How fhalab/spgen works
Project type and purpose
• This is a research‐oriented code repository (“SPGen”) that has two main pillars:
– remote_generation: implements and trains sequence‐generation models for signal peptides (Transformer, seq2seq, heuristics).
– experimental_validation: runs downstream analyses (Jupyter notebooks) on the generated sequences and real data to compare performance and produce figures/tables.
• No traditional frontend/backend – it’s a data‐science pipeline with model training + validation.
File structure
• Top‐level:
– experimental_validation/… : Jupyter notebooks, data, scripts for negative‐control analysis, alignment, HMM/VAE/heuristic comparisons, summary statistics, table generation.
– remote_generation/… : two subprojects:
• attention‐is‐all‐you‐need‐pytorch/… : full Transformer implementation (preprocess.py, train.py, translate.py, transformer modules).
• signal_peptide/… : custom code, data and notebooks for one‐hot, tokenization, PyTorch pair training, translator, tools.
– environment YAMLs (validation_env.yml, transformer_env.yml), README.md, LICENSE.
• Patterns: clearly modular: “generation” vs “validation,” each with its own data flow and dependencies.
• Config/build: conda env files, setup.py in the transformer subproject.
README for additional insights
• High‐level description: “generating signal peptide sequences in remote_generation” + “validating experimental results in experimental_validation.”
Key reminders
• Essential data transformations and storage points.
• Environment files as a reproducibility block.