PoT Examples and Conclusions

Loading...

En provenance du cours de Johns Hopkins University

Toxicology 21: Scientific Applications

14 notes

Johns Hopkins University

Toxicology 21: Scientific Applications

14 notes

This course familiarizes students with the novel concepts being used to revamp regulatory toxicology in response to a breakthrough National Research Council Report “Toxicity Testing in the 21st Century: A Vision and a Strategy.” We present the latest developments in the field of toxicology—the shift from animal testing toward human relevant, high content, high-throughput integrative testing strategies. Active programs from EPA, NIH, and the global scientific community illustrate the dynamics of safety sciences.

À partir de la leçon

Module 3

This module will familiarize you with the Human Toxome Project run by Johns Hopkins University in collaboration with several university, industry, and research agencies. Omics technologies, their advantages and disadvantages, as well as concepts of Adverse Outcome Pathways (AOP) and Pathways of Toxicity (PoT) will be discussed.

Introduction - What Is a Pathway of Toxicity, and Why Do We Need Them?6:43

Omics Technologies and High-Throughput Data11:23

Interpreting Omics Data11:27

PoT Examples and Conclusions15:19

Pathways of Toxicity and Adverse Outcome Pathways (AOPs)13:58

Rencontrer les enseignants

Lena Smirnova
Research Associate
Center for Alternatives to Animal Testing
Thomas Hartung
Professor
Environmental Health and Engineering

So welcome back.

In our previous lecture, we talked about ways you can interpret omics data.

And in this lecture, we're going to give an example of using several different

approaches to analyze a very small microarray study.

And then, we're going to talk about some conclusions and

forward directions for pathways of toxicity.

So for this lecture, I'm going to show you an example using all this tools

together and discuss the future of omics technologies.

The study that I'm going to talk about was part of the Mapping

the Human Toxome project.

And this project was an attempt to use microarray and

other high dimensional data to deduct Pathways of Toxicity.

And one of the things we discovered early on, was that inferential statistics were

not adequate, for all the reasons that I discussed in the previous lecture.

You, typically, will either see too few genes,

if you're very conservative with your multiple hypothesis testing.

Or you'll see too many, if you are not conservative enough.

And it's very, very difficult to differentiate the signal from the noise,

if you're dependent on inferential statistics.

The other problem was that most samples were too small for

conventional machine learning.

So machine learning, you'll often see when it's applied to very,

very large data sets of hundreds of thousands of samples.

For example, that's how Google figures out what you're trying to type,

and it's how we can let machines learn and think and

do all the things that they do in our daily lives.

However, most toxicology sample sets do not provide enough data for this approach.

The other problem that we had was that many approaches depended on using existing

pathways.

So for example, I talked about gene set enrichment analysis will see what you're

looking for, but it can't see unknown biology.

And for us, this was very important,

because in toxicology a lot of processes are not well described.

And it's especially important to remember that a lot of toxic responses are often

repurposed pathways from other processes.

So for example, you'll often see developmental

pathways that are sort of turned on because a toxicant is stimulating them.

So we used an approach called Weighted Correlation Network Analysis, and that is,

like all network based approaches,

essentially a guilt by association approach.

Genes that are correlated together or

doing the same thing at the same time are considered to be interacting.

And you use that to draw a de novo network from the high-dimensional data.

So again, it sounds very complicated and there is a lot of math behind it.

But basically what you're doing is trying to figure out which genes do

the same things at the same time and using that data to connect them in a network.

So you have to think about how Facebook connects everybody with likes,

where the gene correlation analysis connects genes together by looking at

their similar patterns.

And the advantage of this, it does not depend on database annotations, and it

treats the underlying biology as a system rather than at a part by part basis.

I'm just going to give you an example of a case study that we use.

NPTP is a very well known toxicant and

the phenotype is similar to Parkinson's Disease.

NPTP is transported into dopaminergic neurons by the dopamine transporter.

It acts to inhibit mitochondrial respiratory chain complex one and

it causes oxidative stress and ATP depletion.

And this triggers a very complicated cascade of molecular pathways and

it eventually leads to cell death.

This gave us a really nice test case, because a lot of the biology is known.

We know that it starts with mitochondrial dysfunction and this eventually leads to

a problem in calcium homeostasis, neural information and apoptosis.

So there was enough biology that we can see if our approach worked or

not, but it was also there's enough pieces to the puzzle missing that we

could put something interesting on the map.

So we found a very small study, mice injected with 30 milligrams per kilogram

of MPTP, so 12 microarrays that control 1 in 7 days.

And this is, by the standard of microarrays, a pretty small study.

Most studies are typically 30 or 40 arrays.

But it had enough data for us to tease out the pathways.

So again, the way to correlation, a network uses gene correlation combined

with clustering to assign genes into the modules.

So again, you're drawing your network based on some correlation metric, and

the next step is to use the clustering,

that I discussed earlier, to assign the genes into modules.

So each module on this map represents genes that

are clustered together based on their similarity and appear to be coregulated or

involved in the same biological process.

So how do we know if these clusters make sense?

How do we check if these clusters are actually reflect

similar cellular functions?

Well, it's very simple.

You can actually check and see if these genes are enriched for

some of the known GO terms.

In this case, we found that these clusters were enriched for GO terms and

it was the terms that you would expect.

So for example, we found the oxidative phosphorylation pathway from K,

we found the GO term for mitochondria.

We found regulation of apoptosis, and we found molecular transport and

protein location.

So again, this reflects a lot of the known biology for Parkinson's Disease.

The interesting thing was, is we had one of the modules of genes in the middle,

the color blue, that didn't have any relevant enrichment scores.

So there was one module that we didn't really know what it was doing.

However, when we looked at the modules to see if

they were correlated with the time of the exposure to the toxicant.

The blue module, the one that we didn't know anything about,

had the highest correlation and it had the most significant P value.

So although we didn't have anyway of understanding exactly

what the biological function of it was, we knew it was very important.

We also double checked using genes set enrichment analysis,

to see if these genes were involved in disease related patterns.

And what we found, using the MSigDB database,

which is run by the Broad Institute, is it what you would expect.

These genes were down regulated or up regulated, depending on the module.

In a similar way, as you see from patients with Alzheimer's,

which has a lot of similarities in terms of phenotype and with Parkinson's Disease.

It's important to keep in mind that this database does not have a large set of

genes from Parkinson's Disease.

So again you're not going to see that in this example.

So gene set enrichment analysis,

as nice as it is, is only going to be able to see the data sets that it has.

So in this case, it has Alzheimer's but it does not have Parkinson's Disease.

So the network is all well and good, but

it's not exactly what we want for a pathway of toxicity.

It just tells us that thee genes might be doing the same thing at the same time, and

they're involved in the same biological process.

But it doesn't actually give us what we want,

which is a genetic regulatory network.

And it's important to understand the difference, there.

The network diagrams just show two genes that are interacting and

don't give you any idea of what they're doing.

They're often called hairball diagrams, because of the lack of specificity.

A genetic regulatory network is much more precise, because it shows you exactly what

the gene is doing, and describes its molecular function.

So in order to take our network and turn it into a genetic regulatory network,

we decided to use transcription factors from text

mining to understand the observed gene regulation patterns.

So we compared the de novo correlation network to known/predicted interactions

with what is called the Fantom database.

And the Fantom database is a very useful database that just has a lot of data for

various high throughput experiments.

Focusing on both known and predicted transcription factors,

as well as perturbation experiments, and what's called, SIRNA,

which is a way to block the effects of RNA.

So what we've found was that the high-throughput data actual confirmed what

we were observing in our De Novo Network.

We found that SP1, which was a transcription factor that was considered

very important by our text mining and by the motifs that were in our genes,

was also known to interact with JUN and AT4.

Both of which are known to be involved in Parkinson's.

And the nice thing about the Fantom database is that it gives you several

different types of data combined in one database.

So you can also see, for example, that SP1 has a protein DNA interaction with JUN and

JUN also has a protein DNA interaction with itself.

So this gives you important information because you now SP1 can trigger JUN,

but also that JUN can amplify that signal because it stimulates itself as it

looks like STAT1 does.

The other important thing about this, is that it gives us information on,

because there's a lot of perturbation experiments,

the dash lines are showing you where there is a perturbation experiment that

indicates that something is, instead of stimulating, is dampening it.

And that's important information to know.

So again, this is backing out to look at the 10,000 foot view.

We found that when we analyzed the data set,

we found that SP1 interacted with an overwhelming number of genes.

And this was true if we looked at high throughput experiments, but

it was also true if we only focused on published interactions.

Again, these are two different data sets.

If you look at high throughput experiments,

that'll give you a lot of data,

but you have to keep in mind that a lot of data is just spurious or accidental.

So that was why we restricted it to focus only on all published interactions,

because that's higher quality data and

gives you a little more confidence in the connections.

So again, this was a nice study, because we were able to add SB1 to the map

of known genes that are affected in MPTP's pathway of toxicity.

As well, it also captured a lot of the known biology and

we were also able to put a couple of smaller transcription

factors on the map as well, for example.

So just to sum up, deducting a pathway of toxicity from high-throughput

data is possible, but it requires a lot of work.

If I had analyzed this data set just using inferential statistics I would have gotten

five or maybe ten genes.

If I had used a very strict multiple hypothesis correction, and

I would have gotten maybe a few hundred more,

if I had used something that was a little bit more forgiving.

But either way,

I would not have been able to capture the transcription factor network.

So here, we used a fairly advanced algorithm,

called Weighted Gene Correlation Network Analysis, that combine network based

approach with cluster analysis to drive a network from the data.

And then divide the data into modules of genes that have similar function.

And it showed us what we expected to see.

We've found genes that were involved in oxidative phosphorylation, for example,

but it also gave us a bunch of transcription factors that we didn't know.

The annotation based analysis showed us that we had captured all

the biological processes that you would expect.

The gene set enrichment showed us that we had several neurodegenerative gene

signatures.

So again,

this acted as a really good reality check to make sure that the data made sense.

This is very important.

An annotation-based analysis is not necessarily a great way to

discover something new, but it is a very good way to make sure that your data

is telling you what you think it's telling you.

And in order to expand that from the known signatures, we used transcription factor

analyses, text-mining, and then other databases of high-throughput data,

as well as published data to turn our network into a pathway of interactions.

So this is just a very small attempt to expand what we know about MPTP toxicity,

but I think it gives you an idea of how complicated it is to take data and

turn it into a pathway.

I just want to add a cautionary note about being very careful when you're jumping to

conclusions from high dimensional data.

This is an example of a fish in an FMRI and the fish is being put

through the same experiments that we put humans through when we put them in FMRIs.

So it's shown pictures to stimulate the visual cortex.

And then, they use the same algorithms that they use to normalize the data and

test to see if there's any parts that are being stimulated,

compared to the background.

And as you can see, in this fish that's being shown pictures

of whatever a fish is interested in, the visual cortex is clearly being stimulated.

However, the joke in this, is that the fish was actually dead.

And this was a way of demonstrating that a lot of the algorithms that were used to

normalize this very, very high dimensional data,

were extremely prone to artifacts and

could give you the illusion of significant results when none were actually there.

Unless you think that the fish had some sort of after death ability to think about

things.

And if this is true for FMRI, it is most certainly true for microarrays.

And so it's important to be very,

very cautious when you're interpreting any one microarray experiment.

And that is why it is always a good idea to check and

see if these interactions are the pathways that it's predicting,

if there's some basis for that in the published literature.

So what about the future directions?

Well, we've been talking about mostly transcriptomics, but it's kind of

important to remember that transcriptomics only really captures a very,

very small amount of what's going on in the cell.

If you think about it, looking at the genome or looking at the set of MRNAs in

a cell, is really only giving you a very, very small picture of what is happening.

On the other hand, if you're just looking at the metabolome,

that's not going to tell you very much about the underlying regulation.

So if we really want to do systems biology,

we have to learn to look at all of these processes at the same time to be able to

all the way from the genome to the transcriptome to the metabolome.

And understand each part and how it contributes to

the part that's closer or farther from the phenotype.

And this is important because few chemicals only act to perturb a gene or

a metabolites.

So you have to understand the chemical's effects at all levels if you really

want to be able to describe the pathway of facticity.

But combining information from omics technologies is very difficult, but

probably very important for the future.

And there's a couple of stumbling blocks for this.

It requires very careful experimental design,

because you might not have the effects happening at the same time.

Your metabolomics might change a lot faster, or

a lot slower than the transcriptomics.

And the protomics might change more softly or more drastically than either of those.

So a sample size that works for transcript that makes maybe small or

too big for the other level of experiment.

And then, of course, you also have the problem that the more data you have,

the more you have the correct for multiple hypothesis testing.

So it is not always the case the more data is better.

It's better to think very carefully about whether the data you're going to add is

going to increase your understanding or not.

And a lot of the technologies, for example proteomics and metabolomics,

have a much higher signal to noise ratio compared to microarrays.

So think carefully before you just add on more data or

integrate more data if you're really going to be able to improve your understanding.

That wraps up my discussion of pathways of toxicity.

Thank you for joining me.

[MUSIC]

Explorer notre catalogue

Rejoignez-nous gratuitement et obtenez des recommendations, des mises à jour et des offres personnalisées.

Coursera propose un accès universel à la meilleure formation au monde, en partenariat avec des universités et des organisations du plus haut niveau, pour proposer des cours en ligne.

© 2019 Coursera Inc. Tous droits réservés.

Télécharger dans l'App Store

Disponible sur Google Play