RNA-seq Analysis Montreal and St. John's 2026 | 04: Enrichment analysis

Okay, good morning everyone. So I hope you still have some energy left for the last module. So we are going to finish with the enrichment analyszis. So we're going to see the different approach to try to extract knowledge from different um database to make sense of the differentially express question you have from your uh DC2 analyszis. So the learning objectives of this module uh is to be able to explain the principle of enrichment analyszis. So we are going to see two different ones uh over representation analyszis versus gene set enrichment analyszis. I'm going to explain the different uh key databases that exist. So the keg genontology reactto and migdb so it's not an overview of uh everything that exists but it's the main ones that are used uh by the tools. Um during the lab we are going to run uh either ARA or GSE enrichment analysis. We are going to see how to interpret the different results and how to uh make nice plots from these results. So just a reminder uh so since we started the workshop we have processed the raw data we have mapped to the reference genome uh we used a cancer counts and yesterday you just did the differential analyszis. So now you have a list of genes and you want to know which uh pathway or which uh molecular function is related to these differentially expressed genes. So that's what we are focusing here. So the question uh uh the way to do that you you could say okay I'm going to look at each gene individually. Uh why it's not a good idea to do that? Uh because you will focus only of course on individual genes. uh it will be very long long uh way to do uh it will be hard to interpret uh these results and you will miss the biological context. So the better approach is to use the whole list of differentially expressed genes or the whole classification you just did. So that's why we uh they developed enrichment analyzis. So it will integrate all the genes into uh biological pathway molecular function uh into u into maps into reaction. So there are different ways to organize this data. Um also uh it will uh provide you more interpretable uh biological results. So it will summarize the information for you. So the first step is called an over representation analyszis. So it the goal is to test if a specific pathway is over represented in your list of genes. Uh so the input is a pre-selected gene list. So for example the one that differentially expressed. So you did uh D2 analyszis and the output from DC is all the genes and an associated adjusted p value and the lock for change from from that you need to decide which genes you consider being significant using either the adjusted p value or a combination of adjusted p value and the lockful change. So maybe you will end up with uh uh 1,000 genes that are differentially expressed and this will became the input of your over represent uh representation analyszis. So as I said you will choose some cutup to define your list of GGs. So that's why this type of approach are a little bit biased because you selected a subset of GS. So it may be sensitive to arbitrary colors. uh on the other hand you have gene set enrichment analyszis called GSA and uh they identify uh coordinated pathway level changes. So it's to see in a whole if the pathway is up her regulated or downregulated and in this case the input is a little bit different. It's the a ranked gene list. So you take your result from DC2 with all the genes that have been cified in your experiment and you rank them not just the significant genes all of them. So to uh rank them you have different option we are going to uh see it later here also we are going to play with that in the labs but the the idea is to use the adjusted p value and the log fold change to rank your list of genes. So the advantage of such approach is that you don't need an arbitrary threshold to decide which genes I need to include in my test. You take all of them. Um the main pitfall is that it's more sensitive to uh uh for small but consistent uh signal and it's more computationally intensive but still manageable with your computer. So the main question is why you should choose aura versus gene set enrichment analyszis. Um you can choose an enrichment analyszis when you have a clear list of significant genes. For example, you played with different value for uh for your threshold and you see that you did not change much the number of differentially expressed genes. So you know that even with playing with this value that doesn't change much the set of gene you are going to put in uh your analyszis. Um also as I said an uh over representation analysis is uh very fast and simple. It's um it might be easier to uh understand. So if you expect a strong signal uh in your uh trans transcrytoic changes uh this type of approach might be more uh uh suitable uh for your research. On the other end if you choose gene set enrichment analyszis um as I said you need to use all the gene as an input. Uh so it will uh allows you to uh track subal changes in your transcriptto. So if you know that your phenotype is not a whole change of your transcriptum but maybe something that is more subtle uh using GSCA may be my more appropriate. Um so also to be uh here transparent GSCA was developed initially for human and mouse. Uh so the gene set are very well described for these two model organisms. If you are using a non-model organism uh it might be better to use an over representation analysis. So also depending on the organism you work with uh GSE might not be the best option but in real life what people are doing they're using the two options. Uh sometimes you tried uh over representation and you didn't get what you expected to see. So you tried gin set enrichment analyze. So uh that's uh sometimes the reality. Uh so you have here some guidelines but I will I will advise uh if it's applicable to your research area maybe use the two approaches to see what is the strong signal uh that you get from these two approach. So in uh to be able to perform this test this uh test you need databases created databases. So there is a lot of uh databases available. Here I just presented the main ones but sometimes depending on your organisms you can have uh databases that is dedicated to your organisms and you will be able to get the data from these databases and put it in in the tools that perform the over representation analysis. So for example if you are working with ecoli I know there are dedicated databases for these organisms. So you can take the information from that and perform your test. So the main ones are the keg database uh the cyto and cyclopedia of genes and genomes uh the reacttoon that uh will give you reaction instead of just molecular functional biological process the genontology and the mig database so it's a signature database so for the first one the keg database uh the main objective is to uh provide created biological pathways like uh metabolism, signaling, disease uh that integrate different information. So it can be genes, proteins, metabolites, uh biochemical interaction and reaction. So here on the right is an example of the information that is stored in the keg database and they put the information uh under a map. So using a map you will see uh how the re reaction uh goes you know all the enzymes all the genes the proteins that are involved in this uh pathway. So during the lab we will see how to use this map and put the information from our experiment to see which genes are uh um more expressed or less expressed uh in our experiment. So it will be easier for you to see where the changes occurs in the map. Um using keg you have some advantage of using this resource. So it's well it's wellruct structured. It's a huge database with a lot of content and it's uh standardized pathways. So you can have the map for example for the glycolysis for the human and then for the mouse and then for the other organisms. So they have uh define the same pathway in different organisms. So it's widely used. So in most of the paper you will see someone using keg for uh enrichment analyszis. So it's more easy to compare your results with other studies. Um also the advantage is that they produce maps. So uh for a human it's very easy to uh look at this map and better understand uh what is the password but there is some limitation and big ones for example uh there are access restrictions uh not everything is fully open um and for some uses you need a license to be able to use keg uh for non-academic uh use days for example you will need a license also sometimes the password can be uh very simplified, not context specific and incomplete for some uh model organisms. The main pitfall for me as a user of keg is uh that it's not uh frequently updated. So it's not the last updated version of all the knowledge out there but still for the uh most uh common pathways uh the information is real reliable but for the last discoveries they are not fully integrated yet. Um and also uh the annotation quality varies between organisms. So it's well described for human mouse growila or you know the main model organism and when you play with something that is not model uh you may find some uh error in what they have imputed but still it's a good resource. So um more recently uh they develop the react database and uh it's the main idea as with keg but the focus is on human and instead of having map that are you know closed it's uh reaction so they regroup the same information but in a different uh way um so it's again a structured database where you can find biochem biochemical reaction metabolic pathways, signaling pathways and um reaction related to cell cycle, immunity, cell death. Well, you will see exactly the same um re u pathway that you can see in keg but in a different uh manner. It's a more upto-date. So, uh that's why we are uh uh in reality using both keg and react to have a better view of uh these u pathways. So on the right here on the presentation you see an next it's just a snapshot of a reaction. So you see that on the uh on the left and on the right the reaction is continuing. So they are it's like they are all linked all together. You know it's not just a map that is alone in the organisms. So again you can see the impact. If I have more expression of one gene, you can localize on the reaction where this is happening and have a better idea of the impact on your biological process. So the uh of course there is some advantage and limitation of using reactor. The the friends is that it's highly creative. So a lot of people are working to make it the most upto-date. It's very detailed. uh when you are going to look at a reaction you will see a lot of information with link to different publication so uh it will help you having uh trust in this uh resource uh so that's a little bit different than in the keg uh resources um also there is a good support for multiple uh species so they have identified this reaction in the human and then they did a huge work to translate this to other organism uh using the autotology. So you find autoglo genes in other species. So you can transfer the information and the annotation to the other species. Uh so sometimes it's less intuitive to look at reaction than a map. So uh if you are more familiar with keg, it might be a little bit difficult at the beginning to switch from this representation we had here on the right from something here more complicated. But uh with the time I think it's a better way to look at the reaction. Um the main pitfall is that it's uh human centric. So initially all the work was done for the human uh genomes. So it can be less accurate for other non-model organisms and as I said as it's reaction uh the bandaries of the reaction are less uh standardized than in uh keg maps where you know you have a box and everything is in it. Uh the next uh resources is what we call the genomelogy. I will detail more this resource because uh it's also available in other resources. So we see a lot of gontology terms. Uh so it's a very old project uh almost uh 30 years of existence. Um and the idea is to structure the description of genes and genes products uh using control and structure vocabulary. What it means that if you take a gene that has a function in one organism and you look at at another organism and you have the gene that have the same function, you will use the same terms to describe this function. So there is no uh discrepancy between uh the the name of the function. So that's what they call the ontology. So they proposed three different uh ontology. So the researchers uh that's where you know the expert in model organisms uh work very hard to describe the content and the function the processes of their model organism. So initially the work was done for uh the drosophil the mouse and uh the yeast and then uh know with time they have develop uh the same resources for the other model organisms. So it's a big consortium still ongoing and the the uh the database is still growing and adding new content. So it's very alive. So they have a dedicated uh website. I want to mention it because you know we are going to play with it during the uh the labs with R but their website is also working. So you can put your list of of genes and perform enrichment analysis directly on their website. So still an option if you're still not available a u uh if you are still not u uh if it's still not possible for you to do it by yourself. So uh when they uh described the the the onthology they decided to put it in a graph and we call it a directed asylic graph. So I know it's big words but it just mean one thing is that the terms are related. Okay if I take this example I have organisms and I have mitochondrian. So mitochondrian is linked to type cytoplasm and mitochondrian is part of the cytoplasm that just this is the content of the database. So you have terms and you explain the link between these terms on in the other example we have the mitochondria that is an organal you have different links okay so you can be part of something or you are something just to make this the most simple so we have a list of terms and the list of real and the list of relationships between these terms so when we uh talk about graph we have the term and the ages that's the basic of uh the the graph so the nodes are the terms and the ages are the relationships between your terms so they have divided the ontology in three main ontologies I'm pretty sure you already have seen that cellular component biomolecular function and biological process so you have three main ontology at the top and then they described all the uh content with your with the terms. So I put here three example for the cellular component very uh easy the mitochondria is a cellular component. So when you are going to perform a novel representation analysis you are going to uh test if from your list of differentially expressed genes are they related to the mitochondria that's the question with the molecular function you can have something very detailed here for example with an activity of an isomease uh and for the biological process it can be more broad for example these genes are involved in cell division so Again, when you are going to perform the over representation test, you are going to test if your uh gene sets is uh over represented in these terms. So if I took here the three main ontology at the top, you have the general terms. So cellular component, molecular function and biological process. And if you go down in the hierarchy of terms, you will have more specific terms. Okay? So it's uh it's structured. It's not a mess of different terms. You have the general one at the top and the specific one at the bottom. So of course for the specific one, you may have just two or three genes that are linked to these terms because it's very very specific. In this example, it's negative regulation of eye pigmentation. I don't expect to have 100 genes involved in this uh biological process. Okay. So once you go down in the hierarchy, you will have less genes related to these terms and that's very important. So where we when we are going to do the enrichment analyszis we are going to filter the results because it will give us terms that are significant but for some of them we just have for example uh DNA binding if I told you that DNA binding is enriched in your experiment you will not go very excited okay that's too large okay too general so we will filter the results to have terms that are not too general or not too specific. Okay, that's something we will play again during the lab. But that's important to understand that there is a hierarchy between your your last resource is the uh mig DB uh so it's a signature database uh and it was specifically uh built to support gene set enrichment analyszis. So they are built together. Okay. Without the GSCM method we don't have M6DB and it's in both ways. So what is offering M6DB? They offer the collection of gene sets. So they list all the genes related to one uh function or pathways for example. So they have divided this collection uh in different uh subsets. So we have the age uh collection for hallmark. Uh it's a high quality gene sets. So it's very reliable. Uh you can have trust in these gene sets. They came from uh different resources. you have the C2 for example created uh pathways that will include in include uh information from the keg the reactto uh database so that's why here uh because of the existence of these resources they were able to design collection for MCTV because these databases were already uh available in C5 you will be able to perform gene set enrichment analysis for geneontology and C7 7 is an example of uh imunologic signature. So if you are not interested at all in im immunology, you will not test this collection because there is no point for you. So the main difficulty for people is to know which collection to use. But you need to look at the description and see if it applies to your question or not. If you're not in imunology, ignore C7 and just maybe test H, C2, and C5. Sorry about that. What's the meaning of see why you see what to >> Oh, I don't know just a way to name it. I I'm sure there is a story behind that but I I cannot collection for me. Is it collection one? Collection two, but I don't have the answer. That's they just regroup these uh by collection. So um so it's the advantage of M6DB. It's like a large diverse collection. Uh it includes uh gontology, keg, reactum. So they they include the uh very well-known databases. Uh this resource is uh really optimized for gene set enrichment analysis. So that's why it's very interesting and it's regularly updated. So I talk about this a lot, but it's very important when you are using a database, please check when was the last update. If it was five years ago, we are way beyond that. It's just too old. You know, a lot of knowledge has happened between in five years. So, please check when was the last time the resource was updated. Um, and the limitation of of course uh is that it's only for mouse and human at the time, but still you can use gene set enrichment analyszis using your own gene sets. Okay. If you have uh resources where they have listed listed genes that you know you know they are involved in the same uh reaction or pathway you create you can create your own reference and then use GSA to perform your gene set enrichment analysis. So that's something we used you know uh when we are developing new signature for example uh we uh create these uh gene sets and we tested them with DACA. So it's not only limited to what is available on MC database. You can provide your own list subjects. Um I think the limitation is like also it's uh it might be a little bit difficult to interpret the results from GSCA. I hope I will help you on the on that point in the next next slides. So how to perform over representation analysis. So there is different uh key steps I'm going one by one and we are going to do them during the labs. So the first one as I said is to define the list of differentially expressed genes. So by using at least the adjusted p value lower than 0.05 but you can also decide to be more stringent. If you have too many genes at 0.05 you can say okay 0.01 01 because I cannot give all my genes to enrichment because everything will go uh will go back significant. So you can test uh these values during the lab we are going to play with the adjusted p value to see the impact on the on the results. Then you choose a database or not. You can try on all the resources or if you really want to just see the gontology you can uh select the database you want to test. Then what is uh what the test is doing it will count how many uh genes from your list are associated with each pathway. Okay. And what we know is we know that in the genome we know exactly for this term how many genes are linked to this same pathway. So then we will compare these numbers. So from my list I have maybe 10 genes associated with this pathway but in the genome I have in total 20 genes associated with this same password. So the question is if I random uh if if by random selection I take the same number of genes do I have less or more genes annotated with this pathway. So our gene from my uh from this pathway over represented in my list or not. I just want to mention that there there is another test that is available. It's under representation in my list. In the this workshop we are only doing over representation but the option is available in the tools we are going to use. Why I don't present that is because it's less intuitive for the interpretation. the under representation might be uh less more complicated to interpret but still that a test that is possible to do. So the question will be all are the the genes from this pathway less represented in my list than in the reference genome in total. Sorry. What's the meaning of the less >> less than by chance? If I have all the genes and all the terms and I took uh exactly the same numbers as uh of gene that I have in my list, do I have more or less genes annotated with this term by chance? It's just a random selection. So I have a list of gene. Let's say 100 genes. I think I have an example just I have an example it would be easier. Yeah. Um so with the test of course we will have uh p value we will correct for multiple testing you have different option uh we will see it also with the tools we are going to use but usually uh we you not usually we need to correct for multiple testing. So we have adjusted p value then as I said we need to filter you know to help you interpret the results. So we will filter the results with two generic or two specific terms and then uh we will uh try to find biological relevance. So based on your knowledge on these data sets based on the phenotypes you see on your cells or in your experiments you will be able to say this is something that I was expecting that makes sense. You know as a bioinformatician if I don't have in context on your experience I'm not able to say if is if it is interesting or not. you will be the one the the the user will be the one that can say if it's interesting or not. And the last step is the visualization. So yes, we will have tables but we can make a lot of different figures. So we will see during the labs how to highlight these uh enriched terms. So I put an example here easier. So if we take all the genes in the reference genome. So here you have the dots. So it's all the genes uh you have uh a subset of genes in blue you know that are defined uh by these pathways they are all involved in the pathways and the question is from my list of differentially expressed genes how many are in these pathways that's the first question so in this example we have seven genes that are differentially expressed that are in this pathway here okay and the test what is it doing? I have here other significant genes they are not involved in this pathway they are involved in something else. So I have here some uh genes that are involved in the pathway not significant and I have genes involved in the pathway that are significant. If I test, you know, if I I make a random selection of a number of genes, will I have uh seven genes uh from this pathway? You know, if I take from all of the genes available, will I have more or less than what I have here? I have seven out of I don't know, maybe 30. So that's the main question. Do I have more than expected by chance? Is it clear? say some less reference >> less you have uh for example if I if I make a random selection you have you have a a sack you have a bag with all these uh genes and I take 10 of them how many I will have with these terms by chance just random selection so you have if I have just two and in my experiment I have seven I have more than by chance so that mean that it's enriched in my experiment Because when I take from all the genes available in the genome, I have less than that. You get it? I think playing with it will be uh might might be helpful here. So um yes, so from this test we have a lot of information. So we will know how many ter how many genes were annotated with these terms. what is the intersection between how list and what is available in the genome and we will look at these different uh columns. So I just put the main question I have at the platform you know for people using or trying to use these approaches. How you select uh the list of differentially expressed in you know what are the basic rules. Um so usually uh we need to use the adjusted p value and sometimes you want to add another uh threshold so it can be the lo change. So maybe you want to focus more at the beginning on the main changes. So you can put another threshold saying okay I'm interested in on all that is significant but also in all the main changes. So I want look for change to be higher than uh one or two. It depends on on your data sets. Again there is no gold rules because for some experiment the changes will be very small but for some of them you will have a huge ch change of transcriptto. So you will have to put uh stringent uh threshold to be able to just have a um restricted number of gs. So again it's a balance between sensitivity and specificity. So how many DEGs are acceptable? So just think about your organisms. So that's a rule that can apply for example for mouse and uh human. Again if you have smaller genome uh be careful. Okay this is not the the rule for for them. So if you think about the human genome you have uh 20,000 genes. So if you take 5,000 genes for a test you know it's too many. So the rule is more about 100 to 1,000 genes you know that you put in this differentially in this test test. If you have too few genes uh you you you will have a low statistical power. So nothing will be significant. If you have too many then you lost the specificity and all the terms be begin uh significant. So you know it's a balance again between the the two. So if you're working with a small uh small organisms, look at how many genes you have in total in your organism and make and take something that makes sense, you know, based on on the number of genes you have in total in your organisms. What if I have too many genes? So I use adjusted p value lower than 0.05 and I ended up with uh 3,000 genes. So as I said it's uh too many. So you can put a uh more stringent uh adjusted value for example 0.01 or if you look at the distribution of your for change you can see maybe a break in the distribution. So you can decide that you want to take everything that is before these breaks. So using some representation of your uh log distribution of your uh log for change and adjusted p value you can make some decision decision on where you need to put your threshold if it's too complicated to get uh a good idea of you know how I define this list maybe you need to switch to gene set enrichment analyzers okay uh it might be easier because in GSA you don't have to put a threshold you use all the gen. So there's a lot of tools available for for over representation or under representation analysis. Uh so I put the list here. For example, uh you have a lot of our package. We are going to use uh cluster profiler and uh G-profiler for that. Uh so cluster profiler is uh widely used flexible and is also able to uh perform gene set enrichment analyszis. So you have one package that can perform both approaches. Uh top go is the one that is more focused on gontologies uh and with different uh um graphs. So it can be interesting to have a different representation react pa uh that is very specific uh if you have uh human uh data for example and you are looking at reaction in particular react pa might be more appropriate for you. I also put web- based tools uh because you are new in the field for most of you. Uh so uh all of these tools have uh websites where it's very easy to just put a list of genes and have a test done. So uh the main one is a GP profiler. So we are going to use this package during the lab. Uh it's a very very good resource. So it's a project that is supported by uh uh the European consortium called Elixir. So every time you see elixir on a website linked by you can trust this resource. That's a good uh a good one. Uh so the website is very nice, very user friendly. There's a lot of tools. So we are going to use only uh over representation analysis. But you can also convert your uh gene ids from one resource to another one. That's sometime an issue for uh peoples. Uh you can also uh convert some uh snips ids to genes. So there's a lot of things you can do with the G profiler. So very very nice resource. Uh I put David uh resource the oldest from all of us my my known. So it's the historical uh tool uh still widely used but it's made like uh yeah it's not the most upto-date uh resource for uh enrich R. Uh there is a lot of uh visualization uh that can be you know interesting depending of your uh um s sensitivity to uh graphs. So sometimes uh you might switch this one and you also have painter. So it's an old one but still uh used at the time. So it's a lot of tools to do that. We were not going to use all of them but uh still using the the main ones. Uh I also want to uh mention here uh some platforms. Uh the first one is a string database uh that provide network analyszis and enrichment analysis. Um so sometimes people are confusing this resource. So it's not protein interaction like phys only physical protein interaction. The string database you put a list of genes or proteins and it will gives you all the interaction between these proteins. So it can be physical or because the two proteins have been co-ited in the paper or they are co-expressed or they are next to each other in the genome you know there is different type of interaction. So if you have a list of differentially expressed genes I really encourage you to put them in the string database and see what is already known you know so you will not spend time on uh trying to prove that the interaction is already known. it will help you to already decide okay all of this is already well described I will focus on this one but because there is no information on these ones so very useful and based on the network so you will have a graph where each node is a proteins it will perform also an enrichment analyszis so you'll have the genontology and keg reactto uh and will highlight the graph based on this enrichment so if we have time at the end of the lab I would encourage you to use the string database to actually see how it works So it's only online for this one. The last one is called Metascape. I see this one a lot in papers. Uh now I advise you to be very careful uh because it's a black box. We don't know exactly what is done uh with this tool. It's only online. It's very pretty, very uh visual. Uh but the resource is not uh uh we cannot trust the pulling the resource. Okay, but you know it makes nice graph but be careful when you are using this one. So I have an example more more specific here. So in this example we have 50 genes uh that are differentially expressed in an anoric experiment. We have one go term that is called the immune resp response and in the reference genome it it is associated with the 300 known genes. The intersection size between our list and this term is 10. Okay from our 50 genes turn 10 are annotated with immune response. So that's what we call the intersection size. The effective domain size that's a term you will see in G4 fiber for example is uh two uh it's 20,000 genes. So how many genes are in the reference team. So the enrichment test it's a epgeometric test uh will be used to compare the genes we observe the 10 genes with uh what one would expect by chance if I pick uh 50 genes from my whole data sets how many I will expect that are annotated with the immune response and the answer is 0.75 by chance in my case I have 10. So the answer is that we have an over representation of this term in our experiment because by chance I will see maybe one gene annotated with this term and here we have 10. We can also compute the fold of uh enrichment. We will see it in the in the labs. So that's here that's very simple you know you just compare what we have from a random selection. So that's a basic output from G profiler you will have each uh row will be a term with the source here. So we have terms from keg from gontology cellular component and from reactor we have the term name the description the term ID adjusted p value minus lock and adjusted p value so that's uh that's the way we are representing the data uh the higher is better so when you are drawing bar plot for example we are going to use the minus log 10 of the adjusted p value it will be a nicer on the graph. We have the term size. So the number of gene in the reference genome annotated with the ter this term the query size that's how many genes we have in our list of differentially expressed genes and the intersection between the two. So from the 20 differentially expressed genes five are annotated with oxidative postcorrelation. That's the intersection between my list and this list here. And uh we will have here the list of genes the five ones that are here. So then if I am interested in this uh keg uh map I can take these results here. I have the list of the five ones that are uh enriched in my experiment and I can go to the keg uh database and look at the map to see where these genes are situated in the map. It's okay for the output. So the uh first output we will uh use in the lab is called a manatan plot uh where you have on the xaxis the different uh database. So we have genontology with molecular function biological process and cellular component we have the keg reactto and other uh datab bases. I will not uh describe this here but on the gp profofiler website you will have more information on what is here. So for example, one is for microRNAs for example. So it's dedicated databases. On the y-axis it's a minus lock 10 adjusted p value. So it helps you to see if there is a significant term. So higher as I said higher is better. So all the terms that are at the top are significant. So that's the overview of your test. But then you will have to go uh in more in detail for each of them. So for example you can take all the term that are significant and draw the bar plot with here the minus log 10 of the adjusted p value in the x axis and in the y- axis it's each term that is enriched. So usually in papers you will see this type of figures you know. So you know I often I have the question can I remove some of them because they are not meaningful for my paper >> redundant. >> Yeah or redundance. So if you have redundants you can use the term uh size we have here. You can use the term size for example to reduce to more more uh um not too general or too uh specific terms and also you know it's a figure in a paper so you have the right to just put what has a meaning to you for you but in the supplementary data you will put all the list but of course in just the figures on main figures you can just uh list what is significant and what makes sense in your experiment but you will have a full table available to uh the reviewers. I know the item shoes right now relative to level. >> Mhm. >> Too big. >> Yeah. And I don't have the expertise on that. But if it's too big, uh usually the uh the next terms will be also significant. So you just reduce your list and you will see something related to this term significant in your results because all the term are tested individually. So you have the general ones and the one below but that still belong to the same ID. Okay. So if this one is significant that means that the other ones you know that are more specific will also be significant at some point. So all of them will be listed. So from all the results you have you can remove the two general and two specific or not you know if the two specific one are what interested you you you keep them but you know you cannot put protein binding for example that's doesn't mean anything you know. So just my the the the the key message here is that you can filter this test you know to make it uh to to make sense of your biological experience or question. Oh sorry um was the question? Uh >> so the item in which it currently the term actually is. >> Okay. >> So yeah because this is a simple carbon metabolism that's the term is big. >> Yes. So carbon >> need narrow down. So that's what >> so the question is the for example the carbon metabolism term is too general. So can we go to uh a more specific term? So my answer was that in the uh results from deep failure for example uh I'm I'm pretty sure that terms linked to this one will also be significant at some point. So yes you can decide to keep only the more specific ones and remove uh carbon metabolism or you can also just make uh one bar plot with everything related to to carbon metabolism. you know if it's the main result of your experiment yes go for it and just put a subset but in the supplementary data all the old table will be available if if it's needed okay any other question yes >> I think we can just >> yeah is there any reasons why you didn't include um in January analysis like the IPA the program >> oh because it's not free >> okay so that's >> the only reason so the yeah the question is why why we didn't include IPA or you know there are a lot of software that are not free so as a workshop here we don't have this uh anymore I think it's very expensive usually you need we had it at the ICM in the past and it was a shared license between uh three institution to be able to get it So you know it's very very expensive. So IPA is another uh software uh that can you know help you give uh do all these enrichment analyszis and make nice graph. So I have seen it in the past. I'm pretty sure now it's still very nice but uh yeah it's not a free resource. That's that's why but they rely on everything that is public you know they have their own information you know created but they also rely on gontology keg and the react reactto and all of these resources because you cannot publish with just saying I use IPA you know IPA is based on these resources yes uh can >> yeah on slack first can you put your question on slack >> the question on slack is um on site 25 you have about uh 0.05 to 0.5% of the total number of genes as what you select for the good amount of deegs. Is that a rule that you're uh implementing that it's about 0.5% of the total number of genes based on your species >> when you uh here on this one? I don't see the numbers. >> Can you repeat this? Sorry. So um if you say that about 50 to 2,000 is about if you do as a percentage of the total of number of genes that is in your species it looks to be about 0.05% 2.5%. >> Yeah. Yeah. Yeah. 0.5. Yes. If it's the numbers here that's a good idea. It'll just have the same uh proportion as I said here. Uh maximum 1,000 genes for a genome with uh 20,000 uh uh genes. That's a good rule. Yeah. Other question. So the next question is I've seen the term functional enrichment analysis in a few papers which include OMIX. Is this a synonym to a gene enrichment analysis like GSEA when it comes to RNA? So these are so over representation analysis is sometimes called functional enrichment analysis but uh the the over representation analysis and GSC analysis are two different things. So that's here I just explained right now the over representation analysis and it will give you insight on the function of your genes but the gene set enrichment analysis will also give you insight on the functional analysis. So functional analysis is just a broad terms and inside you have two main option over representation analysis where you give a list of differentially expressed genes or gene set enrichment analyszis where you give signature genes and you test if there is over representation or under representation of these gene sets in your uh data. Other questions that was your question. Okay. Um yeah so that's uh one example uh that we will uh do during the lab. So you can do a back plot uh very basic here and here I put another example uh with uh the package called cluster profiler. Uh so in this example it's basically the same as with uh gp profofiler but we include also uh another information that is the count. So the x-axis is uh the number of genes that are annotated with these terms in our list of differentially expressed genes and the bars are colored by the adjusted p value. So that's the main difference. So instead of having just the adjusted p value, you know, here I don't have any idea if I had uh 10 uh genes or 20 genes from my exper my experiment. So it's better to have the two information in the same graphs. On the last example, uh it's uh again the same idea is to combine the different information to make it uh more realistic. So we have a dot plot and uh for each uh pway or each enrich terms uh we have what we call in the xaxis the gene ratio and the gene ratio is the count or the number of genes annotated with this term in my list over the total uh oh sorry oh over the total number of gene in my list whatever the the annotation was. Okay. So this give you give us the ratio. So it will highlight if this term uh is the main uh aspect in our list of deeg or not. So here from what what I see maybe the top here are the ones driving the the the the function and these one are maybe just anecdotic uh terms. So we will see this with the cluster profiler. Um this last resource is called path view. Uh so it's another layer you can add to keg. So uh let's say we do a um over representation analysis and we see that one keg uh map was uh was significant. What you can do is to draw the map. So with past few you're able to say I want to draw this map. you have a map ID and I want to highlight these genes with this information. So this is very simple. We're again going to do it in during the lab. So here the example is put glucans in cancer. I used the uh log for change of the genes that are differentially expressed and the software will highlight here all the genes that were differentially expressed with their change. So for me it's very uh it's easier to interpret what is going on in the map. So I see that most of the gene are down reggulated but here we have one that is upregulated. No. So based on your knowledge in cancer you will be able to understand what is going on in your experiment. >> Yeah. >> Okay. Uh the question is if uh pass view is working for non-model organisms. So it's working very well for human and mouse of course but uh the past view is available for all the keg maps that are in the database. So in keg they have translated this uh maps to other nonod organisms but it's not the most created ones. It's still work but be careful because I'm pretty sure the the maps are not uh full you know you are missing some information but it can be exploratory you know you just use what is already known >> and see if there is any trend you see on the map >> okay >> but it's available you put your gene IDs yeah >> you say which organism you used >> you you select the keg map and it will draw the map >> inside of the pass view >> inside pass view >> yeah okay great thanks >> it's the same in G profiler or cluster profiler you say which organism you're using you put the gene ids and it will give you the enrich terms >> good thanks >> you will try it >> I will try that >> a question from slack >> uh could we have a correlation between different genes with keg maps >> correlation >> I'm not sure I understand the Wait, wait, wait for the microphone. >> Okay. I mean if we have for example uh many genes that are differenti expressed with this kegma map. Could we for example know that we have two different gene that could have an impact on our I mean we cannot explain just one function with one gene. we can explain one function with different genes that could be correlated. >> Yeah. So I I I show you one example of pass view. So one application is to use the list of differentially expressed genes and to highlight them in in the map here. If you have a list uh from another experiment that is not already seleated or magic for example, you can also use pass view to look at this. But I'm I'm not sure I understand the the the the >> the meaning of your question. Like you put a list of genes here, not just once one gene. >> I mean here the um yellow gene is that the uh uh expressed genes, right? It's the level of expression of of here the the the scale is the log for >> okay the idea is to explain the functions of this gene in the metabolic pathway right >> so the idea is to see how this pathway so the uh proteins in cancer is affected so we know we have differentially expressed genes linked to this >> and we know that this kek map was significant So this was enriched. So we have more than expected. So by having the lock for change put on the map here in for example we have a lot of gene that are down reggulated you might uh you might track you know this gene on the map and uh see the function related to this to understand how this is affected. >> Okay. So could we for example for say that we have um two different gene that could have additional effect on the same function or not. >> Yeah. It doesn't call you you know you need to read the map to see you know what what is activation what is inactivation. Sometimes with the arrow you know if it's an activation or inactivation of some genes. You need to interpret this after you know you to read it and have some knowledge on these genes you know to know if is it a transcription factors uh is it an an enzyme you know and and try to make sense of this. I just wanted to show you that you can have different layer of information uh including these maps and your level of expression but it can be also your map and proteomics your map and I don't know whatever experiments you did and you have a list of genes you can use it to draw this and try to make sense of what happened. >> Okay, thank you. >> We can talk about this later maybe. Okay, thank you. >> Other question? No. So the next approach for functional characterization is gene set enrichment analysis. So the first uh step is to rank your gene list. Uh you can use different metrics. Um most of uh people are using two different information. So a combination of adjusted B value and the sign of the lock for change. So based on that your um u most expressed genes will be at the top of your list and the less expressed genes will be at the bottom of your list. So we rank them. Then you choose a database. So we need a list of gene sets. Okay. We want to test. So you can use uh the MC database that have collection based on gontology keg or other specific collections. Then uh you are going to walk along your list of genes and uh you are going to track at on at which pathway they belong to. So I take the first gene of my list and I look in MC database in what in which gene sets it belongs to and I go down the same way. So each time uh I found a gene including one pathway, I increase the score for this pathway. Okay? So I look at the first one, it's in the pathway I'm interested in, I increase the score and I go uh I I I I continue in in my list. So when the gene is not in the pathway, I decrease the score and again and again. Okay? So it's like you walk on your in list and you build like this the score. At the end of this process when you have reached the bottom of your list you compute the enrichment score uh that will reflect the accumulation of uh at the top or at the bottom of uh enrichment. And the main question will be are genes from this pathway systematically enriched at the top or at the bottom of the rank list. Then using uh permutation we are going to compute a normalized score from this enrichment score. We are going to correct them for multiple testing. Again we will work with adjusted p value as always and we will filter the results to focus only on the uh more informative gene set or to remove redundance. Of course after there is a last layer where you are going to interpret the results. uh so you will have biological relevance and you look at uh if the uh the pathway or the gene set is upregulated or downregulated and we will have a lot of visualization uh tools uh to see this enrichments. So I detail the steps but for you it will be uh a black box. You will input a rank uh list and then it will output for you enrichment score normalized enrichment score and you will uh go to visualization and and interpretation. So here it's an example. So on the left we have a list of rank genes uh by for example logful change. So the uh more express are at the top at the top at the bottom the uh downregulated genes. So we look at the first one it's involved in uh the gene sets. So it's in blue. The next one is also in blue because it's involved in the same uh gene sets. So the score is increasing you know because at the bottom we have a lot of genes. At the top we have a lot of genes with involved in these gene sets. In the middle it started to decrease because there is nothing you know related to these gene sets and then again at the bottom we see some downregulated gene that are uh associated with these gene sets. So usually where you are going to look at this type of graphs you will see the same trend. The score is increasing at the beginning then the score is decreasing because in the middle where there is no change you expect this not to be significant and it can increase at the end again because the uh most downregulated genes uh are again related to the gene sets on uh here on the xaxis uh you have a presentation or where each bar is a gene and it will be uh highlighted when the gene is involved is uh in the gene set you tested. So you should have an accumulation of bars either at the left or at the right if it the if the gene set is significant. So the main question where you are performing gene set enrichment analysis how I ranked uh the genes. The gold standard is to use the uh lo the adjusted p value combined with the sign of the log for change. So this way uh the most significant genes that are up regulated that are the top and the the uh most significant gene that are down reggulated are at the bottom. So the sign is very important. Thus the logful change by itself is not sufficient. We need also to rank the significant genes either at the top or at the bottom. Uh what about multiple testing? So that's a tricky question. In get enrichment analysis, even if you go to the website MC database or GSCA, they will tell you that you can use an SDR 0.25. So uh because of uh you testing a lot of gene sets uh uh they said that you can use this type of threshold for exploratory results. So everything that is below 0.05 is significant and you can trust these results. But sometimes you have something interesting at under 0.25 and you can still explore this avenue uh even if it's not significant. So by itself they tested a lot the the tools and they uh they know you know that this is uh the case. So sometimes you have interesting result even with the FDR below 0.25. To my knowledge it's one of the only exception uh of using an adjusted p value of uh 0.25. So in most of the tools it's always significant when it's lower than 0.05. But here there is some relaxing around the adjusted D value. So to perform uh GCA uh there is different tools as I said uh our package uh we are going to use crystal profilier uh for the um over representation analyszis but I also include uh fast GSSE FGSC for gene set enrichment analysis. uh I wanted you to test two different uh packages. Uh FGSCA is very uh interesting because it's very fast. Uh testing uh different gene set takes a long time in computing. Uh so with this package we are able to do it very fast. Also there is other approach included in Lima. So Lima is uh statistical package you can use for differential analysis for example. But there is a lot of thing that Lima can do. So that's why I included here uh because there is uh different uh another application of lineup. Uh also in the future if you're more comfortable there is a standalone a desktop application for GSCA uh it's a little bit old school uh but still working on the on our computers. So if you are more more uh familiar uh you can use the application from the broad institutes but be very careful because by default they have selected a lot of option uh in the software. For example uh they will assume I think that your uh gene list is not ranking uh is not sorted uh but uh in in most of the case you you will input a list that is already sorted. So please be careful with the option that are already selected when you open the software and it's widely used in duplication. Most of the biologists are able to use the this application. So the main output of GSCA is uh the visualization plots you see here on the right. Uh so I just want to demystify this because we see this a lot in papers and sometime we just understand what we are seeing here. So on the xaxis we see the uh black bars here. So the black bars are uh each gene will appear here in the bar if it was in the gene set. So remember we are walking through the g ranked list and each time we see a gene involved in the gene set we highlighted them. So here we see that we have uh more genes uh that are at the top of our list. Then the green uh the green curve show shows you how the enrichment score is going. So we see that it's increasing at the beginning, decreasing in the middle and nothing is going up again in this case. So there is no enrichment for this gene set at the bottom of the list. So most of the gene set is at the top of our list. So it's an up regulation of these gene sets. Everything is here. So this is our peak. This is what we called uh the enrichment score. This is the maximum score we have reached in this test. Uh I think it's I think it's uh I hope this is uh clearer for you. So during the labs you will be able to draw these uh maps and we will see different trends. For some of them the arrangement will be at the bottom or uh a little bit at the top a little bit at the bottom but still enrichment. >> Yes with the microphone or in Slack or microphone. >> Thank you. So I saw this figure before >> in my own data and also in the publication. So what does this exactly to tell the audience? >> Okay. So here if I see that in a paper >> this gene set I don't know it's apoptosis. >> Yeah. >> Okay. In this example is enriched in upregulated genes. >> So you have an upregulation of apoptosis in your experiment. >> So there's no down reggulate. >> It's no down reggulation. If it was done regulation, we will see the enrichment score the score going up again at the bottom. But here all the genes are mostly at the top of your list. All the thick bar here >> or mostly at the beginning of your list. >> So it's between zero and two 2500, right? Maybe those are the most significant up regular genes. >> Yes, in this case. But no I I would say maybe the uh 600 first ones all these ones here. >> Okay. >> Because after it's decreasing so I see that the the majority is very at the top >> but we don't know that exactly the names based on this figure but no based on the figures we don't know right >> the no based on the figure we don't know the gene names. >> Yeah. >> But then when you are doing that you are going to extract the genes >> related to apoptosis. you're going to draw a heat map or uh I don't know a bar plot whatever to highlight this in another way for example or you're going to confirm some of the genes uh either by qPCR or doing other experiments uh so it's just the beginning you know it's just a way to have an idea where I should go I have uh 600 genes and I I don't know what is the function you know so this will help you guide you in your research Okay. Thanks. >> Thanks. Okay. Any other question? No. Actually, one question uh from how many order of magnitude different GSC enrichment plot would you get for your world DG list? Sorry about how many plots >> how how many different GAC enrichment plot would you get for your world DG list? >> Uh so it's not from a DG list it's from all your all your genes just to make clear that you are going to use all of your genes. Uh yeah, I don't have an answer to that because most of the time you will get none using the adjusted B value lower than 0.05. Um but you are not going to get uh hundreds of uh GSCA plots. So yeah, it really depend on your experiment if you had a strong signal or not. Uh did you have good samples? No had no outlier you the right normalization it's totally dependent on your experiment so I don't have a right answer but you will not be uh overwhelmed by by thousand of results that's maybe the only thing I can tell on that we will see it with the example you will not have a lot of significant results other question or no so I can finish the present presentation I had one slide but uh so the take-h home message again uh so the functional characterization the term you can see in the papers so it will help you transform your gene list into biological meaning uh by identifying the process and the pathway they represent uh so please use different tools different approach different resources to uh get the the the the more information from what is already uh known and we help you try to focus on something that is not maybe well described or well known instead of focusing on something that everybody knows. Uh if I can say that. So I think it's all for me right now. No question. Good. Thanks.