BINN
Summary¶
Hartman, Scott, Karlsson, et al. (2023) present BINN, an open-source Python package that automatically converts the Reactome pathway database into a sparse, biologically informed neural network for a given proteomics dataset, then trains it to classify clinical subphenotypes and uses SHAP (Shapley Additive Explanations) to attribute predictive importance down to individual proteins and pathways. Applied to septic acute kidney injury (AKI) subphenotypes and COVID-19 severity from blood plasma proteomics, and to acute respiratory distress syndrome (ARDS) aetiology from Olink proteomics, BINNs outperformed five standard ML baselines while directly yielding biologically interpretable pathway rankings.
Method¶
- Graph-to-network conversion: the Reactome pathway database's biological entity graph (proteins, pathways, higher-level processes) is subsetted to the proteins present in a given dataset and "layerized" into a sequential sparse neural network — each node retains its biological identity (a specific protein, pathway, or process).
- Small, sparse networks: because connectivity is constrained by curated biology rather than fully connected, resulting networks have only thousands of trainable parameters (AKI-BINN: 6.7k; COVID-BINN: 1.6k, versus millions for typical dense architectures).
- Interpretation: SHAP values quantify each node's contribution to the classification, adjusted by the logarithm of the node's reachable-subgraph size to correct for bias toward highly-connected pathway nodes; results are visualized as Sankey diagrams tracing importance from proteins up through pathways to top-level biological processes.
Performance¶
- Septic AKI subphenotypes (141 samples, 728 proteins): ROC-AUC=0.99±0.00, PR-AUC=0.99±0.00, accuracy=98.6±2%, true-positive/negative rates of 94±2%/100±0%.
- COVID-19 severity subphenotypes (687 samples, 173 proteins): ROC-AUC=0.95±0.01, PR-AUC=0.96±0.01, accuracy=87.5±3%, true-positive/negative rates of 87±2%/92±1%.
- Benchmark comparison: BINNs outperformed SVM (RBF kernel), K-nearest-neighbor, random forest, LightGBM, and XGBoost on both datasets under 3-fold cross-validation — all baselines still achieved AUC>0.75, indicating the tasks were tractable overall, but BINN's biologically constrained sparsity gave it a consistent edge.
- Biological interpretation: SHAP-ranked nodes in the AKI-BINN concentrated on metabolic processes (lipid metabolism, PPAR-α signaling), consistent with sepsis's known systemic metabolic effects beyond pure immune dysregulation, while the COVID-BINN emphasized immune system and programmed cell death pathways.
- Cross-platform generalization: BINN was also applied to Olink-platform proteomics distinguishing ARDS aetiologies, demonstrating the framework generalizes across proteomics quantification technologies (mass-spectrometry-based and affinity-based).
Availability¶
Open-source Python package: github.com/InfectionMedicineProteomics/BINN.
See Also¶
- Biologically Informed Neural Networks — the shared design philosophy across BINN, P-NET, and expiMap.
- P-NET — the earlier cancer-genomics application whose architecture BINN's authors cite as direct inspiration.