Team: Amateur
Team members: Qiao Liu, Wanwen Zeng, Chencheng Xu
Project URL: https://github.com/kimmo1019/JAE
In brief, we built an autoencoder for joint embedding (JAE). Each modality will first be SVD transformed and concatenated together (denoted as x). The major difference from standard AE is that we incorporated the information from cell annotations (e.g., cell label, cell cycle score, and cell batch) to constrain the structure of latent features. We desire that some latent features (c) predict the cell type information, some features predict the cell cycle score. Noticeably, for feature (b), we want it to predict the batch label as randomly as possible to potentially eliminate the batch effect. z has no constrain at all to ensure the flexibility of neural network.
In the pretrain stage, JAE was trained with exploration data where the cell annotation information (cell type, cell cycle phase score) is available. In the test stage where the cell annotation information is not available, we only minimize the reconstruction loss of the autoencoder with a smaller learning rate (fine-tune).
Feel free to contact liuqiao@stanford.edu if you have any problem in the JAE model.