🧬 Ensemble Deep Learning Model for Protein Secondary Structure Prediction using NLP Metrics and Explainable AI
U. Vignesh, R. Parvathi, K. Gokul Ram
DOI: https://doi.org/10.1016/j.rineng.2024.103435
Published under: Creative Commons License (Open Access)
This repository presents an Ensemble Deep Learning framework for Protein Secondary Structure Prediction (PSSP), integrating Natural Language Processing (NLP) metrics and Explainable AI (XAI) techniques.
Our model leverages the linguistic nature of amino acid sequences, bridging bioinformatics and language modeling. The approach enhances interpretability and accuracy, outperforming individual models with an exceptional validation accuracy of 94.41% and ROUGE-L score of 0.6766.
- 🔗 Integration of Multiple Sequence Alignment (MSA) with deep neural architectures.
- 🧩 Ensemble Learning combining Sequence Processing and Sequence-to-Sequence predictors.
- 📈 Exceptional accuracy — Training: 94.47%, Validation: 94.41%.
- 🧮 ROUGE-L metric adopted from NLP for evaluating structural integrity of protein sequences.
- 🧬 Explainable AI (LIME & Integrated Gradients) to interpret biological feature importance.
- ⚖️ Data balancing and augmentation mitigate class imbalance across multiple datasets.
Proteins are essential macromolecules that perform critical biological functions. Accurate prediction of protein secondary structures is foundational to understanding protein folding and function.
This study proposes a novel ensemble deep learning model that fuses NLP evaluation metrics with bioinformatics data, enabling robust and interpretable protein secondary structure predictions.
Through integrated explainable AI techniques, our model identifies influential amino acid residues and sequence motifs, providing transparent insights into the biological reasoning behind model predictions.
- Datasets: PS4, CB513, CASP12, TS115.
- Sequence normalization, trimming, and encoding using DSSP8 classification.
- Applied DIVAA (Diversity in Set of Aligned Amino Acids) scoring for amino acid distribution analysis.
- Bias reduction via oversampling and data augmentation techniques.
- Embedding layer + Bidirectional LSTM + Dense output.
- Captures biochemical interrelations among amino acids.
- Embedding + Bidirectional LSTM + Time-Distributed Layer.
- Predicts probability distributions across multiple secondary structure classes.
- Combines SPP and ESSP outputs using an Averaging Layer.
- Enhances accuracy, reduces bias, and stabilizes learning.
- LIME (Local Interpretable Model-Agnostic Explanations): identifies the most influential residues.
- Integrated Gradients: quantifies contribution of each amino acid to predicted structure.
- Enables biologically interpretable insights into model decisions.
| Component | Framework | Description |
|---|---|---|
| Programming Language | Python | Core model development |
| Deep Learning Framework | TensorFlow / Keras | Training and architecture design |
| Data Handling | NumPy, Pandas | Preprocessing and normalization |
| Visualization | Matplotlib, Seaborn | Result plots and XAI visualization |
| Hardware | NVIDIA RTX 3080 GPU | High-performance model training |
| Model | Train Accuracy | Val Accuracy | Train Loss | Val Loss | ROUGE-L |
|---|---|---|---|---|---|
| Sequence Processing Powerhouse | 93.93% | 93.93% | 0.1708 | 0.1708 | 0.6470 |
| Enhanced Seq-to-Seq Predictor | 93.92% | 93.94% | 0.1707 | 0.1704 | 0.6478 |
| Harmonized Ensemble (Proposed) | 94.47% | 94.41% | 0.1569 | 0.1585 | 0.6766 |
| Model | Accuracy |
|---|---|
| PSI-Blast (Baseline) | 75% |
| JOE (Ensemble of Experts) | 83% |
| CNN | 85.3% |
| Temporal CNN | 84.9% |
| Our Ensemble Model | 88.54% (Q8) |
✅ Our model consistently outperforms traditional architectures, while maintaining interpretability and biological consistency.
LIME analysis reveals feature importance maps that highlight key amino acids influencing structure formation:
- Identifies critical residues driving helix or sheet formation.
- Detects biases and underrepresented classes in datasets.
- Provides biological interpretability for domain scientists and model developers.
This transparency bridges the gap between AI predictions and biological reasoning, strengthening model trustworthiness.
| Dataset | Type | Description |
|---|---|---|
| PS4 | DSSP8 | High-quality annotated protein sequences |
| CB513 | Q3/Q8 | Benchmark dataset for secondary structure classification |
| TS115 | Q3/Q8 | Dataset for model generalization |
| CASP12 | Q3/Q8 | Critical assessment dataset for structure prediction |
- Optimizer: Adam
- Learning Rate Scheduler: Custom Warm-Up Scheduler
- Batch Size: 10–20
- Splits: 60% Training, 20% Validation, 20% Testing
- Loss Function: Categorical Cross-Entropy
- Metrics: Accuracy, ROUGE-L, Validation Loss
- ✅ Achieved state-of-the-art accuracy in secondary structure prediction.
- ✅ Integrated NLP metrics (ROUGE-L) for evaluating biological sequences.
- ✅ Applied Explainable AI (LIME, Integrated Gradients) for interpretability.
- ✅ Balanced datasets using oversampling and augmentation.
- ✅ Designed for scalability and reproducibility in bioinformatics pipelines.
- Expand the model to tertiary structure prediction via transformer-based architectures.
- Integrate Graph Neural Networks (GNNs) for residue interaction modeling.
- Deploy an interactive web dashboard for protein sequence prediction and visualization.
- Collaborate with biological research centers for experimental validation of predictions.
If you use this work, please cite:
Vignesh, U., Parvathi, R., & Gokul Ram, K. (2024).
Ensemble deep learning model for protein secondary structure prediction using NLP metrics and explainable AI.
Results in Engineering, 103435.
DOI: 10.1016/j.rineng.2024.103435
Protein Structure Prediction • Deep Learning • Explainable AI (XAI) • NLP Metrics • LIME • Ensemble Learning • Bioinformatics • Drug Discovery
For queries, collaborations, or academic discussion:
📩 gokul.ram.kannan210905@gmail.com
🧪 School of Computer Science and Engineering
VIT Chennai, India
⚙️ “Bridging biology and AI — one sequence at a time.”