126a0653f0
Move misc/tutorials/ to documentation/ and add index readme. Update rescoring.md: add quick-start examples, paper links for all methods, add Pocketeer to supported methods list. Fix stale links in README.md (tutorials path, local-env.sh typo).
2.7 KiB
Training residue and pocket score transformers
As part of prediction output, P2Rank generates raw scores from the interval [0,∞) for all pockets and residues.
The raw score is a value in score column in *_pockets.csv and *_residues.csv.
Additionally, those files contain columns probability and zscore which contain raw scores transformed by pre-trained transformers.
probability: raw score [0,∞) -> [0,1]zscore: raw score [0,∞) -> (-∞,∞)
These transformers are model/parametrization dependent and need to be trained/calibrated for each model on some calibration dataset. HOLO4K was used for this purpose for the default P2Rank pocket prediction model.
To train transformers for a new model defined in newmodel.groovy execute the following command:
./prank.sh eval-predict -c newmodel.groovy ../p2rank-datasets/holo4k.ds \
-visualizations 0 \
-train_score_transformers '(ProbabilityScoreTransformer,ZscoreTpTransformer)' \
-train_score_transformers_for_residues 1
Output: Parameters of the trained transformers will be stored in score and residue-score subdirectories of the output directory.
It is also possible to train pocket score transformers for rescoring models:
./prank.sh eval-rescore -c rescore_2024.groovy ../p2rank-datasets/holo4k.ds \
-visualizations 0 \
-train_score_transformers '(ProbabilityScoreTransformer,ZscoreTpTransformer)' \
-train_score_transformers_for_residues 0 # not possible for rescoring models yet
After transformers are trained, they should be adequately renamed to reflect an association with the new model and configured in newmodel.groovy:
/**
* Path to a JSON file that contains parameters of a transformer from raw score to "z-score calculated from the distribution of true pockets" (pocket.auxInfo.zScoreTP).
* Use path relative to distro/models/score.
*/
zscoretp_transformer = "newmodel_zscoretp.json"
/**
* Path to a JSON file that contains parameters of a transformer from raw score to "probability that pocket with a given score is true pocket" (pocket.auxInfo.probaTP).
* Use path relative to distro/models/score.
*/
probatp_transformer = "newmodel_probatp.json"
/**
* Path to a JSON file that contains parameters of a transformer from raw score to "z-score calculated from the distribution of all residue scores".
* Use path relative to distro/models/score.
*/
zscoretp_res_transformer = "residue/newmodel_zscore.json"
/**
* Path to a JSON file that contains parameters of a transformer from raw score to "probability that residue with a given score is true (binding) residue".
* Use path relative to distro/models/score.
*/
probatp_res_transformer = "residue/newmodel_proba.json"