Benchmarking and optimizing microbiome-based bioinformatics workflow for non-invasive detection of intestinal tumors

Area of Science:

• Microbiome research
• Machine learning applications in medicine
• Bioinformatics

Background:

• The human gut microbiome is increasingly recognized for its association with various disease states.
• Machine learning (ML) holds significant promise for developing novel disease detection tools based on microbiome data.
• Variability in ML workflows, including feature types, preprocessing, feature selection, and algorithms, impacts predictive performance.

Purpose of the Study:

• To systematically evaluate and optimize machine learning methods for classifying colorectal cancer and adenoma using gut microbiome data.
• To benchmark a large number of analytical pipelines to identify optimal strategies for disease detection.
• To develop and validate a robust, generalizable framework for microbiome-based intestinal disease detection.

Main Methods:

• A comprehensive evaluation of 6,468 unique analytical pipelines was performed using 4,217 fecal samples.
• Whole Genome Shotgun (WGS) and 16S ribosomal RNA gene (16S) sequencing data were analyzed.
• Model performance was quantified using the area under the receiver operating characteristic curve (AUC) with dual validation (cross-validation and leave-one-dataset-out).

Main Results:

• Whole Genome Shotgun (WGS) data generally outperformed 16S sequencing data for disease classification.
• Species-level genome bin, species, and genus features showed the highest discriminatory power for WGS data.
• Amplicon Sequence Variant-based features were optimal for 16S data; Wilcoxon rank-sum test and data normalization improved performance.
• Ensemble learning models, specifically eXtreme Gradient Boosting and Random Forest, were the top-performing classifiers.

Conclusions:

• An optimized Microbiome-based Detection Framework (MiDx) was developed based on comprehensive evaluation.
• The MiDx framework demonstrated robust generalizability on an independent dataset.
• This provides a systematic and practical framework for future 16S and WGS-based intestinal disease detection.