How evo-design/protein-dpo works
Project type and purpose
• This is an open-source, command-line protein inference and scoring toolkit (ProteinDPO) built in Python.
• It provides two primary functions:
• Sampling new protein sequences given a structure (sample.py)
• Scoring mutant protein sequences against experimental fitness data (score.py, score_complex.py)
File structure
• Top-level scripts:
• sample.py – entry point for structure-conditioned sequence sampling
• score.py – entry point for scoring a single mutation dataset
• score_complex.py – extension for more advanced scoring workflows
• Support files/directories:
• weights/ – holds pretrained model artifacts (ProteinDPO weights, vanilla ESM-IF1 weights)
• example/ – sample CSV inputs
• environment.yml – Conda environment specification
• README.md, LICENSE – documentation and licensing
• No “src” directory: all logic lives in the root scripts or imported modules from the installed esm package.
• External dependency: the ESM inverse-folding library (installed via pip from GitHub).
README for additional insights
• Describes a two-stage pipeline:
a. Sampling (structure → new sequence proposals)
b. Scoring (sequence dataset → likelihood/fitness scores)
• Lists command‐line arguments and expected file formats (PDB, CSV)
• Specifies where to download and place model weights
Detail without over-splitting
• Logical components (script, model, data, env)
• Avoid down-to-the-line-of-code granularity; focus on modules and interactions
• Every input, process, and output is represented clearly