Applying random forest classifiers to single-cell RNAseq data
HTML-код
- Опубликовано: 11 июл 2024
- Learn how to apply machine learning to single-cell data. Random forest is a powerful machine learning classifier and a great tool for analyzing single-cell RNAseq data. In addition to predicting classifications, you can extract the gene importance from the model as a way to identify genes that describe your populations. Here I use several examples to show you how to use the random forest model in Python to do single-cell analyses.
Notebook:
github.com/mousepixels/sanbom...
Reference:
www.nature.com/articles/s4158...
tabula-sapiens-portal.ds.czbi...
0:00 - Intro
1:10 - Basic RF usage
5:05 - Classifying cells in other data
12:15 - Classifying cells in same data Наука
Thanks for your helpful and technological video! And looking forward to some videos about scATAC seq.
Sometime in the future! I have a few more planned before that and not enough free time. But one day!
Love your work..
Thank you!
Thank you so much!
You're welcome!
Cool video again. Would you be able to make a video on neural network applied to scRNAseq
Thank you so much, amazing video. Can you please tell me where can I get this kind of dataset to try this?
Any single-cell paper should have a data availability or equivalent section that contains links to the raw data or counts tables. Or you can search something like NCBI geo directly. Or you can look at the list of publications on the 10x genomics website.
Amazing video! Are there any existing pre-trained models which we can directly use to auto-annotate cell types given cell clusters?
Not sure about pre-trained models. There are simple models like SingleR or CellTypist. But if you have a reference dataset you can train a model with SCANVI. I have a video on that
@@sanbomics Thanks!
This is amazing! So helpful! I'm looking at applying some of these to publicly available data. How would this workflow change for k-nearest neighbour classification ? What would one need to change to do this ?
I haven't tried KNN for classification in single-cell, but neighborhood graphs are used all the time for unsupervised sc clustering. I'm not sure how well KNN would work without dimension reduction first but you could definitely try it. But dim reduction, like PCA, will require processing of your train/test together. Maybe there is a better way to do dim reduction but keep the train/test independent. RF is pretty flexible with the number of features. TLDR, I don't know, you should try it with only the variable features and see how accurate it is. Please let me know because I am curious!
@@sanbomics I tried it how can I send you the code ? Trying to do a ROC curve with it as well but the kernel keeps dying (even when I'm running it on the cluster).
You can upload it to a public github repository. Were you able to fix it? (sorry i just saw this, I don't get notifications for responses to my response)
That was cool. Do you use your PC to run ML tasks, or are you using HPC systems?
Usually just my PC, but sometimes an AWS EC2 with Nvidia GPUs. Simple models like RF don't take much processing power at all. My PC is decently beefy too with Nvidia gpu, 128 gb memory, 24 cpu
🌹 【promosm】