PathTIGR: A pathway topology-informed graph representation learning framework for immunotherapy response prediction

---

This repository contains source code and data for PathTIGR.

1. Introduction

PathTIGR is a Pathway Topology-Informed Graph Representation learning framework that systematically integrates biological pathway network topology knowledge with genome variation information for immunotherapy response prediction. PathTIGR employs a three-component design: (1) pathway graph encoder with multi-head attention embedding pathway topology knowledge and cancer genomic variants to pathway activity representation; (2) transformer module capturing pathway regulatory dependencies, (3) multilayer perceptron synthesizing pathway-level representations to predict immunotherapy response. This architecture enables PathTIGR to capture complex molecular interactions underlying immunotherapy response while maintaining biological interpretability.

2. Design of PathTIGR

Figure 1: Overall architecture of PathTIGR

3. Installation

• Python (version 3.9.21)
  • PathTIGR relies on Python (version 3.9.21) environments.
• Anaconda

  • Comprehensive installation instructions for Anaconda are available at: https://docs.conda.io/projects/conda/en/latest/user-guide/install/.

  • Following the installation of Anaconda, a virtual environment designated as PathTIGR can be created, and the requisite dependencies can be installed from the GAE_environment.yml configuration file by executing the following command:

  conda env create -f GAE_environment.yml
  

• PyTorch PyTorch requires a separate installation tailored to the specific hardware configuration. The appropriate installation command can be obtained from https://pytorch.org/.

4. Usage

This study trained PathTIGR models for different immunotherapy inhibitors separately. All the codes and data required to execute PathTIGR are provided in this GitHub repository. Please make sure to replace the input data path in the code with your own storage location.

4.1. Code

• The complete code for PathTIGR in the located at folder Code/, with the details of each file as follows:

File	Description
pathway_activity.ipynb	Train the autoencoder on pathway diagrams for generating pathway activity profiles.
predict.ipynb	Training the PathTIGR model for predicting the cancer immunotherapy response.

1. Users may reproduce the PathTIGR model by following the implementation provided in pathway_activity.ipynb, or retrain the model using custom datasets. The principal trainable parameters are specified as follows:

with open("../Data/pathways_adjacency_matrix_without_disease.pkl", "rb") as f:
    pathways_matrix = pickle.load(f)
with open("../Data/Liu/com_data_144.pkl", "rb") as f:
    com_data = pickle.load(f)

pathway_model, patient_feature = train_pathway_model(pathways_matrix = pathways_matrix, features_matrix_list = com_data, attn_type="graph", GAE_epochs = 30000, learning_rate = 0.01, num_heads = 1,
                                   ratio_val = 0, seed = 666, hidden1_dim = 3, hidden2_dim = 1, save_path = '/test_Data', patience = 20,device = devices)

Model Hyperparameters

-attn_type: Controls which type of attention module is used in the encoder to process node features. "self": uses self-attention; "graph": uses graph attention (str, default "self").

-GAE_epochs: The number of training iterations in the GAE model.

-learning_rate: The learning rate of training the the autoencoder on pathway diagrams model.

-num_heads: The number of attention heads.

-ratio_val: Validation set ratio.

-seed: The random seed.

-hidden1_dim: The dimension of the neurons in the first hidden layer.

-hidden2_dim: The dimension of the neurons in the second hidden layer.

-save_path: The model save path.

-patience: Early stopping patience value.

-device: The device where the model and tensors are located.

with open("../Data/Liu//patient_feature_144.pkl", "rb") as f:
    patient_feature = pickle.load(f)
with open("../Data/Liu//pathway_model_144.pkl", "rb") as f:
    pathway_model = pickle.load(f)

pathway_activity = calculate_pathway_activity_3d(pathway_model = pathway_model, patient_feature = patient_feature, latent_dim = 6)

with open(f'../Data/Liu//pathway_activity_144.pkl', "wb") as f:
    pickle.dump(pathway_activity, f, protocol=pickle.HIGHEST_PROTOCOL)

Model Hyperparameters

-latent_dim: The dimension of the neurons in the hidden layer.

2. Users may reproduce the PathTIGR model by following the implementation provided in predict.ipynb, or retrain the model using custom datasets. The principal trainable parameters are specified as follows:

with open("../Data/Liu/pathway_activity_144.pkl", "rb") as f:
    pathway_train = pickle.load(f)
with open("../Data/Liu/patient_response_144.pkl", "rb") as f:
    patient_response_train = pickle.load(f)

pathway_train = pathway_train.view(pathway_train.size(0), -1)
pathway_train = pd.DataFrame(pathway_train.detach().cpu().numpy())
pathway_train.index = patient_response_train.index

with open("../Data/Snyder/pathway_activity_60.pkl", "rb") as f:
    pathway_val = pickle.load(f)
with open("../Data/Snyder/patient_response_60.pkl", "rb") as f:
    patient_response_val = pickle.load(f)

seeds = np.random.randint(1000, 10000, size=10)
for seed in seeds:
    print(seed)
    torch.manual_seed(seed)                                        
    model=train_end_to_end_predict_model_nonEP(pathways_activity = pathway_train, patient_labels = patient_response_train, path_act_val = pathway_val,
                                                         patients_val = patient_response_val, batch_size = 8, num_heads = 1, global_epo = 300,
                                                         num_path = 230, save_best_model_path ='../Model', dropout = 0.3, seed = seed)

Model Hyperparameters

-batch_size: The number of patients for each batch.

-num_heads: The number of attention heads.

-global_epo: The number of total training iterations in the PathTIGR model.

-num_path: The number of pathways.

-save_best_model_path: The model save path.

-dropout: The dropout possibility for the PathTIGR model.

-seed: The random seed.

4.2. Data

• The datasets used to train PathTIGR are partly located in the folder Data/(The complete data can be found at https://zenodo.org/records/18512966):

File	Description
pathways_adjacency_matrix_without_disease.pkl	These biologically relevant pathways, which possess complete graph structure information, encompass important functional modules such as metabolic regulation, immune signaling, and cell cycle control.
Liu/pathway_activity_144.pkl	The sample-specific pathway activity profiles for the Liu Cohort.
Liu/patient_response_144.pkl	The patient's immune response label for the Liu Cohort.
Liu/patient_feature_144.pkl	The feature tensor was saved according to "Pathway → Patient" and aligned to each pathway diagram for the Liu Cohort.
Liu/com_data_144.pkl	The patient's genomic characteristics include mutations, copy number amplifications, and copy number deletions.
Liu/pathway_model_144.pkl	The predicted pathway activity models for the CTLA-4 patients.
Van_Allen/pathway_activity_110.pkl	The sample-specific pathway activity profiles for the Van Allen Cohort.
Van_Allen/patient_response_110.pkl	The patient's immune response label for the Van Allen Cohort.
Van_Allen/patient_feature_110.pkl	The feature tensor was saved according to "Pathway → Patient" and aligned to each pathway diagram for the Van Allen Cohort.
Van_Allen/com_data_110.pkl	The patient's genomic characteristics include mutations, copy number amplifications, and copy number deletions.
Van_Allen/pathway_model_110.pkl	The predicted pathway activity models for the PD-1 patients.
Hugo/pathway_activity_30.pkl	The sample-specific pathway activity profiles for the Hugo Cohort.
Hugo/patient_response_30.pkl	The patient's immune response label for the Hugo Cohort.
Hugo/patient_feature_30.pkl	The feature tensor was saved according to "Pathway → Patient" and aligned to each pathway diagram for the Hugo Cohort.
Snyder/pathway_activity_60.pkl	The sample-specific pathway activity profiles for the Snyder Cohort.
Snyder/patient_response_60.pkl	The patient's immune response label for the Snyder Cohort.
Snyder/patient_feature_60.pkl	The feature tensor was saved according to "Pathway → Patient" and aligned to each pathway diagram for the Snyder Cohort.
best_model_CTLA4.pt	The predicting CTLA-4 immune response models.
best_model_PD1.pt	The predicting PD-1 immune response models.

5. Interpretation of the PathTIGR model

To elucidate the biological mechanisms underlying PathTIGR’s predictions and identify key molecular determinants of immunotherapy response, we employed a dual attention-based importance scoring framework that integrates gene-level and pathway-level interpretability.