How dootmaan/m2ost works
Project Type and Purpose
• This is a research‐oriented deep learning pipeline (not a web app or microservice) for “many-to-one” regression: predicting spatial transcriptomics profiles from digital pathology images.
• It consists of data ingestion/preparation, model definition (M2OST), training orchestration, utility scripts for metrics/visualization, and weight storage.
File Structure Analysis
• Top‐level directories:
• dataset/ – Python Dataset classes for HBC, HER2+, cSCC, plus pre-selected gene arrays (.npy)
• model/ – implementation of the M2OST network (M2OST.py)
• utils/ – helper scripts: FLOPs/parameter counting, feature-selection, visualization (npy→mat, spatial maps)
• weights/ – pretrained model checkpoints (.pth)
• Training entrypoint: train_m2ost_example.py (for HBC; to be adapted per dataset)
• Config/deployment: CUDA_VISIBLE_DEVICES environment variable; log redirection.
Key README Insights
• Describes end-to-end workflow: clone → download/prepare datasets → run training script
• External dependencies: PyTorch, NumPy, Matplotlib (implicit in utils), 7z for encrypted data
• Mentions M2ORT predecessor; weights may be interchangeable after verification.
a. Components
1. External Data Sources
• HBC, HER2+, cSCC repositories (Mendeley, Zenodo, NCBI GEO)
2. Data Preparation Layer
• dataset/*.py classes (HBCDataset, HER2Dataset, cSCCDataset)
• Pre‐selected genes (.npy)
• 7z decryption step (external tool)
3. Feature Selection
• utils/feature_selection_*.py
4. Training Orchestration
• train_m2ost_example.py (scripts)
• Environment: CUDA devices, Python interpreter
5. Model Definition
• model/M2OST.py (PyTorch nn.Module)
6. Utilities
• utils/calculateParamsFLOPs.py
• utils/visualization_*.py
7. Artifact Storage
• weights/*.pth
• Training logs (.log)
• Generated .mat or spatial visualization outputs
b. Relationships & Data Flow
5. “Utilities” can attach to multiple stages:
– After model instantiation: calculateParamsFLOPs
– After training or inference: visualization_npy2mat, visualization_st
c. Architectural Patterns & Principles
• Modular separation: data / model / training / utilities
• Single‐responsibility: each dataset class handles one dataset, each util script one function
• Script‐driven orchestration rather than heavy config frameworks
• Extensibility: add new Dataset subclass or new feature-selection utility without touching core
d. Key Technologies
• PyTorch (training, model)
• NumPy (.npy arrays, pre-selected genes)
• 7z (data decryption)
• Python logging (via STDOUT/redirect in shell)
• Potential plotting libs for visualization (Matplotlib, SciPy I/O for .mat)