Machine Learning Researcher · Cross-modal learning · Reliable prediction · Computational Pathology × Multi-Omics · Seoul
Portfolio: https://geongyu.github.io (English · 한국어 · 日本語)
I develop machine learning models that predict molecular measurements from pathology slides and, in my current industry work, drug response from proteomics. I focus on whether these predictions remain valid at new hospitals and for drugs not seen in training.
I have 5+ years of industry R&D experience. I am an AI Researcher at OMIXAI (formerly RadiSen), where I lead R&D on a multi-omics model that combines proteomics, RNA and H&E to predict drug response (in progress). Before that, at Deep Bio, I built a WSI pipeline and contributed to pre-deployment model validation and regulatory documentation for a medical-AI product submitted to Korea's MFDS.
- MoSPR (preprint 2026, co-first author). Predicts gene expression from H&E using morpho-spatial macrostates and low-rank molecular programs. Ranked 1st of 15 methods in the paper's benchmark on three TCGA cancers (gene-wise PCC, BRCA 0.413). arXiv · code
- Predictability is not substitutability (BIOINFO/GIW ISCB-Asia 2026, accepted poster, first author). One pre-registered protocol with site-disjoint hold-outs and label-shuffle controls was applied to 5 cancers. Of ~15 endpoints, only HNSC HPV status (AUROC 0.959) met the pre-registered confirmation criterion. Most of the others were reported as inconclusive.
- Scientific Reports 2025 (first author). Prediction of 21-gene recurrence-assay risk groups in early-stage breast cancer from H&E alone (n=125, 2 hospitals). Sensitivity L / I / H was 0.86 / 0.75 / 0.53. paper
- G2L (AAAI 2026 Workshop W3PHIAI, oral, co-author). Distillation of giga-scale pathology foundation models into cancer-specific models. arXiv
- KPIs 2024 Challenge (MICCAI 2024). The Deep Bio team placed 2nd in the whole-slide track (glomerular segmentation). Co-author of the challenge report in Medical Image Analysis 2026.
- OMIXAI. Proteomics drug-response model evaluated leave-drug-out, so no test compound is seen in training (Pearson ≥ 0.65; internal R&D, unpublished). Participating researcher in a Korea–Japan Pan-Sarcoma proteogenomics collaboration. Co-led OMIXAI's team entry in the Arc Institute Virtual Cell Challenge.
- Reviewer, ML4H 2026.
† equal contribution (co-first). Full list on Google Scholar. Conference talks and posters are on the portfolio.
- Deep learning: PyTorch · HuggingFace Transformers · PEFT / LoRA · Accelerate · DDP / FSDP
- Pathology: OpenSlide · QuPath · CLAM / weakly-supervised MIL · UNI · CONCH · Virchow · WSI tiling pipelines
- Bio / omics: scanpy · AnnData · proteomic representation learning · single-cell perturbation · ADMET modeling
- MLOps: Docker · Slurm · W&B · MLflow · FastAPI · Git
- M.S. Data Science, Seoul National University of Science and Technology (SeoulTech), 2019 – 2021
- B.S. Information Security, Daejeon University, 2012 – 2019