mGene.web

mGene.web predicts protein-coding gene structures in eukaryotic genomes to support genome annotation and gene discovery.

Key Features:

• Pre-Trained Models: Provides pre-trained models for immediate gene-structure recognition without requiring initial user training.
• Custom Model Training: Allows training on user-supplied datasets to generate organism-specific prediction models.
• Modular Framework: Implements modular components such as promoter prediction tools and splice site predictors that can be used independently.
• Machine Learning Techniques: Employs discriminative machine learning methods, including generalized hidden Markov models (gHMMs) and Support Vector Machines (SVMs), for prediction.
• Proven Accuracy: Validated in an international nematode genome prediction competition and shown to outperform tools such as Fgenesh++ and Augustus at nucleotide, exon, and transcript levels.
• Gene Discovery Potential: Predicted approximately 2,200 genes not present in existing Caenorhabditis elegans annotations, with experimental RT-PCR confirming expression for a substantial fraction.
• Application Across Species: Provides gene predictions for multiple nematode species including C. briggsae, C. brenneri, C. japonica, and C. remanei, enabling identification of species-specific gene inventions.
• Quality Assessment: Predictions have been assessed as the most accurate among several available annotations for the analyzed genomes.

Scientific Applications:

• Eukaryotic genome annotation: Generating gene models for annotation of newly sequenced eukaryotic genomes.
• Novel gene discovery: Identifying candidate genes absent from existing annotations, with candidates validated experimentally by RT-PCR.
• Comparative genomics: Producing gene predictions across nematode species to enable comparative analyses.
• Evolutionary research: Detecting species-specific gene inventions and supporting studies of genome evolution.
• Model customization: Training organism-specific prediction models using user-provided datasets for tailored analyses.

Methodology:

Uses discriminative machine learning, specifically generalized hidden Markov models (gHMMs) and Support Vector Machines (SVMs), within a modular framework that supports promoter and splice site predictors, pre-trained models, and user-provided model training.