AUPRC lower than reported using provided pretrained weights

[Masaaaato]

Hi, thank you for sharing the code and pretrained models.

I tried to reproduce the AUPRC values reported in the paper using the provided pretrained weights from the reproduce/ directory and the notebook gather result-txgnn.ipynb.

To generate the dataset, I used split="complex_disease" as specified in the reproducibility instructions. However, the AUPRC scores I obtain are consistently 0.1–0.2 lower than those shown in the figures of the paper.

I followed the default configuration and dependencies as provided. Could you confirm:

1. Whether the pretrained weights are the exact ones used in the paper?
2. Whether additional steps are required for full reproduction?

Any guidance would be greatly appreciated.

Thank you!

[Catriona-Miller]

Hi, did you ever get it closer to the reported values by following any additional steps? I'm facing the same problem now.