Query regarding inference of trained model and use of mutilated graphs

[kieron15]

Hi
I found the paper very interesting but have some doubts regarding the model and explanation to following points will greatly help in increasing my understanding of the model.

I see that you are training two models, Response prediction model and Perturbagen discovery model.
In some calls to model, mutilate_mutations input argument is used. (e.g. lines a, b)

As per my observation, this input is only used when initial state is 'diseased'. Are you using the mutilate_mutations argument to modify the links in base PPI network to mimic PPIs in disease state?
How do you get this data.mutations for any input sample for the model?

I am also clueless about how one can use the trained model for inference. I guess one needs to run only the trained Perturbagen discovery model during inference. Is this right?
During inference if you have a query vector of disease state and a vector of target/healthy state, what other inputs are needed. How does one get data.mutations during inference?

Can you please clarify how did you run inference for results presented in the paper? It will be very helpful if you can share the inference script, model checkpoints and the sample inputs.

[xianglin226]

Thank you for your interest in our work. Below are our responses to your questions:

Q: The links in the base PPI network are meant to mimic PPIs in the disease state?
A: Yes.

Q: How do you obtain data.mutations for any input sample for the model?
A: Please refer to the section “Disease-gene information” in the Supplementary Information.

Q: I guess one needs to run only the trained Perturbagen Discovery model during inference. Is this correct? During inference, if you have a query vector of the disease state and a vector of the target/healthy state, what other inputs are needed? How does one get data.mutations during inference?
A: Yes, only the Perturbagen Discovery model needs to be run during inference. The data.mutations input is disease-specific. In our study, these were provided as part of the processed data. For new datasets, one would need to compile or collect the disease-associated genes corresponding to the diseased samples.

Q: It would be very helpful if you could share the inference script, model checkpoints, and sample inputs.
A: We have provided notebooks that demonstrate how to train PDGrapher and perform predictions, as well as scripts for preprocessing perturbation data and PPI/GRN networks. The model checkpoints will be shared soon, and we will notify you once they are available.