DRVI

DRVI#

DRVI [Moinfar and Theis, 2024] (Disentangled Representation Variational Inference, Python class DRVI) is an unsupervised deep generative model that learns an interpretable, disentangled latent representation of single-cell omics data.

The advantages of DRVI are:

  • Disentanglement: DRVI disentangled cell identity, signaling pathways, stress responses, developmental trajectories, and perturbation effects into distinct factors.

  • Interpretability: Each factor (latent dimension) can be linked back to genes via the built-in interpretability analyses.

  • Built on scVI: DRVI subclasses SCVI and only swaps the decoder, so it inherits the full SCVI interface — covariates, batch embedding (batch_representation="embedding"), size factors, minified mode, scArches transfer learning, out-of-core training via a datamodule, and the standard likelihoods (plus DRVI’s log-space pnb).

Method background#

DRVI is a variational autoencoder (VAE): the integrated, disentangled representation corresponds to the latent-space embedding of the cells. It shares the encoder, prior (N(0, I)) and reconstruction likelihoods with VAE. Disentanglement is induced in the decoder rather than through the prior.

Disentanglement via an additive split decoder#

The latent vector of dimension K is mapped into n_split_latent independent groups (“splits”). Each split is decoded independently and the per-split outputs are aggregated to form the final reconstruction parameters. Because every split can only influence the output through this shared aggregation, the model is encouraged to allocate independent factors of variation to different latent dimensions.

Two aggregations over the split dimension are available:

  • logsumexp (default): aggregation in log space, which makes the decoder additive in the rate.

  • mean: average of the per-split parameters.

Two latent-to-split mappings are available (n_latent should be divisible by n_split_latent):

  • split_map (default): the latent is reshaped into n_split_latent chunks and each chunk is passed through a learned per-split linear projection.

  • split_mask: latent is split into chunks and padded with zeros so, each split i keeps only its own chunk.

Likelihoods modeled in log space#

DRVI models the generative parameters in log space, which is what makes the logsumexp aggregation a genuine additive decoder (summing the per-split contributions of the rate). In addition to scvi’s nb, zinb and poisson, DRVI adds:

  • pnb — parametrized negative binomial: the same mean as nb (library * softmax) but parametrized and evaluated in log space via LogNegativeBinomial, which is numerically stable and composes the per-split log contributions directly.

  • normal / normal_unit_var — a Gaussian whose mean is modeled directly (no library/softmax), with the per-gene variance modeled in log space (fixed to one for normal_unit_var); intended for continuous, e.g. log-normalized, input.

Interpretability#

Because the decoder is additive over splits, the contribution of each split to the reconstructed expression of every gene can be measured directly. DRVI exposes per-split parameters in an inspection mode and provides methods to quantify these contributions both in-distribution (over the observed cells) and out-of-distribution (by traversing each latent dimension), and to extract the top genes per latent dimension. See set_latent_dimension_stats(), calculate_interpretability_scores() and get_interpretability_scores().

Scope: core model only#

Only the core DRVI model and its core interpretability methods are included in scvi-tools (the model, the additive decoder and log-sum-exp pooling, the log-space likelihoods, and the interpretability methods.).

The utility functions for latent-space visualization and analysis are not part of scvi-tools. For those, install the full drvi-py package:

pip install drvi-py