Detecting Abiotic Stress Responses in Two Wild Oak Species using Untargeted Metabolomics: Machine Learning Approaches
- 주제(키워드) Plant untargeted metabolomics , abiotic stress , Quercus mongolica , Quercus serrata , machine learning
- 주제(DDC) 570
- 발행기관 아주대학교 일반대학원
- 지도교수 Sangkyu Park
- 발행년도 2026
- 학위수여년월 2026. 2
- 학위명 석사
- 학과 및 전공 일반대학원 생명과학과
- 실제URI http://www.dcollection.net/handler/ajou/000000035736
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Predicting species-specific metabolic responses to abiotic stress is critical for understanding oak adaptation under climate change. This study compared stress-induced metabolic responses in two ecologically distinct oak species—Quercus mongolica (montane) and Q. serrata (lowland)—using complementary untargeted GC/MS and LC/MS metabolomics. Seedlings were exposed to drought (reduced irrigation), heat, and salt stress under controlled greenhouse conditions, and differentially expressed metabolites (DEMs) were identified for each species and platform. Across stress treatments, LC/MS generally provided stronger discrimination than GC/MS and detected substantially more DEMs, with the contrast especially pronounced under drought in Q. mongolica. In GC/MS profiles, drought responses differed markedly between species: Q. mongolica exhibited relatively limited changes, whereas Q. serrata showed pervasive downregulation (97%), consistent with a metabolic suppression strategy. Cross-species concordance analysis further indicated that heat responses were highly conserved among shared DEMs (>99% directional agreement), while drought responses diverged strongly (<30% concordance), suggesting habitat-linked differences in water-use adaptation. Because MS-based metabolomic profiles are high-dimensional and nonlinear, we used machine learning to classify stress states and to test whether greenhouse-derived metabolic signatures transfer to wild populations. Models trained on experimental data achieved high internal performance (AUC > 0.87), but stress probabilities for wild samples varied across models and remained difficult to interpret even after feature selection, limiting ecological inference. Environmental validation using field–climate correlation filtering retained only 13–34% of laboratory-identified DEMs, underscoring that ecological relevance is a critical criterion for field application of stress markers. Overall, LC/MS was more effective for capturing stress-associated signatures, whereas GC/MS provided complementary resolution for primary metabolic differences across species, highlighting the importance of rigorous field validation when translating laboratory-derived biomarkers to natural populations.
more목차
1. Introduction 1
2. Materials and Methods 3
2.1. Experimental plant materials 3
2.2. Abiotic factors treatment 3
2.3. Wild sample collection 5
2.4. Metabolite extraction and derivatization 5
2.5. Untargeted LC/MS and GC/MS analysis 5
2.6. Environmental data 8
2.7. Statistical Analysis 8
2.8. Differentially expressed metabolites analysis and environmental validation 9
2.9. Machine learning models for stress prediction in wild samples 9
2.10. Impact of feature selection on model robustness and transferability 10
3. Results 13
3.1. Metabolic discrimination analysis under stress conditions 13
3.2. Differentially expressed metabolites (DEMs) analysis 18
3.3. Comparative analysis of DEMs across stress conditions 20
3.4. Comparative analysis across species 22
3.5. Environment-validated stress markers 25
3.6. Machine learning models and wild predictions 25
3.7. Feature set comparisons and feature importance analysis 28
4. Discussion 33
References 38
Appendix 42
국문요약 48
