Welcome to SCRIBE’s documentation!

Warning

Project Status and Usage Restrictions

This project is currently in a pre-release state and is made available for viewing and evaluation purposes only.

At this time:

  • You may view and evaluate the code and documentation

  • You may not use, copy, modify, or distribute any part of this software

  • You may not incorporate this code into other projects

The software will be released under the MIT License following the publication of the associated pre-print. Until then, all rights are reserved.

SCRIBE (Single-Cell RNA-Seq Inference using Bayesian Estimation) is a Python package for analyzing single-cell RNA sequencing (scRNA-seq) data using variational inference based on Numpyro—a Jax-based probabilistic programming library with GPU acceleration. It provides a collection of probabilistic models and inference tools specifically designed for scRNA-seq count data.

Features

  • Multiple probabilistic models for scRNA-seq data analysis

  • Efficient variational inference using JAX and Numpyro

  • Support for both full-batch and mini-batch inference for large-scale data

  • Integration with AnnData objects

  • Comprehensive visualization tools for posterior analysis

  • GPU acceleration support

Available Models

SCRIBE includes several probabilistic models for scRNA-seq data, all documented in detail in SCRIBE Models for Single-Cell RNA Sequencing:

  1. Core Model: Negative Binomial-Dirichlet Multinomial (NBDM)

    • Models both count magnitudes and proportions

    • Accounts for overdispersion in count data

    • Forms the foundation of SCRIBE’s modeling approach

  2. Zero-Inflated Negative Binomial (ZINB)

    • Handles excess zeros in scRNA-seq data

    • Models technical and biological dropouts

    • Includes gene-specific dropout rates

  3. Negative Binomial with Variable Capture Probability (NBVCP)

    • Accounts for cell-specific mRNA capture efficiency

    • Models technical variation in library preparation

    • Suitable for datasets with varying sequencing depths per cell

  4. Zero-Inflated Negative Binomial with Variable Capture Probability (ZINBVCP)

    • Combines zero-inflation and variable capture probability

    • Most comprehensive model for technical variation

    • Handles both dropouts and capture efficiency

All these models can be extended to mixture variants as documented in Mixture Models to account for heterogeneous cell populations.

Indices and tables