Interpretable Self-Supervised Prototype Learning for Single-Cell Transcriptomics

Learn cross-batch metacells using interpretable self-supervised prototype learning to denoise and preserve biological structure in single-cell data — without using labels.

📄 Read the paper (ICLR 2025 - LMRL Workshop)

Overview

Single-cell RNA-seq data is often noisy, sparse, and affected by batch effects, which can obscure meaningful biological insights.
scProto is an interpretable self-supervised prototype learning method that learns biologically meaningful prototypes and decodes them into metacells — compact, denoised representations of cell populations across batches.

• Learns cross-batch metacells that reflect biologically meaningful cell groups
• Trained to preserve biological structure and cell-cell relationships in the embedding space while mitigating batch effects
• Fully label-free, requiring no annotations

Key Features

• Interpretable prototype learning and metacell decoding across datasets
• Embedding space that maintains biological topology and local cell relationships
• Enhances single-cell analysis by denoising gene expression and overcoming data sparsity

Model Architecture

scProto builds upon the CVAE architecture from scPoli, designed for interpretable reconstruction of gene expression, and combines it with a self-supervised prototype learning strategy based on SwAV (Swapped Assignment between Views).

The model is trained end-to-end to learn:

1. Prototypes via SwAV-style self-supervised contrastive learning
2. Metacell reconstructions by decoding prototypes using the CVAE decoder
3. Coverage of rare cell types via a Propagation loss

This unified design allows the model to aggregate similar cells, preserve cell-cell structure, and denoise gene expression, all while being fully unsupervised.

Method

We use SwAV, a self-supervised contrastive clustering method, to learn prototypes that represent transcriptionally similar groups of cells.

Key Components:

• SwAV Loss (per-batch averaged)
  Prevents batch-specific prototype collapse by computing prototype assignments within each batch and averaging the loss across batches

• CVAE Decoder (from scPoli)
  Ensures that each prototype can be decoded into a metacell, preserving biologically relevant expression profiles while supporting interpretability

• Propagation Loss
  A min-max objective to ensure that rare cell types are assigned to at least one prototype

Together, these components optimize the composite loss:

$$ L_{\text{scProto}} = L_{\text{batchSwAV}} + \lambda_1 \cdot L_{\text{propagation}} + \lambda_2 \cdot L_{\text{CVAE}} $$

This enables scProto to:

• Learn interpretable cross-batch prototypes
• Preserve biological structure in the embedding space
• Denoise sparse gene expression
• Improve rare cell type representation