How adibvafa/codontransformer works
Project type and purpose
• This is an open-source Python library and command-line tool (CodonTransformer) for codon optimization. It provides both training (pretrain.py, finetune.py, SLURM scripts) and inference (CodonPrediction, CodonJupyter) pipelines. It integrates a pretrained BigBird transformer model from Hugging Face, processes genetic data, and offers evaluation utilities.
File structure
• Top-level scripts and configs:
• pretrain.py, finetune.py (model training)
• Makefile, setup.py, pyproject.toml, requirements.txt (build/test/deploy)
• slurm/*.sh (SLURM job submission for HPC)
• demo/ (sample datasets)
• Core package: CodonTransformer/ subpackages:
• CodonData (data preprocessing)
• CodonPrediction (model loading & inference)
• CodonEvaluation (metrics calculation)
• CodonUtils (constants, mappings, helpers)
• CodonJupyter (notebook widgets & formatting)
• src/ holds shipped artifacts (tokenizer JSON, organism pickles, Excel template)
• tests/ implements unit tests for each subpackage
README for additional insights
• The README outlines use cases: interactive Jupyter/Colab, batch inference template, API usage for predict_dna_sequence, finetuning instructions.
• It highlights dependencies (PyTorch, Transformers, Hugging Face), external data sources (Kazusa, NCBI) and output evaluation metrics.