Conditional Protein Conformation Generation Pipeline

This document describes an advanced deep learning pipeline for training a conditional generative model. Unlike a standard generative model, this pipeline can be guided by a novel, unseen protein structure to generate new 3D conformations that are consistent with that specific structural template.

The workflow is divided into two main phases:

Training Phase: Learn the fundamental relationships between protein structures and their dynamic conformations using a dataset of known proteins and their structural variants.
Inference Phase: Use the trained models to generate new conformations for a new, unseen protein structure provided by the user.

Core Components

This pipeline uses a set of specialized scripts and configuration files.

Training Scripts

chebnet_conditional_setup.py: The main script for training the foundational autoencoder (HNO Encoder and Decoder). It processes multiple protein systems and their artificially generated structural variants to create a rich training dataset.
- Config: param_conditional.yaml
conditional_diffusion.py: Trains the core generative model. It's a conditional diffusion model that learns how a specific structural signature (z_ref) influences the distribution of dynamic conformations (pooled_embeddings).
- Config: param_conditional_diffusion.yaml

Inference Scripts

generate_from_new_pdb.py: The first step of the inference pipeline. It takes a new PDB file, uses the pre-trained HNO Encoder to calculate its unique z_ref conditioner, and then uses the pre-trained diffusion model to generate a set of corresponding latent embeddings.
- Config: param_generate_from_new.yaml
decode_novel_latents.py: The second inference step. It takes the latent embeddings generated by the previous script and uses the pre-trained Decoder to translate them back into full 3D atom coordinates.
- Config: param_decode_novel.yaml
h5_to_pdb_novel.py: The final utility script. It converts the generated 3D coordinates from the HDF5 format into standard PDB files for visualization.
- Config: No YAML file; configured via command-line arguments.

Workflow: How to Run the Pipeline

Follow these steps in order. Ensure all paths and model parameters in the YAML files are correctly configured before running each script.

Phase 1: Training the Generative Models

This phase only needs to be done once to train the models.

Step 1: Train the Foundational Autoencoder This script trains the HNO Encoder and the Decoder on your dataset of protein systems.

python chebnet_conditional_setup.py --config param_conditional.yaml

Key Outputs:
- checkpoints/hno_checkpoint.pth (Trained Encoder)
- checkpoints/decoder2_checkpoint.pth (Trained Decoder)
- latent_reps/reference_conditioners.h5 (All z_ref conditioners)
- latent_reps/pooled_embeddings.h5 (All dynamic embeddings)

Step 2: Train the Conditional Diffusion Model This script uses the data from Step 1 to train the generative model. It learns the mapping from a z_ref to its corresponding conformations.

python conditional_diffusion.py --config param_conditional_diffusion.yaml

Key Output:
- conditional_diffusion_output/checkpoints/cond_diffusion_latest.pth (Trained Diffusion Model)

Phase 2: Inference - Generating Conformations for a New PDB

Once the models are trained, you can repeat this phase for any new PDB file you want to analyze.

Step 3: Generate Latent Embeddings for the New PDB Provide your new PDB file. This script will generate the latent embeddings for it.

python generate_from_new_pdb.py \
    --config param_generate_from_new.yaml \
    --pdb /path/to/your/new_protein.pdb \
    --output novel_protein_latents.h5

Key Output:
- novel_protein_latents.h5: An HDF5 file containing the novel_z_ref_conditioner and the generated_pooled_embeddings.

Step 4: Decode Latent Embeddings into 3D Coordinates This script converts the latent embeddings from Step 3 into physical coordinates.

python decode_novel_latents.py \
    --config param_decode_novel.yaml \
    --input novel_protein_latents.h5 \
    --output final_coords_for_novel_protein.h5

Key Output:
- final_coords_for_novel_protein.h5: An HDF5 file containing the final generated_coords.

Step 5: Convert Final Coordinates to PDB Files The final step. This script creates the viewable PDB files from the coordinate data.

python h5_to_pdb_novel.py \
    --input final_coords_for_novel_protein.h5 \
    --template_pdb /path/to/your/new_protein.pdb \
    --output_dir generated_pdbs_for_novel_protein/ \
    --max_frames 50

Key Output:
- generated_pdbs_for_novel_protein/: A directory containing the final PDB files (e.g., generated_novel_frame_1.pdb, etc.).