• Intallation Guide
• Benchmark Data
• Documentation
• Examples
• License
• Acknowledgements

Dataset construction

• dataset: File with positive peptides inFASTA orCSV file. It can also contain negative peptides in which case the files should contain the labels (0: negative or 1: positive) either in the header (FASTA) or in columnY (CSV).
• --balance: IfTrue, it balances the datasets by oversampling the underrepresented label.
• --autosearch: IfTrue, it searches for negative peptides.
• --autosearch_tags: Comma separated list of tags that may overlap with positive activity that are going to be excluded from the negative peptides.
• --autosearch_proportion: Negative:positive ration when automatically drawing negative controls from the bioactive peptides database (Default: 1.0).

Output

• --outputdirdir: Output directory (Default:./apml_result/apml_result).

Protein Language Model

• --plm: Protein Language Model for computing peptide representations. Available options:esm2-8m,esm2-35m,esm2-150m,esm2-650m,esm2-3b,esm2-15b,esm1b,prot-t5-xxl,prot-t5-xl,protbert,prost-t5. (Default:esm2-8m). Please note: Larger Models might not fit into GPU RAM, if it is necessary for your purposes, please create a new issue.
• --plm_batch_size: Number of peptides for which to batch the PLM computation.(Default: 12).

Dataset Partitioning

• --test_partition: Whether to divide the dataset in train/test splits. (Default:True).
• --test_threshold: Maximum sequence identity allowed between train and test. (Default: 0.3).
• --test_size: Proportion of data to be assigned to evaluation set. (Default: 0.2).
• --test_alignment: Alignment algorithm used for computing sequence identities. Available options:peptides,needle. (Default:peptides).
• --splits: Path to directory with train and test splits. Expected contents:train.csv andtest.csv.
• --val_partition: Whether to divide dataset in train/validation folds.
• --val_method: Method to use for creating train/validation folds. Options available:random,graph-part. (Default:random)
• --val_threshold: Maximum sequence identity allowed between train and validation. (Default: 0.5).
• --val_alignment: Alignment algorithm used for computing sequence identities. Available options:peptidesneedle. (Default:peptides).
• --val_n_folds: Number of folds (Default: 10).
• --folds: Path to directory with train/validation folds. Expected contents:train_{fold}.csv andvalid_{fold}.csv.

Model Selection and Hyperparameter Optimisation

• --config: Name of one of the pre-defined configuration files (seeautopeptideml/data/configs) or path to a custom configuration file (see next section).

Other

• --verbose: Whether to display information about runtime (Default: True).
• --threads: Number of threads to use for parallelization. (Default: Number of cores in the machine).
• --seed: Seed for pseudorandom number generators. Controls stochastic processes. (Default: 42)

• dataset: File with problem peptides inFASTA orCSV file.
• --ensemble: Path to the a file containing a previous AutoPeptideML result.
• --outputdir: Output directory (Default:./apml_predictions).
• --verbose: Whether to display information about runtime (Default: True).
• --threads: Number of threads to use for parallelization. (Default: Number of cores in the machine).
• --plm: Protein Language Model for computing peptide representations. Must be the same as used to train the model. Available options:esm2-8m,esm2-35m,esm2-150m,esm2-650m,esm2-3b,esm2-15b,esm1b,prot-t5-xxl,prot-t5-xl,protbert,prost-t5. (Default:esm2-8m). Please note: Larger Models might not fit into GPU RAM, if it is necessary for your purposes, please create a new issue.
• --plm_batch_size: Number of peptides for which to batch the PLM computation.(Default: 12).

The experiment configuration is a file inJSON format describing the hyperparameter optimisation search space and the composition of the final ensemble. The first level of the file is a dictionary with a single key (ensemble ormodel_selection ormodel_selection) and a list of search spaces for the hyperparameter optimisation. For each model within theensemble list,n different models will be trained one per cross-validation fold; in the case ofmodel_selection, only one of the algorithms will comprise the final ensemble; in the case ofmodel_selection, only one of the algorithms will comprise the final ensemble.

Each experiment requires the following fields:

• model: Defines the ML algorithm. Options:KNearestNeighbours,SVM,RFC,XGBoost,LGBM,MLP, andUniDL4BioPep. More options will be added in subsequent releases and they can be implemented upon request.
• trials: Defines the number of iterations for the hyperparameter optimisation search.
• optimization_metric: Defines the metric that should be used for directing the optimisation search. Always, the metric will be calculated as the average across then cross-validation folds. For the metrics available all of the binary classification within the list in thescikit-learn documentation are supported (Default: Matthew's correlation coefficient, MCC).
• hyperparameter-space: List of dictionaries that defines the hyperparameter search space proper. Each of the dictionaries within correspond to a different hyperparameter and may have the following fields:
  • name: It has to correspond with the corresponding hyperparameter in the model implementation. Most of the simpler ML models use thescikit-learn implementation,LGBM uses the Microsoft implementation (More information onLGBM Repository) andUniDL4BioPep uses the PyTorch implementation (More information onUniDL4BioPep PyTorch Repository), though for this model hyperparameter optimisation is not recommended.
  • type: Defines the type of hyperparameter. Options:int,float, orcategorical.
  • min andmax: Defines the lower and upper bounds of the search space for typesint andfloat.
  • log: Boolean value that defines whether the search should be done in logarithmic space or not. Accelerates searches through vast spaces for example for learning rates (1e-7 to 1). It is not optional.
  • value: Defines the list of options available for a hyperparameter of typecategorical for example types of kernel (linear,rbf,sigmoid) for a Support Vector Machine.

There is an example available in thedefault configuration file.

Please check theCode reference documentation