Bioinformatics and Integrated Omics Systems
A computational biology group at the University of Ulsan College of Medicine. We take genome, transcriptome, proteome, phosphoproteome, single-cell and spatial measurements from the same patients, model them jointly rather than one layer at a time, and then test whether what the model finds survives a cohort it was not trained on.
Research
The work has two halves that feed each other. Characterisation measures what is present in a tumour, at each level, as completely as the material allows. Interpretation asks what those measurements mean together, which of them move under treatment, and whether any of it predicts what happens to the patient.
Characterisation
Mutation, copy number and mutational signatures across a cohort, and how they differ between patient populations.
Tumour, immune and stromal compartments separated out of what bulk tissue averages away.
Where a cell sat in the tissue, and which neighbours it was close enough to signal to.
Protein and phosphoprotein evidence tied back to the genome and transcriptome that produced it.
Interpretation
Models that hold several assays in one space, so a genomic change can be followed into protein.
Primary and recurrent tissue from one patient, read as a sequence rather than two snapshots.
Stratification that has to reappear in a cohort we did not train on before we believe it.
Studies designed with clinicians, released afterwards as software other groups can run.
Methods
Three shapes most of our projects end up taking. Drawn for this page as keys; published examples follow below.
Publications
Selected papers. Figures are reproduced under CC BY where the article licence allows it.
An integrative proteogenomic analysis of longitudinal glioblastoma pairs. A highly proliferative state at diagnosis gives way, at recurrence, to neuronal transition and synaptogenic pathways, marked by post-translational activation of WNT/PCP signalling and BRAF kinase.
Single-nuclei sequencing of matched primary and recurrent meningiomas, following how tumour cells change between operations and how they interact with the immune cells around them.
IGFBP5 identified as a ligand for ROR1 that drives invasion of glioblastoma stem-like cells. The screen that found it intersects genes up in invasive stem-like cells with genes prognostic in TCGA lower-grade glioma.
Mutation, copy number, RNA, protein and phosphoprotein data from the same tumours, organised into seven molecular subtypes of medulloblastoma.
Brain · 2021
Radiogenomics
imaging and expression read together, to find what drives infiltration along myelin
A radiogenomic screen for genes that govern how glioma cells infiltrate along myelin, which identified the Nogo receptor NgR1 as a modulator of myelin-associated infiltration.
The genomic landscape of PTEN alteration in a Samsung Medical Center glioblastoma cohort, validated in TCGA, and the functional behaviour of its hotspot missense mutations.
Glioblastoma stem cells from 69 patients, classified by temozolomide response in the dish; the classification separated progression-free and overall survival in the patients the cells came from.
Also
Projects
Three questions the group is pursuing, each growing out of published work. The drawings are schematics of the question, not results.
Project 01 · Tumour evolution
A recurrent tumour is not a new tumour. Which of its features were selected by the treatment, and which were already there at diagnosis, below what the first assay could see?
Project 02 · Single-cell and spatial
Counting cell types tells you what is present. It does not tell you that the immune cells never got past the rim — for that, the coordinates have to be kept.
Project 03 · Precision oncology
Almost any integrated model separates the patients it was fitted on. The part worth publishing is the run against held-out data, and that run happens once.
Software
A method that cannot be run by somebody else is not finished. Our tools are released for other groups to use.
Repositories at github.com/ablancsong. Shareable processed data go to GEO and Zenodo; patient-derived data follow the approved governance route.
Working stack
Nobody is expected to arrive knowing this. The first months of a project are spent learning the parts of it that project needs.
People
Principal investigator
Works on integrating multi-omics cancer data and getting it to say something a clinician can use — from the genome of glioblastoma in Korean patients to the proteogenomics of medulloblastoma and the single-cell history of meningioma.
Joining
If the work above is the kind of thing you want to spend a few years on, please write.
The group takes students from biology, from statistics and from computer science, and it does not much matter which door you come through; whichever half you are missing gets learned in the first year, with help. What is hard to teach is the reflex to stop when a number looks too good and go back to find out why, so that is mostly what I look for.
In practice you would spend your first months on one dataset, learning the tools that dataset needs, and by the end of the year you would own a question that is yours rather than mine.
There is no deadline and no form. Send an email with three things: the part of the research that interests you and why, one analysis you did yourself including what went wrong with it, and a CV or transcript. Every message gets an answer.
— Harim Koo
Open now
Write to
Find us
울산대학교병원 아산의학관
University of Ulsan College of Medicine
Collaborations
Clinical and computational groups working on CNS tumours, gastric cancer and proteogenomics.
Open science
Code under an open licence, data deposited on publication, and preprints posted before review.