🧬 Structure-Based HMM Profiling for Kunitz-Type Protease Inhibitor Domains

Building structure-informed Profile HMMs for Kunitz/BPTI-type protease inhibitor domain detection, comparing sequence-based (MUSCLE) vs. structure-based (PDBeFold) approaches using 2-fold cross-validation on Swiss-Prot data.

---

📋 Table of Contents

• Overview
• Repository Structure
• Objectives
• Installation
• Methodology
• Results
• Contact

---

📌 Overview

This repository contains the code, datasets, and results for a project focused on developing structure-informed Profile Hidden Markov Models (HMMs) for the detection and annotation of Kunitz/BPTI-type protease inhibitor domains. Two HMMs were built:

• Sequence-based HMM — from MUSCLE multiple sequence alignment
• Structure-based HMM — from PDBeFold structural alignment

The models were evaluated using curated datasets from Swiss-Prot, with performance assessed via 2-fold cross-validation. The structure-based approach demonstrated superiority in capturing remote homologs (peak MCC: 0.997 vs. 0.991).

---

📁 Repository Structure

HMM_KunitzDomain/
│
├── Data/                      # Input datasets
│   ├── Raw/                   # Original downloaded sequences (Swiss-Prot, PDB structures)
│   └── Processed/             # Curated datasets, alignments, HMM models, evaluation files
│
├── Docs/                      # Project documentation
│   └── report/                # Final project report (PDF)
│
├── Script/                    # Custom scripts for analysis
│   └── performance.py         # HMM performance evaluation (MCC, TPR, PPV)
│
├── environment.yml            # Conda environment specification
├── requirements.txt           # Python dependencies
├── .gitignore                 # Files to ignore in Git
├── LICENSE                    # MIT License
└── README.md                  # This file

---

🎯 Objectives

• Construct profile HMMs from sequence-based and structure-based alignments of Kunitz domains.
• Evaluate model performance on positive and negative datasets using classification metrics, addressing annotation incompleteness in databases like Pfam.
• Demonstrate the advantage of incorporating structural information for improved sensitivity to remote homologs.

---

⚙️ Installation

Option 1 — Conda (recommended)

git clone https://github.com/MahanBalooei/HMM_KunitzDomain.git
cd HMM_KunitzDomain
conda env create -f environment.yml
conda activate hmm-kunitz

Option 2 — pip

git clone https://github.com/MahanBalooei/HMM_KunitzDomain.git
cd HMM_KunitzDomain
pip install -r requirements.txt

External tools required: HMMER ≥ 3.3.2, MUSCLE v5, MMseqs2, CD-HIT ≥ 4.8.1, UCSF ChimeraX (for structural visualization)

---

🔬 Methodology

Data Preparation

Positive set — Retrieved from UniProtKB/Swiss-Prot (release 2025_03):

• Annotated with Pfam PF00014, InterPro IPR036880, or Prosite (PS00280, PS50279)
• 368 high-confidence Kunitz proteins after filtering, split into two folds (pos_1, pos_2)

Negative set — All Swiss-Prot proteins lacking Kunitz annotations

PDB structures — Pfam PF00014, resolution ≤ 3.5 Å, length 45–80 aa, clustered at 90% identity with CD-HIT

Contamination removal — BLASTp to filter sequences ≥95% identical to structural seeds

Multiple Structural Alignment (MStA)

Structures aligned using PDBeFold (SSM), exported in FASTA format. All structures showed the canonical Kunitz fold: two β-strands, central α-helix, inhibitory loop, three disulfide bridges.

Multiple Sequence Alignment (MSA)

Generated using MUSCLE v5 on the same representative sequences for fair comparison.

HMM Construction

hmmbuild sequence_based.hmm muscle_alignment.ali
hmmbuild structure_based.hmm pdbe_alignment.ali

HMM Search

hmmsearch -Z 1000 --max --tblout pos_1.out structure_based.hmm pos_1.fasta
hmmsearch -Z 1000 --max --tblout neg_1.out structure_based.hmm neg_1.fasta

Performance Evaluation

for i in `seq 1 12`; do python Script/performance.py set_1.class 1e-$i; done

Computes confusion matrix, Q2 (accuracy), MCC, TPR (sensitivity), PPV (precision) across E-value thresholds (10⁻¹ to 10⁻¹²).

---

📊 Results

Model	Peak MCC	Optimal E-value
Structure-based HMM	0.997	10⁻⁵ – 10⁻⁶
Sequence-based HMM	0.991	10⁻⁵ – 10⁻⁶

Key findings:

• E-value separation: Positives scored 10⁻³⁶ to 10⁻⁶⁰; negatives scored 1–10
• Cross-validation: Near-perfect — 0 false positives, 0–2 false negatives per fold
• Misclassification: Apparent false positives were genuine Kunitz domains missed by Pfam
• Domain-level scoring outperformed full-sequence scoring for multidomain proteins

See the full report in Docs/report/Mahan_Balooei_LAB1_Report.pdf for figures (MCC curves, confusion matrices, E-value distributions).

---

📧 Contact

Mahan Balooei

• 🏛️ Department of Pharmacy and Biotechnology, Alma Mater Studiorum – Università di Bologna
• 📧 mahan.balooei@studio.unibo.it
• 🔬 ORCID: 0009-0006-5358-0784

---

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.