Usage (Local)

The ProMaya inference script can be executed directly on a local workstation for analyzing individual protein pairs. For batch processing, we strongly recommend using the Nextflow pipeline (see Cluster/HPC Usage).

Basic Usage

The primary script for predicting interactions between a single protein pair is scripts/predict.py.

python scripts/predict.py \
    --config config/config.yaml \
    --protein-a data/proteins/1ABC.pdb \
    --protein-b data/proteins/2XYZ.pdb \
    --checkpoint models/promaya_v1/best_model.pt \
    --output results/predictions/ \
    --gpu 0 \
    --gradcam

Command-Line Arguments

Argument	Required	Default	Description
--config	Yes	-	Path to the config.yaml file defining model parameters
--protein-a	Yes	-	Path to the PDB file for Protein A
--protein-b	Yes	-	Path to the PDB file for Protein B
--checkpoint	Yes	-	Path to the trained model checkpoint (.pt)
--output	Yes	-	Directory to save the prediction results
--gpu	No	-1	GPU device ID (e.g., 0). Set to -1 to use CPU.
--threshold	No	0.5	Decision threshold on probability
--max-atoms	No	384	Max atoms kept per protein
--max-residues	No	192	Max residues kept per protein
--max-surface-points	No	128	Max surface points kept per protein
--top-k-interactions	No	15	Top residue-residue pairs to report
--gradcam	No	False	Flag to compute Grad-CAM attributions and generate hotspot visualizations

[!TIP] Performance Note: Generating Grad-CAM attributions (--gradcam) involves a backward pass and extensive visualization rendering. Running without this flag roughly triples the throughput per pair, making it ideal for fast, probability-only initial screening.

Python API Integration

You can integrate ProMaya directly into your Python scripts for custom workflows:

import torch
from src.models.promaya_skeleton import ProMaya
from src.utils import load_config

# Load configuration and model
config = load_config('config/config.yaml')
model = ProMaya(config)
model.load_state_dict(torch.load('best_model.pt'))
model.eval()

# Assuming preprocessed feature tensors are loaded
# protein_a, protein_b = load_features(...)

with torch.no_grad():
    interaction_score, outputs = model(protein_a, protein_b)

probability = torch.sigmoid(interaction_score).item()
print(f"PPI Probability: {probability:.4f}")