ProteinPrompt

ProteinPrompt predicts protein-protein interactions from amino acid sequences to identify potential binding partners for applications such as drug target identification.

Key Features:

• Machine learning algorithms: Two complementary methods are implemented: a random forest classifier using auto-correlations from seven amino acid scales, and a graph neural network approach.
• Consensus prediction: Results from the random forest and graph neural network are integrated into a consensus prediction to improve accuracy and reduce error likelihood.
• Comprehensive database: Predictions are made against a thoroughly filtered database of known protein-protein contacts compiled from multiple repositories.
• Proteome-scale scanning: The system supports scanning large sequence libraries, including the human proteome and other organismal proteomes, to discover potential binding partners.
• Performance evaluation: On a redundancy-filtered test dataset excluding sequence pairs similar to training, the random forest achieved accuracy 0.88 and AUC 0.95, the graph neural network achieved accuracy 0.86, and the consensus prediction achieved accuracy 0.89.

Scientific Applications:

• Drug discovery and target identification: Predicting interactions to identify novel therapeutic targets.
• Off-target assessment: Identifying potential unexpected bindings that may indicate side effects.

Methodology:

The workflow compiles known protein-protein interactions from multiple sources into a filtered database, computes sequence features including auto-correlations from seven amino acid scales, applies a random forest classifier and a graph neural network to predict interactions against the database, and compares both methods' outputs to produce consensus predictions.