DNA Methylation in Aging and Cancer

May 26, 2020

Information

Morgan Levine, PhD

Yale Cancer Center Grand Rounds | May 19, 2020

ID5243

To CiteDCA Citation Guide

00:00Morgan is an assistant professor of pathology
00:03and Epidemiology at the school of Madison.
00:06She's a member of the combined
00:09program in computational biology,
00:11environ fanatics, as well as the Center
00:14for research on Aging and her work.
00:17Her multidisciplinary work has
00:18really been integrating new
00:20methods of statistical genetics,
00:22computational biology,
00:23mathematical demography to develop,
00:25sort of a new high dimensional mix
00:28approach to aging in both humans and
00:31animal models and applying those.
00:34Efforts to a variety of major
00:36chronic disease,
00:37most notably cancer,
00:38and so Morgan really pleased to hear
00:40about your work and looking forward to talk.
00:43Thank you so much.
00:49OK, maybe we can see that yes.
00:55And let me make it bigger on my screen.
01:02OK, um, so today I'm going to talk
01:04about some of my work on in developing
01:08biomarkers using DNA methylation data
01:10to study aging and diseases like cancer.
01:17Why isn't it? I'm so I usually like to
01:21kind of remind people what the biggest
01:23risk factor for most major cancers is,
01:26and I like to illustrate this often
01:28using something like lung cancer.
01:30So a lot of times when asking students what
01:32the biggest risk factor for lung cancer is,
01:35they'll say something like cigarette smoking,
01:37which we know increases the risk.
01:39The incidence and death from lung
01:41cancer by about 15 to 30 fold.
01:43But in reality,
01:44aging itself is actually much bigger risk
01:46factor for developing lung cancer, so for
01:49individuals who are 25 to 29 years old.
01:51About one in 200,000,
01:53you have about one in 200,000 chance
01:55of Belton lung cancer, however.
01:58Nearly 400 and 100K,
02:01so it UH-80 full increase risk for the
02:05OR 800 fold increases for those 75 to 79.
02:09And this is the case across
02:11a wide variety of cancers.
02:13We see, UM, in general,
02:15an exponential increase with age in both
02:17incidents in mortality risks from cancer.
02:20And you know,
02:21some people have thought that this
02:23is just commit probability with time.
02:25So at the longer you live,
02:27the more time and the more likely
02:30they will develop cancer.
02:31But really,
02:32what we think is that it's actually
02:34the molecular.
02:35Another changes that accompanied
02:37the aging process that are
02:39actually playing a causal role.
02:41In the ideology of major
02:42diseases like cancer,
02:43so I like this kind of New Yorker Cartoon,
02:46which says you're deliberately putting
02:49yourself at risk avail help by being over 65.
02:52So one thing that my lab is really
02:54interested in is can we actually try
02:56and quantify some of these aging
02:58changes that might underlie risk
03:00for things like cancer or other
03:02major chronic diseases?
03:04And so this is where kind of
03:05biomarkers of aging come in.
03:07Uh, so aging is.
03:09Not an observable,
03:10it's it's this latent concept.
03:12So it's actually hard to define.
03:14But biomarkers can actually serve as
03:17useful proxies that we can estimate
03:19the agent Ness of a cell or tissue,
03:22or on the whole Organism level.
03:24So they serve a variety of purposes.
03:27They can be used as clinical trial
03:30endpoints for interventions to try
03:32and slow the rate of aging there.
03:34You can also be used for basic
03:37biology to understand aging.
03:39And also for risk stratification
03:41and the goals in developing some of
03:43these biomarkers is that you should
03:45have a biomarker that differentiates
03:47between a 20 year old an 8 year old,
03:50which is pretty easy.
03:51You can even use facial image to do that,
03:54but probably the harder thing is,
03:56can you actually differentiate
03:57risks among individuals of the
03:59same chronological age?
04:00So can you identify who might be aging
04:03faster or slower and then in turn,
04:05does that have implications for the
04:08risk of a future morbidity mortality?
04:11So most of the biomarkers in my lab
04:13works on a more epigenetic biomarkers
04:16and specifically involved in DNA methylation,
04:18so I like to think of the meth alone as
04:21kind of the molecular operating system
04:23it instructs else how they should
04:25behave and respond is involved in a
04:28number of different cellular processes,
04:30but a really interesting thing that
04:33was pointed out more than I think
04:3530 years ago is that there does
04:37seem to be genome wide patterns.
04:39Um that emerge in terms of
04:41changes in Maculation with aging.
04:43So you gotta change net in the maculation
04:47landscape as a function of age.
04:50And based on this, uh,
04:51a number of labs,
04:53including ours who developed what
04:54we call these epigenetic clocks.
04:56So because they have been very precise,
04:58age changes that have been observed.
05:01We actually use machine learning
05:02to predict the age of a sample
05:05based on the DNA methylation level.
05:07So you can take a sample from whole
05:09blood from tissue in a cell culture,
05:12and we often measure metalation at
05:1410s of thousands to now up to 850,000
05:17different CP G sites across the genome.
05:20And then what people have done is applied
05:22supervised machine learning methods
05:24to actually develop age predictors.
05:26So most of the early clocks were trained
05:29to predict things like chronological age,
05:31the first Clock being published in 2011.
05:34However, more recent clocks have actually,
05:36which we call the second generation at.
05:39The generic clocks were
05:40developed to predict age coral.
05:42It's so not chronological age,
05:44but things like mortality or physiological
05:47processes that change with aging.
05:48So mostly that was.
05:50Our Clock is one of the second
05:52generation clocks inside the John Clock.
05:55And the second generation clocks actually
05:57tend to be much better predictors of
06:00future disease and mortality risk.
06:03Uhm,
06:03but first I just want to show kind of how
06:06these clocks look across different tissues.
06:09So this is an example of five different
06:12epigenetic clocks in a variety of
06:14different tissue are fluid samples.
06:16On the X axis,
06:18I'm showing chronological age on the Y axis.
06:21Is this predicted at the genetic age?
06:24These two clocks by Horvath were
06:26actually trained using multiple different
06:28issues simultaneously pulled together,
06:30so that's why you get much better
06:33agreement across the tissues in
06:35terms of their predicted ages,
06:37whereas the other three clocks are
06:39actually all trained in whole blood,
06:42but still do predict still do show.
06:45Very heists age correlations.
06:46In other tissues, and actually,
06:48if we were to show this within tissue,
06:52a lot of these age correlations
06:54are above .8 two point 9.
06:57But the interesting thing is you also,
06:59if you actually took the time to
07:02map these colors out is kind of
07:04these divergent issues tend to be
07:06samples from brain or these tend to
07:08be non bring samples and we actually
07:11think that it's important to have
07:13differences in Appleton at age
07:15between tissues because we all know
07:17to choose don't age at the same rate.
07:19So we actually shouldn't be forcing
07:21similar epigenetic gauges across tissues.
07:26And then we can actually also
07:28show that epigenetic age is also
07:30differentiates normal tissue from tumor.
07:32But that is not the case
07:34across all the clocks.
07:36It tends to be the case across
07:38these second generation clocks,
07:40where we can see that in the normal
07:42tissue you get significantly lower
07:44epigenetic age compared to the tumor,
07:46and these are all adjusted
07:49for chronological age.
07:50Um, so on our Clock and also the Clock
07:52by Yang Show these differences across
07:55a variety of different tissue types.
08:00So one question that we've
08:02really been dealing with is,
08:03you know all these clocks for
08:05developed to predict the same thing.
08:07To capture this kind of epigenetic or
08:10metalation based change with aging.
08:12Yet they seem to be perhaps
08:13capturing different parts of
08:15this epigenetic aging signals.
08:16So basically,
08:17can we identify the individual
08:19components and decompose the
08:21signal to adapt to figure out
08:22what the different parts are and
08:24how they map onto disease risk?
08:26So this is kind of an illustration
08:28of taking the clocks apart.
08:30And then figuring out which each
08:32part of the Clock is doing.
08:35So the way that we did this is we
08:38applied something called WG CNA,
08:41so it's a weighted network analysis
08:43and we did this a cross using six
08:47different issue in fluid datasets.
08:49So we had uh samples from dermis,
08:52epidermis, breast dorsolateral
08:53prefrontal Cortex Colon, an full blood.
08:55And the goal here was to identify Co
08:59maculation modules that are shared
09:01across all these tissue or sample types,
09:05and from this we were able to identify
09:0816 of these Co maculation modules
09:10using Skeggs from the clocks which
09:13word starting with about 1600.
09:19I'm so the next thing we did is we actually
09:22looked at how these different modules
09:25are impacting the overall Clock scores.
09:27So in this I've color coded all the
09:3016 modules and you can see that in
09:33our Clock and this Clock by Hannum a
09:36large proportion of this is actually
09:38driven by this yellow module,
09:40whereas the two clocks by Corvette seem
09:43to have relatively similar proportions in
09:45contributing to the overall Clock score.
09:48But the interesting module that
09:49I'm actually going to talk about
09:52today is this Brown module,
09:53which actually is shown in most
09:55of these clocks and has a pretty
09:58similar proportion of about uhm.
10:0010 to 15% in each of the
10:02clocks to the overall signal.
10:06So the other thing we can do is not
10:09just look at what proportion of the
10:11clocks is explained by each module,
10:13but whether what their capturing
10:15is actually the same signal.
10:17So this is all the modules,
10:19but I'm going to really focus
10:21just on 2 for right now,
10:23so basically this is the part of
10:25each Clock that that's represented
10:26by Stevie jobs in this Brown module.
10:29And what you can see is that for these,
10:32epigenetic clocks have really similar
10:34or high agreements in terms of
10:36their epigenetic age signal here.
10:38However, just a contrast this
10:39on this purple module,
10:41you can see that in in two of the clocks
10:43what the proper module is contributing
10:46to is considered accelerated aging,
10:48whereas in the other two
10:50clocks or three clocks,
10:51it's considered decelerated aging.
10:53So this is an example of a module is
10:56differentially waited and might be
10:58contributing to differences in the
11:00performance by the various clocks.
11:02But for the rest of the talk,
11:04I'm going to focus on this Brown module,
11:06which seems to be the one that's
11:08most important in terms of cancer.
11:12So now what we can do is we can look at
11:15instead of looking at the entire Clock score,
11:18look at the individual modules.
11:20So is there a part of the clocks for this?
11:23Actually driving this kind of these
11:25associations that we're seeing?
11:26So in this case I'm looking at just
11:29the part of our Clock that's captured
11:31by CP GS in this Brown module.
11:34So this is just 21 CP GS over all,
11:36and what we can see is we can kind of
11:39recapitulate the finding with the tumor
11:41versus normal across these different issues.
11:44However, in this case it's actually
11:46more significant when we're just
11:47considering this Brown module.
11:51We can also look up this is in normal
11:53breast tissue and we do see that this
11:56module is significantly correlated with
11:58age in normal breast, suggesting that.
12:02Perhaps as women age,
12:04their breasts as she develops.
12:05The more of this accelerated aging phenotype
12:09which could predispose them to cancer.
12:12And this is actually, uhm,
12:14what we can observe when we
12:16look at this is all data from
12:18normal breast tissue from women,
12:21either with or without breast
12:23cancer prior to treatment.
12:24This is a collaboration with others at
12:27Yale and we validated this in the original
12:30study and then also in another study.
12:33Or you can see that women with breast cancer,
12:36their normal tissues seems to
12:38be epigenetically older when
12:40we look at this Brown module.
12:42And women without breast cancer.
12:44And this is all age matched our age
12:47adjusted and adjusted for things like BMI,
12:50smoking another potential confounders.
12:55Uh, we also had a really small
12:58data set where we had, uhm,
13:00this Brown module measured in tumors
13:03and we had information on survival,
13:06so this is a data set with
13:09only 51 samples an over.
13:12I totale I are over 3471
13:15person Montes or 20 deaths.
13:18And what you can see,
13:20we need to validate this given
13:23those small sample where we do
13:25see that this Brown module 1
13:28standard deviation increase in
13:29this module it's associated with
13:32about 2.25 fold increased risk of
13:34mortality over this time period,
13:37and that's adjusting for things like age,
13:40race, ethnicity, tumor grade,
13:41ER and also chemotherapy tree.
13:47So I went looking more specifically
13:49at what's in this Brown module.
13:51Um, these are the individual CP
13:53GS in the Brown module and we
13:56can actually relate each CVG to
13:58some of the outcomes I discussed.
14:01So this first column is whether it
14:04differentiates in normal breast tissue,
14:06women with breast cancer versus controls.
14:08The second column is whether it
14:11can differentiate breast tumors
14:13from normal breast tissue and
14:15the third column is the survival.
14:18I'm finding and basically what we can see is.
14:22There's about a group of 12 CP GS for
14:26which hypermethylation so increased
14:28maculation in these 12 CP GS is
14:32associated with either cancer and
14:34normal tissue or or tumor versus
14:38normal or lower survival rate.
14:41And from the these are the jeans that
14:44these DVD's are in an there actually
14:47almost all in promoter regions in
14:49these jeans and we can use just ease
14:5212 to estimate an overall score.
14:54So we use PCA across these three
14:56samples and we can take PC one of
14:59those 12 jeans and follow up with that.
15:04So the other thing is that we also
15:07find that these 12 genius seemed
15:09to have specific characteristics,
15:11so they seem to be associated with
15:13polycomb group targets and also HT
15:15K27 trimethylation occupancy and see,
15:17and they tend to be ensues.
15:1912 pound jeans.
15:20So this is these 12 selected jeans.
15:23These were all the jeans that were
15:25in the original ground module and
15:27these are all the CP GS that we have
15:30measured in all of our samples.
15:33So about 20,000 CP GS over also.
15:35This is kind of the background.
15:37So about um 65 to 70% of them
15:41are orange juice 12 pound jeans,
15:44about 50% are Co locating with H2K27
15:47trying Appalachian and similarly
15:4950% with Polycom group targets.
15:53And Interestingly,
15:54this Association is actually not news,
15:57so there's some dating back about
16:0013 years of evidence that these
16:03polycomb mediated methylations does
16:06seem to be important in cancer and.
16:10Basically,
16:10Polycom group proteins are
16:12involved in repression of jeans
16:14that are required for salt.
16:16A stem cell differentiation.
16:19Um,
16:20so finally we also looked at
16:22these in non breast cancers again,
16:26so this is in colorectal cancer
16:29and again we find using this 12
16:32PPG DNA methylations score that
16:34we can significantly differentiate
16:37normal tissue from cancerous tissue.
16:41And Lastly, probably to me,
16:43the most interesting thing is
16:45we can look at this.
16:47A trustee PG score in completely
16:50normal tissue across a bunch
16:52of different tissue types.
16:53And basically we see really strong
16:56correlations with chronological
16:57age across all of these.
16:59So in brain whole glide colon,
17:01dermis,
17:02an epidermis which to me suggests
17:04that these might be changes that are
17:07naturally occuring with aging and
17:09that that might be predisposing.
17:11Some of these tissues to tumor Genesis.
17:14I'm so something that we're really
17:16interested now is in terms of
17:18kind of a primary or secondary
17:21prevention approach.
17:22Can you identify people who are
17:24scoring higher for their age then we
17:27would expect an are those boots are?
17:29Are those people more at risk
17:31of developing cancer in these
17:33specific tissues down the road?
17:35The other thing we're interested
17:37in is comparing across tissues.
17:39So are people who seems to be
17:42aging faster in blood also aging?
17:44Faster and something like breast or colon.
17:49And then last, um, basically,
17:51we also looked at this using a cultured
17:54fiberglass and basically we have,
17:57uhm, the early passage controls.
17:59We haven't immortalized transform fiberglass
18:01where you can see an acceleration of
18:04this epigenetic score immortalized,
18:06and we also looked in cellular senescence.
18:09So on pigeon induced, in essence,
18:12an replicative senescence,
18:13and these are near near senescence that
18:16were passage together so prohibitive.
18:19But they, uh, show high snacks and story,
18:22associated beta gal.
18:23And basically what you can see is
18:26compared to the early passes cells.
18:28We can recapitulate this
18:31Indies cultured fiberglass.
18:33So In conclusion, uhm,
18:34there are different kinds of DNA
18:36methylation changes in aging that are
18:38captured in the different epigenetic
18:39clocks and by deconstructing then
18:41we can start to understand the
18:43functionality of the signals that
18:45are captured in these clocks.
18:47And specifically,
18:48the Brown module seems particularly
18:50interesting in terms of cancer.
18:52Is one of the biggest shared signals
18:55across all the epigenetic clocks and
18:58a distinguishes tumor versus normal
19:01in a variety of different issues.
19:04Uh,
19:05differences to normal breasts
19:06are also observed for women with
19:09cancer versus those without,
19:11and the signal from these from the model
19:14and tumors associated with survival.
19:17We can that also narrow it down
19:20to $12.00 that are really driving
19:22the signal in this Brown module
19:25there mainly capturing promoters,
19:27TPG island hypermethylation that tend
19:29to be marked by Polycom extricate
19:3227 trimethylation and sues 12.
19:34We can observe acceleration in culture,
19:37fiberless, appan,
19:38immortalization transformation
19:39and also so there's no sense.
19:42But to me out again,
19:44really interesting thing is that we
19:46actually see linear changes in this
19:48signal across the adult range in
19:50a bunch of different issues which
19:52actually might be informative.
19:53So overall,
19:54I think this may represent an opinion
19:57about genetic aging change that
19:59explains the increase cancer risk.
20:02With that I want to acknowledge people
20:05in my lab and also my collaborators,
20:08both at Yale.
20:11And elsewhere, as well as my funding.
20:15Working, thank you.
20:16That's a terrific presentation
20:18in a really interesting work.
20:20And we actually have a number of
20:23questions that have been put forth
20:25on the chat or let me just run
20:28through a few Dan Demayo ask you
20:31make see that people have recently
20:33described meth elation of RNA M RNA.
20:36Specifically, does that change as well in
20:38the context of what you've been describing?
20:43So we haven't looked at that here.
20:45I know people are looking at that, um,
20:47there's a group at Harvard who is actually
20:50looking at that in terms of aging,
20:52but it for now what I'm discussing
20:54here is just CG metalation in DNA.
20:58Um, one another question sort of.
21:00Have you looked at this in the
21:02context of progeria patients,
21:04which is sort of a really interesting
21:06question as it relates to aging,
21:09is curious if if you are folks
21:11she worked with it worked
21:13in that space and so we we've
21:15looked at the overall Clock scores
21:17in progeria and not all of them,
21:20but some of them do show
21:22acceleration in fridge area.
21:24We haven't looked at this specific modules
21:26for the Brown module or the 12 PPG.
21:29Part of the Brown module in progeria,
21:31but that is actually an interesting
21:33thing and progeria something we
21:35we have plans to look at all the
21:37different modules to see if there
21:39are certain parts that are that are
21:41picking that up because again some
21:43clocks seem to pick up the progeria
21:45acceleration whereas others don't.
21:47Thank you Marcus has a
21:49question which as you can see,
21:51he said for the for the 12 CP GS that
21:53you've identified their individual
21:55basis as opposed to islands in
21:58any variation of those sites.
22:01Uhm, I actually haven't looked at
22:03whether there snips um at those sites,
22:05so they are individual CP GS,
22:07so 12 individuals seeking.
22:08Geez, what we're interested now
22:10is actually looking at the whole
22:12region and looking at it like
22:14variation across the regions,
22:16but we haven't done that yet.
22:18But yeah, I should go back and
22:20actually look at whether they're
22:22adjacent snips that would be.
22:26One question I have is, uhm, you know.
22:29Looking at your data and realizing
22:32that beyond aging there are,
22:34you know many sort of behaviors,
22:37environmental exposures for lack of
22:39a better phrase that drive cancer.
22:42Breast colon, certainly. And should have.
22:46Is there an opportunity to study sort of,
22:49uh, the behavior of of these individuals
22:52overtime that would drive the signature
22:54in a way that you know they are sort of.
22:58They have a greater component of that.
23:02At Methylations signature that not
23:04only is reflective of promoted aging,
23:06but increase risk of cancer. Yeah,
23:09so we can see we have UM shown in
23:11the overall Clock scores that you
23:14do get accelerated at genetic age
23:16in Association with things that we
23:19think of as normal risk factors,
23:21so cigarette smoking obesity I need in
23:23some socioeconomic factors seem to map
23:26onto differences in these aging rates.
23:28We haven't looked again specifically
23:30at this module, although I will
23:32say from some of our other work,
23:35it seems like the Brown module is
23:37not particularly picking up smoking.
23:39But that might just be when
23:41measured in blood,
23:42whether it is in long or or some other
23:45samples that might be different,
23:47whereas it seems more like that purple
23:50module that it didn't really go into.
23:52It's actually picking up
23:53more of those smoking,
23:55and the influence was smoking
23:57in when measured in blood.
24:00Another question is that the methyl
24:02lation that of the 12 jeans in breast
24:05and with regarding the breast in memory
24:07you can obviously the questions you
24:09can see is that breast tissue is.
24:13A combination of various cell types
24:15and have you narrowed down sort
24:17of the epithelial, fibroblast,
24:18other cell types with regard
24:20to what you're finding. Yeah,
24:22so unfortunately we just have bulk samples
24:25so we can actually narrow it down to
24:28which cell type this is coming from,
24:30but I think because breast is
24:32so heterogeneous we actually the
24:34age correlation with our measures
24:36actually much weaker and breast,
24:38I think because it's a little bit
24:40confounded by the cell composition.
24:42However, you know,
24:43part of the reason we did to
24:45follow up in the culture fiberglass
24:47was to make sure we weren't just
24:50capturing something about cell
24:51composition changes with aging.
24:54And the other interesting thing
24:56is that at least the Brown module
24:58seems to be pretty conserved across
24:59cell and tissue types,
25:01so I don't think it is picking up
25:03something from a specific tissue type.
25:05It it would be interesting to look
25:07at epithelial versus fiberglass
25:08and see if one of those is driving
25:10the signal more than the other,
25:12but right now we don't have that data.
25:15And then the last question before
25:17we break is if you looked at
25:19expression of of the individual jeans
25:22a particularly as they relate to
25:25potentially classic tumor suppressor
25:26genes or other typical mechanisms.
25:30I'm so that is the follow up that
25:32we're actually doing right now,
25:33so everything I showed today is either on
25:36the first part of the talk is impressed.
25:38The second part is in progress,
25:40so it's kind of early days still on this.
25:42But yeah, our goal is then
25:44to move to expression.
25:45We have looked at human protein at listen.
25:48Do see some associations in terms of.
25:51Answer and expression in the jeans in
25:54our 12 CG set so we are optimistic
25:57that we'll see differential expression.
26:01Well thank you were or just now at
26:03the top of the hour and I want to
26:06thank Morgan and Marcus for two superb
26:08talks that it really elucidated.
26:10Gray science being conducted
26:12at our Cancer Center.
26:13Thank you all for joining us again for
26:15virtual grand rounds and look forward
26:17again to seeing you all again next week.
26:21Great. Thanks, thank you.