Integrative Analysis and Diagnostic Modeling Using Genetic, Environmental, and Neuroimaging Markers in Children with Autism Spectrum Disorder = 자폐 스펙트럼 장애 아동에서 유전·환경 요인과 뇌영상 지표를 활용한 통합적 분석 및 진단 모델링

국문 초록 (Abstract)

배경: 선행 연구들은 다양한 모달리티를 이용하여 자폐 스펙트럼 장애(autism spectrum disorder, ASD)의 바이오마커를 규명하고자 하였으나, 병태생리적 바이오마커가 아직 확립되지 않아 임상 진단은 여전히 행동 관찰에 주로 의존하고 있다. ASD는 유전적 및 환경적 위험요인의 복합적인 상호작용에 의해 발생하며, 이러한 요인들이 구조적·기능적 뇌 이상, 특히 기능적 연결성의 변화를 통해 증상으로 이어질 가능성이 널리 받아들여지고 있다. 그러나 ASD 연구의 현실적인 제약으로 인해 다수의 위험요인을 동시에 통합적으로 분석한 연구는 제한적이다. 이에 본 연구는 (1) ASD와 환경적·유전적·신경발달 요인의 연관성을 분석하고, (2) 이들 위험요인 간의 상호작용 효과를 평가하며, (3) 머신러닝 기반 진단 알고리즘에서 이들 요인의 상승적 또는 가산적 효과를 검증하고자 하였다.
방법: 3–10세의 ASD 진단 아동과 연령이 일치하는 정상 대조군 아동을 대상으로 환경적 요인(프탈레이트, 비스페놀류, 3-PBA, 납, 카드뮴, 망간), 유전적 요인(자폐 취약 유전자 CNTNAP2, NRXN1, OXTR의 단일염기다형성, SNP), MRI 특징(ROI 간 기능적 연결성 및 그래프 이론 기반 지표), 임상적 요인(IQ 및 자폐 증상 중증도)을 측정하였다. 이들 요인과 ASD 간의 연관성은 회귀분석을 통해 평가하였으며, 이후 서포트 벡터 머신(support vector machine, SVM) 분석을 통해 이들 변수의 조합이 진단 정확도 향상에 기여하는지를 검증하였다. 모든 회귀분석에서는 연령, 성별, 출생연도, IQ를 공변량으로 통제하였다.
결과: 환경적 요인 중 비스페놀 S(BPS, p = 0.022)와 망간(Mn, p = 0.026)은 ASD군(n = 148)과 대조군(n = 26) 간 유의한 차이를 보였으며, 분석된 SNP들의 대립유전자 분포에서는 유의한 차이가 관찰되지 않았다. K-CARS 2 점수를 종속변수로 한 단변량 회귀분석에서는 카드뮴 노출만이 FDR 보정 후에도 유의한 연관성을 보였다(q = 0.041). 환경-유전 요인 간 상호작용은 BPA–rs1045881 (NRXN1), BPA-rs2254298 (OXTR), BPA-rs7794745 (CNTNAP2) 상호작용만이 K-CARS 점수와 유의미한 연관성을 보였지만, 같은 유전자 내 다른 SNP까지 그 영향이 지속되지는 않았다.
기능적 연결성 분석에서는 Cd와 OXTR이 부분적으로 네트워크 특이적, 진단 의존적 영향을 나타내는 것으로 보이며, OXTR–DEHP 상호작용이 좌측 측두극과 하전두회, 좌측 측두극과 안와전두피질 간 연결성과 음의 상관관계를 보였으며, NRXN1–DEHP 및 NRXN1–카드뮴 상호작용 또한 다수의 기능적 연결성과 연관되었다. 그러나 이러한 연결성 변화는 ASD 진단과 직접적인 연관성을 보이지는 않았다.
SVM 분석에서는 포함된 위험요인의 수가 증가할수록 진단 정확도가 전반적으로 향상되었다. 환경 요인 단독 모델의 정확도는 70.8%, 유전 요인 단독 모델은 56.2%, 환경·유전 요인 결합 모델은 73.0%였으며, 모든 요인을 통합한 모델에서는 95%의 최고 정확도를 보였다. 진단 정확도 차이에 가장 크게 기여한 변수로는 좌측 섬엽–좌측 담창구, 우측 대상회 전부–우측 담창구, 우측 시상–우측 소뇌 제8영역 간 최단경로 길이(shortest path length)와 함께 망간, 카드뮴, BPS가 확인되었다.
결론: 본 연구에서는 중금속, 특히 카드뮴이 ASD 관련 지표와 강한 연관성을 보였으며, 비스페놀류와 NRXN1, OXTR의 상호작용이 ASD 증상과 연관될 가능성이 시사되었다. 기능적 연결성 연구에서도 진단 의존적으로 양상이 달라질 수 있는 가능성을 나타냈다. 개별 상호작용 효과는 일관되게 유의하지는 않았으나, 환경·유전·뇌영상 요인을 통합한 머신러닝 기반 분석에서는 진단 정확도가 유의하게 향상되었다. 이는 ASD 바이오마커 연구에서 다차원적 통합 접근의 중요성을 보여주는 결과이다. 향후 더 많은 대상자, 좁은 연령대 설정, 정교한 공변량 통제를 적용한 후속 연구와 함께, 각 위험요인의 병태생리적 기전을 규명하기 위한 in vitro 및 in vivo 연구가 필요할 것으로 사료된다.

다국어 초록 (Multilingual Abstract)

Background: Previous studies have attempted to identify biomarkers of autism spectrum disorder (ASD) using various modalities; however, in the absence of established pathophysiological biomarkers, current clinical diagnosis still relies primarily on behavioral observation. ASD is widely understood to arise from complex interactions between genetic and environmental risk factors, which may contribute to structural and functional brain abnormalities, particularly in functional connectivity. Despite this, simultaneous investigation of multiple risk factors has been limited due to practical challenges in ASD research. Therefore, this study aimed to (1) investigate associations between ASD and multiple environmental, genetic, and neurodevelopmental factors; (2) examine potential interactions among these risk factors; and (3) evaluate their synergistic or additive effects on diagnostic performance using machine learning, specifically support vector machine (SVM) models.
Methods: Children aged 3–10 years diagnosed with ASD and age-matched typically developing controls were recruited. Environmental risk factors (phthalates, bisphenols, 3-phenoxybenzoic acid [3-PBA], lead, cadmium, and manganese), genetic risk factors (single nucleotide polymorphisms [SNPs] of autism susceptibility genes, including contactin-associated protein-like 2 [CNTNAP2], neurexin 1 [NRXN1], and oxytocin receptor [OXTR]), magnetic resonance imaging (MRI) features (region-of-interest to region-of-interest [ROI to ROI] functional connectivity and graph-theoretical metrics), and clinical variables (intelligence quotient [IQ] and autism symptom severity) were assessed. Associations between these variables and ASD were examined using regression analyses. Subsequently, SVM models were trained to evaluate whether combinations of these factors improved diagnostic performance. Age, sex, birth year, and IQ were included as covariates.
Results: Among environmental risk factors, bisphenol S (BPS; p = 0.022) and manganese (Mn; p = 0.026) levels differed significantly between the ASD group (n = 148) and controls (n = 26). No significant differences were observed in allele distributions of the analyzed SNPs. In univariate regression analyses with the Korean Childhood Autism Rating Scale, Second Edition (K-CARS-2) as the dependent variable, cadmium exposure was significantly associated with ASD symptom severity after false discovery rate (FDR) correction (q = 0.041). Interactions between BPA and rs1045881 (NRXN1), rs2254298 (OXTR), and rs7794745 (CNTNAP2) were significantly associated with K-CARS-2 scores; however, these associations were not consistently observed across other SNPs within the same genes.
Regarding functional connectivity, cadmium exposure and OXTR variants demonstrated partially network-specific, diagnosis-dependent differences. In addition, the OXTR–DEHP interaction was associated with reduced connectivity between the left temporal pole and the inferior frontal gyrus, as well as between the left temporal pole and the orbitofrontal cortex. Interactions involving NRXN1 with DEHP and cadmium were also associated with multiple functional connectivity patterns. However, these connectivity alterations were not directly associated with ASD diagnosis.
In SVM analyses, classification accuracy improved with the inclusion of additional risk factors. Models incorporating environmental factors alone, genetic factors alone, and both together achieved accuracies of 70.8%, 56.2%, and 73.0%, respectively, whereas the model integrating environmental, genetic, neuroimaging, and clinical variables achieved an accuracy of 95%. The most influential contributors to classification performance included shortest path lengths between the left insula and left pallidum, right anterior cingulate and right pallidum, and right thalamus and right cerebellum (lobule VIII), as well as environmental variables such as Mn, Cd, and BPS.
Conclusion: Heavy metals, particularly cadmium, demonstrated strong associations with ASD-related outcomes. Potential interaction effects between environmental risk factors, such as bisphenols, and autism susceptibility genes including NRXN1 and OXTR were suggested. Functional MRI findings further indicated diagnosis-dependent alterations in brain connectivity. Although interaction effects were not consistently observed across analyses, the integrative use of environmental, genetic, and neuroimaging features substantially enhanced diagnostic performance in machine learning models. These findings underscore the importance of multimodal integrative approaches for ASD biomarker discovery. Future studies with larger sample sizes, narrower age ranges, improved covariate control, and complementary mechanistic in vitro and in vivo investigations are warranted.

목차 (Table of Contents)

• Table of Contents
• Abstract i
• Table of contents v
• List of Tables ix
• List of Figures xi
• List of Abbreviations xii
• Chapter1. Introduction 1
• 1.1 ASD & biomarker study 1
• 1.2 Environmental risk factors: Environment disrupting chemicals/ heavy metals and ASD 2
• 1.2.1 Phthalate 2
• 1.2.2 Bisphenol, 3-phenoxybenzoic acid 3
• 1.2.3 Manganese, lead, Cadmium 4
• 1.2.4 The current status of studies on environmental risk factors 5
• 1.3 Autism susceptible genes and ASD 5
• 1.3.1 Contactin-associated protein-like 2 / Neurexin1 5
• 1.3.2 Oxytocin receptor 7
• 1.4 The need of multi-modal, multi-variable study of ASD 7
• 1.4.1 The current status of ASD biomarker study 7
• 1.4.2 Lack of Gene-Environment interaction study 8
• 1.4.3 fMRI feature incorporation into interaction studies 9
• 1.5 Purpose of this study 10
• Chapter 2. Methods 12
• 2.1 Study design and participants 12
• 2.2 Measurement of creatinine, phthalate, bisphenol, 3-PBA 13
• 2.2.1 Creatinine measurement 13
• 2.2.2 Phthalate 14
• 2.2.3 Bisphenol Measurement 15
• 2.2.4 3-PBA Measurement 15
• 2.3 Measurement of heavy metal 16
• 2.4 Limit of Detection 17
• 2.5 Adjustment of urinary concentrations 19
• 2.6 Measurement of genetic data 19
• 2.7 Measurement of autistic traits and intellectual quotient 20
• 2.8 MRI imaging protocol 23
• 2.8.1 Preprocessing and denosing for fMRI interaction analysis 23
• 2.8.2 Preprocessing for fMRI data used to extract graph theory features for SVM analysis 25
• 2.9 Statistical analysis 27
• 2.9.1 Associations between risk factors and ASD 27
• 2.9.2 Group division 28
• 2.9.3 fMRI analysis for interaction study 29
• 2.9.4 fMRI graph theory feature extraction for SVM analysis 29
• 2.9.5 SVM analysis 29
• Chapter 3. Results 31
• Part 1. Interaction effect of environmental and genetic risk factors on brain functional abnormalities and ASD 31
• 3.1 Participant Characteristics 31
• 3.2 ASD group versus control group 31
• 3.3 Regression analysis of risk factors and ASD 36
• 3.4 Interaction effect between environmental and genetic risk factors 36
• 3.5 Group interaction 45
• 3.6 Interaction effect of environmental and genetic risk factors on ROI-ROI connectivity 48
• 3.6.1 ASD vs control 48
• 3.6.2 Environmental and Genetic risk factors on ROI-ROI connectivity 48
• 3.6.3 Interaction effect on ROI-ROI connectivity 49
• Part 2. Diagnostic algorithm with environmental, genetic risk factors, and brain functional connectivities and their synergy 71
• 3.7 ASD diagnosis algorithm with SVM 71
• 3.7.1 Environmental risk factors only 71
• 3.7.2 Genetic risk factors only 71
• 3.7.3 MRI feature only 71
• 3.7.4 With Multiple risk factors 72
• Chapter 4. Discussion 78
• Part 1. Interaction effect of environmental and genetic risk factors on brain functional abnormalities and ASD 80
• 4.1 Environmental risk factors’ influence on ASD 80
• 4.1.1 Cd 80
• 4.1.2 Mn 80
• 4.1.2 BPF and BPS 81
• 4.1.3 Phthalate 82
• 4.1.4 Contribution compared to others 82
• 4.2 Genetic risk factors’ influence on ASD 83
• 4.3 Interaction between environmental and genetic risk factors. 87
• 4.4 Brain functional abnormalities of ASD 89
• 4.5 Interaction effect on fMRI brain imaging of ASD 90
• 4.6 Conclusion of Part 1 92
• Part 2. Diagnostic algorithm with environmental, genetic risk factors, and brain functional connectivities and their synergy 93
• 4.7 Synergistic effects on ASD diagnosis algorithm 93
• 4.8 Considerations on SVM analysis 95
• 4.9 Conclusion of part 2 97
• Chapter 5. conclusion 98
• References 101
• Abstract in Korean 107