Project Overview
This project implements comprehensive FAIRification of the ELIXIR 3D-BioInfo Benchmark for Protein–Protein Interfaces, transforming the original dataset into a FAIR-compliant, machine learning-ready resource with enhanced metadata and interoperability features.
Dataset FAIRification
Converted 1,677 protein interfaces with structured Bioschemas metadata, including Schema.org, EDAM, and Gene Ontology annotations.
- Structured JSON-LD metadata
- Bioschemas Profile 1.0 compliance
- Domain-specific ontology integration
Machine Learning Integration
Complete ML pipeline with structural feature extraction and advanced cross-validation for protein interaction prediction.
- GroupKFold cross-validation
- Structural feature extraction
- Multiple ML model comparison
FAIR Evaluation
Comprehensive FAIR assessment comparing multiple repositories with detailed improvement roadmap.
- Findable: 85/100
- Accessible: 80/100
- Interoperable: 90/100
- Reusable: 95/100
Integrated FAIR + ML Workflow
Our approach combines FAIR data principles with machine learning readiness in a seamless pipeline:
Data Collection
1,677 protein interfaces from ELIXIR 3D-BioInfo benchmark
FAIR Transformation
JSON-LD metadata with Bioschemas and Croissant standards
Feature Extraction
Structural features from PDB/mmCIF files within 10Å distance
ML Pipeline
GroupKFold CV with cluster-aware validation
Evaluation
Comprehensive metrics and feature importance analysis
Key Features & Innovations
MLCommons Croissant 1.0
Standardized ML dataset packaging for seamless integration with machine learning frameworks and automatic data loading.
Structural Feature Extraction
Automatic extraction of interface residues within 10Å distance from PDB/mmCIF files using BioPython.
Cluster-Aware Validation
GroupKFold cross-validation using sequence clusters to prevent data leakage and ensure realistic evaluation.
Comprehensive Evaluation
Multiple ML models (Random Forest, SVM, Logistic Regression) with detailed performance metrics and visualization.
Machine Learning Performance
Summary of model performance from our integrated ML pipeline:
Random Forest
Accuracy: 0.827 (±0.012)
ROC-AUC: 0.898 (±0.008)
Best for: Feature importance analysis and robust performance
Support Vector Machine
Accuracy: 0.812 (±0.015)
ROC-AUC: 0.885 (±0.010)
Best for: High-dimensional feature spaces
Logistic Regression
Accuracy: 0.798 (±0.018)
ROC-AUC: 0.872 (±0.012)
Best for: Interpretable models and fast predictions
Key Insight
Random Forest achieves the best overall performance with 82.7% accuracy and 89.8% ROC-AUC, demonstrating the effectiveness of structural features for PPI classification. The cluster-aware GroupKFold validation ensures these results are generalizable to new protein families.
Quick Start Guide
1. Get the Data
git clone https://github.com/biofold/ppi-benchmark-fair.git cd ppi-benchmark-fair
2. FAIR Metadata
python scripts/ppi_benchmark_fair.py \ --fetch-pdb-metadata \ --output bioschema
3. Run ML Pipeline
python scripts/ppi_ml_croissant.py \ --local bioschema \ --extract-pdb-features
Citation and License
Dataset Reference Publication
Schweke, H., Xu, Q., Cazals, F., et al. (2023). Discriminating physiological from non-physiological interfaces in structures of protein complexes: A community-wide study. Proteomics, 23(17), e2200323. https://doi.org/10.1002/pmic.202200323
License
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.