A narrative review of how exposure in the built environment affects human health. Zenodo, DOI 10.5281/zenodo.22012930.
Beyond Tissue of Origin: Quantifying the Added Predictive Value of Genomic Features for Anti-Cancer Drug Response
An end-to-end machine-learning study of 8 cancer drugs across ~700 cell lines, joining GDSC drug-response data (Wellcome Sanger) with DepMap genetics (Broad Institute). The question: does reading a cancer’s genes predict how well a drug will work, better than knowing which organ the cancer came from?
- Result
- Genes beat the organ-only baseline for 7 of 8 drugs, with correlations of 0.42–0.76 on a locked test set. The models also rediscovered known biomarkers unprompted: BRAF for dabrafenib, TP53 for Nutlin-3a, BCL2 for venetoclax.
- Rigour
- Locked 20% test set, no leakage, bootstrap 95% intervals, and SHAP explanations checked against known biology.
- Built with
- Python · LightGBM · SHAP