HD-AE (the Hilbert-Schmidt Deconfounded Autoencoder) is a package for producing generalizable (i.e., across labs, technologies, etc.) embedding models for scRNA-seq data. HD-AE enables the training of "reference" embedding models, that can later be used to embed data from future experiments into a common space without requiring any retraining of the model. Please see our preprint for further technical details.
- Train HD-AE models to embed scRNA-seq data from different sources into a common lower-dimensional space
- Share your pretrained models and reference sets of embeddings with collaborators
- Alternatively, you can download a pretrained HD-AE model and use it to embed your data and compare with previous reference datasets
HD-AE depends on the scanpy, PyTorch, and pytorch-lightning packages. HD-AE was originally built with scanpy v1.7.1, PyTorch v1.8.1, and pytorch-lightning v.1.2.7, though it should work with future versions of these packages barring any major changes.
The simplest way to get started with HD-AE is to clone this repo and run
pip install .
HD-AE and all dependencies should install in 5 minutes or less. HD-AE was originally designed on a workstation running CentOS 7.9, but should work with any operating system with a PyTorch implementation (i.e., Windows/Mac/Linux). When training new models, having a GPU will speed up the training process considerably, though a GPU is unnecessary if only using a pretrained model.
See the notebooks folder for examples of using HD-AE.
If you find HD-AE useful in your research, please consider citing our preprint.