We ask a narrow question carefully: what in a tumor — before treatment, or in how it changes under treatment — actually carries information about how a patient will respond? Our answers come from patient-derived models, clinical biopsies, liquid biopsy and the gut microbiome, and we report the negative results alongside the positive ones.
Two patients receive the same cancer treatment. One tumor disappears; the other barely changes. Our goal is to find the measurement that tells them apart in advance — and to be equally clear about when no such measurement exists in the data at hand.
Getting there takes two things that are usually pursued separately. The first is material: living models of individual patients' tumors, and clinical cohorts measured deeply enough that a finding can be checked against a second, independent layer of evidence. The second is method: models that can combine those layers, and the discipline to audit what they learn from.
A signature that separates responders in one cohort and fails in the next is not a weak biomarker. It is usually a sign that the question was asked at the wrong moment, or of the wrong measurement.
We have spent a decade building the material — patient-derived organoids and cell lines across colorectal, pancreatic, breast and brain tumors, characterized against the tissue they came from. Having built the models, we now use them to ask what they can and cannot predict, and we hold the answers to the same standard either way. When our own evidence shows that a baseline signature does not reproduce, that becomes the finding rather than a result to be reframed.
The same toolkit is not confined to the clinic. Comparative and functional genomics answer questions about environmental bacteria and plant roots just as well, and we take those on where the biology is interesting — because the methods travel even when the subject does not.
Tumors grown outside the patient, as organoids and as continuous cell lines, then characterized against the tissue they came from — which mutations survive the transition, which expression programs hold, and how sensitivity to drugs and radiation can be measured directly. The models are biobanked and distributed, so they outlive any single study.
Our deepest case is rectal cancer, where chemoradiation precedes surgery and the response is graded directly. Using paired biopsies taken before and after therapy, organoids irradiated ex vivo, and a re-analysis of every suitable public cohort, we test where the response signal actually lives — and found that baseline expression carries much less of it than the field assumes.
A tumor leaves traces well beyond itself. We measure the gut and oral microbiome, the peripheral blood transcriptome, and plasma cell-free DNA in the same patients before any treatment, to learn what a first liquid biopsy is really reporting — tumor burden, or the state of the host carrying it.
Cross-attention models that fuse mutations, expression, protein and whole-slide histology to predict drug sensitivity — together with the less glamorous half of the work: checking whether the retrospective outcome labels such models learn from mean what they are assumed to mean, and reporting the drug classes where a simple linear baseline wins.
The same comparative and functional genomics, applied where the biology is interesting rather than clinical: mining livestock-waste metagenomes for bacteria that precipitate carbonate and could turn waste into building material, and reading plant root transcriptomes to test whether a polyphenol acts in planta on the proton pump it inhibits in a test tube.
Soon-Chan Kim works where clinical oncology data meets computational biology. Before joining the University of Suwon in 2025, he spent a decade at the Cancer Research Institute and Korean Cell Line Bank of Seoul National University College of Medicine, establishing and characterizing patient-derived cancer cell lines and organoids across colorectal, pancreatic, breast and brain tumors — work that produced most of his published record.
His current research applies AI-driven multi-omics analysis to treatment response and computational drug-efficacy screening, and he continues to collaborate closely with surgical, radiation-oncology and pathology groups at Seoul National University Hospital. He also serves as Academic Director of the Korean Cell Line Research Foundation, and teaches machine learning at the University of Suwon.
Graduate and undergraduate positions are open. Students who join will be listed here — see Join Us for what the work looks like and who it suits.
Scientific Reports 15, 44165 · 10.1038/s41598-025-27866-1
Breast Cancer Research 27, 66 · 10.1186/s13058-025-02032-7
Scientific Reports 14, 25526 · 10.1038/s41598-024-76737-8
Cellular Oncology 47(5), 1627–1647 · 10.1007/s13402-024-00939-5
Scientific Data 10, 448 · 10.1038/s41597-023-02365-y
Advanced Science 9(5), e2103360 · 10.1002/advs.202103360
Scientific Reports 10, 6801 · 10.1038/s41598-020-63812-z
Pancreas 48(10), 1285–1293 · 10.1097/MPA.0000000000001420
Translational Oncology 12(9), 1185–1195 · 10.1016/j.tranon.2019.05.019
Translational Oncology 11(5), 1232–1243 · 10.1016/j.tranon.2018.07.014
BMC Cancer 18, 594 · 10.1186/s12885-018-4514-3
These are first-author papers. A full list, including collaborative work and book chapters, is on the University of Suwon faculty page and in PubMed.
Pictures of the group, the workspace and the things we get up to away from the screen. This page is new — photographs are being collected and will appear here.
We work with real clinical data, which means the interesting part of the job is rarely the model. It is deciding what a number means: whether two cohorts can be compared, why a sample count changed between analyses, whether a signal survives correction, and when the honest answer is that the data cannot tell.
You do not need a biology degree to start, and you do not need to be a strong programmer to start. You do need to be willing to become both, and to be uncomfortable when a result looks too good.
Projects available now range from organoid and biopsy transcriptomics, to multi-modal deep learning on pathology slides, to comparative genomics of environmental bacteria. We will match the project to what you want to learn.
To apply, email a short note about what you would like to work on and why, with a CV or transcript. Undergraduates interested in a research internship are welcome to write as well.