- Add Nextflow pipeline (main.nf) with Harbor container image - Add nextflow.config with k8s/k8s_gpu/standard profiles - Add params.json for TRS/WES parameter discovery - Add Dockerfile, entrypoint.py, meta.yml from original implementation - Update paths to use /omic/eureka/Pocketminer/ convention - Update .gitignore to allow params.json
294 lines
9.1 KiB
Python
294 lines
9.1 KiB
Python
#!/usr/bin/env python3
|
|
"""
|
|
PocketMiner Entrypoint - Command-line wrapper for cryptic pocket prediction
|
|
|
|
This script wraps the PocketMiner xtal_predict.py functionality with a proper
|
|
command-line interface for Nextflow/Docker integration.
|
|
"""
|
|
|
|
import argparse
|
|
import json
|
|
import os
|
|
import sys
|
|
import numpy as np
|
|
from pathlib import Path
|
|
import warnings
|
|
|
|
# Suppress TensorFlow warnings
|
|
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '1'
|
|
|
|
# Import PocketMiner components
|
|
sys.path.insert(0, '/workspace/gvp/src')
|
|
|
|
try:
|
|
import tensorflow as tf
|
|
import mdtraj as md
|
|
from models import MQAModel
|
|
from util import load_checkpoint
|
|
from validate_performance_on_xtals import process_strucs, predict_on_xtals
|
|
except ImportError as e:
|
|
print(f"Error importing PocketMiner modules: {e}", file=sys.stderr)
|
|
print("Please ensure the GVP repository is properly cloned and models are available.", file=sys.stderr)
|
|
sys.exit(1)
|
|
|
|
|
|
def load_model(model_path, dropout=0.1, num_layers=4, hidden_dim=100):
|
|
"""Load pre-trained PocketMiner model"""
|
|
# Model architecture from original PocketMiner (must match checkpoint exactly)
|
|
model = MQAModel(
|
|
node_features=(8, 50),
|
|
edge_features=(1, 32),
|
|
hidden_dim=(16, hidden_dim), # (16, 100) for pocketminer checkpoint
|
|
num_layers=num_layers,
|
|
dropout=dropout
|
|
)
|
|
|
|
# Load checkpoint
|
|
opt = tf.keras.optimizers.Adam()
|
|
load_checkpoint(model, opt, model_path)
|
|
|
|
return model
|
|
|
|
|
|
def make_predictions(pdb_file, model, model_path, output_folder, output_name, debug=False):
|
|
"""Make cryptic pocket predictions for a PDB structure"""
|
|
|
|
# Load structure using mdtraj
|
|
try:
|
|
struc = md.load(pdb_file)
|
|
strucs = [struc]
|
|
except Exception as e:
|
|
raise ValueError(f"Failed to load PDB file {pdb_file}: {e}")
|
|
|
|
# Process structure to get features
|
|
X, S, mask = process_strucs(strucs)
|
|
|
|
# Get predictions using PocketMiner model
|
|
predictions = predict_on_xtals(model, model_path, X, S, mask)
|
|
|
|
# Extract predictions for the single structure
|
|
# predictions shape: (batch, max_length)
|
|
pred_array = predictions[0] # First (and only) structure
|
|
mask_array = mask[0] # Corresponding mask
|
|
|
|
# Convert TensorFlow tensors to NumPy arrays explicitly
|
|
if hasattr(pred_array, 'numpy'):
|
|
pred_array = pred_array.numpy()
|
|
if hasattr(mask_array, 'numpy'):
|
|
mask_array = mask_array.numpy()
|
|
|
|
# Ensure arrays are NumPy (in case they weren't TensorFlow tensors)
|
|
pred_array = np.asarray(pred_array)
|
|
mask_array = np.asarray(mask_array)
|
|
|
|
# Get only valid (masked) residues
|
|
valid_residues = mask_array > 0
|
|
pred_valid = pred_array[valid_residues]
|
|
|
|
# Save outputs
|
|
output_path = Path(output_folder)
|
|
output_path.mkdir(parents=True, exist_ok=True)
|
|
|
|
# Save binary predictions (full array with padding)
|
|
pred_file = output_path / f"{output_name}-preds.npy"
|
|
np.save(pred_file, pred_valid)
|
|
|
|
# Save human-readable predictions
|
|
txt_file = output_path / f"{output_name}-predictions.txt"
|
|
np.savetxt(txt_file, pred_valid, fmt='%.4f')
|
|
|
|
# Calculate summary statistics
|
|
cryptic_pocket_score = float(np.mean(pred_valid))
|
|
high_confidence_residues = int(np.sum(pred_valid > 0.7))
|
|
medium_confidence_residues = int(np.sum((pred_valid > 0.4) & (pred_valid <= 0.7)))
|
|
|
|
# Save debug features if requested
|
|
if debug:
|
|
np.save(output_path / f"{output_name}_X.npy", X)
|
|
np.save(output_path / f"{output_name}_S.npy", S)
|
|
np.save(output_path / f"{output_name}_mask.npy", mask)
|
|
|
|
# Cluster high-confidence residues
|
|
pocket_clusters = cluster_residues(pred_valid, threshold=0.5)
|
|
|
|
# Generate summary JSON
|
|
summary = {
|
|
"cryptic_pocket_score": cryptic_pocket_score,
|
|
"high_confidence_residues": high_confidence_residues,
|
|
"medium_confidence_residues": medium_confidence_residues,
|
|
"total_residues": len(pred_valid),
|
|
"pocket_clusters": pocket_clusters,
|
|
"output_files": {
|
|
"predictions_npy": str(pred_file),
|
|
"predictions_txt": str(txt_file)
|
|
}
|
|
}
|
|
|
|
summary_file = output_path / f"{output_name}-summary.json"
|
|
with open(summary_file, 'w') as f:
|
|
json.dump(summary, f, indent=2)
|
|
|
|
return summary
|
|
|
|
|
|
def cluster_residues(predictions, threshold=0.5, min_cluster_size=3):
|
|
"""
|
|
Cluster high-scoring residues into spatial pockets
|
|
|
|
Returns list of clusters with residue indices and average scores
|
|
"""
|
|
# Ensure predictions is a pure NumPy array
|
|
if hasattr(predictions, 'numpy'):
|
|
predictions = predictions.numpy()
|
|
predictions = np.asarray(predictions)
|
|
|
|
high_score_idx = np.where(predictions > threshold)[0]
|
|
|
|
if len(high_score_idx) == 0:
|
|
return []
|
|
|
|
# Simple sequential clustering (assumes residues are ordered by sequence)
|
|
# More sophisticated spatial clustering would require 3D coordinates
|
|
clusters = []
|
|
current_cluster = [int(high_score_idx[0])] # Convert to Python int
|
|
|
|
for idx in high_score_idx[1:]:
|
|
idx = int(idx) # Convert to Python int
|
|
if idx - current_cluster[-1] <= 2: # Allow 2-residue gaps
|
|
current_cluster.append(idx)
|
|
else:
|
|
if len(current_cluster) >= min_cluster_size:
|
|
# Use NumPy array indexing for safety
|
|
cluster_indices = np.array(current_cluster)
|
|
cluster_score = float(np.mean(predictions[cluster_indices]))
|
|
clusters.append({
|
|
"residue_indices": current_cluster,
|
|
"size": len(current_cluster),
|
|
"average_score": cluster_score
|
|
})
|
|
current_cluster = [idx]
|
|
|
|
# Add final cluster
|
|
if len(current_cluster) >= min_cluster_size:
|
|
cluster_indices = np.array(current_cluster)
|
|
cluster_score = float(np.mean(predictions[cluster_indices]))
|
|
clusters.append({
|
|
"residue_indices": current_cluster,
|
|
"size": len(current_cluster),
|
|
"average_score": cluster_score
|
|
})
|
|
|
|
# Sort by score
|
|
clusters.sort(key=lambda x: x['average_score'], reverse=True)
|
|
|
|
return clusters
|
|
|
|
|
|
def main():
|
|
parser = argparse.ArgumentParser(
|
|
description='PocketMiner: Predict cryptic binding pockets in protein structures'
|
|
)
|
|
|
|
parser.add_argument(
|
|
'--pdb',
|
|
required=True,
|
|
help='Input PDB file path'
|
|
)
|
|
|
|
parser.add_argument(
|
|
'--output-folder',
|
|
default='.',
|
|
help='Output directory for results (default: current directory)'
|
|
)
|
|
|
|
parser.add_argument(
|
|
'--output-name',
|
|
required=True,
|
|
help='Base name for output files'
|
|
)
|
|
|
|
parser.add_argument(
|
|
'--model-path',
|
|
default='/workspace/gvp/models/pocketminer',
|
|
help='Path to pre-trained model checkpoint'
|
|
)
|
|
|
|
parser.add_argument(
|
|
'--debug',
|
|
action='store_true',
|
|
help='Save debug features (X, S, mask arrays)'
|
|
)
|
|
|
|
parser.add_argument(
|
|
'--dropout',
|
|
type=float,
|
|
default=0.1,
|
|
help='Model dropout rate (default: 0.1)'
|
|
)
|
|
|
|
parser.add_argument(
|
|
'--num-layers',
|
|
type=int,
|
|
default=4,
|
|
help='Number of model layers (default: 4)'
|
|
)
|
|
|
|
parser.add_argument(
|
|
'--hidden-dim',
|
|
type=int,
|
|
default=100,
|
|
help='Hidden dimension size (default: 100)'
|
|
)
|
|
|
|
args = parser.parse_args()
|
|
|
|
# Validate inputs
|
|
if not os.path.exists(args.pdb):
|
|
print(f"Error: PDB file not found: {args.pdb}", file=sys.stderr)
|
|
sys.exit(1)
|
|
|
|
# Check if model checkpoint files exist (model_path is a prefix, not a directory)
|
|
model_index = f"{args.model_path}.index"
|
|
if not os.path.exists(model_index):
|
|
print(f"Error: Model checkpoint not found: {args.model_path}", file=sys.stderr)
|
|
print(f"Looking for: {model_index}", file=sys.stderr)
|
|
print("Please ensure the pre-trained model is available.", file=sys.stderr)
|
|
sys.exit(1)
|
|
|
|
print(f"Loading PocketMiner model from {args.model_path}...")
|
|
model = load_model(
|
|
args.model_path,
|
|
dropout=args.dropout,
|
|
num_layers=args.num_layers,
|
|
hidden_dim=args.hidden_dim
|
|
)
|
|
|
|
print(f"Processing structure: {args.pdb}")
|
|
summary = make_predictions(
|
|
pdb_file=args.pdb,
|
|
model=model,
|
|
model_path=args.model_path,
|
|
output_folder=args.output_folder,
|
|
output_name=args.output_name,
|
|
debug=args.debug
|
|
)
|
|
|
|
print("\n" + "="*60)
|
|
print("PocketMiner Prediction Summary")
|
|
print("="*60)
|
|
print(f"Overall cryptic pocket score: {summary['cryptic_pocket_score']:.4f}")
|
|
print(f"High confidence residues (>0.7): {summary['high_confidence_residues']}")
|
|
print(f"Medium confidence residues (0.4-0.7): {summary['medium_confidence_residues']}")
|
|
print(f"Total residues analyzed: {summary['total_residues']}")
|
|
print(f"\nPocket clusters identified: {len(summary['pocket_clusters'])}")
|
|
|
|
for i, cluster in enumerate(summary['pocket_clusters'][:5], 1):
|
|
print(f" Cluster {i}: {cluster['size']} residues, score={cluster['average_score']:.4f}")
|
|
|
|
print(f"\nResults saved to: {args.output_folder}")
|
|
print("="*60 + "\n")
|
|
|
|
|
|
if __name__ == '__main__':
|
|
main()
|