RNA-seq Analysis Montreal and St. John's 2026 | 03: RNA Expression and Differential Expression

Hello everyone. We are going to start with module 3. Difference expression analysis. As before, if you have any question, just type it in the SL and we will be reading the question so that everyone can hear the questions correctly. And if you see a question that you like, just like it and we will answer the questions that have the most likes first. So what are we going to do in this lecture? So we are going to explain what the goal of a differential gene expression analysis is. We are going to understand the steps in this analysis and also this is more something that you are going to do in the practical but we are going to be aware of what happens if we change anything like if we change the normalization what is going to happen we are going to look at how to interpret visualizations and of course you are going to perform differential g expression analysis in R as before in module one we are going to have some interaction so if you want to connect now you can the QR code or go to www.menti.com and enter this novel [clearing throat] and I will just give you a moment to connect. We have already 24 people connected. So let's start with the first activity. This is kind of a fun activity because we are back from LA so that we all wake up. Just enter three words that come to your mind when you are thinking about RNA and let you see what cloud we get together. [clearing throat] So this will be kind of the screenshot for this [laughs] for this workshop. Expression expression transcriptomic comparison alignment analysis. Cool. So this was just to start with kind of activity and I will just wait a couple more seconds for someone else to enter their words. Okay, we have a cool work cloud. [clearing throat] So what we have been doing so far we are I guess this is exactly the halfway so in the module we look at quality control and training so we use fast QC fast P and multiq and [clearing throat] this this morning we were working with start and with an alignment then what we are going to do now after this module is differential expression analysis and then last module is functional characterization the pipeline there is one step there that is kind of missing right like what happens with quantification ok so what is quantification so quantification is actually also called sumarization so you can hear quantification of sumarization basically referring to the same is to count the number of reads in figure of interest and with feature I mean a predefined genomic region of interest [clearing throat] so for example jeans so in this kind of cartoon toy example we will have somewhere we will have defined these coordinates that are the one that we are interested and the red blocks are read so quantification will say okay from 420 to 830 we have one and from 13 to 1130 we have two reads and will starting create this table and here in this coordinates we have four reads so this is basically quantification there are two kind of standard tools for doing quantification one is feature counts and the other isc there are the most available uh the most widely used tools and there is not really a difference if you use one of the other. So this is one of the steps in that they are mostly going to agree. So I have say predefine genomic figures. So what file have you already seen that contains this predefined genomic region that we can use for quantification? You have already looked at one of these files. So what will be the file that we will be using? Correct. What contains or predefine genomic feature that we are going to use are the GTF GH FF file. That's what we need for quantification. Now, how will we use feature counts? Okay. So fit your counts. For example, if you want to try it in the cloud it is already installed. So you will say figure counts. Minus P means per N. So this option could be minus P and we have per Nend data will be minus P. The other option will be minus S for single N. Then minus Ton means what is the feature that we want to use for quantifying? In this case is exon so from the GTF is going to add the reads in the exons okay so we will use that now minus G ID is how we want to summarize so aggregate. So we want the resource to be based on gen ID. Um here for example you could have transcribed ID or other type of ID. The standard is to have the gen ID. So our table is going to have the gen ids as rows and we will have the counts for each of these gen IDs per sample in the columns. Now here minus a is for annotation. So we will enter our GTF file. So this is the file name for the GTF minus output. So this is the file name of the output file. Um you can give anything that is um reasonable. So can. TXT or something. And then you enter all the b files. So here will be all the b files that we have for the HBR and the UVR. So it will be here all the B files. So you will just write all the B files here and that will give you the quantification for all these samples if you want more info here is the the link for the feature C documentation. And if you want to try a different one as I say they are very similar information from HTC so HTC count is the one that you could use instead. Now do we always need to perform a separate quantification? Think about what we have done in module. Do we always need it? The answer is no. We don't always need it. When do we not need it? Is when we do pseuding because Calisto actually gives you a quantification already. So if we are using star as we did in module two then we need quantification we need to use feature and then we go to differential expression analysis but if we are using calist salmon or any of the other pseudo alignment what they give you because they don't give you a bf they don't give you an alignment what they give you is actually a table so if we are using listo and salmon they will do the quantification and we will just go and take the output from those software directly to differential analysis. So the answer is it depends what you were using in step two. If you were using an alignment tool like start or high then we need to do quantification but if you were using a pseudo alignment tool this alignment actually what they give you is a table. So you will not do an independent separate quantification step. So this is a pipeline of a differential expression analysis and I put this pipeline just to we are working with RNA but just to kind of remind everyone that we used to have another source of expression microrays and still of data about microrays. expression analysis micro was used before and a lot of data from micrays nowads almost everyone is doing but data is still there and the technology still exists so as before we get or biological question we do the experimental design isolate the RNAs and then most people will be doing So the RNAC we make the library, we do the sequencing and then we do the mapping and the quantification then we have three options. So we could do hypothesis testing which is this is the statistical testing that we are going to do for differential expression analysis. We will do exploratory statistic of doing a clustering analysis a PCA more about what is going on with samples and then we will do classification. In this workshop we are going to focus on hypothesis testing and we are going to do some exploratory analysis. classification we are not going to touch it is super cool is when you get to identify some gene and then you can use the expression for classifying samples like kind of and actually that's already something that is clinical trial like there is this uh breast cancer kit that kind of just look at the expression of uh set of jeans and decide what kind of um breast cancer type and then it is used to inform the the the treatment that people are going to receive with breast cancer. So classification is basically using the expression after some analysis the expression of gene that are really relevant to the condition and then you train a machine learning based approach like random forest or vector machines for performing classification later. So you just get the expression enter to the classifier and it will be classifying to this type of this type of this type. Okay. something to stress is that we don't end the analysis here. After doing the expression analysis there has to be some biological confirmation right to this for this we need some biological inside so module four when we are going to do look at functional profiling we are going to start doing this biological inside because most people if we get like a list of 200 we don't know what is going on right we have to actually put together into pathways processes and then we know what is going on in the biology later can be used for diagnosis or for therapeutics so the idea with gene expression analysis that is going to inform us for further biological investigation or fur biological uses okay another pipeline. [clearing throat] another visualization pipeline. So so far we started okay so step one to three are from before. So we have the experimental design RNA acquisition that data acquisition then we get the data in millions of shorts with fast files. And then we do the mapping. And then there are different options there for the mapping. This there are some side tracks like usually we are going to get on map. Um that could be kind of a side project why we have on map but most people will just go with the map risk and then we are going to do the summary of quantification. From that we get the table of counts. And the table of accounts is what we need. as input for this step in the lab we are going to get the table of counts from the NCBI why because we were not using the full data set like we were just looking at single chromosome and very few reads so we are going to get the actual table comes from another study so that we can really do a differential expression analysis that is with a whole data set so mapping we did it in module sumarization I just went with the basics of sumarization and then now here in module 3 we are going to do this two steps so we are going to look at normalization and then differential expression testing so after the end of module 3 we are going to have a list of genes that we are to consider differential expression and that is going to be the start of module four so module four is going to be interpretation and looking at pathways processes so that we can have more biological insight into what is going on in this study. Let's start talking about assumptions. Okay. And this is quite important. Most differential expression gene analysis method assumes two things. One is that the expression levels of most genes remain the same between samples in other words that is only a subset of all the genes that are going to show differential expression. The other big assumption is that the overall mRNA levels are mostly similar across samples meaning we got approximately the same amount of sequencing data from all of the samples. Okay. And this will have an impact on the normalization approaches. If if you are working in a setting where most of the are going to be differentially express then they have to be a custom made normalization approach. What situation could be that for example in cancer cancer cells are so disturbed that the expression of all the jeans is crazy. In that case a lot of people will go for example with housekeeping normalization or spike in normalization because really you cannot assume that most of the are not changing between the two groups. Um it's always good to try different approaches. Okay. So kind of in byatics is always good if you try different methods and see whether they agree. So is the same here. So is the same here that you try different methods or even different normalization approaches and so and then you see what is in the agreement. If the result changes maybe something that you can look further to see why is this changing a lot and if they are not if they are mostly in agreement you can consider that the jeans in the agreement will be the one that are more confident. So several method say this gen is differential express this gen is differential express you can almost use it as a boting skin like yeah this is differential express then you trust that that is differential express okay so kind of reminder always when you are going to do a differential expression analysis stop and say are these assumptions valid for my data If yes, continue. If answer is no you might have to look at different more specialized approaches for your data. In most cases the answer is yes but always good to just check and be sure that you are aware of those assumptions. Okay. The first step in the differential expression analysis is removing the features or jeans that have uniformly low rate. This is usually done is not mandatory but will be recommended. Um and we are going to do this in the module. So if you have jeans that are basically no express you can just remove them and it will help with speeding up the analysis because instead of doing00 gen you might just have to do like 15,000. Uh the second thing is that we apply multiple testing correction. Later on we are going to see more about this and the number of genes affects how strict the multiple testing correction is. So the more gen you have in your analysis the lower the p value has to be so that is going to be considered significant. Okay. So if you remove them from the start then there are more chances that you are going to find something that is significant. As I say this is not necessary frequently done and consider that usually good step to do normalization so what we are doing with normalization with normalization we want to remove biases and these biases can come from a difference in sequencing depths between the samples and also difference in the ying length okay and the technicals effects. So there are lots of different normalization techniques here on the right side we are seeing some of them. So RPKM, FPKM this is R per kilo base of MPs. This is fragments per kilobase of MAP RS. Then we have transcripts per million. You were looking at this transcription medium when you were looking at the calisto output. Then we have TMM uh which is tr mean of M values. This is used by the HR program. We have the UQ upper quarti normalization and we have the RLE relative log expression. This is what is used by DC2. Now why do we need to take into account the L? So why we need to what is the difference between having 100 nucleot length and 10,000 nucleot length why we need to take it into account. Okay. Why do we have to take into account? Okay. You imagine you have a really long gen how the sequencing is done that you are going to get from different fragments you are going to get more there right so you are going to get more reads than you have a short uh g and this and how we are actually going to do the differential expression analysis is based on the count of reads that are map to but the is affecting how many do you get there but it doesn't mean that you can have more ris doesn't mean that you are going to have that more highly express just getting more because of the technology makes sense. Then you need to for the L so that it does you have a longer read or a short it just accounted for the same kind of amount of expression and that's why we need to take into account. The other two they are wrong just things that I made up. So and now why does we why does sequence in dep need to be taken into it account? And are reminding sequencing dep is how many reads we are getting. for one for each sample. So why is that? Why do need to take into account sequence in dep when we are doing normalization? So why is because we have more we have 1 million from one sample and we distribute those into all the possible right and then we have 10 and we distribute the same that will affect how many we are counting per feature is completely independent of the the actual expression so we need to take into account because if we have more and we are just taking the raw count it will look like the one that has the deep sequencing is going to have higher expression and it will look like it's not real just going to look like because we have more there so this is the correct one uh and the we are not going to get more reads from a sample that has highly expressed that's a common misassumption we are going to get the same number of but most of those might be for the highly abundant gen so they will be dominant but we are not going to get mostly like is whatever is in our library we are going to get the same number of reads that we are doing the the deep sequencing or the the dep that we are doing is the one that is going to determine how many risks we get instead of the expression of the gene ok so what determines how many risk we get is the sequ the depth of the sequencing that is predetermined before even looking at the okay [clears throat] in a lot of papers you are going to see RPKM FPKM or TPM and they are still used a lot for normalization but actually for the analysis those methods are not really and they are mostly used for the analysis TMM UK and I okay so most standard methods now are going to use the normalization the in black for the analysis and the others were the kind of normalization methods that were proposed there are still sometimes being in use and there are used mostly for visualization but they are not the recommended approaches for doing an analysis Differential expression analysis as usually informatics we have a lot of different tools. So what is differential expression testing? So the goal is that we want to identify which transcribe features and I say transcribe features because you could do the analysis at the gene level or the exon level or any level that you decide but which transcribe features are likely to have change the label of expression between groups due to the biological phenomena. Okay. So here we want to say like due to the biology for technical artifact but due to the biology. normalization technique and on the on the left side column is the method they are using for the analysis so for example HR useM as normalization linear mod statistical testing lima the one that we are going to use is so it has several different normalization techniques although the default is rle that's the default and then the same h it assumes a negative distribution and use generalized linear models for the statistical testing and then we have BCON and C is it has to be you only use it if you using kind of the C suite of tools and also the ground usually you will use it together with string time or alignment free quantification so these two they are more specialized to whatever you use before for mapping while this three you can use it with any count and I know you cannot read this this is from this paper so they did gaining comprehensive biological insight into the transcript by performing a broad spectrum RNA analysis it was published in nature communication in 2017 and what they did and You can see colors yest of different pipelines like they use different aligners they combine with different normalization they combine with different gin expression analysis tool so they try a lot of the different lots of different combinations of pipelines all of this actually and then they compare them and analyze them and they try to get to what they call optimal pipeline with the tools that were Okay. So the green light that I showing with the cursor that contains there are pipelines that contain the red one here they have HR. Okay. And I do n't know what color is that this pinkish color is slot and the blue one is lime. Okay. What color do you see more in the top? like especially here like we see a lot of green here in the top and here we also here right in the top we have a lot of green so green is this here they are mentioning the spearman rank correlation between the results for the gene expression analysis and the actual results and here they are showing the root me square deviation so with both methods this2 was one of the the top ones was achieving the top results okay let me tell me that you look at probably from here to here the difference between the methods is very slight so probably h2 are going to be fine and the same here is a bit more pronounc but yeah but just don't get into the bottom so those ones are the probably the one that we should not choose the other is almost a vertical line almost there is a bit of slow not very dramatic But if we just go for the results here in the top see is really one of the top approaches that you can use for differential expression analysis and that's why we are using it for the module. So we are going to be using this and do we have any questions? No, we are good. So we are going to look at more detail. What is the analysis? So the input is count matrix and you are going to see a count matrix in the in the count matrix when you looking count matrix here in the row you are going to have the identifier that you choose in the feature so in the feature you use gene id then we are going to have here gen id you use transcribed id we are going to have here transcribed So here is the feature identifier that was used for summarization. Then we are going to have a column for each of the samples and the number here actually is the number of reads that was aligned to this feature in this sample. So we are going to that is 14 reads align to this g in sample one. Okay. So that's the how we interpret the count matrix and you also looking atalisto count matrix calisto is TPN so it is already normalized. [clearing throat] Okay, so we get a count matrix and what this two is going to do is the first step is to do size factor estimation. This is used for normalization. Then it's going to estimate the dispersion or variability of the mean of each gene then going to fit a generalized linear model and then going to do a statistical test. I have used the W statistical test. The result of that is going to be a table where we have the log change, the stat statistical value, the P value and the adjusted P value. Okay. And from that we are going to identify the differential express. Monreal energizing up stretch Montreal stand up. Yes, yes. So, stretch, pret, reset. I know after lunch we are having here really long day and then we can reset attention if everyone is awake we can sit down [snort] ok so um and that's good because we are getting into a question and then we are going to get some more mathematical details. Okay. So, which steps in an RNA pipeline can one get account matrix? So, what are the steps that will give us account matrix? Okay, they were more than one option. I think you can choose more than one option. No. Okay. It was my mistake when I said the question. Okay. Suppos to be multiple choice with multiple so quantification of course have Okay. You supposed to choose therefore from quantification or summary and seed alignment. Now so of numbers here we are going to look at what normalation is doing and I like this normalation because quite straight forward so we have counts so here we have on the rose and we have every sample first step is to get the geometric mean so here we are having the geometric mean this number is the geometric mean for each of the ah so that's the first step we get the gen wise geometric mean and then we divide each of the counts by the corresponding geometric mean so for example we will get for this number 807 we will divide it by this geometric mean the first one okay and we get this table here so this is dividing the cons the r by the geometric mean the corresponding geometric mean so after that the step three is to get the median of each column so we are going to get the median per column and that median per column is what is called size factors or scaling factors. So the medium per each of the columns is the scaling factor for each of the samples. The next step is taking this scaling factor and going back to the original uh count table and what we are going to do is divide the rock counts by the scaling factor. So here you will be dividing 807 in the first for a 15 div by the corresponding scaling factor for that column. So 0.77 and table expression analysis for statistical analysis so normalization are steps getting the geometric mean by dividing the rock by the geometric mean then getting the median per column the scaling factor the median per column is the scaling factor and then we take the scaling factor of each sample and divide all the counts by this scaling factor and that new matrix is ​​the normaliz matrix so that's what is doing in the first step so we go back to this one does this first step is size factor estimation also call scaling factor estimation So what is the purpose of this step? The purpose okay let's start with dispersion. What is dispersion? Is the variation around the expected mean abundance per G. So how much variation we are going to get for each G in the abundance? If we will have many biological replicas estimate that because but because we only have like three replicas or four replicas we need to estimate what will be the actual dispersion. Okay. So and then what we want to do is to actually shrink the dispersion meaning we want to reduce this variability. So this is these figures are actually taken from the D paper D paper. So what they do is that the red dots here let's assume those are jeans. So we have the jeans here and they are going to fit a maximum livelihood estimate which is represented here with the red line. So the red line is basically the maximum livelihood estimate of the dispersion of the jeans. Um [clearing throat] here we see something that is very characteristic of the RNA data. In RNA data in the lower counts we are going to see higher dispersion. In other words, low abundance tends to show more variability. Ok and this dispersion starts to go down and then basically get to a fixed point for higher abundance gen in the x we ​​have the mean of normals. So basically the table that we got from here. Okay, this is the mean of normalized we get the mean and is the dispersion estimate. low abundant jeans tend to be more variable than the highly abundant jeans. Okay. So this kind of shape of this cure like going like this is very common in is what you will expect. Okay. So once that we have estimated this red line now all the gen are going to be moved towards this line. So that's the approach. So here graphically we are seeing these blue arrows like kind of moving in the direction getting closer to the red line. So that's what the dispersion shrink is doing and there are different methods to do this dispersion shrink. Now I consider outers like the ones that have here the blue circle around them on will not be moving are going to not be touch because they are kind of not following the the pattern so they will not be touch okay so this is just with very few jeans the real data will look like this [laughs] so we will have thousands of jeans all those black points there we still have the the the red which is the estimate and then the blue points are the new positions of the jeans. So all these blue ones are the ones that were shrink toward the estimate and the dots here on the top are the ones that were considered outliers and they were not moved. So when we are looking and we are going to look at dispersion uh estimate plots in the lab in this type of method we are looking at the blue points kind of follow the red line. Okay. So you have a dispersion estimate method where there is a red line and the blue points are somewhere else. It didn't work. It's not a good dispersion estimation method and there will be some blue point that are just not been touched because they are considered appliers. Do we have any questions so far? Is there any question about the okay so after this there will be the statistical test I'm not going to details with the statistical test but we will basically get a p value a statistical value and a log change for each of the gen and then we are going to get something else. We are going to get adjusted P values. Sometimes also call corrected P values. So what are these adjusted P values? P values ​​must be corrected for multiple testing. What does multiple testing mean? Okay. All the statistical tests were designed for one test. When you repeat this test thousands of times, just by random chance you are going to get values ​​that look really small like small p values. Okay? because all the tests they have a probability of having an error. So there is the error one and error two. So there is probability that they are going to reject the null hypothesis when the null hypothesis is true. And there is the error that they are going to accept the null hypothesis when the npesis is not true. Meaning sometimes they are going to say a gen is differential express when it is not and sometimes they are not going to detect a gen as differential express. Okay. So the statistical test are not perfect. There is a probability that this error occurs. But when you repeat this test thousands of times that probability will increase. So you have a probability of 0.05 multip per 100 of tests then some of the five of those tests are going to be grown. Okay? multiply by 10000 then you are going to have 50 of the you are going to have 500 of the test are going to be wrong and so this means that 500 false positives for example that's a large a number to correct for that we do adjusted so we do the multiple testing correction and basically the multiple testing correction is going to adjust or correct the individual p vouchers to account for all the Bon is very conservative poren expression analysis the one that is used is the discovery rate so that's the one that used are different versions of this one2 by default use the benjamin Fdr method. So what is this fate? These are the basics of the fate. Okay. So we suppose that we have 100 tests and that we want to define as for statistical significance 0.05. kind of intermission about treasures of arbitrary significance there is no reason there is nothing magic about 0.05 we could have 043 we could have 0.03 by convention a lot of people will use 0.05 and 0.01 kind of an arbitrary treasure you can consider the P values ​​more like a ranking system so for example what I'm trying to say is that it's not magic that something that has a P value of 0.01 is statistically differential express and something that has 0.01 one is not very close right so more like a ranking system anyway so we have arbit of significant 0.05 and we have all the P values. So the first step in f discovery rate is that you order the P values ​​from the smallest to the largest and then rank the P value. So the small the smallest P value is going to be rank one. So the smallest P value is going to be rank one and the largest is going to be the rank of the number of tests that we have done 100. And then every single P value is going to be adjusted and the way that adjusted is very so the original P value times the number of tests that we have done div by the range. So this means for the first p value we will have this p value multiply by 100 and div by one and this is the new corrected p value value. For the second value we will have this number 0.005 multiply 100 and div by and is the new p value and you do that for all the p vales and this new column is what we are going to see in the adjusted p vales questions clear okay then is going to be your yes have a question yes Sorry sh me. Yes. What is the question? Actually, we had a question about the plot that you had. Um, if the blue dots don't follow the red line, what could be the cause? The method is not the appropriate for the data. So um we will and we are going to see that in the module in the practice we are going to use three different dispersion methods and we are going to see that one of them is not the appropriate method. So basically if the blue point don't follow the red line the dispersion shrink method that you are using is not the most suitable for the data. Everyone is ready to play. And now you have to calculate the adjusted P value of a P value that is 0.003. It was rank 100 and you have a study that is 15,000. So what will that value be? Okay. So the correct answer will be 0.45. How did we get this? We will multiply 0.0. Let me try again. We will multiply 0.003 000 div 100 okay so we get 0.45 45 exercise. Ready the only difference is going to be the number of question. Now we only have 10000. How does it change the result? He. Oh. So now the adjust value will be 0.045. Okay. So in one case same prop will be 0.45 way to far away from significance and in this case will be 0.045 we are using like a statistical cut of a significance of 0.05 will still be considered express. I [snort] take message from the last question. [clearing throat] So the more in study the stricter multiple testing correction true or false. True. part of multiple testing correction right00000 resour value are going to change quite dramatically because the number of statistical tests that you are going to have to account for are going to be less so easy way to see how filtering [snort] exard Cameron Cameron Montreal. Montreal estimation publication in your question. Okay, perfect. We will have a question here. You will enter here. He will write in the slag. He will write in the slag. Okay. Thank you. Welcome. So we have time for questions. I usually take longer not clear of time for the question. to use the information from thank you how to use the information from estimation for publication and discussion okay in a lot of journal you are going to have to present your visualization okay the dispersion the dispersion usually will call kind of supplementary material is not something that you will put in the unless you are publishing in a dispersion shrink method not going to be in the main publication how you use it use it to show that your analysis was so you present this like you present your PC plot you present your dispersion estimation plot to show that data and your analysis is meaning dispersion estimation method works well with your data that the examples separated and so how you use it this well the first thing that you use this plot for is for your own sanity like a sanity check to know that your sh method is working well and then for publication you can have ited in the supplementary material kind of to show that it was done correctly um this dispersion estimation methods are not something that you see very commonly plot that you see very commonly in the publications are the plot the volcano plot the PCA plot this one is not something that you will see that often but if you want to have it there it will be to show that analysis was you on the slide or so is kind of like to the So this would be just like a visual sign. There's no like how much out of like bar from the line should they go from saying like maybe not as a sound. Is there like a score to the measurement of sword? Okay. So the question is whether we just look at the plot and decide based on the visualization. or there is a score that we can have to say that this person is working fine. The answer is that we look at the plot there is no score. So there is no score to say this method is working fine. So um so the question is how far they have to be to decide that not working. You are going to see a plot in the lab that is that's the anger right like this plot is too far. So if plots kind of follow the red line we say is fine. If the blue dots are somewhere else and your red line is here then we say this is not working. Okay. So more like a visual thing. We will look at it and if the blue points are following the red line we say okay it looks fine. If the blue points are not there or some of the blue point you have the red line and you have the cloud of blue points over here then you will see no they are not working. So have to be like a very clear. It will have to be a very clear mismatch. Yes, for you to say this is not working. Yeah. So, it has to be a very clear same that the dispersion shrinkage is not working and then you will see it. Yes. Is there any minimum level for normalization? Minimum level for normalization. I'm not sure follow that question. Minimum level for normalization. Oh, what is it? No, the person, give me more explanation. I don't really follow the question. So what I'm thinking is that it has to be like at least normalized in a certain way. So for example whether we could have scaling factors are all one and then you don't get to normalize the data is if I'm going correctly. The answer will be this normalization the one that we saw for this is based on the data. So the amount that is been the data is being normalized meaning the amount that each of these numbers is going to change depends on the data. So sometimes you are going to have like for example you look at here sample 3 the scaling factor is 1.1 so the amount of change for column 3 is very small right and maybe for the sample is 0.77 77. So the change in this column they are going to be more drastic. So if if the question is whether there is a minimum in the scaling factors the answer is no. So the scaling factors will vary between samples and there is no minimum of maximum. It will depend on the data. But I'm not sure if I am answering that question correctly. How are the number of determin from the exle when you would analyze only versus whatever the number of r in your count table that's the number of that you are going to consider for multiple testing so the number of that are actually giving to this number of genjusting the multiple adjusting the P values ​​for multiple testing. So we just count the number of rows in our account table those are the number of jeans and those are the number of statistical tests that we will be performing and that's what we will consider. estimation dispersion estimation for dispersion estimation is the criteria so what we okay the go of the dispersion estimation method is to shrink the dispersion basically we want to reduce the variability that's kind of what we are doing here so we estimate the dispersion and Then what we want to do is to shrink this variability to shrink the dispersion. So the criteria is how we are going to reduce this dispersion so that we are not going to remove biological factors. Let me know if I didn't answer that question as you expected or if I take it in another way. What are some tests we can conduct on our account data prior to normalization? or differential expression analysis suspect batch effect to choose the most appropriate method we are going to do that in the practical so you are we are going to visualize the data so that you are going to do is PCA plot uh you can also do a heat map of the road counts and you can also do a correlation analysis so those things. For example PCA plot what we would like to see in the PC plot is that you have one group and the other group they are separated. Sometimes in the PCA plot instead of the groups you are going to see the samples uh group by day of the analysis or other kind of extremes factor. Okay. So in your PCA plot when you are looking at the at your PCA plot if the groups are not separated and you see samples together that they have in common something else like the sequence in things like that. then you have a batch effect. So we will see the BCH effect in the PCA plot visualization a good PCA plot good data is when the samples from the different groups are separated based on the group that you are interested like in the characteristic definot type that is of interest and b effect will be when the samples are being group that are more similar based on other reason like the the lane of the sequencing the day analysis was done things like Okay. So, we are going to for now I have a bit more of time for module 3, you are going to prework. Okay. For the module 3, we are going to be working in our studio. So all of you are going to have to connect to our studio and I will just like to we are going to use two things about module 3. We are going to use a different data that we were using for the assembly and so on. And we are going to get the data the read counts directly from NCBI because we actually didn't complete the whole analysis with the full data set. So we are going to get different data set and now for the R um in R studio is going to be really easy that you just copy and paste in R studio and run it and then you are going to be done like in half an hour. But the point of this lab really is that you understand the steps so that you know how you are going to have to change it when it is your turn to do this analysis on your own. Okay. So not just copy and paste like kind of go through each of the steps they have comments and try to understand what each of the step is doing and what is the output that you are getting. So I'm showing the output here but kind of saying what the output is. So a lot of the questions for example we are going to do a PCA plot that is done before normalization we are going to do the the correlation analysis we are going to filter based on the abundance of the gene so each step just make sure that you are kind of understanding each of the steps and then I have here these questions are more about understanding the output of the The solutions are not here because the idea is that if you cannot get about the solution you ask it instead of just looking at the solution. So really go through all of these questions and try to find the answer and understand what is going with the R code. Okay. So that's just a comment about module 3 and other than that we have extra time. So we have 20 minutes extra time and we will then start with module 3. for break and then start practical for module.