumulation overall
in our body, leading to lowerrisk of type two diabetes.
And as I mentioned,there were previous studies
that were publishedwhere human recombinant GRP
(40:36):
led to actually reducedintake and weight loss.
So we have highlighted in our paperthat GRP can be a new example
or a potential therapy for typetwo diabetes by decreasing or
lowering the fat accumulationin our body.
(40:58):
It's actually a hot topicfor the weight loss.
A lot of companies now are trying to follow this
game of weight-loss strategy.
I think that weight-lossdrugs are really interesting,
and I am looking forwardto people following up
on this.
And I love the use of
the word beautiful, right.
(41:19):
But it's beautifulbecause these layers are all agreeing.
It's giving us a preponderance of evidencethat gives us a lot of confidence,
that gives us confidencein the other results that you see, too.
Right.
So the fact that you've builtthis systematic map
of these potential causalities.
Yeah.
This corroborates, I think, the approachwhich is
(41:42):
certainly beautiful.
At this point,
I would like to thank Mine and Claudia.
We could just stay here talkinglike this for hours or for days on some
amazing paper and amazing data.
I mean, we learned so much today
and I would like to invite if you have toadd something to our audience. Mine, Claudia,
(42:02):
from your perspective, and wheredo you see this going in the future,
that would be great.I'd love to hear your thoughts about.
Yeah, anything you'd like to add?
We'd love to hear it.
No, I think frankly, as I mentioned,the increasing large scale, high
throughput generationof additional layers of biological
data is frankly very excitingand I'm very excited to be in a field
(42:24):
where we are better understandingtheir potential translational capacities.
And with that,
I also would like to thank both
my colleagues,which have enabled all the work that we do
possible, as well as the first instanceof EPIC-Norfolk study
and the past and present team memberswhich have made this study possible.
(42:48):
Anything more from you, Claudia?
Yeah, I justI think I said already, looking forward,
how it'd be very valuableto increase the kind of phenotypic
space in the diseases we can look at.
I think also moving forward,I certainly think the beauty
of these large-scale populationbased studies is one thing.
That's certainly one of the reasonswhy I love sitting here in East London
(43:10):
and being like just a hundred reaches awayfrom one of the largest hospitals
is because the next step really is howdo we enable in clinical proteomics, i.e.
proteomics in patient studies,how do we design that well
and in a way that enables flexibilityof different research questions?
So that's kind of what I want to focus onand which I think is very exciting.
And for that we would need technologiesthat are kind of ready
(43:33):
to move to thatstandard relatively quickly.
And so I think that's an exciting new areaand doing that in a way
that is affordable even within the contextof a national health system.
So it's exciting,It's really amazing times
and working together with talented
people like Mineand other people in our team. just as you
As one of my mentors used to say:
(43:55):
if it's not fun, it's not epidemiology.
I love that.
Well, you
certainly make it look fun,that's for sure.
Now, thank you so much for the opportunityto talk today.
It was a pleasure to have you guys.
All right.
Thanks, everyone.
Thank you.
(44:18):
Thank you for listening to the Proteomicsin Proximity
podcast brought to you by OlinkProteomics. To contact the hosts
or for further information,simply email info@olink.com
Welcometo the Proteomics and Proximity Podcast
where your co-hosts, Cindy Lawleyand Sarantis Chlamydas, from Olink Proteomics
talk about the intersection of proteomicswith genomics for drug target
discovery, the application of proteomicsto reveal disease biomarkers,
and current trends in using proteomicsto unlock biological mechanisms.
(00:26):
Here we have your hosts, Cindy and Sarantis.
Hello, everyone.
Thanks so much for joiningProteomics in Proximity.
I'm one of your co-hosts, Cindy Lawley,and I have with me my other co-host.
Hello, everybody.
I'm Sarantis Chlamydas, happy to be here.
Very good.
So so today we are joined by Mine Kopruluand Claudia Langenberg.
(00:51):
We are talking today abouta wonderful paper that came out in Nature
Metabolism in March [2023] called "Proteo-genomic links to human metabolic diseases."
Very exciting paper. Byby Claudia Langenberg and Mine
Koprulu standards,a modest sample size but amazing findings,
so I think it's really enabling.We'll dig into that,
(01:13):
but first, let's introduce our guests.
Sarantis,do you want to do the honors?
Absolutely, we'll start with Mine.Mine is a PhD
student on Gates Cambridge scholarshipand supervised by Dr.
Langenberg.
Before her PhD, she received a Bachelorof Studies in Human Genetics at UCL [University of College London]
and then subsequently she completed
(01:35):
a Masters of Philosophy in GenomicMedicine at University of Cambridge
working in genomics data and
proteomics data and dealing with population studies
and the UK Biobank data with us.
And we are really looking forward to hearingabout your project and the paper about proteogenomics.
Yeah,
and digging into how you gotwhere you are.
(01:57):
Anything you want to add to that, Mine,that you want people to know going in?
No, I think that's a very clearintroduction and it's lovely
to be here and I'm looking forwardto sharing a bit more about
both our journeyand our recent publication.
Awesome, fantastic.
So then I have the privilege and and the
(02:18):
opportunity to introduceClaudia Langenberg.
Now, Claudia is probably someonewho needs no introduction, but
the I think the impact
that she's had in the publicationspace is huge.
Last I saw, she had I think this is in
2021 over300 peer reviewed publications at her
(02:40):
very young age and and I think
over 40,000 citations at that point.
And that's already several years old.
So this is athis is a world-renowned
scholar that is working with Mine,
And of course, many of her publicationsare in
journal serieslike Nature, JAMA, Lancet,
(03:02):
very, very
prestigious and very impactful,working with massive datasets.
So I think of her as a large scalepopulation epidemiologist,
such a big word,but it integrates many of the omics
and I think she keeps an eyeon the technology advances to be able to
build systematic approachesto understanding human biology.
(03:25):
So Claudia, I think now you're in you'reat Queen Mary, University of London.
I think when I met you originally,you were at the Berlin Institute of
of Health and Charité.
So I'd love for you to just tell usa little bit
about your position nowand your affiliations, if you don't mind.
Well, thank you so muchfor the extremely kind introduction.
(03:48):
And I think the most complimentary thingwas about the young age,
which is not only a complimentbut a lie,
so thisI think, puts everything in perspective.
And yes, also I should say the workthat we've done that
Mine is a studentat the University of Cambridge
where I've had a dual affiliationand a long career.
And prior to me arrivingat the Berlin Institute of Health,
so originally be Germanyand training as a doctor in Germany.
(04:12):
I came to the UK and I was in Londonoriginally where I did my Ph.D.
in epidemiology and a Master's of hygiene.
And before that I went to Cambridge
to focus more specifically on genomicsand then later other omics.
And so yeah, it's a huge privilegeto be here
and it's an even bigger privilegeto work with talented people like Mine
and others in our team.
So it's a real team effortand a lot of fun there to do that.
(04:36):
So it does seem like you have fun.
And I'll tell you what, wheneverI talk to Mine and when I've seen her
present,she's always got a big smile on her face.
So I would love to dig into your journeyto end up in Claudia's lab,
Mine.
Do you mind giving us a little backgroundand how you got interested in science?
(04:57):
I mean, feel free to startwherever you want.
Yeah, sure.
I've always been very interestedin science, and especially biology.
But especially I always wanted to dosomething with my career, which improves
life, supports and contributesto the society in a meaningful way.
And I've said givenI was quite interested in biology,
human geneticsespecially quite interested me
(05:21):
because a lot of the sciencesare quite established in the sense
that, you know, thehuman genome is very recent compared
to many other scientific fieldsand there's so much unknown.
And I think that is especiallywhat interested me
as well as its potential to, you know,improve the current healthcare system
and the clinical translational potentialof the human genetics findings.
(05:45):
So that was my inspirationto go into that.
And unlike some other people,I think I've been in that quite
narrow specific path ever since then.
So I did my bachelor's at UCL
as Sarantis mentioned in human genetics -it was just strictly.
And after that I did my master'sin University of Cambridge
(06:05):
in genomic medicine, which is where I gotintroduced to bioinformatics.
So I did my Masters projectat Sanger Institute working on UK Biobank
as well as whole genome sequencing datafor each population isolates
at least
in that lab at the time.And after that I moved back to back home,
(06:26):
which is Turkey for me to work on raredisease genetics.
So I spent two years back in Turkeyat a
universityresearching rare disease genetics.
So that was more studyingwhole exome sequencing data
from consanguineous marriages
or consanguineous parents havingdisabled children trying to better
(06:50):
pinpoint the exact genetic changesthat caused such severe, rare disorders.
And in the meantime, while I was backhome, I was thinking about my PhD
and what I might want to pursueand multi-omics integration.
So better understanding the different "omic" layersand how that can actually improve
our understanding of genomic studiesand their clinical translation
(07:14):
was what interested me.
So I got in touch with Claudia
and we had a brief interviewand that was the beginning
of such a beautiful collaboration,at least on my part.
And then starting alittle bit about your paper -
What were you able to understand about the data,about this EPIC-Norfolk study?
(07:36):
The EPIC-Norfolk cohort. What are the characteristicsof this cohort and what makes it so special?
Sure, the EPIC-Norfolk studyis a population cohort
that was first establishedin beginning of 1990s,
so it has approximately 30,000
participants that were followed up.
Phenotypic research in terms of different
(07:59):
behavioral and sort of healthcharacteristics for several years,
as well as linkage to their health records.
And we had proteomics samples
measured from 3000 ofEPIC-Norfolk participants.
But Claudia, if you have anything moreto add on the EPIC reference study ...
I think I can say one of the opportunitiesthis study has offered
(08:21):
and I think the beautyof such perspectives cohorts such as EPIC-Norfolk,
and UK Biobankand so on, things that were set up
in the past is that we benefit from the samplesthat were stored at baseline
in liquid nitrogen.
And those samples are so valid
because anything you can do in the futureand now you measure 3000 [proteins].
And you measure 5000, you measure tenand more thousand proteins.
So anything we usenow we won't be able to use in the future.
(08:44):
So we're always very cautious
of that and hencecautious with our sample use.
So it's really important that proteomictechnologies now use such a little sample.
So you could have used more samplein the past to measure three proteins.
So to be able to [measure] thousands
in such a small amount of bloodis absolutely amazing.
(09:05):
And to have the opportunity, you know,
PIs when our PIs of thosecohorts had the foresight of setting this up.
Nick Wareham in Cambridge and before thatKay-Tee Khaw, others who let us study.
And other cohorts that we hadthe privilege of using the samples for.
It's really,really amazing to be able to do that.
That's one thing.
But the other thingthat's important is as time
elapses, these codesbecome even more valuable because,
(09:27):
you know, sadly people have events, theythey develop diseases, they die.
So if you then look backwardsbecause you have samples stored,
provides the opportunity to have suchan efficient design of doing it,
you don't have to measureeveryone. But you can have a kind of a
very nifty design in choosingpeople who have developed a disease, i.e.
(09:47):
new onset disease, butthe sample was stored before they had it.
So you avoid thiskind of reverse causation by which disease
impacts your proteome and henceyou can't really just associate.
This is what we exactly didin this context.
Mine's study was one of the projectsthat we did on context of that design.
But you can also look at
different diseasesbased on the people who developed them
(10:09):
and then studied. Use that baseline sampleand then you use a big control cohort
of people who serve as the controlsfor many different diseases.
It's called a nested case cohort study.
And it's a really beautiful design,but that can only be used
if you had the foresight of setting upa large prospective cohort like that.
So we were very gratefulto all the participants.
EPIC-Norfolk and the PIs, of course,who's enabled the use of that,
(10:31):
and also coming to the costof some of these kind of very,
you know,
informative molecular technologysuch as yours is.
Of course they're not cheap
and hence that meansif you can use a design that minimizes
the number of samplesyou have to use for given purpose,
of coursethat's incredibly useful for us. So
and I think I just
(10:52):
wanted to comment on this,this liquid nitrogen aspect, this ability,
I think this foresight -I work with a lot of groups
that do a lot of different thingswith large population health studies.
There are very few cohortsthat had that much
attentionto reducing pre-analytical variation
and tracking itas is documented in this cohort.
(11:16):
So that's exciting to see.
I think it's maybe less importantfor proteomics
than metabolomics as an example,but you work in all of that.
And so yeah,I just wanted to underscore that point.
Yeah, sorry, Sarantis, please go ahead.
Okay.
I will like Mineto summarize the take-home messages
of your paper.
and what's so excitingabout this paper.
(11:40):
Yeah.
First, and just alsoclarify a bit [about] the paper,
I was the first author,
but of course it was a team workand everyone contributed a lot
making this work possible,including the EPIC-Norfolk participants
of course.
Just to summarize the paper,
I think I could just briefly saywe looked at the sort of systematically
(12:01):
linking genetic variants, blood proteinlevels as well as disease risk data,
to be able to pinpoint causal genesand proteins that underlie diseases.
So just to give a bit of background sinceenablement
of the genotyping technology,little studies have been conducted
(12:22):
associating genetic variantswith different disease risks or
disease susceptibility.
And today 200,000 genetic variant disease
associations have been establishedand are publicly available.
Looking at such numbers and figures,
we would assume we knowthe biological basis for all diseases.
(12:44):
However, we all know that's not at allthe case, and I think that's
where proteomics and differentsorts of omic layers come into play:
in helping us better understandthe disease mechanisms of
actually what's happeningunderlying the diseases in our body.
So in this studywe had samples from 3000 individuals
(13:06):
measuring three approximately 3000 blood protein levels.
So we first looked at the geneticregulation of the cis regions.
So the cis regions are the protein-encoding regions
and sort of flanking regions aroundthe gene for the protein target itself,
because of our moderate samplesize, as we've already mentioned.
(13:28):
So we looked at the cis genetic regulationor different blood
protein levels,some of which had never been targeted
before, given the recentOlink platform [Olink Explore 3072]
and then we have used that knowledge,that sort of protean genomic knowledge,
to better understand
causal genes of proteins thatunderlie diseases in a systematic manner.
(13:49):
So we have firstlooked at shared genetic regulation
for different disease outcomes and bloodprotein level regulation,
pQTLs as we call it - protein Quantitative trait loci.
And we identified 224 targets
that regulates 500,
(14:11):
approximately 500 different traits.
And we also refined the causal genesor proteins
for 40% of the previously establishedgenetic risk loci,
which was which sort of highlightedthat even moderately sized
proteogenomic studies can contributeto our covering of novel biology
(14:36):
for consanguineous risk loci that werepublished in the literature.
And finally, we looked at the convergence
of the pQTLsstudies and the regulation
wherethe rare variant gene burden analysis.
So, on one hand, comparing the lossof function of the genes with that sort
(14:56):
of genetic regulation of the proteins.
And so I wanted to click back on this: the disease associations (15:00):
undefined
that were enabled by genetics. I think the
the ability to make those associations
in small sample sizes was likelow hanging fruit in the early days
of that sort of GWAS era,which I think of maybe early 2000s, right?
(15:21):
2005, 2006
that's when those things startedreally ramping up discoveries
and then they became harder and harderto make those associations
as those really strong,
strong associationswere discovered and documented
and then larger and larger populationswere needed to make those associations.
(15:45):
I think what you're saying
is that these other layers are helping you
to to do morewith those more modest populations.
And Claudia already made the pointabout the costs of layering
on metabolomics or proteomicsor these additional omics.
Can you say something about what
(16:05):
what that enables and what you thinkwill happen in the future there?
Yeah, of course.
So basically, as I mentionedand as you've mentioned already, thousands
of genetic variants,a disease associations are being made
and those were very valuableand contributed to our understanding
of diseases to some extent.
However, the majority of those associationsfell into noncoding regions of the genome,
(16:28):
meaning it was quite difficultto actually interpret which causal gene
or protein were acting incausing the disease.
Hence, it was quite difficultto understand the pathways
and the mechanisms.
And, well, ifwe don't have a functional target,
it is difficult to actually buildmore effective and safer
therapies or repurpose existing ones.
(16:50):
So in that sense, having the initial layerof the proteomics states actually
helps us to pinpoint a functional entity
that plays a role in the disease.
And so
I said there's
certain statistical methodsthat help us better understand
the shared genetic regulationof both the disease risk
(17:13):
and the blood protein levelsor abundance of certain proteins.
And we see strong statistical evidencefor a shared genetic
regulation done within a large regionof many candidate genes.
We can actually refinethat to a single candidate gene or protein
for particular diseasesin a systematic way
which can then be usedfor as I mentioned, intervention
(17:37):
or more targetedtherapy, safer therapies.
So in that sense, even moderately sized,more regular pQTL studies
can really help us better understandthe disease and pathways that are involved
and also build more effective therapies.
I think we're on a path.
I think this is a path for discoverythat is, I think just going to ramp up.
(17:59):
I think very similarly to what we've seen,those discoveries enabled by genetics.
I'm really excited about that.
I always think about
these different diseasesthat might share pathways in common,
almost like a Venn diagramand understanding the complexities of
of why certain proteins or protein
(18:22):
pathways are related to like,I don't know, caspase,
I think is like showing upin breast cancer and asthma. It's
an example. I thinkthose diseases are so disparate to me.
I wonder if youthink that proteins or pathways
that are showing up as causaland I think you make a good point about
(18:42):
why causality is so importantfor therapies.
Do you expect those same proteinsto be causal in other diseases?
Like what doyou think about that?
You're much more versedin seeing the data.
So I think that is sort of the beautyof our study design
that we sort of approachwith different layers
(19:02):
of biological datastarting from the genome
all the way to the phenomein a hypothesis-free manner.
So doing that and looking at the datasystematically without any sort of prior
hypothesisallows us to see what the data tells us.
So as you mentioned, certainly in lookingat what we have discovered in our papers,
(19:23):
that genetically anchored proteogenomics studies so the proteomics
can help us discovermolecular hubs.
All associations of proteinswith diseases that we wouldn't
have predicted otherwise from justthe general literature or prior knowledge.
And that is actually quite interestingto follow up
because, yeah, the data uncoverssomething that was unexpected.
(19:47):
But likewise we can also seequite specific protein disease pairs,
which can also allow usto have quite specific therapies
for potentially not interveneddiseases before.
But just one more thing to addis that we
currently do not have not a very full completepicture
(20:08):
of the proteomics, so we're only ableto do that for the proteins
we are able to target.
So for future directionsyou asked about,
having a more complete ideaabout the full landscape of proteomics
would be ideal to better understandthose molecular hubs
that we have just talked about.
But so maybeI can add one more thing to this,
(20:31):
because as Mine's workother proteogenomics work
than we previously did in the pasthas shown is exactly the beauty is that, as
you scale up the number of proteinsand as you scale up
the number of diseases,you can look at this kind of proteo-
genomic approach that doesn't requirethat each of these layers is measured
in the same sample size.
You can utilize the power thatyou have across any biggest genomic study
(20:54):
for diseases that you have to geta reaction, a proteogenomic study.
And that is the beauty of looking exactlylike you say, Cindy,
at the overlap of a specific gene to proteinacross any disease that you can look at
because the community is sharingdata openly or the summary statistics
for each of these diseases is sharedopenly for most diseases. Cancer is
(21:16):
lagging behind sadly,a little bit in terms of the openness.
But for many diseases that is there,it enables us to do exactly
that (21:23):
to draw a whole map from gene to protein to disease
and is that link coincidental
or is possibly causaland a rare genetic signal?
And that's such a good wayof prioritizing what you then use for
experimental work downstream
because of coursethis is in the computational approach
(21:44):
and not an outcome of proof,but it's such a good and data driven way
of prioritizing
links betweendiseases, links for where we have already,
you know, specific drug targets, new ones,potentially adverse effects.
So it's a really versatile wayof looking at all of that.
And then just maybe to add one thingagain, as you said, is
(22:06):
you can kind of ramp this upas you increase the number of proteins
or as you increase the density
of your genomic arrayor the coverage of rare variants
and by sequencing,
and so on. Of coursean important part of really making
this more usefulis increasing the phenotypic spectrum.
(22:27):
And that really is only possibleif we move away from diseases. We all
study opportunistically to diseasesthat you can't really easily measure
or nobody's interested in, but they're stillreally important for patients and a lot
less headway has been made in termsof understanding that genetics
releasing the summary statisticsfor those studies, and that's really where
(22:48):
huge studies come in that have electronichealth record data from GPs [general practitioners],
from hospitals, from death certificates,and bring all of this together.
And that's why, for example, UK Biobankbut also Fingen
and many other endeavors around the worldcan really help us
to not just increase the molecular sidebut the phenotypic side.
(23:09):
And that is so important to alsolink diseases which have not really been
so much in the center of attentionand but need to be.
Can you give an example of one of those?
Are we thinking rare disease here?
So I think it can be
across the frequency spectrum,it can be across specialties.
I think these diseaseexamples that I would choose,
(23:30):
others that possibly are not as easily
diagnosed, are not severe enoughto always require
hospitalization,because I think most people around
the world have tried to gettheir hospitalization data
I see decoded is a relatively easy,but the data, for example, in the UK
that comes from England, from primary carerecords, is harder to map,
(23:51):
is kind of bit more diversity in termsof systems, the data structures and so on.
So diseases that are predominantly managedand diagnosed in
primary care are morewhere we need to move towards.
Oh, that's so interesting.
Yeah. I have a bit of a
technical question here.
How do you make the selection
(24:11):
of false positivesand how do you define
these in your analysiscompared to
a lot of technologies, for example, that you applied before?
You're
asking about false positives? Yes.
So in our analysis, we try to be careful
given we're workingwith data and diseases, as you mentioned.
(24:33):
So in terms of our analysis,we always use quite a rigorous
threshold to reportwhat is statistically significant.
So in terms of our statistical threshold,we always called for genome wide
significance times the number of proteinsthat we are including in the study
to sort of minimize the error, includingthousands of targets in this case.
(24:59):
So we usually go for quitea rigorous statistical threshold
as well as goodQC before we feed in the data,
of course.
So that's great.
So some of these pQTLs
may fall in non-coding regions.
(25:20):
Do you have any idea about the type of these regions?Maybe have a promoters there,
I don't know the answers.Do you have an idea and can you map these regions?
So we mapped
all the variants in terms of whatthey're predicted to be in the genome.
So we look at the proportion of proteinaltering variants, those variants
(25:40):
that fall into the proteincoding region itself and alter
the shape or the structure ofthe protein.
And the percentagethat fall into non-coding regions.
However, there's quite limited knowledgeabout the functional characterization
of the non-coding variants themselvesacquired from particular groups studying
(26:01):
particular genes and more sort of cellularor functional models.
So in terms of our work so farwe have only computationally annotated
the predicted consequence of the variants.
However, of course further follow upof analytes and what exactly those
non-coding variants do in termsof cis regulation of the protein itself.
(26:23):
That's great.
So I think a lot of studiesand then from the literature,
as you as you mentioned, for differentproteins from different
mice knockouts,
you can have an ideaabout how they are regulated.
Is there any plan to follow up, like moremice knockouts
in the future to followsome of these targets or collaborations
(26:45):
in this respect?
If you can share this?
I mean, we of course always welcome
any sort of collaborationson the functional characterization.
The idea of doing these studiesis not to put in a long Excel sheet,
as we have done so far
in terms of our supplementary table,but rather genuinely understand
the biological mechanismsthat underlie diseases
(27:06):
and how that can contributeto clinical translation
of some of the findingsand the sort of longer runs.
So, with that, an open callto anyone who finds any of
that targetsthat we highlight interesting.
We are of course
very keen to functionally characterizeand better understand the mechanisms.
I thinkin the paper you mentioned something about -
(27:28):
I was just looking for the reference,something about cellular models
and the history ofof demonstrating functional associations
or understandinga function around cellular models
and how this approach
can be a complementary wayto add an understanding of function.
Do you mind talking a little bitabout that,
(27:50):
like where this fitsinto our traditional approaches?
Like where it might fit inin understanding
what we're learning from single cellor spatial
work that some of those technologiesare really advancing right now?
I'm curious your thoughts on those.
Of course.
So what we do in terms of our teamis more bioinformatics.
(28:14):
So what we work with is what the data and
sort of predictionsand better understanding the computational
and statistical conclusionswe can draw from the data.
However, as we all knowand as you've highlighted,
they are complementary waysin better understanding these.
And basically we are very interestedto see whether
(28:35):
what we observe computationallyand statistically actually
translates into biology,starting from cellular models,
building up to animal modelsand sort of human biology.
So what we're observing almost
quite isolated
looking at singularsort of targets and singular
genetic variants. It will be quite interestingin whether our conclusions hold or
(29:00):
sort of variations fromour conclusions we see in the different models.
So I definitely agreethat they're complementary
and there's ways of integrating thoseknowledge to build a more comprehensive
or sort of holistic understandingof the biological mechanisms.
Oh, I love it.
I love the backgroundyou have both in an understanding
(29:22):
of epidemiologyand your mad bioinformatics skills.
Where does machine learningfit into all of this?
Is it something you've -
yeah, I'll just stop with the question.
No, of course, machinelearning and artificial intelligence
is certainly fieldsthat are very quickly developing
and are grabbing a lot of attentionat the moment. So what
(29:46):
we use are considered basic machinelearning models themselves.
But in terms of employing machinelearning or artificial intelligence to
sort of prettier genomics for theseor more biological studies at the moment
is rather challengingbecause as we all spoken about, this large
high triplet biological datais recently being generated
(30:09):
and they are so recent,but we're spending a lot of time
better understanding what they meanand what they actually are.
And frankly, the biological datathe way they are actually violates
a lot of assumptions that machinelearning and artificial intelligence
models make.
So I think in the future,they can certainly be very useful,
but we certainly need to be cautiousand better understand
(30:33):
thembefore feeding them into the models,
I personally say.And maybe I can add something to this.
So I'm currently sitting inwhat's called the DERI.
So when I came to QMU, or QueenMary University of London last September,
I had this opportunity to set up thisnew institute for precision healthcare.
And it's kind of quite uniquebecause it's a cross functional institute.
(30:55):
So it really draws in a very broadrange of expertise and so [there are] big plans
and developments for a new life sciencesbuilding, and that's all quite away.
So we have to wait for this.
So we need an interim space and we're verylucky and it's actually not coincidence
WhereI'm sitting right now is called the DERI.
It's the only other cross functionalinstitute that QMU has and DERI stands
(31:16):
for the Digital Environment ResearchInstitute, which is focused on AI.
And so it's not just in healthcare, it'sAI across a broad range of applications.
You know, from games - we're sitting onthe second floor with the games people,
which my my kids think I have the best jobin the world.
Actually what they don't know
when they have conversationsabout board games is
(31:37):
I have not heard of a single one of them,so I can't mention really anything else.
But anyway, having said that, this is theenvironment in which it can flourish.
So we focus on healthand on healthcare data
and DERI hasas part of its remit as well.
So it's really importantto bring this together.
And, you know,going away from,
(31:58):
you know, this pitch onhow it's important, the concrete thing
that we're already doing isthe molecular data, which is so complex,
you need efficientand unbiased ways of data reduction.
That, at the moment, is most of the usefor machine learning in our
arena of work.
(32:18):
So where we kind of try to, let's say,
predict diseases,how are you going to prioritize?
Nobody wants to go into the clinicand measure 3000 proteins
to predict a disease.
Do you even have any benefit of
what's the incremental valueof measuring proteins
once you have a core set of tenmost important ones?
(32:39):
So those are all questionsthat machine learning can help us to
to address more systematically,and that's incredibly useful.
So is a steep learning curve for us.
This one, it's not our bread and butter.
We're not the methodological expertsof developing it,
but being in an institutelike this one, for example, means
that we're only very a short way removed
(33:00):
from the people who do develop thesemethods and that we at the right time
can employ them, can test other useful,could they be biased.
And so that's really, reallyyeah, it's a great opportunity, I think.
Well,I think those people that are developing
those methods aren't going to havethe biological know-how to know
when they've overtrained those methodsand are going off on a spurious tangent
(33:21):
so the context I think is essential.
I think Mine makes that point.
Yeah, exactly.
It's this exactly.It's a team effort, right?
And you need the people, havethe samples, and the clinical knowledge.
You need the people who havesome other kind of bioinformatics,
computational and you need the,you know, method developers.
So it's beautiful.
Nobody can do it all.
(33:41):
So, you know, I think it's a great.
It's CQ, right?
It's collaboration quotient, right?
You have your IQ, your EQand CQ. From my experiences,
once you develop your collaboration quotient,
which you're teaching your studentsand your postdocs and all of that
if they don't already have it,
Claudia, is once you have it,you can't go back?
(34:03):
I don't think
once you experiencethe joy and the ability to work together
cross-functionally, it's reallyremarkable what can be accomplished.
Well, coming back to the question that
Sarantis was asking,and I think it's a very good one
and Mine has said so.
The problem that we have encounteredis, you know, is Mine
has come upwith these beautiful candidates.
And it's almost there's two thingsthat are important to kind of what
(34:27):
she said
so well, which is that now using humansas the model organism
is a great way of prioritization.That's number one.
But the second one is kind of, you know,if you come from a Bayesian framework,
you really want to find out what'syour greatest chance of success?
How do you increase your priorof this having success?
And given that the experimental procedureis expensive, it's lengthy
(34:48):
and it can go wrong in so many ways,
It is really a great opportunity.
to prioritizeon the basis of human proteomics
as Mine has said.The problem is: how do you get people
who have the experimental set upto come to our results and take it?
(35:10):
Because that's not as easy as we thought.
We think, "Oh, here!"We're so excited to see it.
We go to the world expert in thisand bring them our example,
it's hard.
It's hard to motivate peopleto step out of what they're already doing
and focus on your finding.
It's hard if you have somebodywho's focused their whole life
on a given pet protein to say,how about this one?
(35:32):
So we need help,I think, to try and learn
how we engage the relevantclinicians and molecular biologists
and other people to also use ourresults rather and at least,
you know,complement what they're already doing
in order to be able to follow some of itup, because we do depend
on the functional valueand the expertise of those people.
(35:55):
So how do we reach out?How can you help us
to reach out to the relevant peopleto really make it worthwhile?
What are the titles of those people?
I mean, what are their jobs?
You know, like,I know there's more than one, but
I think of translational,
certainly translational scientistor implementation scientist
or maybe those are pieces of it.
(36:17):
But you know, who are the onesthat we need to
listen to this?
So I think it depends on the stageof translation
that you're at. For really earlyexperimental work,
it's a very different set of peoplethan it would be for the people
if you have a different kind of work
where you really are ready to maybe,possibly already consider an initial trial,
(36:38):
then you would need a relevant clinicianor a head of that clinical department
to enable that kind of set up within that,let's say, hospital.
So I think it really dependson what level you're talking.
The initial level I think iswhat's the experimental follow up
and validation, because what we do isyeah, the statistics and probability an an error,
but it's still an associationwhere we're under no doubt we do not prove causality.
(37:01):
And so that kind of functionalfollow-up is crucial.
And those are the people at the first stage.
Think actually you don't knowthe cause of some of these diseases,
you don't knowthe cause of the effect, right?
You don't know if this is what you seein the protein level. Is it a cuase
or the effect of the disease?It could be both. The last comment
to me I really like when
you see the co-localization with eQTLsfrom different tissues
(37:24):
like does what we see in the plasma proteome
reflect what happens in the tissues?
And do you have any comment
on how you see these correlateand what is the distribution
of the singular featureto the plasma proteome?
You have some examples,some ideas on that?
And a shout out to gnomAD and GTEX
(37:46):
I will say fantastic resources. Yes.
No, we are very gratefulfor all the publicly available resources,
ranging from the eQTL studiesall the way to the GWAS summary statistics,
which have made our work possible as well.In terms of the eQTL overlap,
that is rather challengingin our field in the sense that now
(38:07):
we see, for example, in our study
that only approximately 4% of the pQTLsthat we see
are in close linkage disequilibrium
that we deal with an eQTL
and certain examples,we see beautiful overlap, right?
Like we find an examplewith a disease for that
(38:28):
act in a particular tissueand we see co-localization,
quite strong co-localizationwith a tissue of interest
where the story becomes quite beautifuland everything makes sense.
However, there is also the flip of the coin
where we see in some examplesquite limited overlap
between the pQTL and eQTL data.
And although our current paperhasn't really focused on that,
(38:51):
I think that is something that weas a community need to better understand
where there is overlapand where there's a lack of
and what might be the underlyingreasons that would be.
I think this is a great
but to mention the really interestinglink to type two diabetes
that you found in here that that we wantsomeone to follow up on, as Claudia mentioned.
(39:16):
Do you want to just summarizethat really quickly, that pathway?
Of course,one of the examples that we have noticed
or we have highlighted in our paperis the gastrin-releasing
peptide, or GRP for shortand its link with type two diabetes.
So what we saw in our paper is that
(39:36):
beautifully different layersof biological data,
so evidence from my studies,evidence from human data as we have been talking
about and different sort of similar modelsall overlapped with the same conclusion.
So what we've identifiedwas that the higher levels of GRP in human
(39:56):
plasma were co-localizing
with lower risk of type two diabetes.
And when we integratedin the different sort of
intermediate layers,which are different body sets and
distribution traits,so where we actually accumulate
fat in our bodyas well as the overall fat,
(40:16):
what we observedwas that higher levels of GRP
was leading to less fataccumulation overall
in our body, leading to lowerrisk of type two diabetes.
And as I mentioned,there were previous studies
that were publishedwhere human recombinant GRP
(40:36):
led to actually reducedintake and weight loss.
So we have highlighted in our paperthat GRP can be a new example
or a potential therapy for typetwo diabetes by decreasing or
lowering the fat accumulationin our body.
(40:58):
It's actually a hot topicfor the weight loss.
A lot of companies now are trying to follow this
game of weight-loss strategy.
I think that weight-lossdrugs are really interesting,
and I am looking forwardto people following up
on this.
And I love the use of
the word beautiful, right.
(41:19):
But it's beautifulbecause these layers are all agreeing.
It's giving us a preponderance of evidencethat gives us a lot of confidence,
that gives us confidencein the other results that you see, too.
Right.
So the fact that you've builtthis systematic map
of these potential causalities.
Yeah.
This corroborates, I think, the approachwhich is
(41:42):
certainly beautiful.
At this point,
I would like to thank Mine and Claudia.
We could just stay here talkinglike this for hours or for days on some
amazing paper and amazing data.
I mean, we learned so much today
and I would like to invite if you have toadd something to our audience. Mine, Claudia,
(42:02):
from your perspective, and wheredo you see this going in the future,
that would be great.I'd love to hear your thoughts about.
Yeah, anything you'd like to add?
We'd love to hear it.
No, I think frankly, as I mentioned,the increasing large scale, high
throughput generationof additional layers of biological
data is frankly very excitingand I'm very excited to be in a field
(42:24):
where we are better understandingtheir potential translational capacities.
And with that,
I also would like to thank both
my colleagues,which have enabled all the work that we do
possible, as well as the first instanceof EPIC-Norfolk study
and the past and present team memberswhich have made this study possible.
(42:48):
Anything more from you, Claudia?
Yeah, I justI think I said already, looking forward,
how it'd be very valuableto increase the kind of phenotypic
space in the diseases we can look at.
I think also moving forward,I certainly think the beauty
of these large-scale populationbased studies is one thing.
That's certainly one of the reasonswhy I love sitting here in East London
(43:10):
and being like just a hundred reaches awayfrom one of the largest hospitals
is because the next step really is howdo we enable in clinical proteomics, i.e.
proteomics in patient studies,how do we design that well
and in a way that enables flexibilityof different research questions?
So that's kind of what I want to focus onand which I think is very exciting.
And for that we would need technologiesthat are kind of ready
(43:33):
to move to thatstandard relatively quickly.
And so I think that's an exciting new areaand doing that in a way
that is affordable even within the contextof a national health system.
So it's exciting,It's really amazing times
and working together with talented
people like Mineand other people in our team. just as you
As one of my mentors used to say:
(43:55):
if it's not fun, it's not epidemiology.
I love that.
Well, you
certainly make it look fun,that's for sure.
Now, thank you so much for the opportunityto talk today.
It was a pleasure to have you guys.
All right.
Thanks, everyone.
Thank you.
(44:18):
Thank you for listening to the Proteomicsin Proximity
podcast brought to you by OlinkProteomics. To contact the hosts
or for further information,simply email info@olink.com
Popular Podcasts
24/7 News: The Latest
The latest news in 4 minutes updated every hour, every day.
Therapy Gecko
An unlicensed lizard psychologist travels the universe talking to strangers about absolutely nothing. TO CALL THE GECKO: follow me on https://www.twitch.tv/lyleforever to get a notification for when I am taking calls. I am usually live Mondays, Wednesdays, and Fridays but lately a lot of other times too. I am a gecko.
The Joe Rogan Experience
The official podcast of comedian Joe Rogan.