BioSieve

BioSieve is a Python toolkit for preparing biological sequence datasets for machine learning.

It covers two main workflows:

• Redundancy reduction — remove near-duplicate sequences before training using sequence, embedding, descriptor, or structural similarity
• Leakage-aware splitting — partition datasets into train/val/test (or k-folds) with strategies that respect biological structure (homology clusters, sequence distance, groups, time)

Installation

BioSieve supports Python 3.11+.

pip install git+https://github.com/kren-ai-lab/biosieve.git

Install optional extras as needed:

• minhash for approximate Jaccard-based deduplication (minhash_jaccard strategy)
• faiss for GPU-accelerated embedding similarity search (embedding_cosine strategy)

pip install 'biosieve[minhash] @ git+https://github.com/kren-ai-lab/biosieve.git'
pip install 'biosieve[faiss] @ git+https://github.com/kren-ai-lab/biosieve.git'

The mmseqs2 reducer and homology_aware splitter require the MMseqs2 binary to be available in PATH.

Tip

You can install MMSeqs2 easily with pixi: pixi global install -c bioconda -c conda-forge mmseqs2.

Quick Start

Redundancy reduction

Remove near-duplicate sequences using k-mer Jaccard similarity:

biosieve reduce \
  -i dataset.csv \
  -o dataset_nr.csv \
  --strategy kmer_jaccard \
  --mapping-output mapping.csv \
  --report-output report.json

Pass parameters via a YAML file:

biosieve reduce \
  -i dataset.csv \
  -o dataset_nr.csv \
  --strategy kmer_jaccard \
  --params params.yaml

# params.yaml
kmer_jaccard:
  threshold: 0.8
  k: 5

Override a single parameter inline without a file:

biosieve reduce -i dataset.csv -o out.csv --strategy kmer_jaccard --set kmer_jaccard.threshold=0.9

Dataset splitting

Split with a leakage-aware strategy:

biosieve split \
  -i dataset_nr.csv \
  -o splits/ \
  --strategy homology_aware \
  --params params.yaml

# params.yaml
homology_aware:
  mode: precomputed
  clusters_path: clusters.csv
  member_col: id
  cluster_col: cluster_id
  test_size: 0.2

Strategies

Redundancy reduction

Strategy	Description	Extra needed
exact	Remove exact sequence duplicates	—
identity_greedy	Greedy reduction by sequence identity	—
kmer_jaccard	Greedy reduction by k-mer Jaccard similarity	—
minhash_jaccard	Approximate k-mer Jaccard via MinHash LSH (fast)	biosieve[minhash]
embedding_cosine	Cosine similarity on precomputed embeddings	biosieve[faiss] (optional)
descriptor_euclidean	Euclidean distance on numeric descriptor columns	—
structural_distance	Graph-based reduction on precomputed structural edges	—
mmseqs2	Homology clustering via MMseqs2	mmseqs binary

Splitting

Strategy	Description
random	Random train/val/test split
stratified	Stratified by a categorical label column
stratified_numeric	Stratified by a numeric label column (binned)
group	No group appears in more than one split
time	Chronological split by a time column
cluster_aware	Group split using a precomputed cluster column
distance_aware	Test set selected as farthest points in embedding/descriptor space
homology_aware	Group split derived from MMseqs2 clusters or precomputed clusters

All strategies also support k-fold variants (random_kfold, stratified_kfold, group_kfold, stratified_numeric_kfold, distance_aware_kfold).

Outputs

Every run produces consistent artefacts:

• Reduced or split CSVs
• Mapping CSV (removed_id, representative_id, cluster_id, score) for reduction runs
• JSON report with strategy name, effective parameters, and reduction/split statistics

Learn More

• examples/README.md for runnable scripts and config files

License

GPL-3.0-or-later. See LICENSE.

Acknowledgements

Built on top of scikit-learn, pandas, NumPy, and optionally datasketch, FAISS, and MMseqs2.

Developed by KREN AI Lab at Universidad de Magallanes, Chile.