How byungukp/masif_pmp works
Project Type & Purpose
• A research-grade, command-line bioinformatics toolkit written in Python plus shell/Slurm scripts.
• Implements several MaSIF variants (site prediction, PPI, ensemble learning, ligand binding, PMP/IBS).
• Designed for HPC execution (Slurm scripts) on large protein surface datasets.
High-Level Component Identification
a. Data Storage
– “data/…” directories: raw inputs (PDB lists), TFRecords, precomputed models.
– Disk volumes or shared filesystem on HPC.
b. Orchestration Layer (Shell & Slurm Scripts)
– *.sh and *.slurm scripts under each pipeline: data preparation, training, inference.
– Acts as the glue invoking Python modules on the cluster scheduler.
c. Core Python Library
– source/data_preparation: PDB download, triangulation, TFRecord creation.
– source/geometry, source/gif_descriptors: feature‐computation routines.
– source/input_output: PDB I/O, mesh handling.
– source/default_config: global settings and hyperparameters.
– source/masif_modules: neural network layers, data readers, training wrappers.
– source/masif_site, source/masif_ensemble, source/masif_ligand, source/masif_ppi_search, source/conf_ensemble, source/triangulation, source/pmp_ibs: pipeline‐specific entry points.
d. Machine-Learning Models
– TensorFlow/PyTorch serialized model files under data/.../nn_models and transformation_training_data/models.
e. Visualization & Plugin
– data/*/color_site.sh and source/masif_pymol_plugin: post-hoc coloring of sites in PyMOL.
Relationships & Data Flow
1) Raw PDBs & lists → Data preparation scripts (source/data_preparation) → triangulated meshes + TFRecords.
2) TFRecords → Feature computation (source/geometry, source/gif_descriptors) → descriptor files.
3) Descriptor files + default_config → Model training (source/masif_modules, pipeline-specific trainers) → serialized NN models.
4) Trained models + new structures → Prediction scripts (source/…_predict.py, *.sh) → score files.
Architectural Patterns & Design Principles
• Layered, modular design:
– Bottom layer: data I/O and mesh handling
– Middle: feature computation
– Top: model training/inference
• Separation of concerns: distinct modules for each MaSIF variant
• HPC integration via Slurm scripts external to the library
• Config-driven behavior via default_config
k. Label external dependencies outside boundary: Python runtime, TensorFlow, Slurm scheduler, PyMOL.