Here I test if chatGPT with the GPT-4 model can do basic single-cell RNA analysis. In short, the results are impressive.

So we did the basics of single cell sequencing without writing any code or using any of our biological knowledge.

We're going to do things a little different today.

I'm sure you've all heard or even tried ChatGPT but recently they published the new model GPT-4 which is even better than the previous ChatGPT model.

So we're going to see if GPT-4 can do single-cell sequencing and by that I mean the analysis.

I'm gonna pretend I have these 2 samples that were processed by 10x and that now I don't know what to do except that I want to use scanP for analysis.

So I'm just gonna input a quick prompt explaining my samples and what I want to do.

I have 2 samples that were processed from CellRanger, lung1 and lung2, but I also want to pretend that I don't know what I'm doing at all so I'm going to need help installing Python and Jupyter Notebook.

Maybe it'd give you different directions if you said you were using Windows.

me some pretty basic instructions for installation which aren't ideal.

I'd rather install it in a package manager like Honda.

And Ubuntu probably already has Python 3.

But anyways, I'm just going to use my Python installation.

But let's actually see what it does for analysis.

So it looks like it imports the right modules.

I'm going to have to update these paths here.

I was hoping it would give me the actual paths to the matrix files with n, long1 and long2.

It cacognates the files and then does some basic pre-processing.

So let's start with this and then see if we can make it go further.

So I wanted to prompt it again because I realized I forgot to specify that the data were from mice and I wanted to specify that these are the output directories from CellRanger so I wanted to know if it could find the actual correct path to the matrix directory within them.

It knows the filter feature matrix is within the outs folder.

And it also gives me the option to use the h5 file, which would work as well.

The data loaded fine with the path it gave me.

Of course I had to modify it slightly because it didn't know I was in this directory.

We have 22, 000 cells, 32, 000 genes, and it's just doing some basic pre-processing here.

So I've noticed 1 mistake but I told it it's mouse but it's still using capital MT for mitochondrial.

So if you look at the ADATA.obs, we're gonna get 0% mitochondrial.

Let's just see, I know I could fix this easily just by lowercasing this, but let's see if we can get chadgpt4 to do this.

I'm just going to tell it the issue that every cell has 0 counts.

I said mouse to begin with it might have done fine.

But I told it that every cell has 0 counts for mitochondrial and it knew the exact problem.

So now we actually have the correct calculations for mitochondrial counts.

So let's just keep on going through what it gave me earlier for the pre-processing.

All right so it's doing the basic pre-processing steps here normalizing every cell to 10, 000 counts, converting to log, finding highly variable genes.

Saving the raw and then subsetting on the highly variable genes.

This almost looks identical to the standard workflow on the scanp tutorial.

It realized that there were different batches but it didn't do any correction.

Maybe I can prompt it afterwards but let's just continue through this acting like I don't know what I'm doing.

This is where it left off so let's see if we can get it to do some clustering and maybe even identify cell populations based on their marker genes.

So always remember to be nice to the AI just in case it takes over the world 1 day.

it can help me with clustering and finding marker genes.

to copy these over to my notebook, but a couple points like it's pretty nuanced, like it knows that you can use the laden or the lovane algorithm here, but it picks the laden by default.

Maybe they do need to be integrated though, so I might ask it at the end.

Let me see if I can get it to give me fewer clusters.

I'm just going to tell it I have too many clusters based on what I would expect biologically.

Of course I know we would just need to change the resolution but let's see if it can figure that out.

It did exactly what I was hoping it would do.

Just gave me a lower value for the resolution and the clustering.

So I'll just put that in there and rerun that.

And I'm wondering if it's smart enough to know an objective way to pick the resolution.

So there are different methods to do that instead of just picking a random resolution.

So it actually gave me a very interesting example here using this silhouette score.

I've never used this before but if this works this could be a nice simple little method.

Let's see what actually happens if I run this.

So it's taking every resolution from 0.1 to 1.5 and 0.1 increments and passing that to this compute silhouette which returns this silhouette score.

So it calculated all those scores and also gave me code to plot it.

So it looks like it's just gonna pick the highest silhouette score.

I don't know if this is actually that great.

This might not be a viable method for single cell analysis, but we're just gonna go ahead and calculate the UMAP and plot it.

0.1 is pretty low, usually stop around 0.3, but it might work for this data set.

These are lung data we probably have around this many major populations.

The lung has a lot of cell types but I doubt I picked them all up in these data.

I'm going to go ahead and recalculate the cluster markers with the code it gave me earlier.

And GPT-4 will eventually be able to take image input.

So I'm going to ask it to give me a way to print these so that I can give it the top 10 markers for the 13 clusters here.

I just told it I re-clustered and I wanted a way I can copy and paste the clusters so that I can input them here into chat-gpt4.

You know, every single time I see it give me output it blows my mind.

It knows exactly where the markers are saved within the A data so it has a very good idea of the structure of AND data.

So I'm just gonna go ahead and copy this over.

So it just says this data frame of the top 5 markers.

I don't know exactly why maybe I can just copy this.

Let's try top 10 and then I'm going to re-specify that these are lung data.

All right so it gave me some like macrophages is definitely going

Endothelial cells for example we see cdh5 here, bcam1.

So these are definitely endothelial cells.

So I asked it to help me label them with abbreviated names and it gave me exactly that.

The dictionary and then the mapper here, so I'm just going to copy those in.

It also gave me another plotting function here so let's see what happens.

we look at AData.ombs Even though the laden column is a 0, 2, 6, and other numbers, they're actually strings.

But let's see what happens if I tell it that all the cell types were mapped as an AN.

So it got it got it wrong the first time, but it knew exactly what the issue was here.

me a new dictionary, where they're actually strings.

So I'm just going to copy this new dictionary and replace the old 1.

We could manually verify with known canonical markers of these different cell types, but I want to know if it can do that.

So can we double check using canonical markers?

All right, so it gave me this dictionary of canonical markers, and then it gave me this dot plot scanp function.

And we're going to run into the problem where some of these aren't going to be in the data which could arise because maybe the gene name doesn't translate to what they actually used.

You know there's different naming conventions or maybe they're just not in the data they got filtered out.

So I'm just going to give it this error and see how it responds.

We could just take them out but I want to see what GPT does.

a little function to check if the gene exists in the data.

I already know it doesn't because that error wouldn't show up if they did so it would just be redundant but basically it's just saying find alternative markers or remove them so I'm just gonna go ahead and listen to it and just remove those from the dictionary it gave me.

Alright, so I removed them and then I'm just going to run this again.

So for example, if we look at endothelial cells here, these 3 endothelial cell genes are expressed in these endothelial cells, but if we look up and down they're not expressed in much other than these lymph endothelial cells.

Fibroblasts, for example, collagen, so these definitely look like fibroblasts.

So it looks like it did a really good job just based on the top 10 marker genes.

They might not be perfect and I'm sure we could find a few discrepancies but I'm just gonna take it at its word right now.

So we have this well-labeled UMAP, but let's say I'm completely stumped.

I don't know what to do for any downstream analysis.

So I'm going to have it recommend me some downstream analysis.

So it suggested 7 different things and some of these are actually pretty good.

Obviously, we would do this differential expression between lung 1 and lung 2 samples, you know, it suggests gene regulatory network.

Pseudotime analysis for this data set, it's probably not that great, but it doesn't really know the background that well, So I don't blame it for that.

You know, cell-cell communication, and then of course pathway analysis on those differentially abundant genes.

Of course, we don't have spatial data, and this doesn't seem that interesting.

So it makes a new column that labels the sample and cell type and it just uses the built-in rank genes group to do differential expression using locoxin.

I'm not going to do that It's just going to be redundant with this, which is what I really want.

It's not a data frame, but they're giving it to me in a dictionary, which is

ScanP does have a instead of plotting it function, it has a data frame function.

You can just get the differential expression results in a data frame which would be nicer than this but let's just go with it.

I really don't like this because it just gives you the cell group and then the differentially expressed genes.

if they're up or down, what the p-value is, so let's ask it to do this in a different way.

I don't like this, can you make it in a data frame format instead.

So it gave me this again it must not know about the scanP method to just make a data frame but this looks like it will work.

So if we look at this top 20 for each testing condition or the actual differential expression data frame which has the top 100 in it for each.

Okay let's see if we can do pathway analysis on this now.

I've never used this GPR profile API, but let's go ahead and try to use it.

I usually use enrich R and it looks like it's going to use go keg and whatever this react is and then it knows again to use musk muskless because I told it

we're using mice and it's doing it for every group

who we got an error without trying to figure it out myself.

Okay it just told me to do this and then replace it so let's see.

Yeah it looks like it just got itself confused because it changed earlier long1 to control and then long2 to treated but in this it has treated and control.

So if we save be treated and I don't want to run all those again.

So instead of printing it, let's just actually look at the data frame.

Here, so we have our pathway analysis now for any of these groups.

So if we want to look at what happened in our T cell treated group, we could do that.

So it did pathway analysis, and we could filter this on Go or keg.

So great, it seems like it worked perfectly fine for pathway analysis.

And now here, anybody wants to see a simple pathway analysis method in Python, here you go.

And I'm not going to do anything too advanced like this because this is going to take a while.

Seeing it takes a while to run, I don't feel like making this a long tutorial.

But, You know, self-self-communication is interesting.

So let's see if it knows how to do this in Python.

It tells me which package I need and it gives me all this code to run it.

It gave me all this code for this package I've never used before.

I'm not even gonna prompt it, I'm just gonna paste the error.

It seems that cell phone has been updated, so the most recent version might not work.

So it might not be able to do this but that might just be a versioning thing.

But now I'm sure I could go to this cell phone DB myself and read their docs and figure it out and then reuse some of this code.

All right let's just give this 1 last try here.

And if it doesn't work it doesn't, we just won't do the...

Yeah let's just give up on a cell chat I'm sure we could figure it out pretty easily if we put some of our own human brain power into this but that's not the point instead let's just stop here.

It did a good job with the basic pre-processing and this weird silhouette method that may or may not be the best for this kind of data but it did a really good job of labeling the cells this was probably my favorite and what I thought was the most impressive was a bit of a job it actually did with labeling and then you can find canonical markers very easily as well.

So I'm sure you could ask it to do specific things like plot the go categories or the enrichment results etc but this is pretty impressive.

I'm not sure if it's at the point of taking my job yet, but it's certainly at the level where it will help people greatly when they're doing their own analysis.


Speaker 0: So we did the basics of single cell sequencing without writing any code or using any of our biological knowledge. We're going to do things a little different today. I'm sure you've all heard or even tried ChatGPT but recently they published the new model GPT-4 which is even better than the previous ChatGPT model. So we're going to see if GPT-4 can do single-cell sequencing and by that I mean the analysis. I'm gonna pretend I have these 2 samples that were processed by 10x and that now I don't know what to do except that I want to use scanP for analysis.

So I'm just gonna input a quick prompt explaining my samples and what I want to do.

Speaker 1: All right so I told it

Speaker 0: I have 2 samples that were processed from CellRanger, lung1 and lung2, but I also want to pretend that I don't know what I'm doing at all so I'm going to need help installing Python and Jupyter Notebook. And I just told it I was using Ubuntu. Maybe it'd give you different directions if you said you were using Windows.

Speaker 1: Alright so it gave

Speaker 0: me some pretty basic instructions for installation which aren't ideal. I'd rather install it in a package manager like Honda. And Ubuntu probably already has Python 3. But anyways, I'm just going to use my Python installation. But let's actually see what it does for analysis.

So it looks like it imports the right modules. I'm going to have to update these paths here. I was hoping it would give me the actual paths to the matrix files with n, long1 and long2. It cacognates the files and then does some basic pre-processing. So let's start with this and then see if we can make it go further.

So I wanted to prompt it again because I realized I forgot to specify that the data were from mice and I wanted to specify that these are the output directories from CellRanger so I wanted to know if it could find the actual correct path to the matrix directory within them. And it does. It knows the filter feature matrix is within the outs folder. And it also gives me the option to use the h5 file, which would work as well. The data loaded fine with the path it gave me.

Of course I had to modify it slightly because it didn't know I was in this directory. So we have our data that's concatenated. We have 22, 000 cells, 32, 000 genes, and it's just doing some basic pre-processing here. So I've noticed 1 mistake but I told it it's mouse but it's still using capital MT for mitochondrial. So if you look at the ADATA.obs, we're gonna get 0% mitochondrial.

Let's just see, I know I could fix this easily just by lowercasing this, but let's see if we can get chadgpt4 to do this. I'm just going to tell it the issue that every cell has 0 counts. Oh I'm actually very surprised. This is pretty crazy.

Speaker 1: If I didn't make that mistake and if

Speaker 0: I said mouse to begin with it might have done fine. But I told it that every cell has 0 counts for mitochondrial and it knew the exact problem. So I'll just replace this line in here. That's pretty crazy. So now we actually have the correct calculations for mitochondrial counts.

So let's just keep on going through what it gave me earlier for the pre-processing. All right so it's doing the basic pre-processing steps here normalizing every cell to 10, 000 counts, converting to log, finding highly variable genes. These look like default parameters here. Saving the raw and then subsetting on the highly variable genes. This almost looks identical to the standard workflow on the scanp tutorial.

It realized that there were different batches but it didn't do any correction. Maybe I can prompt it afterwards but let's just continue through this acting like I don't know what I'm doing. This is where it left off so let's see if we can get it to do some clustering and maybe even identify cell populations based on their marker genes. So always remember to be nice to the AI just in case it takes over the world 1 day. But I'm just simply asking it if

Speaker 1: it can help me with clustering and finding marker genes. So at finish I'm just going

Speaker 0: to copy these over to my notebook, but a couple points like it's pretty nuanced, like it knows that you can use the laden or the lovane algorithm here, but it picks the laden by default. Anyways it's just every time I use it

Speaker 1: I'm pretty impressed. Alright so this is what it

Speaker 0: gave me. Let's just run these. Here's the output. Maybe they do need to be integrated though, so I might ask it at the end. I probably have too many clusters.

Let me see if I can get it to give me fewer clusters. I'm just going to tell it I have too many clusters based on what I would expect biologically. Of course I know we would just need to change the resolution but let's see if it can figure that out. And yes. It did exactly what I was hoping it would do.

Just gave me a lower value for the resolution and the clustering. So I'll just put that in there and rerun that. And I'm wondering if it's smart enough to know an objective way to pick the resolution. So there are different methods to do that instead of just picking a random resolution. So it actually gave me a very interesting example here using this silhouette score.

I've never used this before but if this works this could be a nice simple little method. Let's see what actually happens if I run this. So this is pretty impressive here. So it's taking every resolution from 0.1 to 1.5 and 0.1 increments and passing that to this compute silhouette which returns this silhouette score. Let's run that.

So it calculated all those scores and also gave me code to plot it. So it looks like it's just gonna pick the highest silhouette score. I don't know if this is actually that great. This might not be a viable method for single cell analysis, but we're just gonna go ahead and calculate the UMAP and plot it. 0.1 is pretty low, usually stop around 0.3, but it might work for this data set.

These are lung data we probably have around this many major populations. The lung has a lot of cell types but I doubt I picked them all up in these data. So let's just go with this for now. I'm going to go ahead and recalculate the cluster markers with the code it gave me earlier. And GPT-4 will eventually be able to take image input.

It's not at that point yet. So I'm going to ask it to give me a way to print these so that I can give it the top 10 markers for the 13 clusters here. I just told it I re-clustered and I wanted a way I can copy and paste the clusters so that I can input them here into chat-gpt4. You know, every single time I see it give me output it blows my mind. It knows exactly where the markers are saved within the A data so it has a very good idea of the structure of AND data.

So I'm just gonna go ahead and copy this over. So it just says this data frame of the top 5 markers. And then it does this. I don't know exactly why maybe I can just copy this. Let's see what happens here.

Speaker 1: Okay, so if these are

Speaker 0: right, it's pretty incredible. Maybe if we give it the top 10 markers. Let's try top 10 and then I'm going to re-specify that these are lung data.

Speaker 1: All right so it gave me some like macrophages is definitely going

Speaker 0: to be right. I'm not sure about these others. I'll probably have to double check. Endothelial cells for example we see cdh5 here, bcam1. So these are definitely endothelial cells.

Natural killer cells, etc. Let's just have

Speaker 1: it label the cells now.

Speaker 0: So I asked it to help me label them with abbreviated names and it gave me exactly that. The dictionary and then the mapper here, so I'm just going to copy those in. It also gave me another plotting function here so let's see what happens. So it made a mistake here so I know what

Speaker 1: it did. None of these mapped because if

Speaker 0: we look at AData.ombs Even though the laden column is a 0, 2, 6, and other numbers, they're actually strings. They're not, they're categorical. They're not saved as integers. So this 0 really needs to be this. But let's see what happens if I tell it that all the cell types were mapped as an AN.

And this is incredible. So it got it got it wrong the first time, but it knew exactly what the issue was here. So it's gonna give

Speaker 1: me a new dictionary, where they're actually strings.

Speaker 0: So I'm just going to copy this new dictionary and replace the old 1. This is what we have now. We could manually verify with known canonical markers of these different cell types, but I want to know if it can do that. So can we double check using canonical markers? Let's see how it does here.

All right, so it gave me this dictionary of canonical markers, and then it gave me this dot plot scanp function. So let's see how this does. And we're going to run into the problem where some of these aren't going to be in the data which could arise because maybe the gene name doesn't translate to what they actually used. You know there's different naming conventions or maybe they're just not in the data they got filtered out. So I'm just going to give it this error and see how it responds.

We could just take them out but I want to see what GPT does.

Speaker 1: So it looks like it's just giving me

Speaker 0: a little function to check if the gene exists in the data. I already know it doesn't because that error wouldn't show up if they did so it would just be redundant but basically it's just saying find alternative markers or remove them so I'm just gonna go ahead and listen to it and just remove those from the dictionary it gave me. Alright, so I removed them and then I'm just going to run this again. So for example, if we look at endothelial cells here, these 3 endothelial cell genes are expressed in these endothelial cells, but if we look up and down they're not expressed in much other than these lymph endothelial cells. Fibroblasts, for example, collagen, so these definitely look like fibroblasts.

T-cells, CD3, etc. So it looks like it did a really good job just based on the top 10 marker genes. They might not be perfect and I'm sure we could find a few discrepancies but I'm just gonna take it at its word right now. So we have this well-labeled UMAP, but let's say I'm completely stumped. I don't know what to do for any downstream analysis.

So I'm going to have it recommend me some downstream analysis. So it suggested 7 different things and some of these are actually pretty good. Obviously, we would do this differential expression between lung 1 and lung 2 samples, you know, it suggests gene regulatory network. Pseudotime analysis for this data set, it's probably not that great, but it doesn't really know the background that well, So I don't blame it for that. You know, cell-cell communication, and then of course pathway analysis on those differentially abundant genes.

Speaker 1: Of course, we don't have spatial data, and this doesn't seem that interesting. Let's see if we can get

Speaker 0: it to do some of these. So let's do differential expression. So it makes a new column that labels the sample and cell type and it just uses the built-in rank genes group to do differential expression using locoxin. Then it wants to visualize them. I'm not going to do that It's just going to be redundant with this, which is what I really want.

It's not a data frame, but they're giving it to me in a dictionary, which is

Speaker 1: OK, I guess.

Speaker 0: ScanP does have a instead of plotting it function, it has a data frame function. You can just get the differential expression results in a data frame which would be nicer than this but let's just go with it.

Speaker 1: I really don't like this because it just gives you the cell group and then the differentially expressed genes. I don't know

Speaker 0: if they're up or down, what the p-value is, so let's ask it to do this in a different way.

Speaker 1: So I just told it

Speaker 0: I don't like this, can you make it in a data frame format instead. So it gave me this again it must not know about the scanP method to just make a data frame but this looks like it will work. So if we look at this top 20 for each testing condition or the actual differential expression data frame which has the top 100 in it for each. Okay let's see if we can do pathway analysis on this now. Okay this is pretty impressive.

I've never used this GPR profile API, but let's go ahead and try to use it. I usually use enrich R and it looks like it's going to use go keg and whatever this react is and then it knows again to use musk muskless because I told it

Speaker 1: we're using mice and it's doing it for every group

Speaker 0: who we got an error without trying to figure it out myself. Let's just see what this is. Okay it just told me to do this and then replace it so let's see. Yeah it looks like it just got itself confused because it changed earlier long1 to control and then long2 to treated but in this it has treated and control. So if we save be treated and I don't want to run all those again.

It looks like it has all the data here. So instead of printing it, let's just actually look at the data frame. Here, so we have our pathway analysis now for any of these groups. So if we want to look at what happened in our T cell treated group, we could do that. So it did pathway analysis, and we could filter this on Go or keg.

So great, it seems like it worked perfectly fine for pathway analysis. And now here, anybody wants to see a simple pathway analysis method in Python, here you go. And I'm not going to do anything too advanced like this because this is going to take a while. Seeing it takes a while to run, I don't feel like making this a long tutorial. But, You know, self-self-communication is interesting.

I've never done this in Python. I've done this in R. So let's see if it knows how to do this in Python. So This is pretty incredible here. It tells me which package I need and it gives me all this code to run it.

Let's just go ahead and run this. It gave me all this code for this package I've never used before. We get an error. Let's see if it can fix this. I'm not even gonna prompt it, I'm just gonna paste the error.

It seems that cell phone has been updated, so the most recent version might not work. It seems like

Speaker 1: it thinks this is gonna work.

Speaker 0: So it might not be able to do this but that might just be a versioning thing. But now I'm sure I could go to this cell phone DB myself and read their docs and figure it out and then reuse some of this code. It's probably a simple little fix. All right let's just give this 1 last try here. And if it doesn't work it doesn't, we just won't do the...

Yeah let's just give up on a cell chat I'm sure we could figure it out pretty easily if we put some of our own human brain power into this but that's not the point instead let's just stop here. It did a good job with the basic pre-processing and this weird silhouette method that may or may not be the best for this kind of data but it did a really good job of labeling the cells this was probably my favorite and what I thought was the most impressive was a bit of a job it actually did with labeling and then you can find canonical markers very easily as well. So I'm sure you could ask it to do specific things like plot the go categories or the enrichment results etc but this is pretty impressive. So we did the basics of single cell sequencing without writing any code or using any of our biological knowledge. I'm not sure if it's at the point of taking my job yet, but it's certainly at the level where it will help people greatly when they're doing their own analysis.

See all Sanbomics transcripts on Youtube

Can chatGPT do single-cell bioinformatic analysis?