Geongyu Lee · 이건규
Geongyu Lee · 이건규 Machine Learning Researcher · Seoul

Geongyu Lee이건규

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.

Research

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 distributions shift across 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.

Research Interests

  • Multimodal & Representation LearningH&E images combined with proteomics and RNA (MoSPR, drug-response models)
  • Interpretability & Explainable AIPredictions decomposed into global and region-specific programs (MoSPR), class-activation-map audits
  • Generalization & RobustnessSite-disjoint hold-outs across hospitals, leave-drug-out splits, site-confounding audits
  • Uncertainty & Reliable EvaluationWithholding predictions on unreliable regions (MurSS), OOD-aware active learning, pre-registered protocols

What I bring to a team

  • From research to product. Contributed to pre-deployment validation and regulatory documentation for a medical-AI product submitted to Korea's MFDS.
  • Data pipelines. Built whole-slide-image pipelines for multi-hospital studies and a data-transfer system with a partner company.
  • Mentoring and review. Guide early-career researchers and project teams as a Codeit mentor, and review for ML4H 2026.

Experience

OMIXAIFeb 2025 – Present
Seoul · formerly RadiSen

AI Researcher

Cross-modal learning · Multi-omics · Drug response
  • Lead ongoing multi-omics R&D for drug-response prediction using proteomics, RNA and H&E, in an 8-person team (4 proteomics, 4 AI).
  • Developed proteomics-based IC50 models, a canine-cancer drug-recommendation algorithm and ADMET models.
  • Planned and ran OMIXAI's Arc Institute Virtual Cell Challenge entry, from approach to experiments.
  • Built a data-transfer system with a partner company.
Details
  • Lead R&D on a multi-omics model that combines proteomics and RNA with H&E pathology to predict drug response (in progress).
  • Built a proteomics-based drug-response (IC50) model for cell lines, with Pearson ≥ 0.65 on unseen drugs (leave-one-drug-out over 381 compounds; internal R&D, unpublished).
  • Developed a drug-recommendation algorithm for canine cancer and ADMET prediction models.
  • Planned and ran OMIXAI's entry in the Arc Institute Virtual Cell Challenge, from approach to experiments (predicting CRISPR-knockdown response in pluripotent stem cells).
  • Co-first author of the MoSPR preprint on gene-expression prediction from H&E (code released); ran about half of the experiments, including the baseline comparisons.
  • Co-authored the G2L paper (AAAI 2026 Workshop W3PHIAI; accepted paper, oral presentation by a co-author) and a 2026 spatial-proteomics preprint on triple-negative breast cancer.
Deep BioMar 2021 – Jan 2025
Seoul · Guro

AI Researcher

Computational pathology · Model validation for an MFDS submission
  • Built WSI pipelines and breast-cancer recurrence models for a two-hospital study.
  • Developed lymph-node metastasis detection and glomerular segmentation models.
  • Led prostate-cancer metastasis and recurrence-risk modeling (internal, unpublished).
  • Contributed to model validation and regulatory documentation for an MFDS submission, and built Docker-based server-automation tools.
Details
  • First-authored a two-hospital breast-cancer recurrence study (Scientific Reports 2025). Built a WSI pipeline that processed 500+ slides.
  • Developed a lymph-node metastasis detection model and, with the Deep Bio team, the glomerular segmentation model for the KPIs 2024 Challenge at MICCAI 2024 (2nd place, whole-slide track). Co-authored the challenge report (Medical Image Analysis, 2026).
  • Led prostate metastasis and recurrence-risk modeling projects, including lymph-node metastasis detection (internal, unpublished).
  • Presented breast-cancer pathology AI work as first author at AACR (2022, 2023) and USCAP (2022). Co-authored posters on survival prediction in pancreatic and breast cancer.
Nuricon2021
Pangyo

Intern

Computer vision
  • Built a parking-lot fire-detection AI system.

Projects

multimodal learning · interpretability · reliable MLcode ↗ github.com/Geongyu
FlyDiscovery workbench: PARP1 × niraparib structure, docking and affinity evidence
Side project · NVIDIA Agentic AI Hackathon 2026 · 1st Place

FlyGate: Drug discovery and safety monitoring

Off-the-shelf NIM checks · OpenFold3 RMSD 1.0 Å Boltz-2 affinity ρ 0.767 live ↗ code ↗

1st place (team of 5) at the NVIDIA Agentic AI Hackathon 2026, run with FAST CAMPUS, after reaching the top-5 final. 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).

Precision oncology · OMIXAI

Veterinary oncology CDSS & Virtual Cell

Canine drug recommendation top-k ≥ 70% · in-house cohort · internal R&D, unpublished CRISPR perturbation

Developed a drug-recommendation algorithm for canine cancer, intended for a clinical decision support system. Planned and ran OMIXAI's entry in the Arc Institute Virtual Cell Challenge (predicting CRISPR-knockdown response in pluripotent stem cells).

Brain CT hemorrhage segmentation and class-activation maps
Medical imaging · SeoulTech × SK / Ajou Univ. Hospital · 2019

Brain-hemorrhage detection on CT

Segmentation Dice 0.66 Slice accuracy 0.90 held up on external RSNA data (674k slices) code ↗

Industry project at SeoulTech's Data Intelligence Lab (Jun–Dec 2019, PI Prof. Sangheum Hwang), where I was one of two main researchers. Built hemorrhage segmentation (U-Net, 279 training / 50 test patients) and a slice-level classifier (EfficientNet, 523 / 143 patients). 2D beat 3D because the slices are 5 mm thick. On the external RSNA intracranial-hemorrhage set the classifier's ranking performance did not drop, and class-activation maps confirmed it looked at the bleed. 80% of missed slices had 100 or fewer labelled pixels, pointing to label noise rather than model error.

H&E WSI tiling to slide-level risk prediction pipeline
Digital pathology · Deep Bio

Oncotype DX risk groups from H&E

Scientific Reports 2025 · first author · n = 125, 2 hospitals Sensitivity L / I / H 0.86 / 0.75 / 0.53 Specificity L / I / H 0.82 / 0.80 / 0.97

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.

OMIXAI · Korea–US–Japan · National Cancer Center Korea

Pan-Sarcoma: Drug response in human and canine tumors

Participating researcher 2025–2028 Ongoing; results not yet public

The Pan-Sarcoma proteogenomic profiling program runs as two parallel studies: human sarcoma with Kyung Hee Univ. Medical Center, and canine tumors with the Japanese partner. In the canine study, we combine proteomics, H&E images and RNA provided by the Japanese partner to predict drug response and study molecular characteristics. The dataset began with sarcoma and now covers multiple cancer types.

Multi-omics · OMIXAI

RNA–protein analysis for dual-target siRNA selection

Genes 11,238 Median Spearman ρ 0.441 CPTAC GBM, n = 100

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.

Proteomics · OMIXAI

Drug-response prediction from proteomics

Internal R&D Unpublished

Cell-line drug-response (IC50) prediction from proteomics, evaluated leave-one-drug-out over 381 compounds (train on 380, test on the held-out drug, repeated for every drug), so no test compound is seen in training (Pearson ≥ 0.65). Separate work covers self-supervised proteomic representation learning with LoRA/PEFT.

More research and side projects
Additional R&D · Deep Bio

OOD-aware active learning for pathology · ADMET / DTI

ADMET DTI OOD detection Active learning

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.

K-BeautyGate architecture: one Python workflow run on a Vercel web demo and inside an NVIDIA OpenShell sandbox on Brev
Side project · NVIDIA Agentic AI Hackathon 2026 · Main round

K-BeautyGate: A sandboxed K-beauty shopping agent for visitors to Korea

Nemotron + rule checks KO / EN / JA / ZH OpenShell bypass attempts denied 4 / 4 live ↗ code ↗

Built by Team FlyGate (team of 5) for the main round of the same hackathon; FlyGate (above) was our entry in the final and took 1st place. 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.

Machine-vision screw localization with circle detection
Computer vision · Fronttec

Screw defect classification

5 classes <0.3s / image

Defect detection on a manufacturing line (969 images, 5 classes) under a hard 0.3s/image limit. Used machine vision (not DL) to locate the screw and remove the background before classification, which improved accuracy.

Tabular ML · Zigbang

Apartment price prediction

1.6M records Metric: RMSE

Transaction-price modeling for Seoul/Busan from 1.6M records. Re-matched addresses by postal code to fix mislabeled districts and added public-API features (nearby subway, schools) that the raw data lacked.

Data engineering · SeoulTech

Music-chart crawling pipeline & database

6 platforms Daily crawl Schema design code ↗

Crawled daily Top-100 charts from 6 streaming platforms (Melon, Genie, VIBE, YouTube Music, Apple Music, Bugs) at a fixed time, merged them on track + artist and loaded them into a relational DB. Nullable rank columns handle tracks entering and leaving the chart, and nvarchar columns store multilingual titles.

Computer vision · SeoulTech

Adversarial robustness for semantic segmentation

Adversarial training Label smoothing Cut-out code ↗

U-Net segmentation on PASCAL VOC (20 classes, 1,928 images, 4,203 objects). Applied adversarial training, label smoothing and cut-out alone and in combination to compare the robustness–accuracy trade-off. My later focus on robustness in medical AI started with this project.

Publications

Journal papers, preprints and workshopsGoogle Scholar ↗
MoSPR: patches to morpho-spatial macrostates, then low-rank regression to gene expression
Preprint 2026 Under review Co-first author †

MoSPR: Histology-to-Gene Expression Prediction with Morpho-Spatial Macrostates and Low-Rank Molecular Programs

Research summary

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. My part: about half of the experiments, including the comparisons against the 14 baselines, shared with the other co-first author, plus part of the appendix and copy-editing. The idea and main text were led by the first and corresponding author.

Gene-wise PCC BRCA 0.413 · KIRC 0.334 · LUAD 0.358 1st of 15 methods in the paper's benchmark
ML4H 2026 Findings Accepted paper · Poster Co-author · presenter Also accepted oral at BIOINFO/GIW ISCB-Asia 2026

A reliability map for per-gene multiome RNA velocity parameters in single-cell kinetics

Research summary

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 methods 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.

α reproducibility ρ = 0.88 γ ρ ≈ −0.1 pre-registered 6 / 6 passed
KPIs 2024 patch-level glomeruli: PAS image, ground-truth mask and overlay
Medical Image Analysis 2026 Challenge report Co-author

KPIs 2024 challenge: Advancing glomerular segmentation from patch- to slide-level

Research summary

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.

Whole-slide track 2nd place Deep Bio team MICCAI 2024
Paper ↗
Tumor patch risk-score heatmap with high- and low-risk patch regions
Preprint 2026 Co-author

AI-guided deep visual proteomics reveals recurrence-associated morpho-molecular signatures within triple-negative breast cancers

Research summary

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.

H&E risk model AUC 0.77 / C-index 0.77 independent test H&E + 13-protein composite OOB C-index 0.679 → 0.739
arXiv ↗
Encoder embedding-space illustration for supervised contrastive learning
IEEE Access 2021 Co-author

Supervised Contrastive Embedding for Medical Image Segmentation

Research summary

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.

Contrastive learning Scarce / noisy labels
IEEE ↗

Conferences

Posters and talksPresentation archive ↗
BIOINFO/GIW ISCB-Asia 2026 Accepted poster Co-first, corresponding & presenting author

Predictability is not substitutability: a cost-of-substitution framework for H&E-based molecular prediction across 5 cancers

Research summary

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 (breast, lung, colorectal, gastric, head and neck). Predictions for lung, gastric and head-and-neck cancer were sealed by commit before any hold-out result existed. Of the 10 sealed endpoints, 3 were positive controls and 7 were candidates: 1 confirmed (HPV status in head and neck, AUROC 0.959) and 6 undecided, each with fewer than the 25 hold-out positives the protocol required. Lung histological subtype, a control, scored highest (AUROC 0.939) but was excluded after a site-confounding audit. Breast and colorectal endpoints were reported but not counted. With Pseudo Lab. My part: designed the evaluation protocol, re-ran and finalized the analysis (including the endometrial extension), and wrote the abstract and manuscript drafts.

Confirmed HNSC HPV · AUROC 0.959 Excluded lung subtype · V(site, label) = 1.0
AACR Annual Meeting 2023 Poster First author

Predicting protein receptor status from H&E-stained images in breast cancer

Research summaryGrad-CAM over H&E for protein-receptor status prediction

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.

Slide-level accuracy · class-balanced test set ER 74.6% · PR 66.0% · HER2 76.6%
Poster ↗
USCAP Annual Meeting 2022 Poster First author

Automatic histological grading of breast cancer resection tissue

Research summaryClass-activation maps for breast histologic grading

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.

Patch one-vs-rest accuracy G1 / G2 / G3 0.958 / 0.922 / 0.956
Poster ↗
KIIE Fall Conference 2020 Oral First author

Utilizing a contrastive loss to improve segmentation model performance

Research summaryU-Net with feature-wise contrastive loss on encoder embeddings

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.

Liver DSC 0.912 → 0.923 Liver ASD 0.533 → 0.365
Slides ↗

Education

SeoulTech2019 – 2021
Seoul

M.S. in Data Science

Seoul National University of Science and Technology (SeoulTech) · Advisor: Prof. Sangheum Hwang
  • Thesis: “Utilizing Feature-wise Contrastive Loss for Improving Segmentation Performance.”
Daejeon University2012 – 2019
Daejeon

B.S. in Information Security

Leadership & Community

community research · reviewing · mentoring · writing
Pseudo Lab2026 · Season 13
Open research community

Falsify: A Red-vs-Blue Benchmark for Scientific Paper Verification

Research member (Runner) · Ongoing (Oct 2026 – Jan 2027)
  • Red teams plant plausible defects (numeric mismatches, fabricated citations, conclusions the evidence does not support) into open-access biomedical papers; blue teams build gates to catch them. Attacks that get through become the benchmark.
  • Built a pilot problem set: a synthetic study with code, data, run logs and Git history, with 40 claims and an answer key.
  • Wrote 11 guide posts for the project on agents, skills, harnesses, MCP and CLI, multi-agent review, control groups and answer-key checks.
Pseudo Lab2026 · Season 12
Open research community

AutoBioX: AI Agents for End-to-End Bio Research

Research member (Runner) · Completed
  • Member of a nine-person team working on AI agents for biological research.
  • Two studies from the project were accepted at BIOINFO/GIW ISCB-Asia 2026, one as a poster (first and presenting author) and one as an oral presentation (co-author). The co-authored study was then accepted to the ML4H 2026 Findings track, where I present the poster.
ML4H2026
Academic service

Reviewer

Machine Learning for Health Symposium
  • Peer reviewer for ML4H 2026 submissions.
Codeit2025 – Present
Part-time

AI Career & Project Mentor

Mentoring · Project guidance · Career coaching
  • Mentored 30+ early-career AI and data professionals one-on-one on résumés, portfolios and career direction. Guided 8+ project teams (4–6 members each) across 3 cohorts, from scoping and method design to review of deliverables.
BOOKK2026
E-books · co-author

Co-author of two Korean books on AI drug discovery

Wrote the AI and mathematics parts
  • 질병의 신호를 읽어라 (Read the Signals of Disease): AI drug-target discovery starting from omics data. ISBN 979-11-12-34045-0.
  • 약이 될 분자를 찾아라 (Find the Molecules That Become Medicines): AI small-molecule drug discovery starting from public data. ISBN 979-11-12-34044-3.
  • Written with seven OMIXAI colleagues as a voluntary after-hours project. I wrote the AI and mathematics parts of both books.

Tools

Deep Learning

PyTorchHuggingFacePEFT / LoRAAccelerateDDP / FSDP

Pathology Stack

OpenSlideQuPathCLAM / weakly-supervised MILUNI · CONCH · VirchowWSI tiling

Bio / Omics

scanpy · AnnDataSingle-cell perturbationProteomic representationADMET

MLOps & Languages

PythonR · SQL · BashDocker · SlurmW&B · MLflowFastAPI · Git

Funded research

Participating researcher

Pan-Sarcoma Proteogenomic Profiling for Precision Oncology

National Cancer Center Korea · Korea–US Cancer Joint Research Program · RS-2025-02243055
Korea–US–Japan · human sarcoma (Kyung Hee Univ. Medical Center) · canine tumors (Japanese partner)
Apr 2025 – Dec 2028 · ongoing · Participating Researcher

Virtual-Cell CDSS for Veterinary Oncology

IPET (Korea Institute of Planning and Evaluation for Technology in Food, Agriculture and Forestry) · Precision diagnostics for major intractable diseases · RS-2026-25527499
2026 – 2030 · ongoing · Participating Researcher

General-Purpose AI for Cancer Pathology Diagnosis

Lead: Deep Bio · National R&D Program
Program Apr 2021 – Dec 2025 · Participating Researcher, Apr 2021 – Jan 2025

Contact

For research collaborations or roles, email me. Publications are on Google Scholar and code on GitHub.