Deep learning annotation of cell-types with permutation inforced autoencoder
Project description
scMusketeers
Deep learning annotation of cell-types with permutation inforced autoencoder
Summary
Single cell gene expression atlases are now central to explore the cellular diversity arising at the scale of organisms or organs. The emergence of ever larger datasets are benefiting from the rapid development of deep learning technologies in the field. The constitution of large datasets raises several big challenges due to the presence of highly imbalanced cell types or the existence of large batch effects, which need to be adressed in order to annotate properly newer data derived from very small subsets, transfer a model from one dataset to another.
We developed scPermut to learn an optimal dimension-reduced representation, while preserving the information essential to meeting the above-mentioned challenges. The architecture of scPermut is made of three modules. The first module is an autoencoder which provides a reduced representation, while removing noise, and which allows a better data reconstruction. A classifier module with its focal loss can be combined to predict more accurately small cell types. This second module also supports transferring the learnt model to other datasets. The third module is an adversarial domain adaptation (DANN) module that corrects batch effect.
We extensively optimized scPermut hyperparameters, by conducting a precise ablation study to assess model's performance. We show that our model is at least on par with State-Of-The-Art models, and even outperforms them on most challenges. This was more thoroughly documented by comparing the different approaches in 12 datasets that differ in size, number of cell types, number or distinct experimental modes.
We anticipate that the generic modular framework that we provide can be easily adaptable to other fields of large-scale biology.
Tutorial
Access to the tutorial on Google collab
We will see in this tutorial two use-cases:
- Transfer cell annotation to unlabeled cells
- Transfer cell annotation and reduce batch from a query atlas to a reference atlas
Install
You can install sc_permut with Pypi:
$ pip install sc-musketeers
with conda
$ conda -c bioconda sc-musketeers
with docker
Examples
sc-musketeers can be used for different task in integration and annotation of single-cell atlas.
Here are 4 different examples:
- Label transfer between batch
$ sc-musketeers transfer my_atlas --class_key celltype --batch_key donor
TO DO : Add example atlas in the github or Zenodo
Read the CONTRIBUTING.md file.
Project details
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
File details
Details for the file sc_musketeers-0.1.8.tar.gz
.
File metadata
- Download URL: sc_musketeers-0.1.8.tar.gz
- Upload date:
- Size: 94.5 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/5.1.1 CPython/3.12.5
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | ede7a5657da6bd645218e2ab093392b0ed5280ce3e7466f60c04f1802cba498c |
|
MD5 | af1d757584907883e5b92a54abab5119 |
|
BLAKE2b-256 | 9fabe8614b877e6fd4c6a14c843d4b5461ce96aad7bf13ae28befbb1f1eaad85 |
File details
Details for the file sc_musketeers-0.1.8-py3-none-any.whl
.
File metadata
- Download URL: sc_musketeers-0.1.8-py3-none-any.whl
- Upload date:
- Size: 105.7 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/5.1.1 CPython/3.12.5
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 4e84320f807accd54be3d945a6988449ba051cc40cf415f09b4a3adf42a2c694 |
|
MD5 | 4c8cdf09358835d953fa6cb0b0663752 |
|
BLAKE2b-256 | 52d85879873a54e620d1bd57fba54c37b1ca4b96bb4e1ecdfed96a585f36a23c |