🧬 Protein Function Prediction Pipeline (CAFA 6)

This repository features a robust, automated machine learning pipeline designed for the CAFA 6 competition. The system predicts biological functions of proteins (Gene Ontology terms) using advanced sequence-based feature extraction and hierarchical post-processing.

🌟 Technical Highlights

Modular Architecture: Built with a decoupled design, allowing independent development and testing of feature extractors, models, and post-processors.

Biological Logic Integration: Unlike standard ML pipelines, this system includes an Ontology Filter that ensures predictions respect the hierarchical structure (parent-child relationships) of Gene Ontology.

Modern Engineering Stack: Powered by uv, the next-generation Python package manager, ensuring reproducible environments and ultra-fast dependency resolution.

Scalable Development Workflow: Includes an automated Sampling Engine (light_debug.py) that creates a 5% representative subset of massive biological datasets for rapid prototyping and local testing.

🛠 The Pipeline Workflow

The project follows a structured 4-stage execution flow:

Feature Engineering: Extracts numerical representations from raw protein FASTA sequences.

Predictive Modeling: Implements an optimized K-Nearest Neighbors (KNN) approach via a custom SklearnWrapper for high-dimensional data.

Domain-Specific Processing: Applies Taxonomy mapping and GO Propagation to refine scores based on biological constraints.

Result Aggregation: Merges multi-source predictions into a standardized format ready for large-scale evaluation.

Project Structure

├── main.py                 # Full pipeline entry
├── light_debug.py          # Debug with sampled data
├── config.yaml             # Pipeline configuration
├── data_paths.py           # Auto file discovery
├── factory.py              # Dynamic step creation
├── runner.py               # Pipeline executor
├── features/
│   └── sequence.py         # Sequence feature extraction
├── models/
│   ├── base.py             # Abstract base step
│   └── generic.py          # Sklearn wrapper
├── processing/
│   ├── ontology_filter.py  # GO term filtering
│   └── merge.py            # Result merging
└── tmp_data/               # Sampled test data (5%)

Pipeline Steps

Step	Class	Description
1	SequenceProcessor	Extract sequence features
2	SklearnWrapper	KNN prediction
3	OntologyFilter	Taxonomy + GO propagation
4	Merge	Combine results

Configuration

Edit config.yaml to modify:

• Model parameters (n_neighbors, metric)
• Filter thresholds (min_score, top_k)
• File paths

Data Files

The pipeline expects these files (auto-discovered via DataPaths):

File	Description
train_sequences.fasta	Training sequences
train_terms.tsv	GO term annotations
train_taxonomy.tsv	Taxon mappings
go-basic.obo	Gene Ontology
testsuperset.fasta	Test sequences
testsuperset-taxon-list.tsv	Test taxon mappings

Local Testing

Generate 5% sampled data for quick testing:

# Data is auto-sampled from Downloads/cafa-6-protein-function-prediction
uv run python sample_real_data.py