Protein Surface and Interface Patch Analysis

A Python script for the analysis of protein surface topology, residue exposure, and interfacial patch characteristics from high‑resolution PDB structures using molecular surface algorithms, geometric feature extraction, and machine learning. This pipeline delivers comprehensive descriptors for per‑residue and per‑patch properties, allowing for statistical modeling and predictive tasks in structural bioinformatics. Key features include:

• Automated cleaning of input PDB files (removal of non‑canonical ligands and HETATM records) via cleaner.py.
• Computation of Convexity Index (CX) per residue by volumetric sphere sampling.
• Shrake–Rupley solvent accessible surface area (SASA) profiling over variable probe radii (0.2–2.0 Å) for fractal roughness estimation.
• Principal Component Analysis (PCA)‑based planarity quantification of spatial residue patches.
• Classification of residues into surface, interface, and interior based on relative ASA thresholds and interface burial upon complexation.
• Spatial clustering of residues into patches (neighbors within 11 Å filtered by solvent‑vector angles < 110°) and computation of per‑patch descriptors: CX, mean ASA, hydrophobicity, planarity, and roughness.
• Generation of PDB outputs with per‑atom B‑factors encoding computed metrics and CSV summaries of residues and patches.
• PyTorch‑based multilayer perceptron (MLP) for binary classification of surface vs. interface patches, trained on aggregated per‑patch feature sets.

Usage

1. Input Preparation

• Place target PDB files (with extension .pdb) into the in/ directory.

2. Running the Analysis Pipeline

python main.py

Running main.py will:

• "Clean" each PDB (removing extraneous HETATM records).
• Compute residue‐level CX and write {protein_id}_cx.pdb in out/{protein_id}/.
• Classify residues (surface, interface, interior).
• Identify and characterize patches, generating:
  • {protein_id}_patches.csv (per‑patch descriptors)
  • {protein_id}_residues.csv (per‑residue descriptors)
  • PDBs with B‑factors set to patch hydrophobicity, planarity, roughness, and interface maps.
  • High‑resolution plots (*.png) of SASA vs. probe radius and bar‑charts comparing surface/interface patch metrics.

3. Training a Patch Classifier

python train_patch_classifier.py --data-dir out --epochs 100 --batch-size 32 --lr 1e-3

Outputs:

• patch_classifier.pt: trained PyTorch model state.
• Console logs of training loss, accuracy per epoch, and final test accuracy.

Methodology

Convexity Index (CX)

For each atom, internal volume V_int is summed over atom type–specific van der Waals volumes within a sphere (10 Å radius), with external volume V_ext = V_sphere – V_int. Residue CX is averaged atom‑level V_ext / V_int.

SASA Profiling & Fractal Roughness

Shrake–Rupley algorithm computes residue SASA at probe radii from 0.2 to 2.0 Å. Log–log linear regression of surface area vs. radius yields a slope m, and roughness D_f = 2 – m, approximating the fractal dimension of the surface patch.

Planarity via PCA

Atomic coordinates of a residue patch are mean‑centered and decomposed into principal components. The inverse root‑mean‑square deviation of points from the best‑fit plane (defined by the first two PCs) quantifies "flatness."

Residue & Patch Classification

• Surface residues: relative ASA ≥ 25%.
• Interface residues: ΔASA (complexed – isolated) ≤ −1 Å².
• Patches: central residues sampled every 3 positions, neighborhood radius 11 Å, filtered by solvent‑vector angular criteria (< 110°).

Outputs

For each protein {protein_id} in out/{protein_id}/:

• {protein_id}_cx.pdb - CX in B‑factors
• {protein_id}_ip.pdb - interface map (B=0/1)
• {protein_id}_hydrophobicity.pdb, _planarity.pdb, _roughness.pdb - patches encoded in B‑factors
• {protein_id}_residues.csv - per‑residue features
• {protein_id}_patches.csv - per‑patch features
• Visualizations - .png plots of SASA, roughness, planarity, CX, etc.

Results

Preliminary benchmarking on a set of membrane and soluble proteins show clear discriminative power of fractal roughness and CX in demarcating surface vs. interface patches, achieving >78% classification accuracy with a simple MLP.

---

Developed under the supervision of Professor Demet Akten from Kadir Has University.