NetAutoProbit

NetAutoProbit applies a network-based hierarchical Bayesian auto-probit model to predict protein functions by integrating protein-protein association network topologies and accounting for false negative labels in Gene Ontology (GO) annotations.

Key Features:

• Network-based spatial auto-probit extension: Extends the spatial auto-probit model to network-indexed binary processes for protein function prediction.
• Hierarchical Bayesian probit framework: Uses a hierarchical Bayesian probit-based framework to model binary functional labels.
• Latent multivariate conditional autoregressive Gaussian process: Implements a latent multivariate conditional autoregressive Gaussian process to integrate network topology into latent functional similarity.
• Integration of protein-protein association networks: Incorporates protein-protein association networks (binary or weighted) to inform and propagate functional similarity.
• Modeling of false negative labels: Models and corrects for false negative labels in training datasets derived from Gene Ontology (GO).
• Gene Ontology (GO) integration: Defines protein functions using GO terms and uses GO-derived annotations for training and evaluation.
• Evaluation on STRING networks: Evaluated against standard algorithms using weighted yeast protein-protein association networks from the STRING database.
• Extended uncertainty modeling: An extended version that incorporates uncertainty in negative labels yields significant improvements in predictive accuracy.
• Implementation: Implemented in Matlab.

Scientific Applications:

• Protein function prediction in networks: Prediction of protein functions within complex protein-protein association networks, including weighted yeast networks from STRING.
• Annotation uncertainty correction: Modeling and correction of false negative GO annotations to improve functional annotation accuracy.
• Method benchmarking: Comparative evaluation against standard algorithms for network-based function prediction.

Methodology:

Hierarchical Bayesian probit-based modeling using a latent multivariate conditional autoregressive Gaussian process for network-indexed binary processes, integration of binary or weighted protein-protein association networks, explicit modeling of false negative GO labels, evaluation against standard algorithms on weighted yeast STRING networks, implemented in Matlab.