The scVI Generative Model From Counts to Latents

From a Cell’s Latents to Its Counts

scVI describes a cell’s observed gene-count vector as a random outcome generated from hidden cell-specific quantities. The model separates two sources of variation that are easy to conflate: a cell’s relative gene-expression profile and its overall library-size scale. A latent vector \(z_n\) represents cell state in a low-dimensional space, while a positive variable \(l_n\) represents the scale of counts for cell \(n\). The decoder combines the latent state with any recorded batch covariate to produce gene proportions; the library size scales those proportions into expected counts.

The model is a probability distribution, not a deterministic transformation from counts to a unique biological state. Its generative direction begins with latent variables and produces counts. Inference runs in the opposite direction approximately: an encoder uses observed counts and covariates to estimate distributions over plausible latent states and library sizes.

In the standard scVI count model, the latent state has a standard multivariate normal prior. The library size is commonly modeled with a log-normal prior whose parameters may depend on batch. Conditional on the latent quantities and batch, gene counts are modeled with negative-binomial distributions. A gene-specific dispersion parameter \(\theta_g>0\) controls overdispersion.

Here \(f_w\) is a learned decoder network with parameters \(w\). The softmax applied to its \(G\) outputs guarantees that the proportions are nonnegative and sum to one. The negative-binomial parameterization used here has mean \(\mu\) and dispersion \(\theta\), so its variance is \(\mu+\mu^2/\theta\). In this formulation, \(\theta_g\) is gene-specific and shared across cells; other model variants can use different dispersion structures.

The corresponding conditional factorization is:

The product states that counts for different genes are conditionally independent given the latent variables and batch. It does not assert that gene counts are marginally independent: integrating over shared latent quantities can induce dependence. Batch is conditioned on as known information here; the model does not generate the batch label.

What the Decoder Reconstructs

For a given \(z_n\), batch \(s_n\), and decoder output \(\rho_n\), the expected count for gene \(g\), conditional on \(l_n\), is \(l_n\rho_{ng}\). The proportions encode relative allocation across genes, whereas \(l_n\) sets the overall scale. Since the proportions sum to one, the conditional expected total across genes equals the library-size scale:

This is a statement about the conditional expectation, not a constraint that the sampled counts must add exactly to \(l_n\). Negative-binomial sampling allows the observed total to differ from that scale. Also, because \(l_n\) is uncertain during inference, the posterior expected count is \(E[l_n\rho_{ng}\mid x_n,s_n]\). In general this is not equal to the product of the separate posterior means of \(l_n\) and \(\rho_{ng}\), because the two quantities can be dependent under the posterior.

Inference and Reconstruction

The generative factorization specifies how counts arise, but observed data do not reveal \(z_n\) or \(l_n\) directly. scVI uses an encoder to approximate their posterior distribution. In a common formulation, the approximate posterior is factorized as \(q(z_n\mid x_n,s_n)q(l_n\mid x_n,s_n)\). The encoder’s outputs parameterize these distributions; drawing from them gives latent values that can be passed through the decoder. The approximation is a computational device for inference and does not change the generative factorization above.

A reconstruction can mean different things, so its definition should be stated. For fixed latent values, the decoder’s conditional mean is \(l_n\rho_{ng}\). For a reconstruction averaged over uncertainty in the latent variables, the target is the posterior expectation \(E[l_n\rho_{ng}\mid x_n,s_n]\), approximated by averaging decoded conditional means over draws from the encoder’s approximate posterior. This posterior mean is not itself a newly sampled count vector.

Batch and Library-Size Variation

The batch label can enter two places in this model. It can affect the library-size prior through \(\mu_{s_n}\) and \(\sigma_{s_n}\), and it can be supplied to the decoder \(f_w(z_n,s_n)\). These roles are distinct: the prior allows library scales to have different distributions across batches, while the decoder allows the predicted relative gene proportions to depend on batch. Including a covariate does not by itself guarantee that every technical batch effect has been removed or that biological differences are preserved.

Assumptions, Alternatives, and Failure Modes

The factorization relies on several assumptions that matter for interpretation and computation:

• Conditional gene independence: given the latent state, library scale, and batch, the likelihood factors over genes. Residual dependence not represented by these shared quantities is not modeled directly.
• Negative-binomial count variation: the likelihood accommodates variance above the mean through \(\theta_g\). Its suitability depends on whether this count distribution adequately describes the data; the model does not make that true by definition.
• Simplex decoder output: the proportions are nonnegative and sum to one, so the conditional expected total is determined by \(l_n\). The model allocates that scale across genes rather than assigning an unconstrained mean to each gene.
• Prior and covariate specification: the Gaussian prior on \(z_n\), log-normal library prior, and use of batch labels are modeling choices. Incorrect or incomplete covariates can leave systematic structure unexplained.

Alternatives change one or more of these assumptions. A Poisson likelihood fixes the conditional variance equal to the mean, whereas the negative binomial permits overdispersion. A zero-inflated negative-binomial likelihood adds a separate zero-generating component; it is not required simply because the count matrix contains many zeros. An observed-library-size-offset formulation conditions on a recorded total rather than placing a probability model on \(l_n\). That can be a different modeling choice, but it does not provide the same latent library-size distribution.

Useful structural checks are available before interpreting fitted latents. Verify that decoder proportions are finite, nonnegative, and sum to one; verify that the count mean is \(l_n\rho_{ng}\) under the chosen parameterization; and check that each cell’s batch label is correctly aligned with its counts. Extreme inferred library scales, highly concentrated proportions for nearly every cell, or persistent discrepancies between observed and model-implied count patterns can indicate misspecification, optimization problems, or data-processing errors. These checks flag concerns; none alone establishes a unique cause.

• Pitfall: treating \(l_n\) as the observed total. The reasoning “the proportions sum to one, so the counts must sum to \(l_n\)” confuses an expected value with a sampled realization. The correction is \(\sum_g E[x_{ng}\mid z_n,l_n,s_n]=l_n\); the realized sum is random.
• Pitfall: calling decoder output a count reconstruction without qualification. The vector \(\rho_n\) is a vector of proportions, not counts. Multiply by \(l_n\) for conditional mean counts, and average over posterior draws if reporting a posterior mean reconstruction.
• Pitfall: treating batch conditioning as proof of correction. A batch variable in the model allows batch-dependent distributions; it does not guarantee removal of unwanted effects or prove that a latent coordinate has a particular biological meaning.

The essential distinction is between the generative model and its inferred summaries. The model states how a latent cell state and library scale produce a distribution over counts. The encoder and decoder provide an approximate route from observed counts back to latent distributions and count-scale summaries, with uncertainty remaining at each stage.