생물학적 패스웨이 기반 전이학습을 활용한 단일세포 약물 반응 예측 모델 = A Single-cell Drug Response Prediction Model with Pathway-informed Transfer Learning

다국어 초록 (Multilingual Abstract)

Cross-domain alignment between bulk and single-cell transcriptomes remains a major computational challenge because bulk and single-cell RNA sequencing data exhibit substantial domain discrepancies arising from sparsity, technical noise, and differences in transcriptional resolution. These discrepancies hinder the generalization of predictive models trained on bulk transcriptomic datasets when applied to single-cell data. Although recent transfer learning approaches have attempted to address this problem through latent feature alignment, many existing methods rely on gene-level representations that are sensitive to transcriptional noise and provide limited biological interpretability. To address these limitations, this study proposes a biologically informed few-shot label-guided domain adaptation framework for constructing a shared co-embedding space between bulk and single-cell transcriptomes. The proposed framework employs pathway-specific denoising autoencoders to encode transcriptomic profiles into pathway-level latent embeddings, enabling biologically guided dimensionality reduction while mitigating technical noise. To align transcriptomic modalities, the framework applies few-shot label-guided contrastive learning that selectively aligns bulk and single-cell samples sharing the same drug response states within the pathway embedding space. A limited number of labeled single-cell samples are utilized to guide cross-domain representation alignment, while the prediction module is trained using bulk drug response supervision and subsequently transferred to the single-cell domain. In addition, a multi-head attention module integrates pathway-level embeddings to capture pathway importance and interactions associated with heterogeneous cellular responses. Experimental evaluations across multiple benchmark datasets demonstrated that the proposed framework consistently outperformed existing baseline methods, including scDEAL, SCAD, and Celligner-based approaches, in weighted F1-score, AUPRC, and AUROC. Furthermore, pathway-level analyses identified biologically meaningful signaling signatures associated with drug-specific response patterns, demonstrating the interpretability and robustness of the proposed framework for single-cell drug response prediction.

목차 (Table of Contents)

• 제1장 서 론 1
• 제1절 연구의 목적 1
• 제2절 연구의 범위 및 방법 4
• 1. 연구의 범위 4
• 2. 연구의 방법 5
• 제2장 연구에 관한 이론적 배경 7
• 제1절 프레임워크의 구조 7
• 1. 패스웨이 기반 표현 학습 7
• 2. 대조학습 8
• 3. 멀티헤드 어텐션 9
• 제2절 관련 연구 10
• 1. Bulk-to-single 데이터 정렬 10
• 2. 도메인 적응 11
• 3. 전이학습 12
• 제3장 방법론 14
• 제1절 모델 개요 14
• 1. 데이터셋 구성 16
• 2. 패스웨이 기반 임베딩 생성 및 손실함수 정의 19
• 3. Cross-domain 정렬을 위한 손실함수 정의 21
• 4. 약물 반응 예측 및 손실함수 정의 23
• 5. 손실 함수 24
• 제4장 실험 25
• 제1절 실험 환경 25
• 1. 구현 세부 사항 25
• 2. 일반화 성능 평가 27
• 3. 벤치마크 모델의 예측 성능 평가 28
• 제2절 기존 모델과의 성능 비교 29
• 1. 약물 반응 예측 성능 평가 29
• 2. Ablation study 32
• 3. 일반화 성능 검증 35
• 4. 도메인 정렬 성능 평가 37
• 제3절 패스웨이 수준에서의 생물학적 해석 40
• 1. Cross-domain 간의 생물학적 해석 일관성 40
• 2. 멀티헤드 어텐션 결과를 통한 해석 43
• 3. 패스웨이 어텐션의 잠재 요인 분석 46
• 제5장 결 론 50
• 참 고 문 헌 53
• ABSTRACT 57