How hyukjunlim/cheapnet works
Project Type and Purpose
• CheapNet is a research-oriented, full-stack machine learning project aimed at protein–ligand binding affinity prediction using a novel cross-attention mechanism on hierarchical representations.
• The README clearly explains the primary goal: integrating atom-level representations with cluster-level interactions to predict binding affinity efficiently. The design also highlights efficiency and high predictive performance.
Analyzing the File Structure
• The repository is organized around a core module (atom3d) and several experiment/evaluation directories (cross_dataset, diverse_protein, lep).
• datasets: Code for handling data ingestion, preprocessing, and LMDB-related manipulations.
• filters: Contains scripts for filtering and processing protein structure data (PDB, SCOP, sequence-based filters).
• models: Implements different neural network architectures (CNN, ENN, GNN, MLP) used within the binding affinity prediction model.
• protein: Manages protein-specific functionalities (e.g., SCOP and sequence representations).
• splits: Likely used to manage data splits based on sequence or other criteria.
• util: Offers helper functions for logging, file operations, ensemble methods, graph-related operations, transforms, voxelization, and more.
• The additional top-level subdirectories (cross_dataset, diverse_protein, lep) represent different evaluation scenarios or adaptations of the core model into varying datasets. They include training, prediction, evaluation scripts, logging, model weights, and configuration files.
• Other important files (e.g., environment YAML files, Makefile, README, configuration files) indicate how to deploy and run the experiments.
Instructing the Principal Software Engineer
• Input Data flows to Data Preprocessing.
• Preprocessed Data feeds into the Model Architecture.
• With brief notes:
• That the “cross-attention” mechanism is central to the Model Architecture module.
• Legends for abbreviations like CNN, GNN, ENN, etc.
• The core processing modules in the atom3d package, detailing data handling (datasets, filters, splits), modeling (models, protein representations), and supporting utilities.
• The separate evaluation/training modules (cross_dataset, diverse_protein, lep) as distinct clients interacting with the core engine.
• The overall data pipeline from preprocessing through to prediction and evaluation, including key mechanisms (hierarchical aggregation and cross-attention).