LightCUD

LightCUD classifies inflammatory bowel disease (IBD) and distinguishes ulcerative colitis from Crohn's disease using human gut microbiome sequencing data.

Key Features:

• Biomarker Identification: Two small sets of strain-level biomarkers were identified to distinguish healthy controls from IBD patients and to differentiate ulcerative colitis from Crohn's disease, and these biomarkers were used to optimize model performance during pre-training.
• Data Compatibility: Accepts whole-genome sequencing (WGS) data and 16S rRNA sequencing data as inputs.
• Feature Profiling: Assembles and aligns WGS short reads to generate strain- and genus-level feature profiles.
• Diagnostic Modules: Constructs separate 16S-based and WGS-based diagnostic modules.
• Feature Selection: Employs a novel feature selection procedure to identify case-specific features for discrimination models.
• Machine Learning Core: Multiple machine learning algorithms were evaluated and LightGBM was selected as the core algorithm for classification.
• Validation: Model performance was validated using five-fold cross-validation and independent test datasets, demonstrating high accuracy.

Scientific Applications:

• IBD Diagnosis: Discriminates IBD patients from healthy controls using gut microbiome sequencing data.
• IBD Subtyping: Differentiates ulcerative colitis from Crohn's disease at the strain level.
• Biomarker Discovery: Identifies strain biomarkers for studying factors involved in disease development.
• Microbiome-informed Treatment Guidance: Supports investigation of treatment strategies based on gut microbial community changes.

Methodology:

Analyzed whole-genome sequencing (WGS) data from 349 human gut microbiota samples; assembled and aligned WGS short reads to generate strain- and genus-level feature profiles; constructed 16S-based and WGS-based diagnostic modules; applied a novel feature selection procedure to identify case-specific features; trained and compared multiple machine learning algorithms and selected LightGBM; validated models using five-fold cross-validation and independent test datasets.