서브메뉴
검색
Enhancing Drug Repositioning With Interpretable Graph Neural Network Models on Biomedical Knowledge Graphs
Enhancing Drug Repositioning With Interpretable Graph Neural Network Models on Biomedical Knowledge Graphs
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103157
- ISBN
- 9798283478517
- DDC
- 574
- 서명/저자
- Enhancing Drug Repositioning With Interpretable Graph Neural Network Models on Biomedical Knowledge Graphs
- 발행사항
- [Sl] : The Scripps Research Institute, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 118 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-12, Section: B.
- 주기사항
- Advisor: Su, Andrew I.
- 학위논문주기
- Thesis (Ph.D.)--The Scripps Research Institute, 2025.
- 초록/해제
- 요약Drug repositioning offers a cost-effective alternative to traditional drug development by identifying new therapeutic applications for existing approved medications. Biomedical knowledge graphs (KGs), which model complex relationships between drugs, diseases, genes, and biological pathways as interconnected networks, have emerged as powerful tools for drug repositioning by enabling the discovery of hidden therapeutic connections. However, current computational methods leveraging these KGs often lack interpretability, limiting their translation to clinical applications. This thesis addresses two critical challenges in drug repositioning using biomedical KGs.First, we introduce DrugMechDB, a manually curated knowledge graph containing over 5,600 mechanistic pathways elucidating 4,583 known drug indications across 1,552 unique drugs and 1,290 diseases. DrugMechDB provides a gold standard for evaluating drug mechanism hypotheses. This database integrates multiple biomedical sources, with each entry validated by domain experts, filling a significant gap in high-quality, structured mechanistic data. DrugMechDB aims to serve as a comprehensive repository that captures known mechanistic relationships in a standardized format accessible to computational methods.Second, we present Drug-Based Reasoning Explainer (DBR-X), a novel framework that addresses the interpretability limitations of current Graph Neural Network approaches. DBR-X comprises two complementary modules: (1) a link prediction module that identifies potential drug-disease connections by recognizing similar subgraph patterns using Case-Based Reasoning with an attention-based similarity function that captures both structural and semantic relationships, and (2) a path identification module employing a heterogeneous path-enforcing mask with node degree scoring to identify mechanistically relevant connections while eliminating spurious paths. Unlike previous black-box models, DBR-X is designed to provide human-interpretable reasoning chains that can be evaluated by domain experts and potentially align with established pharmacological mechanisms. The quality of DBR-X's mechanistic explanations was rigorously validated through multiple approaches: comparison with manually-curated drug mechanisms in DrugMechDB, evaluation of explanation faithfulness through perturbation studies, and measurement of stability under graph modifications. Case studies on rare diseases further demonstrate DBR-X's ability to identify promising repositioning candidates with supporting biological evidence.This work advances drug repositioning by establishing DrugMechDB as a gold standard dataset for validating AI model predictions and by developing DBR-X, which enhances both predictive accuracy and mechanistic interpretability. Together, these contributions create a framework that can potentially accelerate the translation of computational predictions into clinical applications for diseases with therapeutic needs.
- 일반주제명
- Bioinformatics
- 일반주제명
- Computer science
- 일반주제명
- Biology
- 키워드
- Drug discovery
- 기타저자
- The Scripps Research Institute Computational Biology/Bioinformatics
- 기본자료저록
- Dissertations Abstracts International. 86-12B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017357262
■00520260202103157
■006m o d
■007cr#unu||||||||
■020 ▼a9798283478517
■035 ▼a(MiAaPQ)AAI31998203
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a574
■1001 ▼aGonzalez-Cavazos, Adriana Carolina.
■24510▼aEnhancing Drug Repositioning With Interpretable Graph Neural Network Models on Biomedical Knowledge Graphs
■260 ▼a[Sl]▼bThe Scripps Research Institute▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a118 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-12, Section: B.
■500 ▼aAdvisor: Su, Andrew I.
■5021 ▼aThesis (Ph.D.)--The Scripps Research Institute, 2025.
■520 ▼aDrug repositioning offers a cost-effective alternative to traditional drug development by identifying new therapeutic applications for existing approved medications. Biomedical knowledge graphs (KGs), which model complex relationships between drugs, diseases, genes, and biological pathways as interconnected networks, have emerged as powerful tools for drug repositioning by enabling the discovery of hidden therapeutic connections. However, current computational methods leveraging these KGs often lack interpretability, limiting their translation to clinical applications. This thesis addresses two critical challenges in drug repositioning using biomedical KGs.First, we introduce DrugMechDB, a manually curated knowledge graph containing over 5,600 mechanistic pathways elucidating 4,583 known drug indications across 1,552 unique drugs and 1,290 diseases. DrugMechDB provides a gold standard for evaluating drug mechanism hypotheses. This database integrates multiple biomedical sources, with each entry validated by domain experts, filling a significant gap in high-quality, structured mechanistic data. DrugMechDB aims to serve as a comprehensive repository that captures known mechanistic relationships in a standardized format accessible to computational methods.Second, we present Drug-Based Reasoning Explainer (DBR-X), a novel framework that addresses the interpretability limitations of current Graph Neural Network approaches. DBR-X comprises two complementary modules: (1) a link prediction module that identifies potential drug-disease connections by recognizing similar subgraph patterns using Case-Based Reasoning with an attention-based similarity function that captures both structural and semantic relationships, and (2) a path identification module employing a heterogeneous path-enforcing mask with node degree scoring to identify mechanistically relevant connections while eliminating spurious paths. Unlike previous black-box models, DBR-X is designed to provide human-interpretable reasoning chains that can be evaluated by domain experts and potentially align with established pharmacological mechanisms. The quality of DBR-X's mechanistic explanations was rigorously validated through multiple approaches: comparison with manually-curated drug mechanisms in DrugMechDB, evaluation of explanation faithfulness through perturbation studies, and measurement of stability under graph modifications. Case studies on rare diseases further demonstrate DBR-X's ability to identify promising repositioning candidates with supporting biological evidence.This work advances drug repositioning by establishing DrugMechDB as a gold standard dataset for validating AI model predictions and by developing DBR-X, which enhances both predictive accuracy and mechanistic interpretability. Together, these contributions create a framework that can potentially accelerate the translation of computational predictions into clinical applications for diseases with therapeutic needs.
■590 ▼aSchool code: 1179.
■650 4▼aBioinformatics
■650 4▼aComputer science
■650 4▼aBiology
■653 ▼aBiomedical knowledge graphs
■653 ▼aDrug discovery
■653 ▼aDrug repositioning
■653 ▼aGraph neural networks
■690 ▼a0715
■690 ▼a0984
■690 ▼a0306
■71020▼aThe Scripps Research Institute▼bComputational Biology/Bioinformatics.
■7730 ▼tDissertations Abstracts International▼g86-12B.
■790 ▼a1179
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357262▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
