Skip to content

Getting Started

Installation

1. Create a conda environment

conda create -n splicevi-env python=3.12
conda activate splicevi-env

2. Clone and install

git clone https://github.com/daklab/SpliceVI.git
cd SpliceVI
pip install -e .

This installs all dependencies from pyproject.toml. After installation, from splicevi import SPLICEVI works from any script or notebook.

Data format

SpliceVI expects a MuData object with two modalities:

  • rna — raw gene expression counts (cells × genes)
  • splice — junction usage ratios (PSI values) (cells × junctions)

The splicing modality also requires:

  • junc_counts — integer junction read counts (cells × junctions)
  • atse_counts — total read counts at the ATSE level (cells × junctions)
  • psi_observed_mask — binary mask indicating which junctions are observed per cell

See the SplicingDataset documentation in LeafletFA for utilities to construct this format from raw data.

Training

Using the Python API

import mudata
from splicevi import SPLICEVI

mdata = mudata.read_h5mu("train_data.h5mu")

SPLICEVI.setup_mudata(
    mdata,
    rna_layer=None,           # use .X for raw counts
    batch_key="mouse.id",
)

model = SPLICEVI(
    mdata,
    n_latent=30,
    splicing_loss_type="dirichlet_multinomial",
    splicing_encoder_architecture="partial",
    modality_weights="equal",
)

model.train(
    max_epochs=800,
    batch_size=256,
    n_epochs_kl_warmup=200,
    lr=1e-4,
)

model.save("models/my_run/")

Using the SLURM script

Configure paths and hyperparameters at the top of train_splicevi.sh, then:

sbatch train_splicevi.sh

Extracting outputs

model = SPLICEVI.load("models/my_run/", adata=mdata)

# Joint latent representation
latent = model.get_latent_representation()          # (cells × n_latent)

# Modality-specific latents
z_expr = model.get_latent_representation(modality="expression")
z_spl  = model.get_latent_representation(modality="splicing")

# Imputed PSI values
psi = model.get_normalized_splicing()               # (cells × junctions)