EnsembleACPredict: An Ensemble Machine Learning Framework for Anticancer Peptide (ACP) Classification with LGBM‑Based Feature Importance

---

Project Summary:

Anticancer peptides (ACPs) are short amino acid sequences with selective cytotoxicity toward cancer cells and comparatively low toxicity to normal cells. This project presents EnsembleACPredict, a supervised machine learning framework for ACP classification that integrates multiple classifiers with ensemble learning.

Key contributions:

• Extraction of sequence-level features using Pfeature.
• Application of feature selection & scaling for preprocessing.
• Training of SVM, Decision Tree, Random Forest, LightGBM, XGBoost, and Bagging Classifier.
• Integration using VotingClassifier ensemble (soft voting).
• LightGBM feature importance analysis for biological interpretability.

---

Objective:

• Build a robust ML framework to classify ACPs vs non-ACPs
• Analyze key features using LightGBM importance
• Provide a reproducible model for drug candidate prioritization

---

Dataset

• Source: ENNAACT Anticancer and Non-Anticancer Peptide dataset

• Classes:

  • 1 → Anticancer Peptides (ACPs)
  • 0 → Non-ACPs

• Format: Plain-text sequences (single-letter amino acid codes).

• Preprocessing: Deduplication, removal of non-standard tokens, uppercase standardization.

---

Feature Extraction (via Pfeature)

We computed sequence-level descriptors:

• AAC: Amino Acid Composition
• AAB: Amino Acid Binary
• PCB: Physicochemical Binary
• PCP: Physicochemical Properties
• DPC: Dipeptide Composition
• PAAC: Pseudo Amino Acid Composition

---

Methods

1. Data Preprocessing

   • VarianceThreshold to remove low-variance features
   • StandardScaler (for SVM and distance-based models)

2. Base Models

   • Support Vector Machine (SVM)
   • Decision Tree
   • Random Forest
   • LightGBM
   • XGBoost
   • Bagging Classifier (SVC-based)

3. Ensemble Strategy

   • VotingClassifier (Soft Voting over SVM, LightGBM, XGBoost, Bagging SVC)

---

Results

Model	Accuracy (%)	Precision (%)	Recall (%)	Specificity (%)	F1 Score (%)	ROC AUC (%)	MCC
Ensemble Model	97.48	93.97	82.58	99.34	87.90	97.87	0.87
Bagging Classifier	97.40	94.69	81.06	99.43	87.35	97.49	0.86
SVM	97.32	95.45	79.55	99.53	86.78	97.50	0.86
XGBoost	97.23	92.31	81.82	99.15	86.75	97.70	0.85
LightGBM	97.23	88.37	86.36	98.58	87.36	97.81	0.86
Random Forest	95.39	100.00	58.33	100.00	73.68	97.29	0.74
Decision Tree	93.37	70.23	69.70	96.32	69.96	83.01	0.66

• Best Accuracy: 97.48% (Ensemble Model)
• Best Biological Interpretability: LightGBM feature importance

---

Feature Importance

• Top features identified by LightGBM:

  • PCP_Z4 (Hydrophobicity index)
  • AAC_K (Lysine composition)
  • AAC_C (Cysteine composition)

These features are biologically significant, consistent with ACPs’ membrane-targeting activity.

---

Future Scope

• Integration with deep learning models (CNNs, BiLSTMs, Transformers).
• Deployment as a web tool for biologists & drug discovery researchers.
• Application to other bioactive peptides (AMPs, AVPs, antifungal).
• Hybrid approaches with molecular docking & dynamics simulations.

---

References

1. Garai, S., Thomas, J., Dey, P., Das, D. (2023). LGBM-ACp: An ensemble model for anticancer peptide prediction. Molecular Diversity.
2. Pfeature Manual. (n.d.). Documentation for Pfeature library.

---

👩‍💻 Author

Sujata Sinhababu B.Tech CSE, 4th Year Summer Internship, NIT Agartala (2025)