HAIRpred2

HAIRpred2 is a structure-based computational tool for predicting antibody-interacting residues in human antigen structures. It uses Relative Solvent Accessibility (RSA) combined with physicochemical properties in a sliding window framework with a pre-trained Random Forest model (AUC = 0.78).

🌐 Web Server: https://webs.iiitd.edu.in/raghava/hairpred2

---

Installation

pip install hairpred2

System dependency — mkdssp (required)

HAIRpred2 uses DSSP to compute RSA values from the PDB structure. Install mkdssp:

# Linux / Ubuntu
sudo apt install dssp

# Conda (recommended — any platform)
conda install -c salilab dssp

# macOS (Homebrew)
brew install dssp

# Google Colab
apt-get install -y dssp

Download the model file

The pre-trained model (~253 MB) is not bundled with the pip package. Download it separately:

# Download to the package data directory automatically
hairpred2-download-model

# Or download manually and place in your working directory:
# https://webs.iiitd.edu.in/raghava/hairpred2/download/best_model_random_forest.pkl

---

Usage

Command line

# Basic usage
hairpred2 -i antigen.pdb -c A

# Custom output prefix
hairpred2 -i antigen.pdb -c A -o my_results

# Multiple antigen chains
hairpred2 -i antigen.pdb -c A,B

# Filter buried residues (recommended)
hairpred2 -i antigen.pdb -c A --min-rsa 0.05

# Custom probability threshold
hairpred2 -i antigen.pdb -c A -t 0.4

# Use a custom model file
hairpred2 -i antigen.pdb -c A --model /path/to/model.pkl

Python API

from hairpred2 import run_pipeline

# Basic prediction
results = run_pipeline(
    pdb_file      = "antigen.pdb",
    chain_ids     = ["A"],
    output_prefix = "my_results",
)

# With options
results = run_pipeline(
    pdb_file      = "antigen.pdb",
    chain_ids     = ["A", "B"],
    output_prefix = "my_results",
    threshold     = 0.4,
    min_rsa       = 0.05,
    model_path    = "/path/to/best_model_random_forest.pkl",
)

# results is a pandas DataFrame with columns:
# Residue, RSA, Probability, Prediction
print(results.head())

Individual functions

from hairpred2 import (
    validate_pdb,
    build_residue_dataframe,
    generate_features,
    load_model,
    predict_residues,
    detect_epitope_patches,
)

# Load and validate
validate_pdb("antigen.pdb")

# Build features
df, temp_pdb = build_residue_dataframe("antigen.pdb", ["A"])
X = generate_features(df)

# Predict
model = load_model()          # uses bundled/downloaded model
# model = load_model("/custom/path/model.pkl")  # custom model
probs, labels = predict_residues(model, X, threshold=0.5)

# Detect epitope patches
patches = detect_epitope_patches(df, labels, probs)

---

Output Files

Every prediction generates 5 output files (all sharing the same prefix):

File	Description
<prefix>.csv	Per-residue: Residue, RSA, Probability, Prediction
<prefix>_summary.txt	Statistics + top 10 predicted residues
<prefix>_bfactor.pdb	PDB with B-factor = probability × 100
<prefix>.pml	PyMOL script — colors red/blue + residue labels
<prefix>_patches.txt	Spatially clustered epitope patches

Visualization in PyMOL

# Run in PyMOL
@my_results.pml

# Or color by probability using B-factor PDB
load my_results_bfactor.pdb
spectrum b, blue_white_red

---

Arguments

Argument	Required	Description
-i / --input	Yes	Input antigen PDB file
-c / --chain	Yes	Chain ID(s): A or A,B
-o / --output	No	Output prefix (default: hairpred2_results)
-t / --threshold	No	Probability threshold (default: 0.5)
--min-rsa	No	Minimum RSA filter (e.g. 0.05)
--model	No	Custom model .pkl path

---

Model Performance

Trained on 221 human Ag-Ab complexes (SAbDab), tested on 56 independent complexes:

Metric	Value
AUC	0.78
Sensitivity	0.73
Specificity	0.65

---

Citation

If you use HAIRpred2, please cite:

Mehta N., et al. (2026) HAIRpred2: Human Host-Specific Prediction of Antibody-Interacting Residues Using Hybrid Physicochemical and Structural Features. (manuscript in preparation)

Previous tool:

Sahni R., Kumar N. and Raghava GPS (2025) HAIRpred: Prediction of human antibody interacting residues in an antigen from its primary structure. Protein Sci, 34(8):e70212

---

License

MIT License — free for academic and non-commercial use.

© 2025 Raghava Lab, IIIT Delhi — https://webs.iiitd.edu.in/raghava/