🧬 DeepFoldChange

A Deep Learning Framework that Emulates Statistical Models for Differential Gene Expression Analysis

---

🌍 Background and Motivation

Traditional RNA-seq differential expression (DE) analysis relies on statistical tools such as DESeq2, edgeR, and limma-voom.
While these methods are robust, they can be computationally intensive, require repeated normalization, and are not easily generalizable across datasets.

DeepFoldChange offers a new perspective — a deep neural network (DNN) framework that learns to imitate the statistical inference process used in DESeq2.
By training on DESeq2-derived log2FoldChange (LFC) values and expression count matrices, DeepFold accurately predicts fold-changes for unseen genes, effectively serving as an AI-driven surrogate model for DE analysis.

---

⚡ Key Features

• 🧠 Deep Neural Emulation — Learns DESeq2-style fold-change behavior directly from count matrices.
• 🧩 Lightweight & Scalable — Uses PCA-compressed expression features for faster training.
• 📊 Model Evaluation Suite — Automatically generates regression, error, and agreement plots.
• 💻 Reproducible Pipeline — Compatible with any RNA-seq dataset that includes count and LFC tables.

---

📁 Repository Structure

DeepFold/
│
├── train_predict_lfc_dnn.py           # Main training + prediction script
├── plot_deepfold_performance.py       # Performance plotting utility
├── requirements.txt                   # Package dependencies
├── lfc_dnn_outputs/                   # Folder containing model outputs
│   ├── holdout_predictions.csv
│   ├── all_predictions_fullfit.csv
│   ├── pca_dnn_lfc_predictor.pkl
│   ├── metrics.txt
│   ├── plot_true_vs_pred_scatter.png
│   ├── plot_error_distribution.png
│   ├── plot_bland_altman.png
│   ├── plot_cumulative_accuracy.png
│   ├── plot_accuracy_vs_threshold.png
│   └── plot_accuracy_bar.png
└── README.md

---

🧩 Model Architecture

DeepFoldChange uses a PCA → DNN regression pipeline:

Counts → CPM normalization → log1p → PCA(≤100) → 
DNN(hidden: 256→128→64, ReLU, Adam optimizer) → Predicted LFC

---

📊 Example Results

1️⃣ True vs Predicted LFC

Pearson r = 0.99, Spearman ρ = 1.00, MAE = 0.10.
Predicted fold-changes align almost perfectly with DESeq2 results, demonstrating that DeepFold reproduces both the magnitude and direction of differential expression.

---

2️⃣ Prediction Error Distribution

Errors are sharply centered at 0, indicating no bias between predicted and true LFC values.

More than 95 % of genes fall within ±0.5 log₂ fold-change error.

---

3️⃣ Bland–Altman Agreement Plot

Differences between predicted and true LFCs remain consistent across the entire range of expression changes — confirming uniform agreement and no proportional bias.

---

4️⃣ Cumulative Accuracy Curve

Over 90 % of genes are predicted within |ΔLFC| ≤ 0.5,
showing high fidelity between DeepFold and statistical estimates.

---

5️⃣ Accuracy vs. Error Threshold

This curve quantifies model accuracy as a function of tolerance.
For example, if the threshold is ±0.25 LFC, the model correctly predicts ~80 % of genes.

---

🧪 Evaluation Summary

Metric	Value	Interpretation
Pearson r	0.99	Linear correlation between predicted and true LFC
Spearman ρ	1.00	Rank-order agreement
MAE	0.10	Average absolute LFC difference
90 % genes	≤ 0.5	High-confidence accuracy

DeepFoldChange achieves statistical-model-level precision, validating its use as a deep learning surrogate for DE analysis.

---

⚙️ Installation & Environment Setup

Option 1: Conda (recommended)

conda create -n DeepFold_env python=3.10
conda activate deepfold_env
pip install -r requirements.txt

Option 2: Manual install

pip install pandas numpy scikit-learn scipy joblib matplotlib seaborn

---

🚀 Running the Pipeline

1️⃣ Prepare input files

• filtered_counts_DEGs.csv – normalized or raw counts (rows = genes, columns = samples)
• resLFC_p_cut.csv – DESeq2 output containing at least a log2FoldChange column.

2️⃣ Train and Predict

python train_predict_lfc_dnn.py   --counts filtered_counts_DEGs.csv   --lfc resLFC_p_cut.csv   --outdir lfc_dnn_outputs

This will:

• train the PCA + DNN model,
• evaluate on a hold-out set,
• fit on all data,
• and generate predictions + metrics.

3️⃣ Plot performance

python plot_deepfold_performance.py

All figures are saved inside lfc_dnn_outputs/.

---

📘 Output Files Explained

File	Description
holdout_predictions.csv	True vs predicted LFC for test genes
all_predictions_fullfit.csv	Predicted LFC for all DEGs
pca_dnn_lfc_predictor.pkl	Serialized model (scaler + PCA + DNN)
metrics.txt	R², MAE, Spearman metrics summary
plot_*.png	Performance and accuracy figures

---

🧩 How to Use Trained Model for New Data

Once trained, you can load the model and predict on new normalized counts:

import joblib, pandas as pd
model = joblib.load("lfc_dnn_outputs/pca_dnn_lfc_predictor.pkl")

new_counts = pd.read_csv("your_new_counts.csv", index_col=0)
predicted_lfc = model.predict(new_counts)

---

📈 Example Interpretation

DeepFoldChange closely approximates DESeq2-derived fold-changes with > 90 % accuracy for |ΔLFC| ≤ 0.5.
The framework offers a reproducible, fast, and model-agnostic solution for high-throughput DE prediction using deep learning.

---

✨ Citation (suggested)

If you use or adapt this repository, please cite:

Debnath, J. P., et al. (2025). DeepFoldChange: A Deep Learning Framework that Emulates Statistical Models for Differential Gene Expression Analysis. GitHub repository: https://github.com/Prokash21/DeepFoldChange

---

🧠 Author

Joy Prokash Debnath
Department of Biochemistry and Molecular Biology
Shahjalal University of Science and Technology, Sylhet, Bangladesh

---

🧾 License

This project is distributed under the MIT License – free for academic and research use.

---