This repository includes our implementations for the final project of the course Deep learning in iomedicine , EPFL.
Group 8 : Hanrong Hu, Nearchos Potamitis, Karim Abi Said.
Our report can be found here.
In this report, we evaluate several few-shot learning models on a newly compiled ATAC-seq dataset of human cells (Zhang et al., 2021), encompassing fetal and adult samples. The dataset includes 30 different tissues from adult samples with 111 cell types, and 16 tissues from fetal samples with the same number of cell types. Overall, over 1.3 million single nuclei were sequenced, identifying approximately 1.2 million candidate CREs. The candidate CREs were classified in to promoter (-200 to 200 bp to transciption starting site), promoter proximal and distal.
🔴 We recommand running the exploration notebook with >32G memory🔴
We explore the dataset in Anndata format by:
- Adding the aligned metadata information.
- Exploring the number of cells, tissues and candidate cis-regulartor elements in adult and fetal samples.
- Exploring the spilts of tissues and the resulting cell type distribution.
There are 2 options to acquire the necessary data:
Download the following files (by clicking on the links) and move them to DLBIO/fewshotbench_v2/data/atacseq/.
If you're already using a google cloud VM instance then the preferrable way would be to directly download the data from our Cloud Storage Bucket to your instance. To do that SSH into your instance and run the following commands in the terminal:
gcloud storage cp gs://atacseq-np/data/atacseq/Cell_metadata.tsv.gz DLBIO/fewshotbench_v2/data/atacseq/
gcloud storage cp gs://atacseq-np/data/atacseq/cCRE_hg38.tsv.gz DLBIO/fewshotbench_v2/data/atacseq/
gcloud storage cp gs://atacseq-np/data/atacseq/matrix.h5ad DLBIO/fewshotbench_v2/data/atacseq/
Our work has been fully integrated with the given "Few Shot Benchmark for Biomedical Datasets" (fewshotbench_v2). If you're not familiar with that take a look at its readme file and make sure to create and activate the conda environment mentioned. Otherwise, the dataset of the main.yaml file has already been set to atacseq and as soon as the data is downloaded the code is ready to be tested! For example you can test the MAML method simply by running the following command in the terminal:
python fewshotbench_v2/run.py exp.name=test_grou_8 method=maml dataset=atacseq
We implemented several data subsets and splits options in the atacseq.yaml file. To test on your sub-dataset of interest:
python run.py experiment_subset.life_stage=Fetal experiment_subset.feature_class=Promoter experiment_subset.subset_fraction=0.1