bitser 0.4.0

Binary Pattern Sequence Recognition for biological sequences

BITSER

BInary paTtern SequencE Recognition

Overview

BITSER (Binary Pattern Sequence Recognition) is a software tool built with the Python language that extracts features segments of each genetic sequence at a local level.

The method for feature extraction utilizes the concept of Local Binary Pattern (LBP), as well as adapted versions of the Texture Unit and Texture Unit Number from the field of computer vision, to obtain informative features from sequences organized in FASTA files.

A k-mer window (default size 9) slides over each genetic sequence, comparing the leftmost nucleotide or aminoacid in the window with the 8 other members.

This tool is targeted for usage by biologists, researchers and other professionals in the field of bioinformatics.

Installation

pip install bitser

After the installation, run bitser --help to see all the available commands.

CLI commands

BITSER offers the following commands:

COMMAND	FUNCTION
metadata	Parse FASTA headers and create metadata.tsv with train/test splits
train	Extract features and train a classification model
predict	Load a trained model and predict classes on new sequences

metadata command

This command must be run on your training dataset directory.

The dataset directory must contain a sequences/ subfolder with the FASTA files.

The command scans all .fasta files in the sequences/ subfolder, parses headers to extract class labels, and creates a metadata.tsv file, which is used for the train and predict commands.

Example structure:

dataset/
├── sequences/
│ ├── class_a.fasta
│ └── class_b.fasta

After executing the command:

dataset/
├── sequences/
│ ├── class_a.fasta
│ └── class_b.fasta
└──metadata.tsv

train command

This command initiates the feature extraction and model training workflow.

Training performs the following steps:

1. Feature extraction using sliding windows;
2. Construction of the training feature matrix;
3. Model training using cross-validation;
4. Saving the trained model.

Parameters

Parameter	Description	Required	Default
--input, -i	Path to the dataset directory containing metadata.tsv and sequences/	✔
--output, -o	Path to save the trained model		model.pkl
--classifier, -c	Classifier algorithm		xgb
--flank, -f	Number of neighbors compared to the reference character in the sliding window		8
--translate / --no-translate	Translate nucleotide sequences to proteins before feature extraction		False
--splits, -s	Number of folds used for cross-validation		10
--repeats, -r	Number of cross-validation repetitions		10
--seed	Random seed for reproducibility		7

Output

• Trained model file (.pkl);
• Training evaluation results stored in the results/ directory.

predict command

Uses a trained model to classify sequences from a testing dataset.

Feature extraction settings must match those used during training.

Parameters

Parameter	Description	Required	Default
--model, -m	Path to the trained model file	✔
--data, -d	Dataset directory containing metadata.tsv and sequences/	✔
--flank, -f	Number of neighbors compared to the reference character in the sliding window, must match value used during training		8
--translate / --no-translate	Must match the translation setting used during training		False

Output

The prediction step generates:

• Classification results;
• Per-class performance metrics;
• Feature importance;
• Prediction report (CSV).

Acknowledgements

• This study was supported by national funds through the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) - Finance Code 001, Fundação Araucária (Grant number 035/2019, 138/2021 and NAPI - Bioinformática), CNPq 440412/2022-6 and 408312/2023-8.