Research

Collaborative science

We work closely with domain experts: clinicians, geneticists and experimentalists. Many current projects concern blood and cancers, for example learning the determinants and effects of tumor phenotypes from high-throughput sequencing, imaging and registry data.

Research themes

Somatic mutagenesis

Deep learning models of the causes and consequences of somatic mutations in cancer.

We model somatic mutagenesis with deep learning. Using data from large cancer sequencing projects (iCAN, Genomics England, ICGC, TCGA), we build integrative models of multilayered data to shed light on the causes and consequences of somatic mutations.

Mutation-Attention (MuAt): deep representation learn…Pan-cancer analysis of whole genomes…Discovery of potential causative mutations in human …BasePlayer muat mutation-attention ionstats

Blood cell phenotypes

Determinants of blood cell phenotype variation, from single-cell imaging and genotyping data.

Together with the Finnish Red Cross Blood Service and collaborators, we work to understand phenotypic variation in blood cells and its determinants. We combine imaging and multiomics data with genotyping and health registers.

DeepIFC: virtual fluorescent labeling of blood cells…Morphological single-cell analysis of peripheral blo…DeepIFC blood-cell-painting-paper

Blood cancers

Germline determinants of hematological malignancies, modeled from high-throughput multiomics data.

We identify germline determinants of hematological malignancies by modeling high-throughput multiomics data, with particular focus on DNA repair deficiencies that contribute to mutational burden in malignancies such as acute leukemias.

Distinct stem cell identities converge into shared e…Pathogenicity evaluation of coding germline variants…Blood has never been thicker: cell-free DNA fragment…PaVaDi ERCC6L2-scRNA

Deep learning for medical imaging

Deep learning models that read pathology slides and brain scans, and how they behave under domain shift.

We develop deep learning models that read medical images: whole-slide prostate histology and amyloid PET scans of the brain. HistoEncoder is a foundation model for prostate tissue, ArcheD predicts cerebrospinal fluid amyloid-beta from PET images, and petVAE identifies amyloid subgroups across the Alzheimer's disease continuum. A second line measures how far accuracy falls when a network trained on one laboratory's slides is applied to another's, and develops training methods that reduce the drop.

HistoEncoder: a digital pathology foundation model f…ArcheD, a residual neural network for prediction of …Spectral decoupling for training transferable neural…HistoEncoder arched petVAE StrongAugment

Funding and support