Differential Expression & Splicing

SpliceVI supports both differential gene expression (DE) and differential splicing (DS) analysis between groups of cells. Both use the same underlying Bayesian framework from scvi-tools — they differ only in what quantity is being compared and how the effect size is defined. The API mirrors MultiVI, with differential_splicing and get_normalized_splicing replacing the chromatin accessibility equivalents.

For full background on the statistical framework, see the scvi-tools differential expression guide.

How it works

Both DE and DS use _de_core from scvi-tools internally. The key idea is:

Sample normalized values from the posterior for each cell in group 1 and group 2 separately, by passing cells through the encoder and then the relevant decoder (expression or splicing)
Compute a per-feature effect size from those posterior samples
Estimate the posterior probability that the effect size exceeds a threshold \(\delta\) (the "change" mode)

The per-feature Bayes factor is then:

\[ \text{BF} = \log \frac{P(|\Delta| > \delta \mid \text{data})}{P(|\Delta| \leq \delta \mid \text{data})} \]

An FDR-controlled call of differential features is made using a target FDR threshold (default 5%).

Differential Expression

DE compares normalized gene expression between two groups using get_normalized_expression (see Imputed Splicing & Expression).

The effect size follows the standard scVI convention — a log-fold change:

\[ \text{LFC} = \log_2(\hat{x}^{(2)}_g + \epsilon) - \log_2(\hat{x}^{(1)}_g + \epsilon) \]

where \(\hat{x}^{(k)}_g\) is the posterior mean normalized expression for gene \(g\) in group \(k\).

de_results = model.differential_expression(
    adata=mdata,
    groupby="cell_type",
    group1="Neuron",
    group2="Astrocyte",
    delta=0.25,
    fdr_target=0.05,
)

Differential Splicing

DS compares junction usage (PSI) between two groups. Because PSI is already on a \([0, 1]\) probability scale, the effect size is a direct difference rather than a log-fold change — analogous to how ATAC-seq accessibility scores are handled in scvi-tools:

\[ \text{effect size} = \hat{\psi}^{(2)}_j - \hat{\psi}^{(1)}_j \]

where \(\hat{\psi}^{(k)}_j\) is the posterior mean PSI for junction \(j\) in group \(k\), computed using the DM posterior mean by default (norm_splicing_function="dm_posterior_mean"). You can switch to the raw decoder output with norm_splicing_function="decoder".

ds_results = model.differential_splicing(
    adata=mdata,
    groupby="cell_type",
    group1="Neuron",
    group2="Astrocyte",
    delta=0.10,
    fdr_target=0.05,
    norm_splicing_function="dm_posterior_mean",   # recommended
)

Output columns

Column	Description
`proba_ds`	Posterior probability of differential splicing
`is_ds_fdr`	Boolean FDR-controlled call at `fdr_target`
`bayes_factor`	Log Bayes factor
`effect_size`	\(\hat{\psi}^{(2)} - \hat{\psi}^{(1)}\) (model posterior means)
`emp_effect`	Empirical \(\bar{\psi}^{(2)} - \bar{\psi}^{(1)}\) from observed PSI values
`est_prob1` / `est_prob2`	Model posterior mean PSI per group
`emp_prob1` / `emp_prob2`	Empirical mean PSI per group (observed cells only)
`n_obs_group1` / `n_obs_group2`	Number of cells with observed data per junction per group

Notes

Very sparse junctions (low n_obs_group1/2) will have wider posteriors — consider filtering before interpretation
Both methods support batch_correction=True to marginalize over batch effects when comparing groups