A Multi-Representation Ensemble Learning Model for Accurate Bitter Peptide Identification
Sarfraz Ahmad1 · Momina Ahsan1 · Muhammad Nabeel Asim2 · Andreas Dengel2 · Muhammad Imran Malik1
1 NUST, Islamabad, Pakistan · 2 DFKI, Kaiserslautern, Germany
iBitter-Stack is a two-tier stacking ensemble that classifies peptide sequences as Bitter or Non-Bitter. It constructs 56 base learners by pairing 7 feature representations with 8 classifiers, selects the top performers by strict quality thresholds, and fuses their soft-probability outputs into a compact meta-dataset consumed by a Logistic Regression meta-learner.
Key design choices:
- Combines deep protein language model (ESM-2) embeddings with handcrafted physicochemical and compositional descriptors
- Dynamic base learner selection (MCC > 0.80, Accuracy > 90%) rather than a fixed configuration
- Meta-level fusion via soft probabilities — avoids feature redundancy from early concatenation
Dataset — BTP640
| Split | Bitter | Non-Bitter | Total |
|---|---|---|---|
| Training (BTP-CV) | 256 | 256 | 512 |
| Independent Test (BTP-TS) | 64 | 64 | 128 |
| Total | 320 | 320 | 640 |
- Bitter peptides collected from peer-reviewed experimental studies
- Non-bitter peptides randomly sampled from the BIOPEP database
- Peptides with ambiguous residues (X, B, U, Z) and 100%-identity duplicates removed
- Additional experiment with 80% identity threshold confirms robustness (Acc: 95.3%, MCC: 0.91)
Feature Representations (7 types)
| Feature | Dimensions | Category |
|---|---|---|
ESM-2 esm2_t6_8M_UR50D |
320 | Protein language model |
| Dipeptide Composition (DPC) | 400 | Compositional |
| Amino Acid Entropy (AAE) | 60 | Positional |
| Binary Profile N/C-terminus (BPNC) | 200 | Positional |
| Amino Acid Index (AAI) | 36 | Physicochemical |
| Global Tripeptide Composition (GTPC) | 125 | Physicochemical |
| Composition-Transition-Distribution (CTD) | 147 | Physicochemical |
Base Learners & Selection
7 features × 8 classifiers = 56 base models, each trained with 10-fold cross-validation.
Classifiers: SVM · Random Forest · Logistic Regression · KNN · Naive Bayes · Decision Tree · AdaBoost · MLP
Selection threshold: MCC > 0.80 and Accuracy > 90%
Eight models qualify and contribute to the meta-dataset:
| Model | CV Accuracy | CV MCC |
|---|---|---|
| ESM_SVM | 85.5% | 0.71 |
| ESM_RF | 83.4% | 0.67 |
| ESM_MLP | 83.6% | 0.67 |
| ESM_LR | 83.6% | 0.67 |
| ESM_ADA | 83.0% | 0.66 |
| CTD_MLP | 81.1% | 0.62 |
| CTD_SVM | 83.2% | 0.66 |
| AAI_RF | 78.5% | 0.57 |
Stacking Architecture
Peptide Sequence
│
├─── ESM-2 ──────────────────────────────┐
├─── DPC ────── [8 classifiers each] ───┤
├─── AAE ───────────────────────────── ├──► Select top 8 ──► 8D probability vector
├─── BPNC ───────────────────────────── │
├─── AAI ───────────────────────────── │
├─── GTPC ───────────────────────────── │
└─── CTD ──────────────────────────────┘
│
Logistic Regression meta-learner
│
Bitter / Non-Bitter + confidence
The meta-learner is Logistic Regression (L2 penalty, max_iter=1500), chosen for its strong empirical performance in ensemble stacking scenarios and computational efficiency at inference time.
| Metric | Value |
|---|---|
| Accuracy | 96.1% |
| Sensitivity | 95.4% |
| Specificity | 97.2% |
| MCC | 0.922 |
| AUROC | 0.981 |
| Predictor | Year | Accuracy | Sensitivity | Specificity | MCC | AUROC |
|---|---|---|---|---|---|---|
| iBitter-SCM | 2020 | 84.0% | 84.0% | 84.0% | 0.69 | 0.90 |
| BERT4Bitter | 2021 | 92.2% | 93.8% | 90.6% | 0.84 | 0.96 |
| iBitter-Fuse | 2021 | 93.0% | 94.0% | 92.0% | 0.86 | 0.93 |
| iBitter-DRLF | 2022 | 94.0% | 92.0% | 96.9% | 0.89 | 0.97 |
| UniDL4BioPep | 2023 | 93.8% | 92.4% | 95.2% | 0.87 | 0.98 |
| Bitter-RF | 2023 | 94.0% | 94.0% | 94.0% | 0.88 | 0.98 |
| iBitter-GRE | 2025 | 96.1% | 98.4% | 93.8% | 0.92 | 0.97 |
| iBitter-Stack (Ours) | 2025 | 96.1% | 95.4% | 97.2% | 0.92 | 0.98 |
iBitter-Stack achieves the highest specificity (97.2%) among all models and a more balanced sensitivity-specificity trade-off than iBitter-GRE (98.4% / 93.8%), indicating better control over both false positives and false negatives.
A real-time prediction interface is available at ibitter-stack-webserver.streamlit.app.
Features:
- Single sequence prediction with confidence score
- Batch CSV upload for high-throughput screening
- Per-base-learner probability visualisation
- Downloadable results
git clone https://github.com/SarfrazAhmad307/iBitter-Stack.git
cd iBitter-Stack
pip install -r requirements.txt
streamlit run src/app.pyiBitter-Stack/
├── src/
│ ├── app.py # Streamlit web application
│ ├── single_sequence.py # Single sequence prediction UI
│ └── batch_sequences.py # Batch upload UI
├── backend/
│ ├── predictor.py # Ensemble inference pipeline
│ ├── model_loader.py # Saved model loader
│ └── embeddings/
│ ├── esm.py # ESM-2 feature extractor
│ ├── aai.py # AAI feature extractor
│ └── ctd.py # CTD feature extractor
├── data/ # Dataset files
├── requirements.txt
└── README.md
If you use iBitter-Stack in your research, please cite:
@article{ahmad2025ibitter,
title = {iBitter-Stack: A multi-representation ensemble learning model
for accurate bitter peptide identification},
author = {Ahmad, Sarfraz and Ahsan, Momina and Asim, Muhammad Nabeel
and Dengel, Andreas and Malik, Muhammad Imran},
journal = {Journal of Molecular Biology},
volume = {437},
number = {24},
pages = {169448},
year = {2025},
publisher = {Elsevier},
doi = {10.1016/j.jmb.2025.169448},
url = {https://www.sciencedirect.com/science/article/pii/S0022283625005145}
}