Seminar & Colloquium
[세미나: 6월 26일(월), 오전 11시] Prof. Kwonmoo Lee, Harvard Medical School
Title
Uncovering Cellular Subtypes with AI-based Analysis of Heterogeneity
Speaker
Prof. Kwonmoo Lee, Department of Surgery, Harvard Medical School
EDUCATION
- 2014 Post-Doctoral Training in Quantitative Cell Biology,Harvard Medical School
- 2010 Ph.D. in Physics (Biophysics), Massachusetts Institute of Technology
- 1998 M.S. in Physics (Statistical and Biological Physics), Pohang University of Science and Technology
- 1996 B.S. in Electronic and Electrical Engineering, and Physics (minor), Magna Cum Laude, Pohang University of Science and Technology
Experience
- 2020 ~ Assistant Professor, Boston Children’s Hospital, Harvard Medical School
- 2014 ~ 2020 Assistant Professor, Worcester Polytechnic Institute
- 2010 ~ 2014 Postdoctoral Fellow, Harvard Medical School
- 2004 ~ 2010 Graduate Student, Harvard Medical School
- 2000 ~ 2003 Research Engineer, Samsung SDS
- 1998 ~ 2000 Research Engineer, LG Cable
- 1996 ~ 1998 Graduate Student, Pohang University of Science and Technology
| Date | Monday, June 26th, 2023
| Time | 11:00 ~
| Venue | 신소재공동연구소(131동) 1층 세미나실
[Abstract]
Diseases and cell subtyping is essential for personalized medicine and understanding disease mechanisms. While there are many new opportunities for subtyping with single-cell data, it is challenging with high-dimensional datasets. Feature selection or learning algorithms can reduce the number of features for downstream analysis for subtyping, but most conventional methods eliminate heterogeneity and collapse feature space, hindering effective subtyping. Our feature embedding using deep metric learning revealed that the features with significant differences in interquartile range (IQR) between known states preserve heterogeneity while maintaining the discrimination. Based on this finding, we developed a statistical method, PHet (Preserving Heterogeneity), to identify heterogeneous-preserving discriminative features for effective subtype clustering. PHet was tested on public datasets of microarray gene expression and scRNA-seq, and it effectively identified disease or cell subtypes and significantly outperformed previous methods. Furthermore, PHet identified novel subtypes of basal cells from scRNA-seq data of airway epithelium and breast cancer stem cells from live cell time-lapse movies.
We also developed a novel deep learning architecture called HoloNet (Holographical Network) for breast cancer cell phenotyping. HoloNet analyzes images from lens-free digital in-line holography (LDIH), which produces cellular diffraction patterns (holograms) with a large field of view that conventional lens-based microscopes cannot offer. HoloNet extracts large features from diffraction patterns, integrates them with small features from convolutional layers, and outperforms other state-of-the-art deep learning methods for the classification of breast cancer cells and provide superior interpretability of raw holograms from breast cancer cell markers, ER/PR and HER2. We also used the HoloNet embedding to reveal breast cancer cell subtypes shared by multiple breast cancer cell types. Identifying such rare and subtle phenotypes of breast cancer cells will enable us to perform detailed analyses of the heterogeneity of cell phenotypes for precise breast cancer diagnosis.
| Host | 도준상 교수 (02-880-1605)