owkin/apholo8-egnn

Apholo8-VN-EGNN

Apholo8-VN-EGNN is an equivariant graph neural network for predicting protein binding sites from three-dimensional protein structures. The model is trained on a mixture of apo and holo structures and is designed to generate multiple spatially distinct binding-site predictions.

Model description

Apholo8-VN-EGNN is based on the VN-EGNN architecture. The protein is represented as a graph of Cα atoms or residues, augmented with residue-level sequence embeddings. The model uses virtual nodes to represent candidate binding-site centers.

During inference, the model predicts:

• The 3D coordinates of the virtual nodes.
• A confidence score for each candidate pocket.
• A residue-level binding-site probability for each protein residue.

For multi-pocket prediction, candidate virtual nodes are ranked according to their confidence scores and filtered using spatial non-maximum suppression to remove redundant predictions. This is done in https://github.com/owkin/apholo-egnn

Training data

The model was trained using a mixture of apo and holo protein structures from:

• scPDB (holo dataset, mono pocket)
• Cryptobench (apo dataset, multi pocket)

Training procedure

The model uses:

• Number of VN-EGNN layers: 5
• Number of virtual nodes: 8
• Residue embeddings: ESM2 3B (D=2560)
• Graph construction: 10Å for dist cutoff
• Optimizer: AdamW
• Learning rate: 0.00001
• Effective batch size: 32
• Maximum training duration: 600 epochs

The training objective combines:

• Virtual-node position loss with repulsion term (penalizes pairs of nodes that come within 3Å in Euclidean distance)
• Residue-level segmentation loss
• Virtual-node confidence loss

Inference procedure

The model produces all virtual-node predictions, the candidates are sorted by confidence and spatial non-maximum suppression is applied before selecting the final pocket centers

Evaluation

The model was evaluated on scPDB and Cryptobench test set using :

• DCC@4Å
• DCA@4Å

How to use

See https://github.com/owkin/apholo-egnn

Credits

Developed at Owkin by Christian Esposito (Owkin), Laura Dumont (Owkin), Rose Hoste (Owkin), Floriane Montanari (Owkin)