Research
Figures: FAK activator graphical abstract ([Comput Biol Chem 2025](../publication/yoon-2025/), © Elsevier), KMU-11342 ([Pharmaceuticals 2026](../publication/jeon-2026/), CC BY 4.0), ACh-gain landscape, Supplementary Fig. S1 ([Cognitive Neurodynamics 2026](../publication/seo-2026/), open access), T-maze and cumulative-reward curves ([Sensors 2024](../publication/seo-2024/), CC BY 4.0), CASCADE ([Bioinformatics 2026](../publication/avila-2026/), CC BY 4.0), SNOMED-CT mapping ([Med Biol Eng Comput 2026](../publication/oh-2026/), open access), symbolic regression vs machine-learning classifiers, ROC ([Biomedicines 2026](../publication/oh-2026-2/), CC BY 4.0).
1. AI-Accelerated Drug Discovery
We combine virtual screening, AI-based property prediction and physics-based simulation to find and explain new therapeutic candidates, then close the loop with wet-lab collaborators. Our FAK-activator study coupled similarity-based screening, docking and deep-learning predictors with molecular dynamics to select compounds with better binding profiles than the reference ligand (Comput Biol Chem 2025). With Keimyung University we characterised two indolin-2-one kinase inhibitors: KMU-11342 suppresses proliferation, migration and stemness of colorectal cancer cells in 2D, 3D spheroid and patient-derived organoid models, with kinase profiling and docking pointing to GSK3β/CDK1 regulation through p53/NF-κB and FoxO1 signalling (Pharmaceuticals 2026); KMU-11361 attenuates rheumatoid arthritis by inhibiting the TAK1–NF-κB–NLRP3 axis (Inflammation Research 2026). Ongoing work targets nervous-system disorders, including selective modulators of the lysosomal ion channel TMEM175 and drug discovery for REM sleep behaviour disorder.
2. Brain Data & Neuropsychiatry
We build interpretable analysis pipelines across scales, from single-cell electrophysiology to population recordings. We showed that a neuron’s transgenic marker can be predicted from its electrophysiological properties alone (Brain Res Bull 2019), and human cortical recordings revealed that the net synaptic drive onto fast-spiking interneurons is inverted towards inhibition in FCD type I epilepsy (Nature Communications 2024). Because every multi-electrode array manufacturer ships its own closed software, we released CASCADE, an open-source Python pipeline that reads recordings from five MEA platforms and computes standard network metrics together with criticality and avalanche statistics (Bioinformatics 2026). We also analyse immune–neural crosstalk in Alzheimer’s disease with immune cell-enriched single-cell RNA-seq (J Neuroimmunol 2025), benchmark end-to-end EEG-based stress detection (IEEE Access 2026), and are constructing in-silico epilepsy models with NEURON and NetPyNE, including mechanistic analysis of Nav1.2 variants linked to developmental and epileptic encephalopathy. Earlier work links neurogenesis-driven pattern separation to mood disorders (J Korean Soc Biol Ther Psychiatry 2020).
3. AI-Driven Clinical Decision Support
Our clinical work favours models that clinicians can read. We fine-tuned ClinicalBERT to map free-text diagnosis spans in electronic medical records to SNOMED-CT concepts and analysed how the latent space separates ambiguous from clearly distinct diagnoses (Med Biol Eng Comput 2026). For liver-transplant recipients we derived an intuitive risk equation for post-transplant bloodstream infection with symbolic regression, and SHAP analysis highlighted EBV/HBV serological markers as candidate risk factors beyond routine laboratory values (Biomedicines 2026). Earlier studies used machine learning to distinguish cytomegalovirus from herpes simplex esophagitis on endoscopy (Scientific Reports 2021), tested whether endoscopists can tell GAN-generated gastroscopy images from real ones (J Digit Imaging 2023), and examined the obesity paradox in colorectal cancer with TCGA cohorts (J Cancer 2023). Our early perspective on ChatGPT in medical education is among the most cited papers on the topic (Anat Sci Educ 2024).
4. Hippocampus × Reinforcement Learning
We test the hypothesis that the hippocampus implements the successor representation and ask how neuromodulators shape its learning rules (Biosystems 2022; Front Comput Neurosci 2025). Simulation studies showed how weight initialisation affects successor-feature learning (Electronics 2023), how predecessor and successor features transfer in noisy T-maze tasks (Sensors 2024), and how both algorithms tolerate noise in one- and two-dimensional environments (Sensors 2025). Most recently we proposed an acetylcholine-modulated predecessor-feature model in which eligibility-trace–gated depression lets an agent keep exploring after reward; in n-arm radial maze simulations it outperforms the standard model within an effective range of modulation that narrows as environments grow more complex (Cognitive Neurodynamics 2026).
“From synapse to silicon, insight travels — and patient care transforms.”