Second Day doing Bioinformatics
Today I wanted to start with the next phase of the project: Testing how well the model traiend with the bone marrow-optimized hyperparameters performs on predicting the cell types based on the gene expression (mRNA) data.
But then I learnt that the specific type of data that was used for finetuning the model, called CITE-seq, is composed of two things: the gene expression data (also called transcriptome) and protein expression data.
CITE stands for Cellular Indexing of Transcriptomes and Epitopes by Sequencing and it uses antibodies to bind specfic proteins on the surface of cells ( --> Epitopes) and then the antibody--surface-protein--cell complex get encapsulated, the cells get lysed and the RNA sequenced (--> Transcriptome)
The Epitope and the Transcriptome can also be indexed by the specific cell they belong to because during the encapsulation process each capsule gets a certain tag.
Usually antibodies that are used in lab environments are corresponding to well known immunce cell surface proteins (CD4+, CD8+ etc.), so CITE-seq data then accordingly is mostly available only for immune related cells.
So the original plan, evaluating scGPT on heart cells, brain cells etc. is no longer possible because there is not the specific CITE-seq data available.
But Claude found another public dataset of lung tumor cells with CITE-seq data that for which i kicked off the experiments today and will report the results tomorrow.
I also learnt a lot today about the aforementioned CITE-seq, the specific metrics used to evaluate the model, the UMAP diagram and also 10x Genomics, a company that is active in the sequencing business nad publishes datasets.
Yeah, thats it, looking forward to continung tomorrow