CRISP

CRISP processes GC×GC-TOFMS 4-D contour images using deep learning to enable untargeted metabolite profiling, contour identification, synthesis, resolution enhancement, and classification for metabolomics studies.

Key Features:

• Deep Learning Architecture: Multiple customizable deep neural network architectures for semi-automated identification, synthesis, resolution enhancement, and classification of contour images derived from GC×GC-TOFMS.
• Aggregate Feature Representative Contour (AFRC) Construction: Constructs unbiased datasets that enhance contrast between test groups and is suitable for small sample sizes.
• Stacked ROIs: Generates deepstacked datasets by identifying and integrating over five regions of interest (ROIs) to improve contour fidelity and resolution.
• Generative Image Synthesis: Uses generative models to create high-fidelity 512×512-pixel contour images trained to a Fréchet inception distance of less than 47.00.
• Classification Performance: Includes a classifier reporting area under the ROC curve (AUROC) greater than 0.96 and classification accuracy exceeding 95%, with reported robustness to column bleed.

Scientific Applications:

• Untargeted metabolite profiling: Rapid, untargeted metabolite profiling directly from GC×GC-TOFMS 4-D contour images.
• Comparative metabolomics in disease: Application to GC×GC-TOFMS contour images from patients with late-stage diabetic nephropathy and healthy controls for comparative analysis.

Methodology:

Constructs AFRCs; identifies and integrates five ROIs to produce deepstacked datasets; employs multiple customizable deep neural network architectures including generative models that synthesize 512×512-pixel images (Fréchet inception distance <47.00); and uses a classifier reporting AUROC >0.96 and accuracy >95% with robustness to column bleed.