please feel free to add/edit
Model for predicting/screening small molecules with ability to achieve specific gene expression profile
(2020.04 - 2020.05)
Performance:
- model selection, and biological meaning interperation;
(2020.01.19 - 2020.02)
ref: https://github.com/google-research/disentanglement_lib
- models: β-VAE, FactorVAE, β-TCVAE, DIP-VAE
- metrics: BetaVAE score, FactorVAE score, Mutual Information Gap(MIG), DCI disentanglement
- dataset: L1000 gene expression subset.
(2020.02 - 2020.03)
model conducted: beta-VAE, AnnealedVAE, beta-TCVAE, FactorVAE, DIP-VAE;
metrics used:
- Independence between the latent variables - total correlation/correlation between variables;
- capacity of latent variable - KL(q(z_i|x)||p(z_i));
- Mutual information between the latent variables and the data variable;
Done by Pengcheng.
Dataset: to check performance on test set 1) from the same distribution with training set; 2) novel cell type/small molecule;
loss: E(logp(yi,xi,zi)|zx,zd) = E(logp(yi,xi,zi|yi^,xi^,zi^)) ~ E(MSE). concat or not before MSE? => equivalent
Evaluation: metrics design
maybe worth trying: Does disentanglement performance correlated with our prediction result?