YSPH Biostatistics Seminar: “Feature Aggregation in Causal Discovery for High-dimensional Data: Application to Targeting the “Gut-Brain-Axis” via the Microbiome Diversity"

September 21, 2023

Information

Jinyuan Liu, PhD, Assistant Professor, Department of Biostatistics, Vanderbilt University Medical Center

September 19, 2023

ID10729

To CiteDCA Citation Guide

00:01<v ->All right, I'm very excited</v>
00:03to introduce our speaker for today.
00:04We have Dr. Meghan Short.
00:06Dr. Short has completed fellowships
00:08at the Glenn Biggs Institute for Alzheimer's
00:10and Neurodegenerative Diseases,
00:12and at Harvard's Huttenhower Lab.
00:14Currently, Dr. Short is an assistant professor
00:17at Tufts University.
00:18Let's give a warm welcome to Dr. Short.
00:31<v ->Hi, everyone, Thank you for being here.</v>
00:34Can you all hear me, okay?
00:35<v ->Sign in if you're registered.</v>
00:38<v ->All right, so, today, I'm going to talk about a project</v>
00:41that I worked on as part of my postdoc
00:43down at UT Health San Antonio
00:46with the Glenn Biggs Institute for Alzheimer's
00:49and Neurodegenerative Diseases,
00:51and I wanted to talk about this as a...
00:56None of the sort of methods that I'm gonna talk about
00:58in this talk are particularly new.
01:01This wasn't sort of a methods development project.
01:04So the sort of main network method I'll talk about
01:08is about a decade old at this point, at least,
01:10but what's nice about it is that
01:13with increasing availability
01:15of high dimensional biomedical data,
01:17it's sort of seeing more use cases,
01:20and it's not something that, at least, I learned about
01:22in my graduate program in biostatistics,
01:24but it's something that I thought
01:26would be good to talk about today
01:28since it's such a useful method.
01:32So let's see if I advance.
01:35There we go.
01:36So I'll start just by giving a quick introduction.
01:39I know that when I was in grad school, I always wanted,
01:43I thought it was interesting
01:44to hear about people's career paths
01:46as I was considering my own.
01:48So I started in biology as a field.
01:53I studied salt marsh ecology as an undergrad,
01:56and then by the end of undergrad,
01:58I was interested in getting more into sort of a human,
02:00more directly human-focused environment,
02:02and so I considered public health.
02:05I learned about statistics
02:06as part of my research in undergrad
02:08and wanted to continue with that so I participated in SIBS,
02:11which is a program that you may be aware of,
02:14and that was my first intro to biostat.
02:16I was a graduate student at Boston University.
02:19I had fortune of working
02:21with the Framingham Heart Study,
02:22which is where the data comes from
02:24that I'll be talking to you about today,
02:26which is a really interesting study,
02:27and I'll get more details on in the few slides.
02:29That was sort of my introduction
02:31to working with epidemiological data.
02:34After grad school, I continued on,
02:36again, to UT Health San Antonio,
02:38and then following that to postdoc at Harvard
02:42looking at developing methods for microbiome analysis.
02:47So if you have any interest in that,
02:49feel free to approach me,
02:51although I'm not gonna talk about that today,
02:54and then as of March this year,
02:57I started as an assistant professor at Tufts Medicine
03:00where I'm working on a variety of projects
03:02but a lot related to sort of omics data
03:06and aging and longevity.
03:12So I'll start today's talk with a bit of motivation
03:15for why network-based analyses we're a good fit
03:18for looking at sort of the proteome in Alzheimer's disease.
03:24So first of all, Alzheimer's disease
03:27is a very prevalent condition.
03:30Many of you may be like me and know some family members
03:33or people who have been affected by it.
03:36It's very common and expect it to be more so
03:40as populations age, and it's a leading cause of mortality,
03:44disability, and poor health among seniors,
03:47and one interesting feature of this disease
03:49is that precursors of it can appear years to decades
03:52before symptoms manifest.
03:55So those precursors can include indicators
03:58that are visible on brain MRIs,
04:01performance on neurocognitive testing, changes in gait,
04:05even changes in sense of smell,
04:08and cerebral spinal fluid markers, such as tau and amyloid.
04:17Because of this, there's interest in being able to find
04:21plasma biomarkers for Alzheimer's disease
04:24and related dementias.
04:25ADRD is a acronym we'll be using sort of throughout.
04:30Because since there are indicators
04:33of sort of pre-disease development
04:35in years to decades before being able to detect those,
04:37either earlier or in a less invasive or expensive way,
04:40is very useful,
04:44and so when I say invasive, I mentioned CSF markers,
04:50such as how an amyloid can predict dementia,
04:53but that involves doing a lumbar puncture
04:56versus something like a blood draw, which is easier to do.
05:01Another good aspect of trying to find biomarkers
05:04is that you can get a sense of biological processes
05:07that are involved in disease development,
05:10and that can hopefully lead to either preventative
05:13or therapeutic interventions.
05:19What makes this difficult?
05:21So in my case, I was looking at proteins.
05:24There are thousands and thousands to select from,
05:27and you get sort of this inherent trade off
05:30between trying to control a false positive rate
05:33for all these multiple tests that you may be performing,
05:36but if you effectively control the false positive rate,
05:39you're going to likely end up with low statistical power.
05:42There's this trade off between...
05:45It's sort of a needle in a haystack.
05:47Another thing that has tended to be true
05:50is that there is not very good replicability across studies.
05:53So one study may find 20 biomarkers
05:57and maybe one or two of them
05:59may replicate in a different study.
06:01So there's a lot of noise that ends up coming through.
06:08The approach that I took in this project
06:10was to use network analysis
06:13to analyze the protein data,
06:17and the motivation there is to try and capture
06:20subtle but consistent variation in groups of proteins.
06:23I'll refer to them as modules during this talk.
06:28In then just a few things, so first of all,
06:30it reduces the dimensionality
06:32of the statistical testing problem that you have.
06:35So rather than testing each protein individually
06:37and having to adjust for all of those multiple tests,
06:40you can sort of reduce the space
06:43to a smaller number of tests
06:46where the proteins within each group being tested
06:49are inter-correlated with one another,
06:52and unlike other dimensionality reduction methods,
06:55something like a principle components analysis
06:57that you may have maybe familiar with,
06:59the network method has sort of a benefit of looking
07:03not just at, say, correlations
07:06or relationships between pairs of proteins,
07:09but, also, at sort of the correlational neighborhood
07:11of what common neighbors
07:13those proteins share in the network.
07:18Another benefit of or sort of way
07:22that we try to get around some of the pitfalls
07:23of proteomic analysis is by focusing on biological pathways
07:29instead of on individual proteins themselves.
07:32So within groups of proteins that we find to be of interest
07:36or possibly associated with dementia outcomes,
07:40we use a tool called over-representation analysis,
07:43which I'll talk about later,
07:45but it essentially tries to pinpoint biological pathways
07:48that may be overrepresented by the proteins
07:51that are found to be associated with the outcome,
07:54and the hope there is to find,
07:56to get sort of insights that are more robust across studies
08:01and, hopefully, address some of the issues
08:03with replicability.
08:08Okay, so that's sort of the motivation for this study,
08:11and, now, I'll talk a little bit about the data.
08:18The data for this study
08:19comes from the Framingham Heart Study,
08:22which has been going on for a very long time.
08:24It started in 1948 in a town of Framingham, Massachusetts,
08:29and at the time they enrolled,
08:31they reached out to two-thirds of the population of the town
08:34to try and enroll them in this epidemiological study.
08:36It was one of the first ones of its kind,
08:39and people would come in for exams every few years,
08:42and they would take all of this information about them,
08:45and then follow them for outcomes.
08:47Cardiovascular outcomes was really
08:49the sort of outcome of interest when it first started.
08:53Over the years, they've then enrolled offspring
08:57of the original cohort participants
08:59as well as grandchildren and third generation,
09:02and then as sort of the demographics
09:06of Framingham have changed over the years,
09:09if you're only enrolling descendants
09:10of people who live there in 1948,
09:12you're not gonna capture that.
09:13So they also have been enrolling omni cohorts
09:15to reflect sort of more diverse populations (indistinct).
09:21Again, they were sort of aiming
09:23towards identifying risk factors
09:25and etiologies of cardiovascular disease,
09:29but as those populations age,
09:31brain health and cognition is also an important outcome,
09:34and so they've measured sort of cognitive outcomes
09:39and incidents of dementia as well, and, of course,
09:41those things are also related to cardiovascular.
09:48For our study in particular,
09:51we were using the offspring cohort,
09:53and at their examination cycle five,
09:55which was in the early 90s, they collected blood samples,
10:00and froze the plasma from those samples,
10:03and years later, when they sort of had
10:06these broader proteomic analysis assays available,
10:11they measured the plasma proteome,
10:14I'll talk about the methods for that on the next slide,
10:18but they did this in about 1,900 participants
10:21who were approximately aged 55 when the blood was drawn.
10:24So this is sort of a middle-aged cohort,
10:26generally, cognitively healthy
10:29and a little more than half women.
10:33The main outcomes that we looked at in this study
10:35are MRI-based measures, so brain MRIs were taken
10:41about 10 years or so, five to 10 years
10:45after the initial blood draws, and those had...
10:51The sort of outcomes that I looked at there are
10:54total brain volume as well as the volume of the hippocampus
10:57and then a measure called white matter hyperintensities,
11:01which is sort of a measure of vascular injury in the brain,
11:06and a reason to look at those outcomes is that
11:10I mentioned there are sort of precursors of dementia
11:13or risk factors for dementia that can be identified on MRI,
11:16those are some of the big ones.
11:19Especially since we had a middle-aged cohort,
11:22you may not see a lot of incident dementia,
11:24and so being able to detect proteins
11:27that are associated with some of those precursors
11:29is a way of getting at this issue.
11:34We did also look at incident dementia.
11:36So we had about 20 years of follow-up,
11:37which is one of the strengths of this,
11:40looking in this particular sample,
11:42and we had 128 incidences of dementia
11:46of which 94 of them were classified
11:48as Alzheimer's type dementia.
11:53We also had a replication cohort.
11:55I mentioned the importance replication,
11:58and so we worked with collaborators
12:00at the University of Washington and their cohort study
12:04called the Cardiovascular Health Study,
12:06which has sites, I think, four different sites around the US
12:09and has measures of the same proteomic platform
12:13and same outcomes that we're looking at in the study.
12:19The assay that we used to measure proteins
12:23is called SOMAScan.
12:24It's by this company called SomaLogic.
12:27They use these single-stranded DNA aptamers
12:29that are designed to specifically bind
12:31to different proteins, and you can sort of tag them
12:35that way and measure their concentrations.
12:38In our sample, the assay had 1,300 proteins,
12:42which that's even sort of becoming dated now.
12:45I think the latest version
12:46has something like 7,000 proteins.
12:48So there's a lot that can be measured with this,
12:51but there is some sort of bias towards, I think,
12:57molecules that sort of have some evidence
12:59of being important in cardiovascular disease.
13:01So it's not an entirely sort of agnostic choice of proteins,
13:06but it does get a pretty wide range.
13:11Okay, so that's a description of the data,
13:15and, now, I want to dig in a bit
13:17to the network methods that we used.
13:20So this is sort of a graphical abstract
13:24from their original paper,
13:28describing this weighted gene
13:29correlation network analysis method.
13:32So that's what WGCNA stands for.
13:34I put gene in parentheses because they've started
13:37dropping that from the name when it gets used elsewhere
13:40because, originally, it was developed
13:42for gene expression data, but it's been found to have use
13:45in other high dimensional data sets as well,
13:48and so in our case, we're using it to analyze proteins,
13:52but the language here makes reference to gene expression.
13:57So just broadly, what this method does
14:01is you get a co-expression network,
14:04and I'll sort of give details on the next few slides,
14:08but the idea is that the network is based
14:10on co-occurrence or correlation in your sample.
14:14So there's not really information coming from outside.
14:17You're not even considering your outcome at all.
14:19It's just looking at the space of the proteins
14:21and which proteins are correlated with one another.
14:26Once you've identified this sort of network matrix,
14:29you use a hierarchical clustering algorithm
14:32to define modules.
14:34It's a little small here, but I'll show a a bigger example.
14:37Basically, you have a dendrogram,
14:39and you see that if sort of proteins
14:42are on this x-axis of this figure here.
14:44I'll do the mouse for people who are online.
14:48You get these sort of bands or groups of proteins
14:51that are highly correlated with one another
14:53and not correlated with other proteins.
14:58So that is where those sort of protein groups come from.
15:02Once you have those, you can use a numerical summary
15:06of each protein group as sort of a feature or a predictor
15:10in a regression or some sort of analysis
15:13to try and relate the modules or groups
15:15to external information.
15:16So that's how we relate our protein groups
15:20to dementia outcomes in this study.
15:24There's also the possibility
15:25of looking at relationships between modules.
15:28So I mentioned the modules in the network
15:32are highly inter-correlated
15:33within the proteins within themselves,
15:36but there may also be some correlation between modules,
15:38and that could be important to look at as well,
15:41and then within modules, you may have
15:44tens or hundreds of proteins, and so trying to figure out
15:47which proteins within those modules
15:50are driving any associations you see
15:52is sort of a final step that can be
15:55useful for getting sort of biological meaning
15:57out of these associations.
16:02So that's a broad overview.
16:03This is sort of a more graphical abstract from our study,
16:08and I'll sort of go through bit by bit
16:11the different pieces of the analysis.
16:15So, again, this WGCNA step is sort of the first step
16:18of getting from this protein expression matrix
16:20where you have sort of your proteins by participants,
16:24and using the sort of correlations in your sample
16:27to come up with these modules of co-expressed proteins.
16:33The first step in doing that
16:35is to make a pairwise correlation or similarity matrix.
16:39So if you have n proteins,
16:40then that becomes an n by n matrix
16:43where each cell is describing
16:45the similarity or correlation
16:47between protein i and protein j in your sample.
16:52You then use this to create
16:54what's called an adjacency matrix, which is,
16:56I'll talk about more in the next slide,
16:58but is sort of a more networky way
17:02of describing the association between proteins,
17:05and then a topological overlap matrix,
17:08which then takes into account
17:10not only the correlation between proteins
17:13but their shared neighborhood, and then, again,
17:16that is what is used to cluster the proteins.
17:23So to get into a bit more detail
17:25about sort of the network construction,
17:30again, you described the network as an n by n matrix
17:33with the number of nodes or genes, proteins, et cetera,
17:36and, in our case, we use to describe the similarity,
17:39just a simple correlation,
17:42absolute value of the correlation,
17:44between a given node i and j.
17:48The adjacency is then a measure of whether or how strongly
17:51the nodes are connected in the network.
17:53So the idea being that
17:56nodes that have very high correlations
17:58are particularly interesting.
18:00Nodes that have moderate to low correlations
18:02are probably not informative
18:03is sort of the the underlying idea,
18:07and so if you look at sort of this figure here,
18:12the correlation or similarity is on the x-axis,
18:15and then the adjacency is on the y, and so if you use
18:19what's called an unweighted network approach,
18:22you pick a threshold value, here, it's 0.8,
18:25and you say that anything with a similarity less than 0.8
18:28is considered to not be a connection in the network,
18:31and everything greater than 0.8
18:33is considered to be a connection.
18:34So it's sort of a binary yes or no.
18:38What WGCNA does that was novel
18:42was to introduce a weighting
18:45where sort of the downside of this unweighted metric is that
18:49if you have a correlation of 0.79,
18:52that could be useful to know, but it counts as a zero.
18:55So you're losing information,
18:57and so what the weighted network does
19:00is it uses a sort of power transformation
19:03to get from sort of the straight correlation
19:07shown in this red line,
19:08and sort of depending on this power value that you use,
19:12you weight more or less towards the higher correlations
19:16in your network, and when you fit this model
19:20or when you sort of build the network, your choice of data
19:24is sort of one of the parameters that you choose going in,
19:27and there's ways to sort of measure
19:30which gives the best fit to the data.
19:38So then once you have your sort of unweighted
19:41or weighted adjacency matrix,
19:45then is the part where you account for shared neighbors.
19:48So this is this topological overlap matrix that is created,
19:52so, basically, this measure omega of connectedness.
19:58The equation, I don't find super sort of intuitive,
20:01but the components are...
20:03This is the sum, so u are, basically,
20:05all of the nodes other than i and j
20:07that you're looking at the connectedness between,
20:10and so you're summing up
20:11the sort of common connection strength between i and u
20:15and j and u as a product.
20:18So if I and J both have a strong connection
20:22to this other node, then that's adding to this term l,
20:27and then these k terms here
20:29are just the individual connections between, no,
20:32each sort of the node i of interest
20:34and other nodes in the network,
20:37but I find sort of the easiest or most intuitive explanation
20:41from this original paper shows that for the unweighted case,
20:46omega is equal to one if the node with fewer connections
20:50has all of its neighbors,
20:51also, has connections of the other node.
20:53So the connections of node i
20:55are a subset of the connections of node j,
20:59and, also, i and j are directly connected.
21:01So that's sort of the most interconnected
21:03that those two nodes can be,
21:05and then the least interconnected they can be
21:08is if they are not connected to one another,
21:10and they don't share any neighbors.
21:11So that would be sort of the zero case.
21:16So this a value can either take on
21:18the unweighted or the weighted case,
21:20and in our sample with WGCNA,
21:23we're using those sort of weighted network connections
21:26that just adds more information
21:28into this topological overlap matrix.
21:36Okay.
21:39So, now, once you have the topological overlap matrix,
21:45again, this measure of sort of interconnectedness
21:48accounting for shared neighbors,
21:51then you can use hierarchical clustering
21:53to divide those proteins
21:57into groups based on their similarity,
22:00and this is the results from our analysis.
22:03So sort of on the x-axis,
22:06you have the different proteins, you have the dendrogram,
22:09which represents the hierarchical clustering
22:11of the topological overlap matrix,
22:14and then you have this dynamic tree cut algorithm
22:20which then defines these clusters
22:22which are shown in colors on the bottom based on the tree.
22:26So you see this huge branch down here.
22:28That's gonna be this black cluster.
22:30There's this other cluster over here in green,
22:33and so there's, again, a few more parameters
22:37that you can use to decide how those cuts are made,
22:40and, in some cases, you can sort of merge branches
22:43that have correlation with one another,
22:45and my general advice
22:48for when you're doing this on real data
22:49is to try different values
22:51and see how robust the network is
22:53to choosing different values because, in our case,
22:56it tended to be pretty consistent
22:59where we saw four modules pretty much regardless.
23:02I think if we merged,
23:04if we really cranked up one of the merging parameters,
23:06we would get to three,
23:07but other than that it sort of stayed put.
23:13Okay.
23:15So the next step is trying to get
23:18a numerical summary measure of the groups of proteins
23:22that we've identified from our network.
23:25So from these modules of co-expressed proteins,
23:28we then use, basically, a principle components analysis
23:33to get what we call an eigenprotein
23:35or it was called an eigen gene in the original paper.
23:39What it is is, essentially, a weighted sum
23:43of the values of each of the proteins in the module,
23:47and the weights correspond to sort of how well correlated
23:50that protein is with the overall module.
23:53So if a protein has a high weight in the module,
23:56it means that it's sort of the most interconnected
23:58in the module or sort of best represents the overall module.
24:04So each person is going to have
24:06an eigenprotein value for each module,
24:16and when we look at the sort of weights
24:18within each of the modules, so just to sort of orient us,
24:22on the x-axis are each of the module eigen genes
24:27or eigenproteins, and then each sort of bar
24:34on the y is a different protein.
24:36In this case, we're only including
24:39proteins that fall into one of the four modules.
24:42There were, also, if you notice on the last slide,
24:45plenty of proteins that didn't fall into any module
24:48and were sort of the extras, so to speak,
24:51and if you were to expand this down
24:54and include more rows with those,
24:56that would sort of show those, but for purposes of this,
25:00we're just including ones
25:01that fell into at least one of the four,
25:04and each of these bars represents a correlation
25:09between the individual protein
25:11and the overall eigenprotein.
25:13So for these blocks of red,
25:15it's sort of the higher weighted proteins
25:17that are within in this example module one,
25:21module two, three, and four, and then you can see,
25:24if you look sort of laterally from these proteins,
25:28it's the correlation of these proteins
25:30with the other modules.
25:31So the idea being we wanna see sort of blocks of red,
25:36and then not a lot of correlation
25:38between the blocks and other modules,
25:40which is what we see.
25:46All right, now that we've constructed our network,
25:50and we've come up with numerical summary measures
25:52for each of the protein groups that we've identified,
25:56that is sort of the input or the predictor
25:59for these associations with outcomes.
26:02So for the MRI measures, which, again,
26:04our total brain volume, hippocampal volume,
26:07and white matter hyperintensities,
26:09we use just a simple or, you know,
26:12linear regression with covariates,
26:14and then a Cox proportional hazards regression,
26:17we use to predict incident dementia
26:20and, specifically, Alzheimer's type dementia.
26:26These are the regression equations.
26:28Again, these eigenproteins are,
26:30they're sort of one for each module.
26:32So we'll run a separate regression analysis
26:35for modules one, two, three, and four.
26:37We adjust for age and age squared, sex education.
26:42APOE is a gene that confers a lot of risk
26:44for Alzheimer's disease.
26:45So it's associated with the outcomes,
26:47and we include it as a covariate,
26:49and then a measure of time lag
26:51between when the blood was sampled
26:53and when the MRI was taken to account for any differences
26:57between people or the time difference,
27:01and for dementia, it's slightly simpler regression equation.
27:06We only adjust for age, sex, and APOE status.
27:13All right, so next, I will show
27:17the results in the Framingham Heart Study.
27:21So from the four modules that we tested,
27:24there were two that we identified to have
27:27some association with outcomes.
27:29The first is module two.
27:31I gave it sort of a name clearance and synaptic maintenance,
27:35and I'll talk about how I arrived
27:37at that name for the module in a bit.
27:40It has 165 proteins in it.
27:44Some of the half weighted proteins sort of give an idea
27:47of which ones are sort of most highly weighted
27:51or sort of most correlated with the eigen protein.
27:56I'll talk about how we got to these
27:59in another slide as well,
28:01but, basically, this is from that
28:02over-representation analysis
28:04where you're trying to identify biological pathways
28:06that are important or overrepresented
28:09by proteins in those modules.
28:12So we have the Axon guidance pathway
28:14was most strongly associated with this module,
28:21and then in terms of relating to outcomes,
28:25total brain volume
28:26was the only significant association that we saw.
28:29So since this is a linear aggression,
28:33effect greater than zero means a positive association.
28:37So we see that for larger values
28:40of the eigenprotein for module two,
28:42we saw larger total brain volume.
28:44So it's sort of a protective effect
28:47since brain atrophy is what is the risk factor for dementia,
28:53and then for incident dementia,
28:55we did not see a significant effect
28:56after correcting our p-values
28:58using a Bonferroni correction.
29:00You'll notice that the confidence interval excludes one,
29:04which would be the null value,
29:05and that's just because that's based
29:06on the non-Bonferroni corrected value,
29:10but after testing for or adjusting for the four modules
29:14that we tested, we didn't see a significant association.
29:18It is nice at least that the direction of effect
29:22is what we would expect
29:23based on our total brain volume association,
29:26which is that higher values of M2
29:31correspond to sort of a lower incident dementia occurrence.
29:38The second module that we found to be associated
29:41with total brain volume was this M4,
29:44which I will call sort of an inflammation-related module.
29:47It had 42 proteins in it.
29:50The highlighted pathway there
29:52was cytokine-cytokine receptor interactions,
29:55so these sort of immune signaling molecules,
29:57and in this case, the association
30:00was in the opposite direction
30:01where higher values of this module for eigenprotein
30:05are associated with lower total brain volume.
30:07So it's sort of a risk conferring module
30:10and, again, similar to what we saw here, not a significant,
30:14sort of an annoyingly borderline association
30:17between this and dementia, but, again,
30:20the direction of effect is what we would expect
30:24based on our observed association with brain volume,
30:29and, also, I'll just mention that I standardize
30:31the eigenprotein so that the effect sizes
30:34correspond to a standard deviation increase in eigenprotein.
30:37So it's a little bit...
30:39One sort of drawback I would say
30:40of these methods is the interpretation
30:43since a standard deviation increase, in this case,
30:47depends entirely on the sample that you're using.
30:49So it's really just sort of a direction of effect
30:54more than anything.
30:56So to try and get at some of, get a better understanding
31:00of how these modules relate to our data
31:03or sort of what may be responsible
31:06for some of the associations we see,
31:08this is a map of the correlations
31:12between different demographic variables
31:15and each of the modules, and I mentioned that we have
31:18a replication cohort as well, the CHS.
31:20So these two bars, sort of the two columns,
31:23show the two different cohorts that were included.
31:28So I put blue arrows to show the covariates
31:31that were included in our regression model,
31:34and you can see that there are some correlations
31:35between, say, sex and the modules,
31:38not really anything with APOE carrier status,
31:42maybe some education associations,
31:44and some associations with age.
31:46So it's good that we adjusted for those in our models.
31:49However, you can also see there are a lot of other factors,
31:53cardiovascular risk factors,
31:54such as systolic blood pressure, BMI,
31:58fasting glucose that have associations with these modules.
32:02So we wanted to see if any of those could perhaps explain
32:05the associations that we saw.
32:10So I'm repeating sort of our standard model here
32:14was what I showed results from previously.
32:17The expanded model that we considered
32:19included a bunch of these risk factors,
32:23basically, something representing BMI,
32:27hypertension, sort of lipid dysregulation, and diabetes,
32:33and I also included smoking as well,
32:37and we also included a measure of kidney function,
32:40which can also be an indicator of cardiovascular disease.
32:45So for module two,
32:48I'm repeating the sort of effects we saw
32:50from the standard model here,
32:53and when you adjust for the expanded set of covariates,
32:56your effect is attenuated by half,
32:58and it's no longer significantly associated.
33:01So with that says, it's either you have
33:04a sort of confounding issue
33:08where the association you're seeing between these proteins
33:12and total brain volume is really just in effect
33:16of sort of poor cardiovascular health
33:20or better cardiovascular health
33:22or you may think that it might be
33:25some sort of mediation effect
33:26where perhaps the risk associated
33:31between the proteins and the sort of total brain volume
33:34could be mediated
33:35by some poor cardiovascular health outcomes,
33:41and then for module four,
33:43again, this sort of inflammation module,
33:45we don't see any real effect attenuation.
33:48Regardless of whether you adjust
33:49for cardiovascular factors or not,
33:52it's still associated with total brain volume,
33:54which suggests it's sort of different mechanism
33:57or lack of compounding between
33:59or based on cardiovascular health.
34:05Okay, so I mentioned
34:08in the sort of initial graphical abstract
34:12that once you find protein modules
34:14associated with your outcomes of interest,
34:16it can be good to look within the proteins of those modules
34:19to try and find sort of subsets
34:21or specific proteins that may be driving the associations.
34:26So for modules two and four,
34:27where we found associations with brain volume,
34:30we wanted to see if we removed proteins one at a time
34:35based on their sort of increasing weight,
34:37so remove the lowest weighted proteins in the modules first,
34:42what sort of happened to the strength of the associations.
34:46So these are both associations with total brain volume.
34:49It's sort of the p-value on the y-axis,
34:53and you can see that as you remove, say, from module two,
34:57the first 20 proteins or so,
34:59you're really not seeing a difference
35:01in the effect of the overall module with total brain volume,
35:05which suggests that those proteins
35:07aren't really impacting the association,
35:11whereas beyond that point, once you start removing proteins,
35:15the association becomes less strong,
35:17and so that's suggesting that those proteins
35:20may have more of an impact on sort of the overall module,
35:25and so for both of these modules, we identified the spot
35:29where sort of the based on the lowest p-value,
35:32which proteins were
35:35sort of the most important in the module.
35:37I wanna emphasize that we didn't use this to...
35:41So for things like dementia, if you were to run this,
35:44since we didn't see a strong association
35:47or a significant association beforehand,
35:50we didn't sort of use that to try and find a subset
35:52that we're significantly associated
35:54because I would call that cheating.
36:01Okay, so the last piece that I'll talk about
36:05in terms of teasing apart associations
36:09or sort of understanding protein within the modules
36:13is this functional enrichment
36:16or over-representation analysis within the modules.
36:20So based on the ones, sort of the significant modules
36:24or significantly associated modules with the outcomes,
36:28there is this software called STRING
36:31that does a few different things, but what I used it for
36:35is doing an over-representation analysis
36:38of biological pathways.
36:41So the idea is that there are annotation databases
36:45for proteins that sort of group them
36:48into biological functions
36:51or pathways that they're involved in,
36:53and the idea is that if you have a module
36:55that has more proteins than you would expect
36:58from a given pathway,
36:59then that's sort of the over-representation piece,
37:02and it indicates that that biological pathway
37:05might be important in whatever functions
37:08the module is carrying out.
37:12So this is just a screen grab of one example.
37:16So this is from module four.
37:18So you can see the annotation database is over on the left.
37:22So KEGG is one of them.
37:24Gene Ontology is another,
37:26and so you have these sort of observed proteins,
37:30and then the background is sort of the total number
37:33of proteins that are in the pathway,
37:36and the idea being that if you were to grab, I don't know,
37:39however many proteins out of the background,
37:41like how many would you expect to be in this module
37:45due to chance, and do we have sort of over-representation
37:49compared to what we would expect?
37:51And so for module four,
37:52the cytokine-cytokine receptor interaction
37:55was the strongest overrepresented pathway,
37:59and then you can sort of look at these others that
38:03have some sort of false discovery rate greater than 0.05,
38:08and so I found the KEGG pathways, personally,
38:11to be the most informative.
38:12Gene Ontology tends to be a lot more specific,
38:15which may be more useful for targeting
38:18certain sort of therapeutic processes
38:21or something like that,
38:22but so depending on the scale that is important to you,
38:25you can sort of use different annotations.
38:31Okay, so the last thing I wanted to talk about,
38:33with the Framingham data in particular,
38:38was sort of getting back to our motivation
38:40for doing a network analysis in the first place.
38:43So the sort of contrast or comparator would be to do
38:47individual protein analyses where you're running
38:49a regression model for each protein that you're analyzing,
38:53and so we did that as a point of comparison.
38:55So for total brain volume, there were like a dozen proteins
38:59that were associated with total brain volume.
39:02One was associated with hippocampal volume,
39:04and two were associated with Alzheimer's disease
39:07at an FDR value of less than 0.1.
39:11So what was interesting,
39:14especially with the brain volume results,
39:16and, again, that was where we had seen
39:17associations with these modules,
39:19some of the proteins that were significantly associated
39:23were from module two and module four and others weren't.
39:29So what I get from that is a few things.
39:32One is that some proteins
39:34that are associated with the outcome
39:36are sort of individually associated
39:39but not sort of detectable
39:41within sort of a larger network of proteins
39:44that are associated with that outcome,
39:46and then the other is that
39:48for those that are within the modules,
39:51we would only be getting information
39:53about sort of a few of the proteins in the modules,
39:56whereas, as we see here,
40:00the associations tend or continue to get stronger
40:03with sort of looking at the broader network
40:06around sort of the most highly weighted proteins.
40:09So you're getting a bit more information
40:10about proteins that may be associated
40:13with total brain volume
40:14and maybe at some of the biological processes
40:17compared to if you're looking at things individually,
40:20but, again, because you're seeing associations
40:22that you don't catch with the modules,
40:23it's sort of important to look at both,
40:25and you get sort of complimentary information
40:28from the two approaches.
40:33So a caveat,
40:36I mentioned issues with lack
40:37with sort of difficulties in replication.
40:40We replicated this analysis
40:42in the Cardiovascular Health Study,
40:44and we did so by taking the same module,
40:47so module two and module four,
40:50taking the same weights from those proteins
40:52and applying them to the protein concentrations
40:56in the Cardiovascular Health Study.
40:59So we didn't do a network reconstruction or anything
41:02in the different study.
41:03We were just seeing if these modules replicated
41:07in their associations with outcomes in a different cohort.
41:10So in this case, it's really not seeing much
41:14in terms of association with both total brain volume
41:18and we also looked at dementia out of interest
41:22since things were sort of close in our cohort,
41:26but, really, we're not seeing much in terms of associations.
41:31Part of the reason for that,
41:33so there are not that many cohorts
41:36that are available that have a large proteomic panel
41:39with the same proteins that we were looking at
41:41as well as MRI and incident dementia outcomes,
41:45and, in this case, the demographics of the cohort
41:48are fairly different from (indistinct) Framingham.
41:51So about 20 years older on average.
41:56I'm just including the sort of first few rows
41:58of our table one, but you can see differences in education,
42:01systolic blood pressure, and the same is true
42:03of a lot of the other cardiovascular risk factors.
42:06So it's a very different cohort,
42:08and digging a bit into the literature
42:10about sort of proteins over the life course,
42:13it's not too surprising that we don't see
42:16the same associations, but it it does sort of,
42:19it's a good cautionary message
42:20about drawing conclusions too far
42:23based on sort of one set of data
42:25or one set of demographics.
42:30Just to put these results in context,
42:32so our module four included
42:36a lot of immune-related signaling molecules
42:38like interleukins, TNF receptor proteins,
42:41which are both types of cytokines, and have been associated
42:44with Alzheimer's disease previously,
42:47in particular, interleukin-1 beta was in our module four,
42:52and it had been found to be elevated
42:53in 80 cases in a meta-analysis.
42:56However, other biomarkers that have been sort of validated
43:00in other cohorts were not identified in our module.
43:08In module two, we saw Axon guidance pathway proteins
43:11including ephrins, netrins, and semaphorins,
43:13which have been associated with AD in previous work,
43:17and complement cascades are also have been associated
43:20with AD probably for the reason
43:22of inducing these immune cells called microglia
43:27in the brain to, basically, eat up
43:31cells in response to amyloid deposition.
43:35So there's some biologically plausible mechanisms
43:37that could be associated with these modules
43:42in Alzheimer's disease,
43:46and the last thing I'll say is talking about some sort
43:49of other ways of approaching this problem,
43:51so as I mentioned, the CHS cohort
43:54has different underlying characteristics,
43:56and so it may well have a different network structure.
43:59So one thing that could be good to do
44:02is to look at sort of consensus modules across the cohorts
44:07where you construct networks in each cohort,
44:09and then look at where the overlaps are,
44:12and you can get sort of a more,
44:14hopefully, more robust network across cohorts,
44:18and then there are other network-based approaches
44:20that can incorporate external information.
44:22So, again, our network approach
44:24was just based on correlation in our dataset,
44:27whereas other methods use sort of those annotation databases
44:33and that sort of thing to construct the networks
44:35and sort of decide how strong the similarities between nodes
44:39or the strength of connections will be.
44:41So that's another approach,
44:43and then the last thing I'll say is that
44:45I'm sort of still using this kind of method
44:48now in work with longevity and aging
44:51and trying to apply it to metabolomics,
44:54so metabolites data in cohorts related to those outcomes.
45:02So thank you all for being here.
45:04Thank you, my collaborators.
45:06This is the folks down at UT.
45:09I'll say that (indistinct).
45:11Thank you.
45:16<v ->Thank you for wonderful presentation.</v>
45:18We're open for questions.
45:20So let's start with people in the room.
45:21Any questions?
45:23<v ->Got one over here.</v> <v ->Perfect, thank you.</v>
45:25<v Audience>Yeah, so my research interest</v>
45:26is about the cancer, and, also,
45:28we're interested in your study.
45:30So I've got some technical issues about this project.
45:35So the first issue that,
45:36how do you do the normalization in your process?
45:41<v ->Yeah, great question.</v>
45:42So yeah, I totally glossed over
45:44all the pre-processing stuff.
45:47So before doing the network construction,
45:51I log transformed the protein concentrations
45:54to reduce stiffness.
45:56There was a standardization within,
45:58there were sort of two phases of runs of protein modules,
46:02so I sort of standardized within those batches,
46:06and then after that, I did a rank normalized
46:11or inverse normal rank transformation to sort of-
46:16(audience speaks indistinctly) <v ->What's that?</v>
46:17<v ->(indistinct) normalization?</v> <v ->Basically.</v>
46:19Yeah, yeah, yeah.
46:20So that was sort of the data pre-processing.
46:23So I think I, you know,
46:26I've thought about sort of the pros and cons
46:28of those things as well and I think my biggest qualm
46:31with the way that I did it is sort of interpretability,
46:34because, yeah, sort of what does it mean
46:37to be at one quantile versus another
46:39where you have this huge dynamic range
46:40of protein concentrations?
46:42<v Audience>So another question is that</v>
46:44I know that in your project,
46:46the modules identification is very important.
46:48So I wonder,
46:53you have talked a little bit
46:54about how to answer the modules,
46:57but so can you explain a little bit more
47:00about how you gonna bring modules from the data?
47:08<v ->I'm not sure, can you say a little bit more?</v>
47:11<v Audience>Yeah, so in your previous pages,</v>
47:13I think you talked a little bit about the clustering
47:17of the modules so that we know
47:18that there are four main modules.
47:22<v ->Yes.</v> <v ->In the whole dataset.</v>
47:24So what is the name of that algorithm
47:28and how it basically work?
47:31<v ->Yeah, so the clustering itself was done</v>
47:36using algorithm called H+.
47:41To be honest, I'm not too sure
47:43about sort of the details of it.
47:45It can use any dissimilarity measure,
47:48which, in our case, comes from the TOM matrix, but-
47:52<v Audience>So this is the algorithm that we separate</v>
47:55the whole proteins into four different modules
47:58so that we can analyze it one by one.
48:00<v ->Yeah, yeah, yeah, yeah.</v> <v ->Yeah,</v>
48:01so I also noticed that
48:07in the weighted protein expression network analysis,
48:13you talk about the beta values.
48:16<v ->Yes.</v> <v ->That you use that</v>
48:20like the soft threshold. <v ->Yeah.</v>
48:23<v Audience>To make the genes to be more important</v>
48:28if that is the thing that you wanna analyze.
48:31So in this process, I want to know how you would make sure
48:35the value of the data in this process.
48:39<v ->So sorry, we have to end 'cause it's 12:15.</v>
48:42I know others have classes and everything.
48:44Maybe you guys can discuss a little bit.
48:46<v ->Yeah, (indistinct), yeah.</v> <v ->Maybe if you have time.</v>
48:48Please, if you're registered,
48:49make sure you signed in on a sign in sheet.
48:51There's three of 'em.
48:52You only have to sign on one of them,
48:54and then one-fourth page reflections will be due
48:57before the next speaker's time to speak.
48:59(indistinct talking)