Random Forest classifier to identify genetic interactions
PhD project code for predicting cell line specific synthetic lethality between paralog pairs
Overview
context_specific_SL_prediction is the code accompanying my PhD work on systematic prioritisation of context-specific paralog pair vulnerabilities in cancer. It implements a random forest classifier that predicts, for each of ~33,000 paralog pairs across 1,005 DepMap cancer cell lines, the probability that the pair is synthetic lethal in that specific cell line.
- Repository: github.com/naroddaldal/context_specific_SL_prediction
- Preprint: Kebabci, Ajmal, Adams & Ryan, Systematic prioritisation of context-specific paralog pair vulnerabilities in cancer — currently under review at Genome Medicine
- Matrix of Predictions: Prediction scores generated by the model
What it does
- Integrates cell-line-specific features (paralog expression, essentiality, copy number, mutation status from DepMap) with cell-line-agnostic features (sequence identity, protein–protein interaction and Gene Ontology network features) into a single classifier.
- Provides two modelling strategies: a contextualised model that refines predictions from a prior context-agnostic classifier, and a full model that jointly trains on all features.
- Ships four cross-validation strategies for realistic evaluation — random splits, unseen cell lines, unseen paralog pairs, and unseen cell-line × pair combinations — plus external validation against Klingbeil, Parrish, Harle, and Gonatopoulos-Pournatzis screens.
- Generates 33.5 million cell-line-specific SL predictions, powering disease-stratified prioritisation across 26 cancer types in the web resource.
Why it matters
Paralog synthetic lethality is highly context-specific, but experimentally mapping tens of thousands of pairs across a thousand cell lines is impractical. The full model reaches ROC AUC ≈ 0.87–0.93 across evaluation scenarios, with precision/recall on independent screens matching the reproducibility observed between independent experimental studies — suggesting predictions are as accurate as current screen data allows. Applied to HER2-amplified breast cancer, the model recovers known synergies (EGFR–ERBB2, AKT1–AKT2, ERBB2–ERBB3) and prioritises novel biomarker-associated vulnerabilities that replicate in HER2-amplified gastro-oesophageal models.