How lpdi-epfl/masif works
Core System Layers and Components
• Data Ingestion & Preprocessing: At the base lie the raw protein structural data and PDB inputs. This module involves external tools (e.g., MSMS, Reduce, PDB2PQR, APBS, BioPython, PyMesh, open3D) used to download protein structures, add protons, compute surfaces and chemical features, and ultimately decompose protein surfaces into patches.
• Deep Learning & Descriptor Computation: Central to the workflow is the neural network module (built with TensorFlow). Here, “geodesic convolution” layers process patches and output surface fingerprint descriptors. This component also branches into sub-applications:
– MaSIF-ligand: For ligand binding pocket prediction.
– MaSIF-site: For identifying protein–protein interaction sites.
– MaSIF-search: For fast scanning and matching protein surfaces.
• Model Training and Evaluation: Includes components for training (with job slurm scripts, GPU support), evaluation (storing outputs such as logits and labels) and caching data (e.g., tfrecords, cached training pairs).
• Visualization & Reporting: The final outputs are visualized via a PyMOL plugin (source/pymol_plugin) and even web assets for rendering images (e.g., molecular surface renderings) and possibly further analysis notebooks.
• Infrastructure & Deployment: Finally, encapsulate the containerized deployment (Docker tutorial) and integration with HPC clusters (SLURM job scripts) for large-scale processing.
Relationships and Data Flow
• Connectors should show how external tools (MSMS, APBS, etc.) are invoked during the data preparation steps.
• Transformation of data into different intermediate representations (PDB files to surface meshes, then to patch files, then to TensorFlow-compatible TFRecords, and finally to neural network predictions).
Architectural Patterns and Technologies
• Pipeline Orchestration & Distributed Execution: Emphasize the use of shell scripts and SLURM files to asynchronously process large datasets on clusters.
• Modular Design: The repository is split into “source”, “data”, “comparison”, “docker_tutorial.md”, and others, showing a modular division between production code (in source/masif_modules, source/masif_site, etc.) and data preparation pipelines.
• Configuration & Reusability: Note the “default_config” directory that provides global variables and options which configure different modules.