🧬 Exon & Intron Classification Using BI-LSTM | Genomic Sequence Analysis with Deep Learning

This project implements a Bidirectional LSTM (BI-LSTM) neural network to classify exonic and intronic regions in human DNA sequences using character-level sequence modeling. It includes a complete ETL pipeline, dataset preprocessing from FASTA to CSV, baseline model comparisons (RNN, LSTM, GRU), and visualization of training progress.

Originally developed for a scientific paper accepted at a bioinformatics conference in Portugal, the project focuses on reproducibility and performance benchmarking using real-world genomic data sourced from the Ensembl Genome Browser.

Table of Contents

1. Model Overview
2. ETL Pipeline (Data Preparation)
3. Final BI-LSTM Model Architecture
4. Baseline Evaluation
5. Data Description
6. Conclusion
7. License

---

Model Overview

The goal of this project is to accurately distinguish between exons (coding regions) and introns (non-coding regions) in DNA sequences using deep learning techniques. Among several models tested, the BI-LSTM architecture showed the best performance in terms of accuracy and generalization.

---

ETL Pipeline (Data Preparation)

The ETL pipeline was designed to transform raw biological data from the FASTA format into a format compatible with RNN-based models.

1. Extract

• Source: Ensembl Genome Browser
• Genes included: ANKRD1, PGK1, B2M, GAPDH, PPIA, RPLA13A, NEB, TTN
• Format: FASTA files containing nucleotide sequences labeled by region (exon/intron)

2. Transform

Each FASTA file was parsed and transformed into a structured CSV file, following these steps:

• Gene label extraction using regex from FASTA headers
• Binary labeling: exon = 1, intron = 0
• Intron masking for gene IDs not present in intronic regions
• Feature engineering: adding metadata such as start, end, length, and sequence
• Cleaning and standardization to create uniform-length input sequences

🛠️ Code: featureExtraction.py

3. Load

• After preprocessing, sequences were tokenized (character-level) and padded to a maximum length of 500 nucleotides.
• Data was split using train_test_split:
  • 80% Training
  • 10% Validation
  • 10% Testing
• Sequences were processed in chunks of 1000 to improve memory efficiency.

---

Final BI-LSTM Model Architecture

The final model was implemented using TensorFlow and Keras, with the following structure:

Layer	Configuration
Embedding	input_dim = vocab_size, output_dim = 32, input_length = 500
Bi-LSTM Layer 1	32 units, return_sequences=True
Dropout Layer	Dropout rate = 0.2
Bi-LSTM Layer 2	32 units
Dropout Layer	Dropout rate = 0.2
Dense Layer	64 units, activation='relu'
Output Layer	1 unit, activation='sigmoid'

Training Details

• Epochs: 60
• Batch Size: 16
• Optimizer: Adam (default learning rate)
• Loss Function: Binary Crossentropy
• Tokenization: Character-level (A, T, G, C, etc.)
• Padding: Post-padding of sequences up to 500 characters
• Evaluation Metrics: Accuracy, Precision, Sensitivity (Recall), Specificity, F1-Score

Visualization

The training process includes real-time visualization of:

• Training vs. Validation Loss
• Training vs. Validation Accuracy

📈 These plots confirm that the model generalizes well without overfitting, respecting the converging graphics between validation and loss.

---

Baseline Evaluation

To validate the BI-LSTM model, it was compared with three other RNN-based architectures: Simple RNN, LSTM, and GRU. All models were trained under the same conditions and evaluated on the same dataset splits.

Model	Accuracy	Precision	Sensitivity	Specificity	F1-Score
Simple RNN	0.6820	0.6216	0.9981	0.3375	0.7661
LSTM	0.9860	0.9810	0.9923	0.9790	0.9866
GRU	0.9960	0.9981	0.9942	0.9979	0.9961
BI-LSTM	0.9980	1.0000	0.9961	1.0000	0.9981

📁 Scripts for each model:

• Simple RNN
• LSTM
• GRU
• BI-LSTM

---

Data Description

The final dataset included 9,971 sequences, evenly distributed between exons and introns, from eight genes. The sequences were balanced in quantity and diversity to ensure generalization.

Gene	Exons	Introns	Total Sequences	Exonic Bases	Intronic Bases
ANKRD1	9	8	17	1,790	7,202
PGK1	37	31	68	9,539	339,889
B2M	40	28	68	10,553	51,222
GAPDH	79	68	147	13,269	21,371
PPIA	80	62	142	34,258	80,547
RPLA13A	123	101	224	29,782	54,231
NEB	844	823	1,667	119,394	1,106,064
TTN	3,822	3,807	7,629	1,247,226	2,273,905
Total	5,034	4,928	9,971	1,469,811	3,885,762

📂 Data files (FASTA and CSV): Available here

---

Conclusion

The BI-LSTM architecture demonstrated superior performance in distinguishing exons from introns, with a final accuracy of 99.80%. This positions it as a strong candidate for use in bioinformatics pipelines, gene structure analysis, and even medical genomics research.

The project emphasizes:

• 🔍 Transparency via open-source code and data
• 🔁 Reproducibility through complete ETL steps
• 📈 Scientific rigor in metric evaluation and cross-model comparison

📝 Notes

This project was developed as part of a research initiative and was accepted for presentation at an international conference in Portugal.
To access the full paper, feel free to contact me via email: arielabadebandeira@gmail.com

One of the main challenges during development was finding comparable research papers with similar approaches and consistent evaluation metrics.

Another significant difficulty was obtaining reliable and publicly available genomic datasets, especially for cross-testing against other models.

Most related studies used in-house datasets and unfortunately did not provide access to their training or testing data, nor the full implementation details.
As a result, it was difficult to establish fair comparisons or assess how well those external models actually performed in practice.

📬 For more information, contact: arielabadebandeira@gmail.com