ML & DataGroup project · University of Helsinki2026Completed
Cancer Type from Mutational Signatures
Predicting cancer type from whole-genome mutation patterns.
- accuracy, 23 cancer types (MLP)
- 85.9%
- accuracy, 23 cancer types (MLP)
- published reference using more data types
- ~90%
- published reference using more data types
Overview
Group project using the PCAWG dataset (2,771 whole-genome samples). We classified 23 cancer types from trinucleotide mutation profiles and compared SVMs, random forests, XGBoost, an MLP and a 1D CNN. The best model came within about 4 points of a published reference, which also used additional mutation types.
What made the difference
- Biology-informed features: signature activities, mutation burden and ratio indices on top of the 96-channel spectrum.
- SVM baseline 80.5% → MLP with engineered features 85.9% (F1 0.854).
- Reference: Jiao et al. 2020, Nature Communications. They reached ~90% with structural variants, copy-number changes and indels, which we did not use.