sxlaep 1.0.4

sxLaep: Rapid Enzyme/Non-Enzyme Prediction using sequence features and XGBoost

sxLaep: a Lightweight and Accurate Enzyme Predictor

sxLaep is an efficient machine learning tool for predicting whether a protein sequence is an enzyme or non-enzyme. It combines multi-physicochemical sequence features with an XGBoost classifier for fast and accurate binary classification.

QUICK LINKS

GitHub Repository

API Reference

cmdline workflow example

INTRODUCTION

In bioinformatics, accurate identification of enzymes from protein sequences is crucial for understanding metabolic pathways, drug target discovery, and functional annotation of novel proteins. Traditional experimental methods for enzyme classification are time-consuming and expensive, making computational prediction approaches highly valuable.

sxLaep provides a machine learning-based solution for enzyme/non-enzyme classification from protein sequences. The tool extracts comprehensive sequence-derived features that capture various physicochemical properties and structural characteristics of proteins:

• Multi-property Pseudo-Amino Acid Composition (Pse-AAC): Captures sequence-order information through hydrophobicity, polarity, and charge properties
• Composition-Transition-Distribution (CTD): Describes amino acid distribution across different physicochemical groups
• Windowed Amino Acid Composition: Captures local sequence patterns by dividing sequences into N-to-C segments
• Sequence Length: Raw protein length as a complementary feature

These features are fed into a pre-trained XGBoost classifier that has been optimized for performance on enzyme classification tasks. The model achieves high accuracy while maintaining fast inference speed, making it suitable for large-scale protein annotation projects.

sxLaep supports both single-sequence prediction and batch processing of FASTA files, with built-in parallelization for efficient feature extraction. The package provides both a simple Python API and a command-line interface for flexible integration into various bioinformatics workflows.

METHODS

Figure 1. The sxLaep pre-screening workflow. The framework acts as a lightweight ”gatekeeper” for massive protein datasets. It extracts alignment-free physicochemical features and utilizes a machine learning classifier to rapidly filter out non-enzymatic sequences. This process significantly reduces computational load for downstream analysis while effectively retaining potential enzymes, including remote homologs (”Functional Dark Matter”)

SOFTWARE

sxLaep (Lightweight and Accurate Enzyme Predictor)

The package is available for Python 3.9+ on Windows, Linux, and macOS.

Latest updates and source code are available at: https://github.com/labxscut/sxLaep

DOCKER

A pre-configured Docker image is provided to run sxLaep in an isolated environment with Python 3.11 and all required dependencies.

Quick Start

Pull the official image directly:

docker pull labxscut/sxlaep:python3.11

Running the Container

Mount your local data directory to the container's working directory (/workspace) and start an interactive shell:

docker run --rm -it -v $(pwd):/workspace labxscut/sxlaep:python3.11 bash

Once inside the container, the sxlaep CLI is ready to use (the sxlaep conda environment is auto-activated).

Build Locally (Optional)

If you prefer to build the image from the provided Dockerfile:

docker build --no-cache -t labxscut/sxlaep:python3.11 .

Then run it using the docker run command above.

INSTALL

Requires Python 3.9+. Dependencies (numpy, pandas, scikit-learn, xgboost, joblib) are pulled in automatically when you install the package.

Recommended: pipx (CLI on your PATH, no project venv)

pipx installs the app into its own isolated environment and links the sxlaep executable onto your PATH—you do not need to create or activate a virtualenv for normal command-line use.

Install pipx once (if needed), and add it into PATH:

pipx ensurepath

Then:

pipx install sxlaep --pip-args="--prefer-binary -v"

If downloads are slow, add an index URL inside --pip-args, for example:

pipx install sxlaep --pip-args="-i https://pypi.tuna.tsinghua.edu.cn/simple --trusted-host pypi.tuna.tsinghua.edu.cn --prefer-binary -v"

Upgrade or reinstall:

pipx upgrade sxlaep
# or: pipx reinstall sxlaep

Optional: pipx install from GitHub (latest main)

Use this when you need commits not yet published to PyPI. You must have git installed and on your PATH (pip clones the repository).

This path is more fragile than PyPI: pipx may appear to hang on determining package name while it installs into a temporary virtual environment to discover the distribution name, then installs again into the real pipx venv. Dependency resolution and xgboost wheels or source builds can take many minutes. Use --verbose so you can see progress.

export PIP_DEFAULT_TIMEOUT=120
pipx install --verbose --force "sxlaep @ git+https://github.com/labxscut/sxLaep.git@main" \
  --pip-args="--prefer-binary -v --default-timeout=120"

To track a different branch or tag, change the trailing @main (for example @develop or @v1.0.0).

Equivalent VCS URL form (some users prefer this):

pipx install --verbose --force "git+https://github.com/labxscut/sxLaep.git@main" \
  --pip-args="--prefer-binary -v --default-timeout=120"

To refresh an existing git-based install, re-run the same command with --force. To go back to the PyPI build, use pipx uninstall sxlaep then pipx install sxlaep as in the recommended section above.

Development from a clone (editable install, still isolated like other pipx apps):

git clone https://github.com/labxscut/sxLaep.git
cd sxLaep
pipx install -e . --pip-args="--prefer-binary -v"

Alternative: pip in a virtual environment (library / notebooks)

Use this when you import sxlaep inside your own project or Jupyter and want dependencies in your venv—not required for the standalone CLI if you used pipx above.

Use a recent pip, then install with binary preference:

python3 -m pip install -U "pip>=24" setuptools wheel
python3 -m pip install --prefer-binary sxlaep

Run the same commands after activating whichever venv you use for that project.

Editable install (optional, for development without pipx):

git clone https://github.com/labxscut/sxLaep.git
cd sxLaep
python3 -m pip install -U "pip>=24" setuptools wheel
python3 -m pip install --prefer-binary -e .

If install is slow or hangs

• First-time resolution and wheel download can take several minutes; keep -v so you see progress.
• pipx install … @ git+https://… (GitHub method above): pipx may sit on determining package name while it clones the repo and runs pip in a temporary venv, then repeats work in the final venv—use --verbose and a generous PIP_DEFAULT_TIMEOUT / --default-timeout (as in the examples).
• Prefer wheels with --prefer-binary (commands above) to avoid long source builds of xgboost.
• On Debian/Ubuntu, if xgboost still builds from source, install compilers and CMake, then retry:
  sudo apt-get install -y build-essential cmake ninja-build

Python API (after install)

See Quick start below for load_model / predict_sequences and run_prediction_pipeline.

QUICK START

Python API

from pathlib import Path

import sxlaep
from sxlaep.model import load_model, predict_sequences

ubj = Path(sxlaep.__file__).resolve().parent / "enzyme_xgb_model.ubj"
model = load_model(ubj)
df = predict_sequences(model, ["MKVLWALIFLLKSAF"])
# columns: pred_label, enzyme_probability

FASTA to CSV in one call:

from sxlaep import run_prediction_pipeline

run_prediction_pipeline(
    model_path="path/to/model.pkl",  # or .ubj / .joblib
    fasta_path="sequences.fasta",
    output_csv="predictions.csv",
)

Command line

Training:

sxlaep train --noenzyme-fasta neg.fasta --enzyme-fasta pos.fasta --outdir results/sxlaep_training

Prediction (writes a CSV):

sxlaep predict --model path/to/model.pkl --fasta sequences.fasta --output predictions.csv

Shorthand (bundled enzyme_xgb_model.ubj in the installed package; no train / predict subcommand):

sxlaep --input sequences.fasta --output predictions.csv
# or: sxlaep -i sequences.fasta -o predictions.csv
# default output if -o omitted: sxlaep_predictions.csv in the current directory

Help:

sxlaep --help
sxlaep train --help
sxlaep predict --help

EXECUTABLES

• sxlaep — CLI: subcommands train and predict, or shorthand sxlaep --input / -i (bundled model) with --output / -o (see Quick start). Use --help for global and subcommand options.

From a clone, optional end-user helper:

• Example FASTAs: tests/enzyme_example.fasta, tests/noenzyme_example.fasta.
• Run cd tests && ./install.sh: downloads missing FASTAs when needed, pipx install or pipx upgrade from PyPI, then runs sxlaep --help and sxlaep --input example.fasta (copy of enzyme_example.fasta in a temp directory), prints a short CSV preview, and reports success (set SXLAEP_SKIP_CLI_SMOKE=1 to skip). No sudo. Override raw URLs with SXLAEP_RAW_BASE / SXLAEP_RAW_REF (default main).
• Developers: from repo root, run pytest tests/.

REFERENCE PARAMETERS

Feature Extraction Parameters

Parameter	Type	Default	Description
lag	int	10	Maximum correlation lag for pseudo-AAC. Controls the length of sequence-order correlation terms. Higher values capture longer-range dependencies but increase feature dimensionality.
weight	float	0.05	Weight applied to sequence-order correlation terms in pseudo-AAC. Balances between standard amino acid composition and sequence-order information.
n_segments	int	3	Number of N-to-C sequence windows for windowed AAC. Divides the sequence into equal segments for local composition analysis.
add_length	bool	True	Whether to append raw sequence length to the feature vector. Provides additional size information that may be correlated with enzyme function.
properties	dict	PROPERTIES	Amino-acid physicochemical property tables used by pseudo-AAC. Default includes: hydrophobicity (HYDRO), polarity (POLAR), and charge (CHARGE).

Command line arguments

Shorthand sxlaep --input (bundled model)

Argument	Required	Description
--input, -i	Yes	Input FASTA path (uses packaged enzyme_xgb_model.ubj). Do not combine with train or predict.
--output, -o	No	Output CSV path (default sxlaep_predictions.csv in the current working directory).
--lag	No	Pseudo-AAC lag — fixed to 10 for bundled model; use sxlaep predict to customize.
--weight	No	Pseudo-AAC weight — fixed to 0.05 for bundled model; use sxlaep predict to customize.
--segments	No	Window-AAC segments — fixed to 3 for bundled model; use sxlaep predict to customize.
--add-length / --no-add-length	No	Append sequence length — fixed for bundled model; use sxlaep predict to customize.
--properties	No	Physicochemical properties — fixed to hydro polar charge for bundled model; use sxlaep predict to customize.

sxlaep train

Argument	Required	Description
--noenzyme-fasta	Yes	FASTA path for non-enzyme (negative) sequences.
--enzyme-fasta	Yes	FASTA path for enzyme (positive) sequences.
--outdir	No	Training output directory (default results/sxlaep_training).
--lag	No	Pseudo-AAC max correlation lag (default 10).
--weight	No	Pseudo-AAC sequence-order weight (default 0.05).
--segments	No	Number of window-AAC N-to-C segments (default 3).
--add-length / --no-add-length	No	Append raw sequence length to feature vector (default: enabled).
--properties	No	Physicochemical properties for pseudo-AAC: one or more of hydro, polar, charge (default: all three).
--test-size	No	Held-out test fraction (default 0.1).
--seed	No	Random seed for reproducibility (default 42).
--n-jobs	No	Parallel workers for feature extraction (default -1 = all cores).

sxlaep predict

Argument	Required	Description
--model	Yes	Trained model path (.pkl / .joblib / native .ubj as supported by load_model).
--fasta	Yes	Input FASTA path.
--output	No	Output CSV path (default results/predictions.csv).
--lag	No	Pseudo-AAC max correlation lag — must match training (default 10).
--weight	No	Pseudo-AAC sequence-order weight — must match training (default 0.05).
--segments	No	Number of window-AAC segments — must match training (default 3).
--add-length / --no-add-length	No	Append sequence length — must match training (default: enabled).
--properties	No	Physicochemical properties — must match training; one or more of hydro, polar, charge (default: all three).
--n-jobs	No	Parallel workers for feature extraction (default 1).

Output file formats

CSV output (sxlaep predict or sxlaep --input)

The prediction CSV includes (among others) identifier columns plus scores:

Column	Description
sequence_id	FASTA record id (first token of the header).
description	Full header line after >.
pred_label	0 = non-enzyme, 1 = enzyme.
enzyme_probability	Estimated probability of enzyme class (when available from the model).

EXAMPLES

Batch FASTA processing (Python)

from sxlaep import run_prediction_pipeline

df = run_prediction_pipeline(
    model_path="enzyme_xgb_model.pkl",
    fasta_path="proteins.fasta",
    output_csv="predictions.csv",
)
print(df[["sequence_id", "pred_label", "enzyme_probability"]].head())

Same task from the shell

sxlaep predict --model enzyme_xgb_model.pkl --fasta proteins.fasta --output predictions.csv

With the bundled model (after pipx install sxlaep, from any directory):

sxlaep --input proteins.fasta --output predictions.csv

NOTES

• The pre-trained model is shipped as enzyme_xgb_model.ubj (XGBoost native format). Training still writes a sibling .ubj when you save *.pkl for backward compatibility.
• The Python package name for installation is sxlaep (for example: python3 -m pip install --prefer-binary sxlaep).
• Sequences containing non-standard amino acids are automatically sanitized (only the 20 standard amino acids are used).
• Short sequences (< 10 amino acids) may yield less reliable predictions.
• For large-scale predictions (>10,000 sequences), consider processing in batches or using a high-performance computing environment.

CONTACT

[e-mail] [e-mail] GitHub: https://github.com/labxscut

CITATIONS

Please cite sxLaep if you use it in your research:

labxscut. sxLaep: a Lightweight and Accurate Enzyme Predictor for High-throughput Mining of Metagenomic Sequences. (2026).