scvi-linearscvi

To extract interpretable gene loadings from scRNA-seq counts, train a linear decoder model using raw counts to reveal gene-to-factor relationships. This approach directly maps latent dimensions to biological gene programs without requiring log transformation.

No, scvi-tools LinearSCVI for factor analysis requires raw scRNA-seq counts without log transformation. Using log-transformed data disrupts the count-aware modeling logic needed to accurately train the linear decoder and derive gene loadings.

The best way to identify drivers of population structure in single-cell datasets is using interpretable factor analysis on raw counts. This method trains jointly across cells to derive latent factors and gene-signature lists that explain biological variation.

LinearSCVI compares favorably to PCA for discovering gene programs by offering count-aware modeling and batch-friendly training. Unlike standard PCA, it uses a linear decoder on raw counts to provide interpretable latent factors directly linked to gene loadings.

Yes, you need to configure the number of latent factors before training the interpretable scRNA-seq model. Initializing an AnnData object with raw counts and setting the latent factor count are required steps before model training and loadings retrieval.

Interpretable factor analysis requires raw counts because the underlying linear decoder model relies on count-aware modeling to accurately capture gene-to-factor loadings. Normalized or log-transformed data violates the distributional assumptions needed for joint training across cells.