🧬 Deep learning-based prognosis and diagnosis using Escherichia coli growth-coupled metabolic sensors

This repository contains the full codebase used for prognosis and diagnosis of COVID-19 based on Escherichia coli growth dynamics, as well as statistical analyses using Generalized Additive Mixed Models (GAMMs).

---

📄 Reference

Preprint Deep learning-based prognosis and diagnosis using Escherichia coli growth-coupled metabolic sensors

---

🧠 Overview

This project explores multiple machine learning and statistical approaches to classify biological conditions from microbial growth data:

• Growth parameter-based models
• Time-series deep learning models
• Context-aware models (FiLM)
• Multistrain fusion strategies
• Statistical inference using GAMMs

All methods are evaluated under a nested cross-validation framework to ensure robust and unbiased performance estimates.

---

🧪 Python Models (Classification)

1. Growth Parameter-Based Classification

📂 Growth_parameter_based_classification/

• Input: extracted growth parameters

• Models:

  • Support Vector Machine (SVM)
  • Logistic Regression
  • XGBoost
  • Soft-voting ensemble

• Main script: growth_parameters_main.py

---

2. Time-Series Classification

📂 Time_series_classification/

• Input: raw growth curves

• Models:

  • 1D Convolutional Neural Network (CNN1D)
  • Temporal Convolutional Network (TCN)

• Optional features:

  • First & second derivatives
  • Channel-wise normalization

• Scripts:

  • time_series_main.py
  • time_series_model.py

---

3. FiLM-Based Context-Aware Models

📂 FiLM/

• Combines:

  • Growth parameters → conditioning signal
  • Growth curves → primary input

• Mechanism:

  • Feature-wise Linear Modulation (FiLM)
  • Parameters generated via MLP

• Models:

  • FiLM-CNN1D
  • FiLM-TCN

---

4. Multistrain Models

📂 Multistrain_models/

Early Fusion

• Combines multiple strains as multi-channel input
• Optional FiLM conditioning per strain

Late Fusion

• Independent models per strain
• Weighted soft-voting ensemble
• Weights learned per fold to avoid leakage

---

📊 Model Evaluation Strategy

• Nested cross-validation:

  • 5 outer folds
  • 3 inner folds

• Patient-level splitting

• Shared splits across all models

• Optimization target: balanced accuracy

• Hyperparameter tuning:

  • Optuna (for time-series models)

---

📈 Statistical Analysis (R - GAMMs)

📂 GAMM models/

Model Selection

• Based on Akaike Information Criterion (AIC)
• Basis dimension tuning via k.check()

Scripts

• GAMM_all_mutants_model_comparison.R
• GAMM_selected_mutants_model_comparison.R

→ Select best model per strain and classification task

• GAMM_all_mutants.R
• GAMM_selected_mutants.R

→ Perform statistical testing:

• Detect differences between conditions
• Identify significant time windows
• Compute confidence intervals

Technical details

• Fitted using mgcv::bam() (v1.9.3)

• Includes:

  • Simultaneous 95% confidence bands
  • Pointwise variance estimation

---

📁 Repository Structure

FiLM/
GAMM models/
Growth_parameter_based_classification/
Multistrain_models/
Time_series_classification/

Each folder contains:

• Model scripts
• Data inputs (parameters or time series)
• Cross-validation splits

---

⚠️ Notes

• RStudio-generated files (.Rproj.user, .RData, .Rhistory) are excluded from version control
• Excel files correspond to experimental datasets used in the study
• Splits are shared across all models to ensure fair comparison

---

📬 Contact

• Paul Ahavi 📧 paul.ahavi228@gmail.com

• Jean-Loup Faulon 📧 jean-loup.faulon@inrae.fr