Deep Learning Microflora Diagnostic Platform • PyTorch FT-Transformer

Early Detection of Colorectal Cancer from Gut Microbiome Data using FT-Transformer

A clinical-grade deep learning diagnostic system leveraging Feature Tokenization and Multi-Head Self-Attention to identify early-stage Colorectal Cancer dysbiosis markers from metagenomic relative abundance profiles.

Test Accuracy

98.33%

Held-out test set (n=120)

Macro ROC-AUC

0.9998

Multi-class discrimination

Early-CRC Sensitivity

95.24%

Stage 1 recall accuracy

Microbial Features

64 Taxa

Centered Log-Ratio (CLR)

Interactive Microflora & Dysbiosis Visualizer

Gut Microbiome Architecture & Key Taxa Indicators

High-resolution metagenomic micro-flora mapping and dysbiosis driver analysis

Metagenomic Relative Abundance CLR Spectrum

CLR Transformed

Fusobacterium nucleatum+2.41 CLR (High Dysbiosis)

Bacteroides fragilis (ETBF)+1.85 CLR (Enriched)

Parvimonas micra+1.54 CLR (Translocated)

Faecalibacterium prausnitzii-1.92 CLR (Depleted Protective)

Composite Microbial Dysbiosis Index (MDI):0.84 (High Risk)

Select Taxon Biomarker to Inspect

Pathogenic / Dysbiotic Driver

Fusobacterium nucleatum

CLR Abundance

+2.41

Adheres to intestinal epithelial cells, promoting inflammation and tumor proliferation in early CRC stages.

SHAP Feature Contribution:High (+0.38)

End-to-End Deep Learning Research Pipeline

Strict scientific methodology from raw reads to clinical decision support

1. Data Ingestion & CLR

Centered Log-Ratio transformation resolves compositional constraint bias in metagenomic abundance.

2. FT-Tokenizer

Numerical features projected to dense vector embeddings with learnable [CLS] token insertion.

3. Self-Attention Encoder

Multi-Head Self-Attention layers model non-linear inter-microbial dysbiosis dependencies.

4. SHAP Biomarkers

Attribution engine isolates Fusobacterium nucleatum & key dysbiotic drivers for clinical validation.