by-campaign-optimizer

Multi-round protein design campaign parameters are optimized by training a RandomForestRegressor on existing scored designs to calculate feature importances and propose data-driven thresholds, design counts, and exploration regions for metrics like ipsae, iptm, plddt, and rmsd.

Next-round campaign threshold optimization requires at least 10 scored designs to function. Once this minimum is met, the optimizer evaluates metrics like cdr3_length and liabilities to output a ready-to-use YAML config with recommended parameters.

To generate a next-round YAML config from scored protein design data, run the optimizer on your existing metrics like plddt and rmsd. It processes the scores and emits an automated config alongside an audit trail in the optimizer metadata.

Yes, scikit-learn and random forest models are used to recommend exploration regions for protein design. A RandomForestRegressor with fixed hyperparameters computes feature importances to identify the most informative feature space regions for sampling.

Multi-round campaign optimization uses fixed hyperparameters for its RandomForestRegressor, meaning there is no automated hyperparameter tuning. This provides stable feature importance calculations but may limit optimization performance on highly complex or noisy datasets.

Yes, by-campaign-optimizer supports metrics like liabilities and cdr3_length for threshold guidance. It processes these alongside ipsae, iptm, plddt, and rmsd to provide data-driven recommendations for your next-round design parameters.