This repository implements an integrated framework for combining competing endogeneous RNA (ceRNA) network data with single-cell RNA sequencing (scRNA-seq) data. Our dual-view autoencoder leverages graph convolutional and attention mechanism to learn robust latent representaitons that capture both ceRNA regulatory interactions and cell-specific expression profiles.

The ceRNA network captures regulatory interactions among mRNAs, lncRNAs, and (optionally) circRNAs that compete for shared microRNAs (miRNAs). To build this network:

• Obtain Interaction Data

  • starBase – interactions among mRNAs, lncRNAs, and miRNAs (CLIP-Seq)
  • LncBase – lncRNA–miRNA interactions
  • miRTarBase – experimentally validated miRNA–mRNA interactions
  • miRcode – predicted miRNA binding sites on lncRNAs
  • (Optional) circBank / circInteractome – circRNA–miRNA interactions

• Filtering Interactions

  • Apply a confidence score threshold to remove low-confidence edges.

• Expression-Based Filtering

  • Remove any RNA nodes with zero expression across all cells in the scRNA-seq dataset.

• Start from a normalized scRNA-seq expression matrix.
• Filter to retain only those RNAs (mRNAs, lncRNAs, circRNAs) present in the ceRNA network.
• Result: every network node has an associated single-cell expression profile.

Usage

python preprocess_scRNA.py \
  --expr data/raw/scRNA_normalized.tsv \
  --network data/processed/ceRNA_interactions_merged.tsv \
  --out data/processed/scRNA_filtered.tsv

A dual-view encoder captures both the regulatory network structure and single-cell expression variability.

1. Build a cell–cell similarity graph (e.g., KNN) from expression profiles.
2. Apply one or more graph convolution layers to aggregate information among similar cells—this denoises and imputes missing values.

1. Treat the ceRNA network as a graph (nodes = RNAs, edges = regulatory interactions).
2. Apply graph convolution layers to aggregate regulatory signals across RNAs.

• Fuse the two views via a graph attention layer.
• Learn edge-importance weights; use a sigmoid activation (rather than node-wise softmax) to allow global thresholding.
• Dynamically prune weak/spurious edges, refining the network topology based on reconstruction performance.

• ceRNA Network Reconstruction
  Inner-product decoder on the latent RNA embeddings. Use contrastive loss comparing positive (observed) and negative (random) edges.

• Gene Expression Reconstruction
  Multi-layer fully connected decoder to reconstruct the scRNA-seq expression matrix.

• Combine ceRNA network reconstruction loss and gene expression reconstruction loss.
• Balance with hyperparameters so neither dominates.
• Include regularization terms (e.g., L2 penalties) to prevent overfitting.

• Periodically prune low-quality edges using learned attention coefficients.
• Ensures the final network reflects true biological interactions.

• For large scRNA-seq datasets, split the cell graph (or cells) into mini-batches.
• Reduces memory usage while learning a unified embedding space.

• Framework: PyTorch + torch-geometric
• Hardware: NVIDIA A100 GPUs (or equivalent)
• Optimizer: Adam (tune learning rate, weight decay)
• Monitoring: AUROC on both network and expression reconstruction; adjust hyperparameters via validation.

• ceRNA Interaction Databases
  starBase, LncBase, miRTarBase, miRcode
  (Optional) circBank / circInteractome

• scRNA-seq Data Repositories
  GEO (Gene Expression Omnibus)
  ArrayExpress
  Single Cell Portal (Broad Institute)

• Annotation Resources
  Ensembl or NCBI RefSeq for gene identifiers and metadata
  KEGG, Reactome for pathway and functional validation

This pipeline provides a complete workflow for integrating ceRNA regulatory networks with single-cell RNA-seq data using a dual-view graph neural network. It covers data acquisition and filtering, model architecture (encoder, attention, decoder), adaptive graph pruning, training strategies, and required resources. By jointly modeling post-transcriptional interactions and cell-specific expression, it uncovers cell-type–specific ceRNA mechanisms and enhances our understanding of RNA crosstalk in complex tissues.