• update the message passing optimizer

• Release metabolism-related network GGSL-V2
• Optimize the prediction only step.
• Optimize the training and testing step.

• Release the initial version and installation manual
• Release Antigen Presentation Pathway
• This version of the algorithm can handle cycles in the graph to some extent.

• Release the initial version and installation manual
• Release Antigen Presentation Pathway

If you have an interested topic, please feel free to open an issue or I can also merge your completed function into main branch

We developed a Message Passing Optimizer-Based Supervised Neural Network to Estimate the Cell-Wise Metabolic Using Single Cell RNA-seq Data. To infer single cell fluxome from single cell RNA-sequencing (scRNA-seq) data, Our framework is powered by three sub-algorithms:(1) scFEA(single cell Flux Estimation Analysis), a self-constrained Graph Neural Network to generate the initial flux values; (2) MPO(Message Passing Optimizer), a Belief Propagation-based message passing algorithm to optimize the initial flux values; (3) SNN(Supervised Neural Network), a neural network-based supervised learning algorithm to learn the predictor between the optimized flux values and the scRNA-seq data.

To benchmark the method, we applied it to the following transcriptomics data. We applied the method to the bulk tissue RNA-seq data of a melanoma data set (GSE91061) collected from patients under anti-CTLA4 and ant-PD1 therapy. In total, we obtained 105 samples from the GSE91061 data set, including 48 PR, 34 SD, and 23 PD patients. We applied the method to this data to compute the sample-wise activity level of the 11 modules. Biologically, we expect the higher level of antigen presentation activity to be associated with a better response. We observed that module 8 (trimming of peptides) and module 11 (T cell level) are significantly associated with responsiveness.

We further adopted stepwise multi-variate logistic regression to identify the top variables and best model in predicting patients’ response to anti-CTLA4 and ant-PD1 therapy. We also included the total T cell level and cytotoxic CD8+ T cell level predicted by deconvolution analysis and MSI/MSS status predicted by gene expression data. The final selected model is $$Response = M_8 + MSI status$$

, as detailed below. Our analysis suggested the activity level of trimming of peptides and MSI status are predictive of the outcome of immuno-therapy.

We further checked how M_8 level varies with respect to responsiveness, MSI status and treatment status. We identified that the level of M_8 is higher in RP and SD patients compared to PD patients in all groups. Interestingly, we found that the level of M_8 shows a significant difference between MSI and MSS patients only for the on-treatment group. Specifically, the PR and SD MSS on treatment patients have a significant increase of M_8 compared to (1) the PR patients MSS on treatment patients and (2) all PD and SD MSI on treatment patients. This observation suggests that increasing antigen presentation activity during anti-CTLA4/PD1 treatment may increase the response for MSS patients. This observation is explainable as the MSS patients who have less neoantigen may demand higher antigen presentations to enable a sufficient T cell recognition.

MPOSNN is implemented by Python3. If you don't have python, please download Anaconda with python 3 version.

• torch >= 1.13.1
• numpy >= 1.23.3
• pandas >= 1.4.4
• matplotlib >=3.6.2
• magic >= 2.0.4
• scikit-learn >= 1.1.1
• networkx >= 2.8.8
• pytorch-lightning >= 1.8.1

Download MPOSNN:

git clone https://github.com/ptdang1001/MPOSNN.git

Install requirements:

cd MPOSNN
conda install --file requirements
conda install pytorch torchvision -c pytorch
pip install --user magic-impute

You can see the input arguments for MPOSNN by help option:

python src/main.py --help
usage: main.py [-h] [--input_dir INPUT_DIR] [--output_dir OUTPUT_DIR] [--geneExpression_file_name GENEEXPRESSION_FILE_NAME] [--compounds_modules_file_name COMPOUNDS_MODULES_FILE_NAME] [--modules_genes_file_name MODULES_GENES_FILE_NAME] [--n_epoch_all N_EPOCH_ALL]
               [--imbalance_loss_limit_all IMBALANCE_LOSS_LIMIT_ALL] [--save_predictions SAVE_PREDICTIONS] [--pca_components_selection PCA_COMPONENTS_SELECTION] [--do_imputation DO_IMPUTATION] [--experiment_name EXPERIMENT_NAME] [--module_source MODULE_SOURCE]
               [--load_checkpoints_dir LOAD_CHECKPOINTS_DIR] [--load_weights_folder LOAD_WEIGHTS_FOLDER] [--n_epoch_scfea N_EPOCH_SCFEA] [--n_epoch_mpo N_EPOCH_MPO] [--n_epoch_snn N_EPOCH_SNN] [--do_train_snn DO_TRAIN_SNN] [--n_train_batch_snn N_TRAIN_BATCH_SNN]
               [--do_predict_snn DO_PREDICT_SNN] [--output_grad_snn OUTPUT_GRAD_SNN]

MPOSNN: A Massage Passing Optimizer-Based Supervised Neural Network Model to Estimate Cell-Wise Metabolic Using Single Cell RNA-seq Data.

options:
  -h, --help            show this help message and exit
  --input_dir INPUT_DIR
                        The inputs directory.
  --output_dir OUTPUT_DIR
                        The outputs directory, you can find all outputs in this directory.
  --geneExpression_file_name GENEEXPRESSION_FILE_NAME
                        The scRNA-seq file name.
  --compounds_modules_file_name COMPOUNDS_MODULES_FILE_NAME
                        The table describes relationship between compounds and modules. Each row is an intermediate metabolite and each column is metabolic module. For human model, please use cmMat_171.csv which is default. All candidate stoichiometry matrices are provided in /data/
                        folder.
  --modules_genes_file_name MODULES_GENES_FILE_NAME
                        The json file contains genes for each module. We provide human and mouse two models in scFEA.
  --n_epoch_all N_EPOCH_ALL
                        The user defined early stop Epoch(the whole framework)
  --imbalance_loss_limit_all IMBALANCE_LOSS_LIMIT_ALL
                        The user defined early stop imbalance loss.
  --save_predictions SAVE_PREDICTIONS
                        Save results. 0=False, 1=True
  --pca_components_selection PCA_COMPONENTS_SELECTION
                        Apply PCA to reduce the dimension of features. 0=False, 1=True
  --do_imputation DO_IMPUTATION
                        Imputation on the input gene expression matrix. 0=False, 1=True
  --experiment_name EXPERIMENT_NAME
  --module_source MODULE_SOURCE
  --load_checkpoints_dir LOAD_CHECKPOINTS_DIR
  --load_weights_folder LOAD_WEIGHTS_FOLDER
  --n_epoch_scfea N_EPOCH_SCFEA
                        User defined Epoch for scFEA training.
  --n_epoch_mpo N_EPOCH_MPO
                        User defined Epoch for Message Passing Optimizer.
  --n_epoch_snn N_EPOCH_SNN
                        User defined Epoch for Supervised Neural Network training.
  --do_train_snn DO_TRAIN_SNN
                        Train the SNN model, 0=False, 1=True.
  --n_train_batch_snn N_TRAIN_BATCH_SNN
  --do_predict_snn DO_PREDICT_SNN
                        Predict the flux values via the trained SNN model, 0=False, 1=True. FYI: If you have already trained the SNN model, SNN saves the model automatically, then you can set --do_train_snn 0 and --do_predict_snn 1 to predict the flux values directly.
  --output_grad_snn OUTPUT_GRAD_SNN
                        Save the gradients on each gene.

1. scRNA-seq data(rows:genes, columns:samples/cells)

2. Pathway data, Adjacency matrix, a factor graph

   • 1:="The parent node(reaction M_i) of a compound",
   • -1:="The child node(reaction M_i) of a compound",
   • 0:="No connection bewteen compound and reaction".

   Please see the graph bellow, you can find it in the directory "inputs":

   

   Please see the adjacency matrix bellow, you can find it in the directory "inputs":

   

3. Antigen Presentation Pathway reactions and the genes. Modules(Reactions) and contained genes, you can find it in the directory "inputs".

   Please see the json file sample bellow:

   

You can find the results in the directory "outputs", please click here to see the output examples(Just for testing, no mearning):

1. "flux_scfea.csv", flux values from scFEA, rows:samples, columns:modules, each entry is a flux value.

2. "flux_snn.csv", flux values from SNN, rows:samples, columns:modules, each entry is a flux value.

3. "flux_snn_grad.csv", the gradients, rows:genes, columns:samples, each value represents the partial derivative of the model with respect to the gene.

4. "Compounds_Modules_FactorGraph_original.png", the visiulization of Factor Graph.

5. "flux_scFEA_MPO_SNN_std_scale_imbalance.png", the analysis of predicted flux values.

   • 5.1. "module wise std":= $\frac{ \sum std(Y_{:,j}^{predicted})}{n}$
   • 5.2. "all mean scale":= $\frac{\sum \sum Y_{i,j}^{predicted}}{M*N}$
   • 5.3. "sample wise imbalance loss":= $\frac{\sum ImbalanceLoss(Y_{i,:})}{M}$

   

** Traning and Predicting example: **

# just copy your data into the directory "inputs"
# The algorithm saves the model weights automatically
# you can get the results in the directory "outputs"


#python src/main.py --geneExpression_file_name "your scRNA-seq data file name" --compounds_modules_file_name "your adj matrix file" --modules_genes_file_name "your modules and genes file" --module_source "your pathway name" --experiment_name "your experiment name"

python src/main.py --geneExpression_file_name TCGA.csv.gz --compounds_modules_file_name ANT2_compounds_modules.csv --modules_genes_file_name ANT2_modules_genes.json --module_source ANT2 --experiment_name FluxEstimation

** No Training, Predicting only example: **

# just copy your data into the directory "inputs"
# you have to train the model at least once befor this predicting only step. We don't offer default trained model. 
# you can get the results in the directory "outputs"

#python src/main.py --geneExpression_file_name "your scRNA-seq data file name" --modules_genes_file_name "your modules and genes file" --module_source "your pathway name" --experiment_name "your experiment name" --load_weights_folder "The folder name you got from the tranning step" --do_train_snn 0

python src/main.py --geneExpression_file_name TCGA.csv.gz --modules_genes_file_name ANT2_modules_genes.json --module_source ANT2 --experiment_name FluxEstimation --load_weights_folder TCGA_ANT2_FluxEstimation_00001 --do_train_snn 0

If you have any questions or problems, please feel free to open a new issue here. We will fix the new issue ASAP. For code questions, please contact Pengtao Dang.

• Pengtao Dang ([email protected])

For any other further questions or requests, please contact the Principle Investigator of BDRL lab.

• Prof. Chi Zhang ([email protected])

PhD candidate at Biomedical Data Research Lab (BDRL) , Indiana University School of Medicine

1. N. Alghamdi, W. Chang, P. Dang, X. Lu, C. Wan, Z. Huang, J. Wang, M. Fishel, S. Cao, C. Zhang. scFEA: A graph neural network model to estimate cell-wise metabolic using single cell RNA-seq data, under review at Genome Research, 2020.