What I work on
Research
My research builds machine-learning systems that turn routine clinical data — pathology slides, omics profiles, medical images — into decisions that matter for cancer patients. A few threads run through all of it: learning from weakly-labelled, high-dimensional biological data; building models that generalize across hospitals and scanners; and pairing methodological novelty with honest, well-powered evaluation.
Computational pathology for gynecologic cancer
The core of my work is deep learning that reads H&E whole-slide images to make clinically meaningful calls. With collaborators across UBC and BC Cancer I have built models for ovarian carcinoma histotype diagnosis (Modern Pathology, 2022) and uncovered a previously unrecognized, prognostically distinct subtype of endometrial cancer (Nature Communications, 2024), and extended these systems to predict response to therapies such as PARP inhibitors and bevacizumab. The recurring challenge is generalization: models that hold up across institutions, scanners, and staining protocols rather than overfitting to a single centre.
Foundation models & representation learning
I develop self-supervised and foundation-model approaches that learn reusable representations of gigapixel pathology images and high-dimensional omics — contrastive cell-level representation learning (Volta, Nature Communications 2024), graph-structured whole-slide representations (GRASP, AAAI 2024), and general-purpose representations for bulk RNA-seq (GeneExpert, under review at Nature Biotechnology). I care just as much about rigorous benchmarking: understanding precisely when large pretrained models help and when simpler baselines are enough.
Multimodal & spatial integration
A growing line of work links tissue morphology to molecular state — for example, predicting spatial gene expression directly from histology images (KAFSTExp, IEEE JBHI 2026). The aim is to bridge what a pathologist sees with what sequencing measures, so that inexpensive, routinely-available images can augment or stand in for costly molecular assays.
Cancer genomics & clonal evolution
My earlier work concerns the statistical reconstruction of how tumours evolve. I contributed computational methods for clonal decomposition and phylogenetics (E-Scape, Nature Methods 2017; clonealign, Genome Biology 2019) and for tracking genomic clones from single-cell and circulating-tumour-DNA data (Nature 2015; Nature Communications 2015). Inferring hidden structure and dynamics from noisy biological measurements still shapes how I approach modelling today.
Flagship project
The UBC-OCEAN Challenge. I co-led the organization of UBC-OCEAN, an international Kaggle competition for ovarian-cancer subtype classification and outlier detection on whole-slide histopathology images. It drew machine-learning teams from around the world and assembled one of the largest multi-centre benchmarks in computational pathology; the resulting analysis of what does and doesn’t generalize is under review at Nature Medicine (2025). Organizing the challenge — curating multi-institution data, defining the task and metrics, and synthesizing hundreds of approaches — reflects how I like to work: open, collaborative, and benchmark-driven.
Methods and results are detailed across my publications.