Interpreting the probabilistic model

Reviewed against scvi-tools 1.5.1. These are interpretation rules, not guarantees that a particular trained model is scientifically adequate.

Likelihood and variational inference

For latent state z, observed data x and technical covariates s, an encoder approximates q(z | x, s) and the decoder models p(x | z, s). The simplified objective is

ELBO = E_q[log p(x | z, s)] - KL(q(z | x, s) || p(z))

Additional latent variables and losses depend on the model. The neural-network weights are usually point estimates, so posterior draws do not capture every source of parameter, preprocessing, reference or sampling uncertainty. Optimizing this nonconvex objective does not guarantee a global optimum or calibrated biology.

SCVI commonly uses negative binomial counts with Var(x)=mu+mu^2/theta, or a zero-inflated NB mixture. A zero-inflation parameter is a mixture probability; it cannot by itself identify technical dropout versus biological absence. High sparsity does not prove that ZINB is preferable to NB. Compare suitable likelihoods using held-out fit, posterior predictive checks and domain knowledge. AUTOZI tests evidence for additional zero inflation under its assumptions; it does not label every zero's biological origin.

Do not generalize the count likelihood to all scvi-tools models. PeakVI uses Bernoulli accessibility; methylation models use methylated/coverage counts; SysVI and CytoVI operate on appropriately transformed continuous values; VeloVI uses preprocessed spliced/unspliced abundances and kinetic assumptions.

Integration and identifiability

Conditioning the decoder on batch encourages removal of nuisance variation from z; it does not mathematically force a batch-invariant latent space. Complete batch/condition confounding prevents identification of their separate effects. Validate both mixing among comparable cell states and preservation of cell-type, condition, donor and trajectory structure that should remain.

Covariates are model inputs, not a list of everything available in .obs. Library-size handling is already part of RNA models. Additional depth, tissue, condition or donor covariates can change the estimand or remove wanted signal. Continuous covariates must have appropriate scale; registration does not promise automatic standardization.

transform_batch decodes a fitted cell state under a different registered batch. This is model-based counterfactual decoding. It is not an identified causal intervention, nor evidence that the model predicts unobserved drug responses.

Transfer, samples and spatial structure

scArches performs model-specific architectural surgery with selective parameter freezing; it is not simply "freeze the encoder, train the decoder." Use prepare_query_anndata/load_query_data and inspect their freeze options. Query feature order, count scale, covariates and reference coverage matter. A classifier can confidently mislabel a cell type missing from its reference; evaluate rejection and uncertainty with held-out biological samples.

MRVI uses sample-unaware u and sample-aware z representations with learned sample-conditioned transformations. It is not defined by the generic additive formula z_shared + z_sample or a simple normal random-intercept hierarchy. Sample distances and multivariate comparisons depend on sample support and design.

DestVI deconvolves spatial expression with a labeled CondSCVI reference and continuous within-type states. Spatial coordinates do not automatically imply a spatial autocorrelation prior; validate proportions against independent tissue information. Tangram mapping is an optimization over cell-to-location assignment weights, not proof of the physical location of an individual sequenced cell.

Posterior contrasts

For RNA change-mode DE, positive LFC means group1 over group2. The returned bayes_factor is on the natural-log posterior-odds scale. Posterior FDP selection is conditional on the fitted model and hypotheses; it is not a conversion to frequentist p-values. More posterior samples reduce simulation noise, not biological uncertainty from a small number of donors. See differential-expression.md for exact API choices.

The logged elbo_validation training metric is a loss (lower is better), while get_elbo() returns an ELBO (higher is better). Compare the same metric on the same held-out cells/features; a training curve alone does not establish convergence, identifiability, generalization or successful integration.

Sources: variational inference, SCVI, AUTOZI, SysVI assumptions, MrVI, DestVI, transfer implementation.