Machine Learning Researcher · Seoul
AI Researcher · OMIXAI
Multimodal Learning · Interpretability · Reliable ML
I build machine learning models that hold up on data they were not trained on, such as new hospitals and unseen drugs. I connect different types of data, including images, proteins and RNA, and check where each prediction comes from before trusting it.
I have 5+ years of industry experience in machine learning R&D, spanning research questions, data analysis, model development, experimental design and scientific writing.
My research interests center on learned representations, the basis of model predictions, and the conditions under which results can be trusted. I study methods for connecting different types of data and building interpretable models, with evaluation that accounts for distribution shifts and uncertainty.
Most of this work used biomedical data, where labels are expensive, data shift between hospitals and models must pass regulatory validation. Methods that hold up under those conditions carry over to other domains, and I want to apply them to new data and questions.
Cell-line drug-response (IC50) prediction from proteomics, evaluated leave-drug-out so that no test compound is seen in training (Pearson ≥ 0.65). Separate work covers self-supervised proteomic representation learning with LoRA/PEFT.
Part of a glioblastoma dual-target siRNA program. Matched transcriptome and proteome for 100 CPTAC GBM patients and tested how well RNA expression tracks protein abundance for each gene. Used 1,000-permutation nulls, multiple-testing correction, sensitivity checks and flags for variability, low expression and missingness to build an evidence table for ranking wet-lab candidates.
Within the Pan-Sarcoma proteogenomic profiling program, we combine proteomics, H&E images and RNA from canine tumors provided by the Japanese partner to predict drug response and study molecular characteristics. The dataset began with sarcoma and now covers multiple cancer types.
NVIDIA Korea Agentic AI Hackathon 2026 entry (team of 5). It links pre-market target binding (FlyDiscovery) and post-market adverse-event surveillance (FlyVigilance) in one agent. I built the FlyDiscovery module and its discovery-workbench UI. The module runs structure prediction, docking and binding-affinity scoring with BioNeMo NIM (OpenFold3, DiffDock, Boltz-2).
Built with Team FlyGate (team of 5). The agent turns a visitor's preferences, ingredients to avoid, budget and time into product picks with reasons, a store route and a Korean card to show shop staff. I wrote the core Python workflow, the web demo and its Vercel deployment, the checks against Korean MFDS and Taiwan TFDA ingredient and recall data, and the guards that refuse risky requests and prompt-injected documents. The same code runs inside an NVIDIA OpenShell sandbox on Brev, where the archived run shows normal model calls allowed and file and network access outside the policy denied.
Developed a drug-recommendation algorithm for canine cancer, intended for a clinical decision support system. Co-led OMIXAI's team entry in the Arc Institute Virtual Cell Challenge (predicting CRISPR-knockdown response in pluripotent stem cells).
Predicted 21-gene recurrence-score (Oncotype DX) risk groups from H&E WSIs alone, with no other clinical inputs. Patch predictions were selected by confidence and majority-voted into slide-level low / intermediate / high risk groups.
ADMET prediction and drug–target interaction modeling for early drug-property profiling, and an active-learning framework that defines out-of-distribution pathology samples to make labeling more efficient. Internal R&D, no public results.
2D and 3D segmentation of hemorrhage regions plus slice-level classification. Validated domain generalization on external public data and used class-activation maps to check classification reliability (U-Net vs DeepLab).
Earlier work (2019–2020) on industrial vision, tabular ML, data engineering and segmentation robustness is on GitHub.
A multi-center retrospective study (n=125, National Cancer Center and Korea University Guro Hospital) showing that deep learning on H&E slides alone can approximate the low, intermediate and high risk groups of the 21-gene recurrence assay.
A linear model that groups WSI patches into morphological microstates, finds spatial macrostates from their adjacency, and maps them to gene expression through low-rank molecular programs. It outperformed 14 baselines on three TCGA cancer types and stayed ahead of them when trained on less data. Because the model is linear, each prediction can be decomposed into global and region-specific programs.
A review of current trends and future directions for AI in urological cancer pathology, from clinical, technical and regulatory perspectives.
Official report of the KPIs 2024 Challenge on glomerular segmentation (Medical Image Analysis, vol. 114, 104234). It covers generalization from patch-level training to slide-level inference. I am a co-author as a member of the Deep Bio team, which placed 2nd in the whole-slide track.
H&E-based AI risk scores were used to select regions for spatial proteomics in 156 triple-negative breast cancer patients. High-risk niches showed proliferative programs and low-risk niches showed immune activation. Combining the resulting 13-protein signature with the image-based risk score improved prognostic performance.
An efficient AI analysis that predicts biochemical recurrence from full multi-section specimens rather than from a single slide.
Distills giga-scale pathology foundation models into smaller, cancer-specific encoders. The distilled encoders keep downstream performance while using less compute.
Uses several input resolutions at once and withholds predictions on unreliable regions through a confidence-based reject loss. It was designed for histopathology data with noisy labels.
A lab paper related to my M.S. thesis. It proposes a pixel-wise contrastive loss defined at the region-annotation level, which stabilizes segmentation when medical-image labels are scarce and unreliable.
Seoul · Nov 17–20, 2026. High accuracy does not make an H&E model a substitute for a molecular test. We defined substitution cost as the deviation from a prespecified treatment routing caused by each misclassification, and applied one pre-registered protocol (UNI + CLAM-SB, site-disjoint hold-outs, label-shuffle controls) to five cancers. Of about 15 endpoints, only HPV status in head and neck met the confirmation criterion. Lung histological subtype scored highest (AUROC 0.939) but was excluded after a site-confounding audit. Most of the rest were reported as undecided rather than negative. With Pseudo Lab.
Seoul · Nov 17–20, 2026. We asked whether the per-gene parameters from multiome RNA velocity methods (transcription rate α, degradation rate γ, chromatin-to-transcription lag) are reliable. Five velocity arms were tested on four axes (cross-method reproducibility, an ATAC-shuffle control, replication in five external multiomes, and comparison with measured synthesis and degradation rates). Only α reproduced across methods, and the γ and lag estimates need orthogonal validation.
Orlando · April 2023 · Abstract #5404. Predicted ER, PR and HER2 status from H&E WSIs alone, without IHC. On TCGA-BRCA (728 cases), a multi-task model used a shared morphology encoder, per-receptor heads and confidence-based patch selection. Patches below 70% confidence were dropped. Color augmentation was applied to reduce sensitivity to stain color variation. With Korea University Guro Hospital.
New Orleans · April 12, 2022. Predicted 21-gene recurrence-score risk groups from H&E WSIs in 125 cases. No high-risk case was misclassified as low-risk. This work led to the Scientific Reports (2025) paper.
New Orleans · April 2022. A histology-based survival model trained on TCGA pancreatic adenocarcinoma (179 cases) was applied to rectal and breast cancer. Performance largely held in rectal cancer and dropped in breast cancer.
March 18, 2022. Predicted breast cancer death risk from H&E WSIs (TCGA-BRCA, 980 slides) using features from a pre-trained prostate diagnosis model. High- and low-risk groups separated in Kaplan-Meier analysis.
March 17, 2022. Classified tumor patches as histologic grade 1, 2 or 3 on 125 pathologist-annotated H&E WSIs to reduce inter-observer variability. With Korea University Medicine.
Advised by Prof. Sangheum Hwang at SeoulTech. Added a feature-wise contrastive loss at the encoder bottleneck, which improved U-Net and UNet++ on CT lung and liver segmentation. This work became the basis of my M.S. thesis.
For research collaborations or roles, email me. Publications are on Google Scholar and code on GitHub.