AutoPeptideML

Main class for handling the automatic development of bioactive peptide ML predictors.

Initialize instance of the AutoPeptideML class

Parameters:

Name	Type	Description	Default
verbose	bool	Whether to output information, defaults to True	True
threads	int	Number of threads to compute parallelise processes, defaults to cpu_count()	cpu_count()
seed	int	Pseudo-random number generator seed. Important for reproducibility, defaults to 42	42

autosearch_negatives(df_pos, positive_tags, proportion=1.0)

Method for searching bioactive databases for peptides

Parameters:

Name	Type	Description	Default
df_pos	DataFrame	DataFrame with positive peptides.	required
positive_tags	List[str]	List of names of bioactivities that may overlap with the target bioactivities.	required
proportion	float	Negative:Positive ration in the new dataset. Defaults to 1.0., defaults to 1.0.	1.0

Returns:

Type	Description
pd.DataFrame	New dataset with both positive and negative peptides.

balance_samples(df)

Oversample the underrepresented class in the DataFrame.

Parameters:

Name	Type	Description	Default
df	DataFrame	DataFrame with positive and negative peptides to be balanced.	required

Returns:

Type	Description
pd.DataFrame	DataFrame with balanced number of positive and negative peptides.

compute_representations(datasets, re)

Use a Protein Representation Model, loaded with the RepresentationEngine class to compute representations for the peptides in the dataasets.

Parameters:

Name	Type	Description	Default
datasets	Dict[str, DataFrame]	dictionary with the dataset partitions as DataFrames. Output from the method train_test_partition.	required
re	RepresentationEngine	class with a Protein Representation Model.	required

Returns:

Type	Description
dict	Dictionary with pd.DataFrame id column as keys and the representation of the sequence column as values.

curate_dataset(dataset, outputdir=None)

Load a DataFrame or use one already loaded and then remove all entries with non-canonical residues or repeated sequences.

Parameters:

Name	Type	Description	Default
dataset	Union[str, DataFrame]	Dataset or path to dataset.	required
outputdir	str	Path were to save the curated dataset, defaults to None	None

Returns:

Type	Description
pd.DataFrame	Curated dataset.

evaluate_model(best_model, test_df, id2rep, outputdir)

Evaluate an ensemble model.

Parameters:

Name	Type	Description	Default
best_model	list	List of models with a predict_proba method.	required
test_df	DataFrame	Evaluation dataset with id, sequence and Y columns.	required
id2rep	dict	Dictionary with keys being the id and the values the peptide representations.	required
outputdir	str	Path were to save the evaluation data.	required

Returns:

Type	Description
pd.DataFrame	Dataset with the evaluation metrics.

hpo_train(config, train_df, id2rep, folds, outputdir, n_jobs=1)

Hyperparameter Optimisation and training.

Parameters:

Name	Type	Description	Default
config	dict	dictionary with hyperparameter search space.	required
train_df	DataFrame	Training dataset with id column and Y column with the bioactivity target.	required
id2rep	dict	Dictionary with pd.DataFrame id column as keys and the representation of the sequence column as values.	required
folds	list	List with the training/validation folds	required
outputdir	str	Path to the directory where information should be saved.	required
n_jobs	int	Number of threads to parallelise the training, defaults to 1.	1

Returns:

Type	Description
list	List with the models that comprise the final ensemble.

train_test_partition(df, threshold=0.3, test_size=0.2, denominator='n_aligned', alignment=None, outputdir='./splits')

Novel homology partitioning algorithm for generating independent hold-out evaluation sets.

This method partitions the provided dataset into training and testing sets based on sequence similarity. It ensures that sequences in the training and testing sets do not exceed a specified sequence identity threshold, resulting in distinct datasets for evaluation.

:example: data = pd.DataFrame({'id': [...], 'sequence': [...], 'Y': [...]}) partitioned_data = train_test_partition( df=data, threshold=0.4, test_size=0.25, denominator='shortest', alignment='needle', outputdir='./data_splits' )

Parameters:

Name	Type	Description	Default
df	DataFrame	Dataset to partition with the following columns: id, sequence, and Y.	required
threshold	float	Maximum sequence identity allowed between sequences in training and evaluation sets. Sequences exceeding this threshold in similarity will not appear in both sets. Defaults to 0.3.	0.3
test_size	float	Proportion of samples in evaluation (test) set. A float between 0 and 1, where 0.2 means 20% of the dataset will be allocated to the test set. Defaults to 0.2.	0.2
denominator	str	Denominator used to calculate sequence identity between pairs of sequences. Options include: - 'shortest': The shortest sequence length. - 'longest': The longest sequence length. - 'n_aligned': The length of the aligned region between sequences.	'n_aligned'
alignment	str	Sequence alignment method to compute similarity. Options include: - 'peptides': Peptide sequence alignment. - 'mmseqs': Local Smith-Waterman alignment. - 'mmseqs+prefilter': Fast alignment using Smith-Waterman with k-mer prefiltering. - 'needle': Global Needleman-Wunsch alignment.	None
outputdir	str	Directory where the resulting train and test CSV files will be saved. Defaults to './splits'.	'./splits'

Returns:

Type	Description
Dict[str, pd.DataFrame]	A dictionary containing the training and testing DataFrames: - 'train': The DataFrame for the training set. - 'test': The DataFrame for the testing set.

Raises:

Type	Description
FileNotFoundError	If the output directory cannot be created or accessed.
ValueError	If an unsupported alignment method is specified.

train_val_partition(df, method='random', threshold=0.5, alignment='peptides', denominator='n_aligned', n_folds=10, outputdir='./folds')

Method for generating n training/validation folds for cross-validation.

Parameters:

Name	Type	Description	Default
df	DataFrame	Training dataset with id, sequence, and Y columns.	required
method	str	Method for generating the folds. Options available: random through sklearn.model_selection.StratifiedKFold or graph-part through graphpart.stratified_k_fold, defaults to random.	'random'
threshold	float	If mode is graph-part, maximum sequence identity allowed between sequences in training and evaluation sets, defaults to 0.5	0.5
denominator	str	Denominator to calculate sequence identity. Options; - shortest: Shortest sequence length - longest: Longest sequence length - n_aligned: Length of the alignment	'n_aligned'
alignment	str	If mode is graph-part, alignment algorithm to use. Options available: mmseqs (local Smith-Waterman alignment), mmseqs+prefilter (local fast alignment Smith-Waterman + k-mer prefiltering), and needle (global Needleman-Wunch alignment), defaults to 'mmseqs+prefiler', defaults to 'mmseqs+prefilter'	'peptides'
n_folds	int	Number of training/validation folds to generate, defaults to 10	10
outputdir	str	Path where data should be saved, defaults to './folds'	'./folds'

Returns:

Type	Description
list	List of training/validation folds