Back to Models

Cell BLAST

Generative Models

Cell Querying via Neural Embedding

RNA

Neural network VAE method for cell searching and annotation via unbiased cell embedding

Publications

Searching large-scale scRNA-seq databases via unbiased cell embedding

Cao et al.2020
Complexity
moderate
Interpretability
medium
Architecture
VAE with Search
Latent Dim
10
Used in LAIOR Framework

Cell-to-Cell Similarity Search

Cell BLAST learns unbiased cell embeddings via VAE that enable accurate cell-to-cell similarity search across datasets

Main Idea

Enable accurate cell annotation by learning batch-corrected embeddings for similarity search with reconstruction

Key Components

VAE Encoder

Learns cell embeddings from expression

Batch Correction

Removes multi-level batch effects

Decoder

Reconstructs expression with uncertainty

Similarity Metric

Computes cell-to-cell similarity with uncertainty

Mathematical Formulation

L = E[log p(x|z)] - KL(q(z|x)||p(z)); Similarity based on probabilistic embeddings

Loss Functions

ELBO
Reconstruction + KL divergence

Data Flow

Query Cells → VAE Encoder → Batch Correction → Decoder → Similarity Search → Reference Matches

Architecture Details

Architecture Type

VAE with Cell Similarity Module

Input/Output Types

single-cellreconstruction

Key Layers

EncoderBatchCorrectorDecoderSimilarityLayer

Frameworks

PyTorch

Tags

vaesearchannotationbatch-correctionrna