Skip to content

Latest commit

 

History

History
113 lines (90 loc) · 5.72 KB

README.md

File metadata and controls

113 lines (90 loc) · 5.72 KB
Raw

AlphaTalk


AlphaTalk incorporates protein structures to grain deeper cell-cell communication insights. The framework is composed of three modules, including

  1. GNN-based LRI predictor in gnn_predictor/
  2. LRI network-based algorithms in network_algorithms/
  3. Spatial transcriptomics filtering in st_filtering/

The first two modules use GNN methods and network-based algorithms to generate novel LRIs for Human and Mouse. The results are stored in lr_novel.csv and lr_novel_mouse.csv, both located in st_filtering/data/. ST filtering is employed to filter reliable novel LRIs based on two criteria: spatial proximity and the ability to activate downstream intracellular signaling transduction.

🚀 Dependencies

Clone the repository and create a new alphatalk environment.

git clone https://github.com/deepomicslab/AlphaTalk.git
cd AlphaTalk
conda env create --file alphatalk.yml
conda activate alphatalk
pip install transformers SentencePiece

🚀 GNN-based LRI predictor


GNN module


This module generates the interacting probability of all combinations of ligands and receptors in df_potential_lr.csv. The results are finally saved in network_algorithms/lr_database_raw.csv. Please refer to the instructions below to generate the results yourself or access the paper version of the results deposited on Google drive.

📁 Data details

  • 📁 data/
    • 📁 source/
      • human_v4/ Predicted human protein structures from AlphaFold DB.
      • human_v4.fasta Amino acid sequences of proteins in human_v4/.
      • human_v4_ProtT5/ Generated ProtT5 embeddings for proteins in human_v4.fasta.
    • 📁 df/
      • lrpairs.tsv LRDB records retrieved from SpaTalk.
      • df_potential_lr.csv Classified ligands and receptors in human_v4/ with "origin" columns indicating whether the protein is novel (not in LRDB). "c": LRDB; "u": novel.
      • df_human_dataset_nfold1.csv The LRI dataset with positive and negative samples.
    • 📁 graph/
      • human_v4_6_alpha.pt A list composed of pyg graph objects.

📄 Train the model and predict novel LRIs

  1. 📁 (Optional) Download predicted human protein structures from AlphaFold DB and extract mmCIF files to data/source/human_v4.
  2. 📁 (Optional) Run generate_data.py to transform protein structures in human_v4/ into pyg graph objects.
  3. 📦 Download the graph objects from Google drive and put it under data/graph/.
  4. ⚙ Run prott5_embedding.py to generate ProtT5 embeddings for proteins in data/source/human_v4.fasta.
  5. ⚙ Run train_gnn_pair.py to train the GNN model.
  6. ⚙ Run get_lr_database.py to Predict the interacting probability for all possible combinations of ligands and receptors.

🚀 Network-based algorithms


Network


This module uses the topological information in LRDB to filter the predicted LRIs in lr_database_raw.csv (deposited in Google drive). The results are saved in lr_novel.csv and lr_novel_mouse.csv under st_filtering/data/. The "type" column indicates whether the ligand, receptor, or both are novel.

  • ⚙ filter_novel.py Filter the predicted LRIs. The intermediate results are saved in result/.
  • 🧬 convert_gene_symbol.py Transform the human genes in the potential LRIs to their mouse orthologs.
  • 📊 analysis.ipynb Plot the results of filter_novel.py.

🚀 Spatial transcriptomics filtering


ST


This module uses single-cell spatial transcriptomics data to filter novel LRIs in lr_novel.csv and lr_novel_mouse.csv under st_filtering/data/.

Perform ST filtering

  • ⚙ ccc.py Perform the ST filtering with the following arguments:
    • st_file (str): Path of the single-cell ST data in AnnData format. The expected format of anndata is described in the notebook in analysis_script/. The example PDAC dataset is given, and others can be downloaded from Google drive.
    • species (str, optional): Species of the ST data; currently only support {'Human', 'Mouse'}. (default: 'Human')
    • tmp_dir (str, optional): Root directory of the results. (default: ./tmp/)
    • job_name (str, optional): Job name. (default: job1)
    • parallel (bool, optional): Whether to use multiprocessing. (default: False).
    • n_core (int, optional): The number of processes; only works when parallel is True. (default: 4)
    • filtering (str, optional): Method for calculating valid expression count threshold; choose one from {'median', 'mean', 'null'}. (default: 'mean')
    • dist_thd (bool, optional): Whether to use the 95% quantiles of distance to filter cell neighbors. (default: True)
    • lrdb (bool, optional): Whether to use LRIs from LRDB in conjunction with the novel LRIs. (default: True)
cd st_filtering
python ccc.py --st_file ./ST_dataset/pdac/pdac_sc_st.h5ad --job_name pdac --parallel
  • 📊 analysis_script/ Analyze the CCC results.