Synthetic Lethality Engine
Identifies druggable vulnerabilities from mutation-pathway interactions
Curated synthetic lethal gene pairs with drug mappings
All unit tests passing including edge cases
Pathway norms calibrated on 9,415 TCGA samples (p95=4.221)
Genes with functional essentiality data from CRISPR screens
Overview
Identifies synthetic lethal drug opportunities by combining patient mutation profiles with pathway context and functional genomics data. When a tumor has lost a specific gene function, certain other genes become essential for survival — creating a druggable vulnerability. The engine classifies mutations as loss-of-function or gain-of-function, scores pathway context using calibrated VAE norms, incorporates DepMap essentiality data for 36 key genes, and filters by cancer-type-specific evidence. Includes structured abstention for insufficient evidence.
Inputs
3 inputsSomatic mutationsVariablePatient mutation list with gene names and alteration types
z_pathway200 (50×4)Calibrated pathway activity scores (z-score + sigmoid normalized)
DepMap essentiality36 genesFunctional essentiality scores from genome-wide CRISPR screens
Outputs
3 outputsSL opportunitiesRanked listRanked synthetic lethal drug opportunities with composite scores
Evidence tiersPer opportunityEvidence level (TIER_1/TIER_2/TIER_3) based on data support
GDSC sensitivityPer drugCell-line drug sensitivity evidence from GDSC
Mathematical Formulation
Weighted combination of SL evidence, pathway context, essentiality, and drug sensitivity
Z-score + sigmoid normalization of pathway L2 norms
Key Features
- LOF/GOF mutation classification for alteration-type-aware scoring
- Calibrated pathway context via empirical TCGA norm statistics
- DepMap CRISPR essentiality integration for target importance
- Cancer-type-specific filtering (validated pairs only)
- Structured abstention when evidence is insufficient
- GDSC drug target mapping for actionability
Key Innovations
- 1VAE-derived pathway context for mutation-specific vulnerability scoring
- 2Empirical calibration (z-score + sigmoid) preserves directionality
- 3Multi-source evidence integration (mutations + pathways + essentiality + drug sensitivity)
- 4Abstention-aware output prevents overconfident recommendations
Hyperparameters
28 curated4 per hallmark (50 hallmarks)TCGA n=9,415, p95=4.221Essentiality score weightingTIER_1 (clinical), TIER_2 (preclinical), TIER_3 (computational)Training Details
Not a trained model — uses curated synthetic lethal gene pair database, calibrated VAE pathway norms (TCGA n=9,415), DepMap CRISPR essentiality scores, and GDSC drug sensitivity data. Cancer-type filtering based on evidence annotations.