Models built for the clinic, not the demo.
Deep-learning research in computational oncology: multi-omic integration and cancer subtyping, evaluated on real patient cohorts.
Multi-omic autoencoder for sarcoma patient stratification
A Transformer autoencoder integrates RNA-seq and 450k DNA-methylation data into a shared latent space. Unsupervised clustering of that space splits 250 patients into two risk groups with significantly different overall survival, a prognostic signal single-omic analysis misses. The separation holds across 100 bootstrap runs (mean ARI 0.92) and stays significant in multivariable Cox regression (HR 1.99, adjusting for age, sex, and histology).
Gated cross-attention network for breast cancer subtyping
A gated cross-attention mechanism fuses gene-expression and mutation features: each modality attends to the other while a learned gate controls how much each contributes. On a held-out test set of 267 patients it classifies the intrinsic molecular subtypes (Luminal A/B, HER2-enriched, Basal) at 92.8% accuracy, and the learned fusion layer separates the subtypes cleanly in t-SNE.
