서지주요정보
AI-based Parkinson's disease subtype classification using label-free Images of hiPSC-derived neurons = 인체유래 신경세포의 비표지 이미지를 이용한 인공지능 기반 파킨슨병 하위 유형 분류
서명 / 저자 AI-based Parkinson's disease subtype classification using label-free Images of hiPSC-derived neurons = 인체유래 신경세포의 비표지 이미지를 이용한 인공지능 기반 파킨슨병 하위 유형 분류 / Seung Ju Yoo.
발행사항 [대전 : 한국과학기술원, 2025].
Online Access 원문보기 원문인쇄

소장정보

등록번호

8044004

소장위치/청구기호

학술문화관(도서관)2층 학위논문

MBCS 25006

휴대폰 전송

도서상태

이용가능(대출불가)

사유안내

반납예정일

리뷰정보

초록정보

Parkinson’s Disease (PD) is a complex and heterogeneous neurodegenerative disorder that poses significant challenges for personalized medicine due to its diverse pathological mechanisms and predominantly idiopathic nature. Current treatment approaches fail to address the underlying mechanisms of the disease, partly due to the lack of robust systems for identifying patient-specific mechanisms. In this study, we propose a machine learning-based classifier for PD subtypes, leveraging in vitro cellular models and a label-free imaging technique. By introducing chemically induced perturbations in healthy cortical neurons, we generated reproducible and versatile models of key PD pathological features, including mitochondrial dysfunction, lysosomal dysfunction, and protein aggregation. The transformer-based model achieved an exceptional classification accuracy of 96% during training with these chemical models. Furthermore, validation using patient-derived neurons carrying the SNCA triplication mutation underscores its translational relevance and potential for real-world applications. We establish a powerful drug screening paradigm that enables highly efficient, patient-specific therapeutic development by distinguishing the cellular and molecular underpinnings of the disease in individual patients.

파킨슨병은 복잡하고 이질적인 퇴행성 뇌질환으로, 다양한 병리 기전과 특발성으로 인해 환자 개개인에게 최적화된 맞춤의료 실현에 어려움이 있다. 현행 치료 방법은 질병의 근본적인 기전을 해결하지 못하고 도파민 결핍을 보충하기 위한 일률적이고 제한적인 약물 치료에 머물러 있다. 본 연구에서는 파킨슨병 세포 모델과 비표지 이미지를 이용한 파킨슨병의 하위 유형을 분류하는 인공지능 플랫폼을 제시한다. 건강한 유도만능줄기세포에서 유래한 대뇌피질 신경세포에 파킨슨병의 대표적 병리적 특징을 화학적으로 유도하여 미토콘드리아 기능 장애, 리소좀 기능 장애, 단백질 응집 등의 세포 모델을 만들었으며 이를 트랜스포머 기반 기계학습 시스템을 통해 학습하여 96%의 뛰어난 분류 정확도를 달성했다. 또한, 파킨슨병 환자의 체세포에서 유래한 신경세포에서 그 유전적 변형(SNCA)의 병리적 특성에 맞는 하위 유형으로 분류하는 것을 검증했다. 본 연구를 통해 각 환자의 세포 및 분자 기전을 구분하여 환자 맞춤형 치료 개발을 가능하게 하는 강력한 약물 스크리닝 패러다임을 제시하며, 이를 통해 파킨슨병 환자 각각의 병리적 특성에 최적화된 맞춤의료를 제시할 수 있을 것으로 기대한다.

서지기타정보

서지기타정보
청구기호 {MBCS 25006
형태사항 iv, 37 p. : 삽화 ; 30 cm
언어 영어
일반주기 저자명의 한글표기: 유승주
지도교수의 영문표기: Choi, Minee
지도교수의 한글표기: 최민이
학위논문 학위논문(석사) - 한국과학기술원 : 뇌인지과학과,
서지주기 References: p. 31-35
주제 image classification
Parkinson's disease
personalized medicine
neurodegenerative disease
stem cell
neuron
iPSC
holotomography
이미지 분류
파킨슨병
맞춤의료
퇴행성 뇌질환
줄기세포
신경세포
유도만능줄기세포
홀로토모그래피
QR CODE

책소개

전체보기

목차

전체보기

이 주제의 인기대출도서

Subtypes of Parkinson's Disease.

Phase contrast imaging and quantitative phase imaging. (A)A comparison between phase-contrast and holotomography images highlights the advantages ofholotomography, a quantitative phase imaging technique. Although the samples presented differ and are not to scale, holotomography provides high-contrast images by representing pixel intensities as refractive indices. This enables the detailed visualiz

Materials for cell culture and imaging.

Chemical details for induced dysfunctions.

Overview of the model pipeline. (A)Aschematic representation ofthe modeltraining process. Preprocessing involves maximum intensity projection of 3D holotomography (HT) images, normalization to a range of[-1, 1], data augmentation, and filtering based on a mean intensity threshold The model architecture is based on Vision Transformer (VIT-L/16; Dosovitskiy, et al. 2021), pretrained on ImageNet21k a

Characterization of generated neurons byimmunocytochemistry. (A) TuJ1 and SATB2 expression confirms neuronal differentiation and cortical specificity (Hoechst: nucleus, SATB2: cortical upper layer neuron, TuJ1: neuron-specific 3-tubulin). (B MAP2 expression confirms neuronal maturation. (MAP2: mature neuron) (C) Intracellulal calcium concentration before and after KCI treatment, indicated by calci

Confirmation of mitochondrial dysfunction in Subtype chemical model. (A) Representative fluorescent images of mitochondrial dysfunction induced by antimycin A (0.1 nM 24 h), with Hoechst and LysoTracker intensity ranges adjusted for visibility. Scale bar: 20 nm. (B) Quantification of normalized mitochondrial membrane potential via mean TMRM fluorescence intensity across puncta. (C) Quantification

Confirmation of lysosomal dysfunction in Subtype II chemical model. (A) Representative fluorescent images of lysosomal dysfunction induced by chloroquine (5uM, 24h), with Hoechstand LysoTracker intensity ranges adjusted forvisibility. Scale bar: 20 nm. (B) Lysosoma activity quantified by normalized mean LysoTracker fluorescence intensity across puncta. (C) Quantification of lysosome volumes measur

Confirmation of protein aggregation in Subtype III chemical model. (A) Representative fluorescent organelle images of protein aggregation induced by alpha- synuclein preformed fibrils (1 ng/ml, 72 h). Scale bar: 20 nm. (B) alpha-synuclein area, normalized to the number of nuclei. p=0.095. N=1, with 2 images (control) and 5 images (PFF). (C) Quantification ofnormalized mitochondrial membrane potent

Two-step training achieves 96% classification accuracy. (A) Illustration of two different methods for training the model, along with representative accuracy. Model-P0 serves as the reference model, trained directly on the base model (a pretrained ViT using the ImageNet dataset). Model-C represents an intermediate model trained on the base model, where classification was performed chemical-wise. Mo

Visualization ofattention maps reveal thekey organelle for prediction. (A) Representative images of each subtype. The attention map represents the region in the origina image to which the model attended. Each organelle mask was generated from live-cell fluorescen images. (B) Colocalization ofthe attention map and the fluorescent organelle mask (OCF), normalized to that of the Healthy class (marked

Model predicts protein aggregation in SNCAx3 carrier neurons. (A) Prediction outcomes of Model-A, trained on the entire dataset for final validation. The model identifies 74% of SNCA images as Protein Aggregation subtype. (B)A confidence heatmap across subimage classifications. The average confidence of 82% for the Protein Aggregation subtype indicates the model's robustness in distinguishing this