Batch Condition Signed with airSlate SignNow

Batch condition signed

um so um one thing in microarray days last year we we started the course with expansion microwave but this year we started with RNA seek with microarrays people do normalization quite a bit um actually now with RNA see if you use salmon or our Sam to calculate a TPM the normalization is done already and so this you don't probably need to worry too much but I kind of just want to very quickly go over what kind of the things that people do with normalization supposedly you have some measurements on genes each gene you have some number and you are looking at two samples or you can look at many many samples right what are the some basic things people do with normalization the very simple thing you do is call the median scaling so supposedly this doesn't have to be rnaseq it could be proteomics or metabolomics you just have some numbers a list of genes or a list of variables right and you're looking at the sample and you try to normalize by the way when we try to normalize one important assumption we make is that most of the things most of the thoughts should be fairly similar between the two samples right that you you have to make some reasonable assumptions there most of the the dots should be similar between the two and so if we were to look at this you might see that sample one has a higher median and it also has a higher spread the things are just you know the range is much much bigger on the y-axis it seems like things are fairly correlated but the mean is different the median is different the variations is different so media scaling is just you subtract each sample by a scaling factor to kind of move their median to be the same and then you multiply by another coefficient so that these two are really under diagonal and then things will be No normalized then you can the numbers for the X and the numbers on the Y will then be more comparable and when would you use something like this for example in the early days when people are using let's just say if in our rnaseq analysis we don't calculate a GPM this first is a cap could happen supposedly we sequence the first sample much much deeper than the second sample and then we just look at all the reads that landed on that on each gene for from the first sample you can see that it will have a big spread and also the numbers in general are bigger and if you didn't sequence deep enough you will get another say another replicate and just have lower dynamic range of the other and supposedly we don't have an algorithm like salmon or our Sam to do the TPM normalization a very simple thing that people initially tried you know this has been tried in early RNA seek data is you calculate the read that lands on each gene you get this you know the number for each each gene on each sample and you just do median scaling it's you know like a reasonable job but you know I think for calculating differential expression it may or may not be the best way but this is certainly a potentially good way to try another one is called the low s and you probably see quite a bit of this sometimes when you try to draw a lot of dots in the in a curve on R and the algorithm will automatically show you another line which is kind of the best fitting to your data so so Louis is is a non linear it's kind of a nonlinear smoothing but you can see here if the data look like this each dot is one gene and x-axis is one sample and y-axis is another sample I would imagine for some of the metabolic profiling you you don't have 10,000 genes you have a few hundred numbers you might look the data might look like this right so and in this case if we still make the assumption that most of the thoughts should be similar between the two samples that lie on the diagonal we can also try to normalize this and the way to do this is in each section we try to divide the data into different chunks in each chunk based on all the data within here we estimate a linear scaling approach rates and then we move to the next section we do a linear move the next section we do a linear and when we try to connect them together the slope of the linear changes and that will give you kind of a nonlinear way to scale things but at the end imagine if you try to force this nonlinear thing into the diagonal then things will be together and there are standard are functions to do this as well you just you say okay I want to normalize second sample to the first sample who is the last it will try to normalize them then the x and y will become haribol for this one you need a reference sample you know I'm normalizing my Y axis by the value of x so you have to decide which one is the reference and which one I want to change right so in this case yeah so the original Y we want to keep the same and the X is changed to X prime based on Y so in this case we're treating the Y as the reference sample and then X prime with Y are more comparable the difficulty comes when sometimes in expression data you don't know which one is a reference sample you just have you know ten disease and ten normal which one do you use as your reference sample it's difficult to find and so another normalization method which is really very widely used in biological system is called Quantum normalization again it assumes that most of the measurements is don't change between the samples and it does something like this so supposedly you have you know again this big matrix though each column is a different sample and each row is a different gene what it does is we create an additional and then in each role we calculate the top percentile the average of the top percentile from each of the samples let's think about another example supposedly we want to look at the GPA of say I believe colleges but supposedly Harvard has very bad gravy inflation a lot of students have 3.8 3.9 GPA s where is another school let's just say UPenn I don't know they grade very harshly right and so their students never really have that high score they their their average is 3.2 and so it's like yeah that the range might be different that the me is different the the range is different can you use median scaling you can probably do that right but which one to use as reference and you know you correct everything to Harvard or you correct everything to you pan now you don't know therefore what you can do is you get all the Ivy League schools together you assume that that's a yeah that's an important assumption you assume that students from all the sudden Ivy League schools are the same right and they roughly what percentage of students their top 1% the Harvard top 1% and you can't have 1% are literally the same level okay on the same league therefore what you do is you take the GPA of the best student that the top 1% student at Harvard that the you can Columbia Princeton or whatever and you take all of their GPA you average them together and create this top 1% students GPA and then you take the top 2% is from 1% to 2% house student from all the seven schools you average them together and then you create that average 2nd percentile here and then average 3rd percent how an average you know 100 percent house is the worst student from every school average together and once you have this you replace the original data with the present house or you know whatever this guy's GPA is you don't care now if it's the top one GPA student or the 1% how I'm gonna replace it with the average from all seven schools back to the original data and you just replace all the data at the end none of the matrix cells retain its original value but then this one you were really forced all the different samples to have the same distribution because that's that that's the distribution you calculate it from this one you'll replace each of the percent health with the average and that's quantum normalization so you can see here we don't really pick a reference sample we just say okay you know this roughly these are all similar schools their students should be the same they should have the same distribution yeah so then you you calculate the overall percentile level and you replace the original data with the percentile and so quantum normalization I would think nowadays if you use some proteomics data metaball metabolomic data it's you know pretty good right now with the RSM or summin Callisto type of algorithm you can probably get just a TPM score and it should be more or less comparable and you don't need to do this but then even after these type of normalization there are two cases with batch effect which means that you there might be some technical variations which are not really biological so supposedly if you do a cancer normal comparison right you did patient all the cancer samples and all the normal samples and you try to clasp her them so supposedly let's just say you have let's do something simple you have 20 tumor samples of the same cancer type and then the adjacent normal because sometimes with surgery the doctor take out a little bit outside tissue just to make sure they have a clean cut they take a little adjacent to a normal tissue and you put both into the profile and they did say 20 pairs of such patients okay and then after that you do the normalization and you start uuu calculates the TPM and you cluster the samples supposedly these are all the same cancer type how would you expect the samples to cluster yes so basically you have 20 pairs of data which is 40 40 samples right 20 disease and 20 normal of the same let's say breast cancer will be right ideally you want the samples to cluster by cancer and normal but I can tell you we have examples where you know some groups were analyzing data they are asking us to help them with an analysis of their data anthem no matter what we do they are like said the data are clustered by two two groups initially initially we just look at how the tumor samples are classified and they are clustered in two groups and we were wondering oh maybe the cancer subtypes and then no they are not subtypes and they finally found out one group of sample were profiling the spring and other group of samples profiling the summer so clearly there are some things that are not biological right and then we asked them can you also put the normal samples in a class er again you see the normal and tumor are clustered together in the spring and the tumor normal are clustered in the summer in a different class er then we will know it's it's a batch effect and you can see it can make it samples not directly comparable I would say they is a pretty extreme case most of the time you may not see that strong class during effect or batch effect they could be pretty subtle but it can still have a huge effect on your final differential expression analysis and the way that the reason that we see this type of differences yeah we mentioned the day and months of the experiment and you might be using different reagents you know enzymes buffers that you made or enzyme you bought from different lots of from the even same company you might be using different mice they might be hosts in different cages or you just bought them from different companies even as the final sequencing yeah so also the lab technician or the postdoc who is doing the experiment maybe is a different technician then then you accounted for is out and even at the end after sequencing the sequencer might be different you send it to a different sequencing facility or something they will give you a different result so all of these can cause potential problems yeah so for for example this is an experiment where people use RNA I to knock out to knock down three different genes to kind of reduce expression level of three different genes and so they have three petri dish or three batches of experiments in here there's a batch one two and three and also the R and C is control versus RNA I whether you knock how the G or not and the third is the cell clone so each of them they also had different cells you know like dish and so um you can see in here ideally what you want to see is that all the controls are together and all the treatments are together but after you do the clustering it's you know maybe control are a little bit more you here and there treatment is a little bit more enriched here but they're just too many other variables in here it's so you have to remove the batch effect and I want to show another example this actually gets published it's a very striking finding where the the experimental group has done RNA seek of many different tissues in human and many different tissues in mouse and then when they run a PCA so this is a three dimensional PCA you can see this projection now in three dimension you can see all the human samples are circles and other balls and all the mouse samples are cones and you can see wow it really separates and the first principal component separates by human and mouse and the shocking result that's actually reported is something like you know samples are really separate by animals or the species than by the organs which means that you know the human brain is closer to the human heart then it's it's to the mouse brain the reason this is kind of shocking is do you know why we even doing experiment implies or worm or or or Mouse is because it's very difficult to do the experiments in humans you can't make them you cannot how the gene in human right it's difficult to do a lot of things in human and we are assuming that Mouse are more or less similar to human therefore if we observe this thing in Mouse in that issue you will recapture it will you will recapitulate what you are you will capture what you see in human and if this conclusion is right it would mean all the experiment that we do in animal models is a complete waste right because you know what do you know about human brain from mouse brain you should take their human foot to do the experiment right so it's it's it's pretty scary to think that after so many years that we would reach a conclusion like this and then a year later you see this being blocked and they look at this in this it's experiment they sequence the RNA in different lanes and in the first Lane it's all human sample in the second lanes on human third Lane seven Lane it is all human samples and lane four and six are almost mouse samples fortunately there's a different run in a different Lane where they have some you know post human Mouse right but after they have removed the batches they know there are some very abilities that are captured in that batch is probably that technician where that sequencer or whatever and finally the tissues indeed a cluster together the human brain and the mouse brain are together now and the human heart and mouse heart are together it's actually kind of embarrassing that something so obvious computational biologists do not think about removing batch effects so it's a definitely a big deal and so anytime that you have say you do it the differential expression experiment when you don't see enough differential gene expression don't go to gene set enrichment right away remember previously we said oh after you do the ecig to you correct from multiple hypothesis testing you have you know a few hundred genes differentially expressed you're happy but actually even if you have a few hundred differentially expressed genes you might want to double-check to make sure they are not complete garbage right you want to make sure and also if you don't have enough differentially expressed genes before you try to do inside enrichment analysis double check your batch in fact and so there are definitely some important messages about experimental design so better technology never eliminates the need for better experimental design or good experimental design and so first of all easy case is we try to be consistent and process all of the samples at the same that would be the easiest case but if you have to do them through different batches you know make sure first of all you try to run you try to record as all these variables as possible and what's the date of the experiment work where the labs the personnel because sometimes actually this project was done in a consortium setting in different places right yeah so the you know the person now the environment all these variables make sure that you have record of those and also if possible try to balance the groups of interest at least include some controls in each batch for example between each laying of samples you put some human sample and somehow samples there that at least you are you will have ways to correct it later on so try to avoid perfect a confounding experiment when the batch groups are perfectly correlated for example if you do all of your treatments in one batch and all of your control in another batch there is no way you can figure out what your control or what your treatment is really doing to the cell because it could be the experimenter it could be the cell culture it could be the sequencer and it's also part of the treatment you just cannot use them up hard and so supposedly you do a tumor normally experiment and if you take all the Lomo samples from one hospital and all the tumor samples from another hospital and you do the expression profiling and you want to compare tumor normal that would be a disaster right you will never know what's the difference between the two hospitals or the difference between cancer and normal yeah so how do you really remove the batch effect so you can imagine supposedly by the way I would say Banshee effect removal right now there are some reasonable way to do it it's not perfect because a lot of these batches and removal algorithms were developed for microarrays with RNA see they kind of works but we don't know for sure that they always work the best so you just have to check so supposedly you have a big expression matrix the columns are the samples and the rows are the genes and we are saying that there is a baseline gene expression which is what you measure it's kind of the average expression across all the samples and then supposedly you are trying to estimate it remember if you are using mimic to do differential expression you are saying okay these samples are tumor and these samples are normal and then we want to ask whether each gene so by the way this is kind of the value you ask for each gene is this gene differentially expressed between the tumor samples and normal samples right and you are trying to ask whether the coefficient is different from the null from the average right and if it's a tumor normal that coefficient is significantly different you will call it a differential gene and so in this column most of the values will be close to zero but occasionally there will be some genes that is upregulated or downregulated right and but you're trying to compare the treatment and control and of course you're trying your best but there are you know all the small variations there are some small noise that's associated with every sample is every gene so this one capture all the remaining small noises so in here the primary variable is the treatment and control or your tumor versus normal but imagine if you have another batch so supposedly it's some are batch one some are patch - you can imagine the batch is another treatment because because of the technician because of the weather the sequencer the reagent or whatever it's creating another effect on all the genes in the cell well not all the genes but like some genes in the cell probably in a coherent way you are trying to estimate how much is that and you can actually do this directly in Lima and just remember Lima can give you some complex designs so we are saying yes there are tumor versus Lola but then we are also trying to estimate what is the batch doing to each of the jeans so you are saying it for example if it's a batch one batch - then you asked image what is the batch creating the difference on g1 vs jin - vs jin story so it's it's going to try to look at all the samples in one batch compared to all the samples in another batch and estimate whether there's a observed slight differential expression between the batches and you want to take that out before you compare the the real difference between your tumor versus normal okay and so you can consider batch effect as another treatment and if you know the batch is ahead of time you can remove it yes so intuitively you considered batch as some kind of treatment effect and you can use hours of lightly metal to remove it they are existing so another algorithm that can remove the effect is called combat this actually works very very well on RNA seed and can capture even nonlinear type of variabilities and it seems to work pretty well and so actually how do you really even see or did identify or correct the batch effect the first thing is we need to first see whether there are batch in fact present or not and that's usually done by clustering you can do this also by PCA this is actually one example if you have all the samples available and each dot is one sample you just ran a PCA you look at the result and see whether so in this case it could be that the samples are colored differently by the day the samples a process or you can separate them out by all the tumor samples coming from one hospital and another hospital and a third hospital and so on right so if after you run the PCA you click you clearly see that the samples clustered in a way that's corresponding to the batch that's a very good indication that this data needs normalization and it needs match in fact removed whereas if after whatever kind of normalization the batches you are seeing is no longer separating the samples or they are kind of all merged together that's a reasonable indication that you are removing out of the batch effect or it could be that you still see them clustered but they are clustered in a more biologically meaningful way such as all the tumors are here and all the normals on the other side and they are no longer they differ by the date of the experiment or the the hospital right then then you you know if you have removed the batch effect the way to detect this is to do the clustering you want to visualize the samples for example using LOC ppm and in order to remove it if you have some really simple batches just to to condition two days right um you can use you can actually also three days is also okay combat can remove those known batches if you know that there are three different days you process the sample so supposedly if you're doing the experiment on cells or a mouse you want to do three replicate you will be better that you treat the cell and get the treated and untreated in day one you treat again in another in another group of cells you get the treatment control in a second day and then a surge of the batch you do treatment control and a third day and then that you can use definitely use combats to remove the batch effect for Lima if you have more complicated examples for example recently we have some samples initially was the same order the RNA it's it's sent to three different centers and each Center also have different days and also the same tumor has freshly processed versus paraffin-embedded tissues even though they are coming from the same tumor and so you can see it's kind of a nested design and we put them in all the different three different labs fresh frozen versus FFPE samples and also different replicates and then you want to do the classroom for that while we use something like Lima to really remove the complex treatment so you are saying that the FAPE process probably did something to the tumor I each sequencing center probably has another effect and then each day we'll have a separate effect and Rima can removed them all separately and then at the end you will get the batch in fact removed lock CPM value and that you can use as your gene expression index and you can use this for your other calculations for gene expression and then to make sure that you have correctly remove the batch effect you might want to do a sample clustering again either ways hierarchical clustering or with PCA to make sure that the samples are now separated by the biological difference not by the hospitals or the day of the experiment by the way if you have simple badge you can also use the ec2 and and also to basically add ec2 now count can do some simple batch in fact removal again consider the batch as another variable when you are trying to compare the sample between tumor normal or treatment and control you say okay I have another variable which is the treatment date and or the batch and so this one will tell you the level of differential expression and give you that the genes that are really different between the tumor and normal or the different conditions but it will not attach the original expression index files you will just get the level differential expression which match effect removed but it will not it will not give you a corrected a lot ppm value whereas if you do Lima you can get the corrected GBM levels at large scale okay so that's a the rough idea for a Bosch effect so PCA sometimes you might do a PCA and the first principal component you can see oh it's a batch that's a very good indication you should remove batch in fact and so sometimes you may not know which variable is really creating the battery but it is the technician wrong or the ragin wrong or they top the experiment or the mice that we bought from different companies which one is wrong so what you can do is for every sample you draw dots right on the PCA and then you you draw each variable with you know like say if we separated by Hospital draw them in different colors do we see them separate if not then it's okay right if you see oh it's really the date of the experiment the early versus late are now clustered then you can use it to help you identify which condition or which batch really created the biggest problem and then you can put that as a variable in DC to help you remove that variable okay but this is definitely important by the way homework to that sample has batch effect you need to remove it otherwise you will not get the correct differentially expressed gene expression okay okay thank you that's all for today