BioTorchby Chris Hayduk RAW PYTORCH · TENSOR PRIMITIVES

MODEL LEARNING PATH

scVI

scVI explains RNA counts using a latent cell state, sequencing depth and a count distribution. An inference network estimates latent variables; a decoder predicts gene-expression parameters for training and posterior analysis.

Loading the interactive page…

A landmark in count-aware variational modeling that separates cell state from library-size and batch effects. Its direct extensions include scANVI, and its probabilistic approach supports successors such as totalVI; this historical lineage, rather than universal benchmark superiority, motivates the featured page.

Chapters

  1. Infer latent variables

    Counts enter an encoder that estimates Gaussian cell-state and, in the original model, log-library posteriors.

    Distinguish inferred posterior parameters from a sampled cell state.

  2. Predict count distributions

    Latent state and batch context determine RNA composition, depth, dispersion and the zero-inflation mixture.

    Turn latent context into count means, dispersion and a zero-inflation mixture.

  3. Fit the variational model

    The objective combines count reconstruction with cell-state and library-prior KL terms.

    Combine reconstruction and prior regularization under a documented reduction.

  4. Analyze the posterior

    Posterior draws support predictive moments, expression comparisons and cell-state summaries.

    Use posterior draws to summarize uncertainty rather than a single point estimate.

All exercises in this guide