PLATO

PLATO applies a positive-unlabeled semi-supervised learning framework to improve somatic variant calling and peptide-level neoepitope identification for prediction of tumor-rejection mediating neoepitopes (TRMNs) in personalized cancer immunotherapy.

Key Features:

• Semi-Supervised Positive-Unlabeled Learning: Implements a positive-unlabeled learning framework (Positive-unlabeled Learning using AuTOml) to model scenarios with limited labeled data and rescore candidate calls.
• Integration with Azure AutoML: Integrates Azure AutoML for hyper-parameter tuning and classification threshold optimization to adapt models across diverse clinical samples.
• Bootstrapping: Incorporates bootstrapping techniques to enhance the reliability of predictions.
• High-Confidence Positive-Call Generation and Rescoring: Generates high-confidence positive calls via stringent filtering of model-based predictions and rescoring of remaining candidates.
• Somatic Variant Calling (Exome sequencing): Processes exome sequencing data to identify somatic mutations relevant to tumor-specific neoepitopes.
• Peptide Identification (MS/MS): Analyzes mass spectrometry (MS/MS) data to identify peptides that may serve as neoepitopes.
• Robustness to Patient-to-Patient Variation: Employs model selection and AutoML-driven tuning to maintain performance across heterogeneous clinical samples.

Scientific Applications:

• Personalized Cancer Immunotherapy: Predicts tumor-rejection mediating neoepitopes (TRMNs) to inform design of personalized cancer vaccines.
• Somatic Mutation Discovery: Identifies cancer-specific somatic mutations from exome sequencing data for downstream neoepitope discovery.
• Neoepitope Discovery from Proteomics: Prioritizes candidate neoepitope peptides from MS/MS data for vaccine target selection.
• Experimental Prioritization: Reduces experimental burden by computationally prioritizing candidate neoepitopes when comprehensive experimental assessment is impractical.

Methodology:

Generates high-confidence positive calls via stringent filtering of model-based predictions, applies positive-unlabeled learning to rescore remaining candidates, integrates Azure AutoML for hyper-parameter tuning and classification threshold optimization, and incorporates bootstrapping to enhance prediction reliability.