TumE

TumE infers ongoing cancer evolution and subclonal selection dynamics from variant allele frequencies (VAF) in single tumour biopsies.

Key Features:

• Variant Allele Frequency (VAF) input: Uses VAF as the primary input encoding evolutionary information from single tumour biopsies.
• Synthetic models of cancer evolution: Integrates synthetic evolutionary models to generate training data for supervised learning.
• Bayesian neural networks: Employs Bayesian neural networks for probabilistic inference of evolutionary parameters.
• Synthetic supervised learning: Applies synthetic supervised learning as the core approach for model training and inference.
• Positive selection detection: Detects signatures of ongoing positive selection within tumours.
• Subclone deconvolution and frequency estimation: Deconvolves selected subclonal populations and estimates subclone frequencies from VAFs.
• Analysis of synthetic and patient-derived data: Validates methods on both synthetic datasets and patient-derived tumour samples.
• Transfer learning and model reuse: Incorporates transfer learning to leverage stored knowledge and provide a library of recyclable deep learning models.
• Improved performance: Demonstrates improvements in accuracy and computational speed relative to existing methods.

Scientific Applications:

• Evolutionary inference from single biopsies: Infers ongoing cancer evolution and subclonal selection dynamics using VAF data from single tumour biopsies.
• Selection analysis: Identifies and characterizes positive selection acting on tumour subclones.
• Subclonal composition profiling: Deconvolves tumour subclonal populations and quantifies their cellular frequencies.
• Tumour heterogeneity characterization: Profiles tumour heterogeneity and evolutionary patterns across samples.
• Resource-efficient model adaptation: Uses transfer learning to reduce data and computational requirements for related evolutionary inference tasks.
• Reuse of pretrained models: Enables reuse of pretrained deep learning models for downstream cancer evolution analyses.

Methodology:

TumE applies synthetic supervised learning on VAF inputs using synthetic cancer-evolution models and Bayesian neural networks, with transfer learning for model reuse and deconvolution-based subclone frequency estimation.