We used the data leakage-free PPI dataset from Bernett et al available on figshare (DOI: 10.6084/m9.figshare.21591618.v3).
For the proteins contained in the dataset, we generated ESM-2 per-token and per-protein embeddings data/extract_esm.py for esm2_t33_650M_UR50D, esm2_t36_3B_UR50D, and esm2_t48_15B_UR50D.
As a baseline, we used a Random Forest Classifier - with full embeddings and PCA-reduced embeddings (400 and 40 components). The associated code is in models/baselineRFC.py.
As an advanced model, we re-implemented the fully connected model by Richoux et al. models/fc2_20_2_dense.py. A version including a Transformer encoder is contained in models/attention.py#L595.
The 2d-baseline model is implemented in models/baseline2d.py.
We extended the 2d-baseline by inserting a Transformer encoder with self- models/attention.py#L90 or cross-attention models/attention.py#L15.
Further, we re-implemented D-SCRIPT models/dscript_like.py. Also, we implemented a version that included a Transformer encoder models/attention.py#L157.
We also re-implemented TUnA models/attention.py#L321.
Hyperparameter tuning was done with wandb. All code necessary for repeating the analyses is found in main.py.
PPIs included in the PDB were identified in data/get_contact.py. Those were filtered for confident predictions made by the models data/find_confpreds_with_structure.py.
Finally, the distance maps were calculated and their correlations to the predicted distance maps were obtained in data/get_cmap.py.
All other visualizations can be found in plots/.