README.md

@@ -84,6 +84,8 @@ To get other biomedical knowledge graphs for comparation, plz refer to [HetioNet
 3/20      Upload data and model weight for "PertKGE identified five new scaffold hits for ALDH1B1"
 4/24      Upload src code and data.
 ```
+## Cite
+Shengkun Ni, Xiangtai Kong, Yingying Zhang, et al. Identifying compound-protein interactions with knowledge graph embedding of perturbation transcriptomics. Cell Genom. (2024).
 
 ## Contact
 nishengkun@simm.ac.cn

src/README.md

@@ -1 +1,79 @@
+## For target inference
 
+If there exist new interested compound with perturbation transcriptomics to conduct target inference, plz follow this step.
+
+### Step1:Processing your own transcriptomics data
+
+This step depends entirely on the format of your data. You can process it using your own differential expression gene (DEGs) analysis code. Generally, two strategies are recommended:
+
+1. Use the "top strategy" to identify DEGs, as described in our paper.
+2. Alternatively, use a standard DEG pipeline, such as Limma, etc.
+
+In either case, the final data format should be as following triples:
+
+```
+<Interested_compound_name  Downregulates  mRNA:ATP1B1>
+# Note: aligning gene name using our map_file
+```
+
+### Step2:Preparing training file
+
+We need to prepare four files:**"cause.txt", "process.txt", "effect.txt", and "test.txt"**.
+
+The simplest approach I recommend is to directly use the files from the directory "../processed_data/target_inference_2/". The only part that needs modification is **adding the triplets identified in the first step to the "effect.txt" file**. 
+
+Then, place "cause/effect/test"  into a new folder, such as "../processed_data/<name>/", where <name> can be any name you choose to name the task.
+
+## Step3:Training stage
+
+Using following cmd to train the PertKGE by replacing  **<name>**:
+
+```
+$ python train_pertkge.py --cause_file "../processed_data/<name>/cause.txt"\
+                          --process_file "../processed_data/knowledge_graph/process.txt"\
+                          --effect_file "../processed_data/<name>/effect.txt"\
+                          --test_file "../processed_data/<name>/test.txt"\
+                          --h_dim 300\
+                          --margin 1.0\
+                          --lr 1e-4\
+                          --wd 1e-5\
+                          --n_neg 100\
+                          --mode 'reproduce'\
+                          --batch_size 2048\
+                          --patients 5\
+                          --warm_up 10\
+                          --processed_data_file "../processed_data/<name>/"\
+                          --save_model_path "../best_model/<name>/"\
+                          --task "target_inference"\
+                          --run_name "new_target_inference"
+```
+
+### Step4:Inference stage
+
+Please follow the steps sequentially as outlined in the `target_inference.ipynb`. Adjust the paths in the user-defined cells accordingly. 
+
+```
+'''
+This section is user defined !!!
+'''
+h_dim = 300
+
+data_path = "../processed_data/<name>/"
+save_model_path = '../best_model/<name>/'
+output_path = "../results/<name>/"
+
+ent_list = ['Interested_compound_name']  # The `ent_list` refers to the list of compounds you wish to infer, which should correspond to the compound names you processed in the step 1.
+```
+
+### Another notation
+
+In target inference, we provide two metrics: "ti_score" and "confidence".
+
+- **"ti_score"** is the score directly predicted by the model.
+- **"confidence"** represents how many other compounds were ranked lower than this compound for the target. This metric can help filter out potential false positives caused by representational bias.
+
+
+
+## ligand virtual screening
+
+The simplest approach is to directly use the model weights we provided for virtual screening.