Variational Inference Methods for Single-Cell Genomics

Demonstration code and notebooks for the paper Variational Inference Methods for Single-Cell Genomics.

This repository reproduces the simulations, single-cell probabilistic inference (scPI) examples, and temporal GP analyses.

Repository structure

01_simulation/
  01_LMM/
    LMM.py
    LMM_simulation.ipynb
  02_scLDA/
    LDA.py
    ScLdaSimData.py
    LDA_simulation.ipynb
    run_paper_cavi_svi_mse.py
    run_paper_pyro.py
  03_GLMM/
    GLMM.py
    GLMM_simulation.ipynb

02_scPI/
  FA.py
  ZIFA.py
  01_computational_time.ipynb
  02_performance_comparison.ipynb

03_TemporalGP/
  GP.py
  utils.py
  01_compare.ipynb
  02_leave_cohort.ipynb

Contents

01_simulation/

• 01_LMM/ — Linear mixed model estimators: EM, PX-EM, MM, and mean-field CAVI.
• 02_scLDA/ — Single-cell LDA with conjugate CAVI/SVI and black-box Pyro AutoNormal SVI.
• 03_GLMM/ — Grouped Bernoulli GLMM estimators: Laplace, PQL, and Pyro VI.

02_scPI/

• FA.py — Factor analysis with amortized VI or non-amortized VI (method="amortized" / "vi").
• ZIFA.py — Zero-inflated FA with classic EM, block EM, or Pyro VI (method="classic" / "block" / "pyro", plus amortized vs non-amortized inference).
• 01_computational_time.ipynb — Runtime benchmarks over cell/gene sizes using the mouse brain 10x matrix at datasets/mouse_brain/datasets/1M_neurons_filtered_gene_bc_matrices_h5.h5.
• 02_performance_comparison.ipynb — Cortex imputation and clustering comparison using expression_mRNA_17-Aug-2014.txt.

03_TemporalGP/

• GP.py — Temporal count models in Pyro: GP_MF, GP_Full-rank, and Indep_MF.
• utils.py — Age standardization and RBF temporal kernel helpers.
• 01_compare.ipynb — Fit the three models on Microglia from datasets/aging_svz_adata.h5ad across gene-panel sizes.
• 02_leave_cohort.ipynb — Leave-cohort experiment: hold out each cohort, then compare missing time-point estimates to full-data baselines.

How to run

1. Install the Python dependencies used by the notebooks you plan to run (numpy, scipy, pandas, matplotlib, and for Pyro-based sections also torch, pyro-ppl; TemporalGP / scPI notebooks additionally use anndata, h5py, and scikit-learn as needed).
2. Place required external datasets under the paths noted above.
3. Open and run the notebooks in order within each folder.