Prediction of protein interaction sites with neural networks

Goal: Predict surface residue in contact with another protein.

Method: Feed-forward neural network trained with standard back-propagation.

Output layer: Single neuron representing contact / no contact

Hidden layer: 4 nodes

Input layer:

- Patches in protein structures (sets of exposed neighbour residues)

11 residue-long window, central surface residue + 10 patch neighbours

- Evolutionary information

Residue coded as a vector corresponding to frequencies in the MSA

Result: 73% average accuracy of correctly predicted interacting residues

Fariselli, Pazos,Valencia, Casadio, Eur. J. Biochem. 02