Explainable Machine-Learned Modeling and High-Throughput Assay for Engineering of Dihydrodaidzein Reductase = 다이하이드로다이드제인 환원효소의 개량을 위한 설명 가능한 기계학습 모델링과 고속처리 분석법

국문 초록 (Abstract)

다이하이드로다이드제인 환원효소 (dihydrodaidzein reductase, DHDR)는 다이드제인 (daidzein)으로부터 유래되어 다양한 생리적 이점을 갖는 (S)-에쿠올 ((S)-equol, EQL)의 생합성 과정에서 핵심적인 역할을 수행하는 효소이다. DHDR의 활성을 향상시키는 것은 EQL 생산성을 높이기 위해 매우 중요하지만, 기존 연구들은 낮은 효율의 분석법과 기작적 이해에 있는 제약으로 인해 어려움이 있었다. 본 연구에서는 DHDR 활성을 체계적으로 개선하기 위해 합리적 효소 공학, 설명 가능한 기계학습(machine-learning, ML) 모델링, 그리고 고속처리 분석법을 통합한 전략을 제시한다.
분자 동역학(molecular dynamics, MD) 시뮬레이션 기반의 합리적 자리 포화 변이 (site-saturation mutagenesis)와 후속적인 조합 변이 (combinatorial mutagenesis)를 통해 기질 결합 및 전환에 큰 영향을 미치는 주요 변이를 규명하였다. 총 65종의 YFP-DHDR 변이를 제작하였으며, 다수의 실험 배치 (batch)에서 재현성 있게 비활성을 측정하여 예측 모델링에 적합한 고품질 데이터셋을 구축하였다. 이 중 L214S 변이는 야생형 DHDR 대비 약 두 배의 활성을 보였다.
이 재현성 높은 데이터셋을 기반으로, 우리는 MD 시뮬레이션에서 도출된 국소 및 전역 수집 변수 (collective variable, colvar)의 작은 세트를 활용하여 설명 가능한 ML 모델을 개발하였다. 본 모델은 본 연구에서 도입한 소프트마스크 (softmask) 함수를 이용함으로써 아미노산 치환과 촉매 활성 사이의 비선형적이지만 물리화학적으로 타당한 관계를 성공적으로 포착하였다. 이를 통해 DHDR 기능에 중요한 구조적 및 동역학적 특징을 규명하고, 효소-기질 상호작용뿐 아니라 DHDR-(S)-다이하이드로다이드제인-NADPH 삼원 복합체의 준-전이 상태 (quasi-transition state) 배열에 대한 기작적 통찰을 제공하였다.
향후 DHDR 활성 데이터셋을 확장하기 위해, 우리는 폴리프로필렌 96-well plate에서 트랜스-테트라하이드로다이드제인 (trans-tetrahydrodaidzein)의 형광을 이용한 고속처리 및 비-HPLC (HPLC-free) 분석법을 구축하였다. 이 분석법은 기존 HPLC 기반 측정과 높은 상관성을 보였으며, 고속처리 스크리닝 (high-throughput screening)에 적합한 우수한 통계적 성능을 나타냈다. 이 간단한 마이크로플레이트 기반 측정법은 기존의 HPLC 분석 과정과 비교하여 분석 시간을 크게 단축시켜 DHDR 변이체를 신속하게 평가할 수 있게 하였다.
종합적으로, 본 연구는 합리적 설계, 해석 가능한 기계학습 모델, 확장 가능한 분석법을 통합한 강력한 효소 최적화 프레임워크를 제시한다. 이러한 전략은 DHDR 및 관련 효소의 공학적 개량을 가속화하고, 궁극적으로 연구 및 산업적 응용을 위한 EQL 생합성을 향상시키는 데 기여할 것이다.

다국어 초록 (Multilingual Abstract)

Dihydrodaidzein reductase (DHDR) is a key enzyme in the biosynthesis of (S)-equol (EQL), a daidzein-derived metabolite with notable physiological benefits. Enhancing DHDR activity is critical for improving EQL production, but previous efforts have been limited by low-throughput assays and a lack of mechanistic insight. Here, we present an integrated strategy that combines rational enzyme engineering, explainable machine-learned (ML) modeling, and a high-throughput assay method to systematically enhance DHDR activity.
Rational site-saturation mutagenesis guided by molecular dynamics (MD) simulations, followed by combinatorial mutagenesis, identified key mutations that significantly affect substrate binding and conversion. In total, 65 YFP-DHDR variants were generated, and their reproducible specific activities were measured across multiple experimental batches, yielding a high-quality dataset suitable for predictive modeling. Among these variants, L214S demonstrated approximately a twofold increase in activity relative to wild-type DHDR.
Leveraging this reproducible dataset, we developed an explainable ML model using a reduced set of local and global collective variables (colvars) derived from MD simulations. The model successfully captured non-linear, physicochemically meaningful relationships between amino acid substitutions and catalytic activity through the use of softmask functions introduced here. It revealed structural and dynamic features critical for DHDR function, providing mechanistic insights not only into enzyme-substrate interactions but also into the arrangements of quasi-transition states in DHDR-(S)-dihydrodaidzein-NADPH ternary complexes.
To facilitate the generation of future DHDR activity datasets, we established a high-throughput, HPLC-free assay using trans-tetrahydrodaidzein fluorescence on polypropylene 96-well plates. The assay showed a strong correlation with conventional HPLC measurements and demonstrated favorable statistical properties for high-throughput screening. This simple, microplate-based measurement substantially reduced analysis time compared with the classical HPLC workflow, enabling rapid evaluation of DHDR variants.
Overall, this work establishes a robust framework for enzyme optimization that integrates rational design, interpretable machine learning, and scalable assay methods. The strategies developed here can accelerate the engineering of DHDR and related enzymes, ultimately enhancing EQL biosynthesis for both research and industrial applications.

목차 (Table of Contents)

• Chapter 1. Introduction 1
• 1.1. (S)-Equol and rate determining step of its pathway 2
• 1.2. Machine learning in the biological space 5
• 1.3. The scope of this thesis 9
• Chapter 2. Rational Site-Saturation Mutagenesis for the Enhancement of Dihydrodaidzein Reductase Activity 13
• 2.1. Introduction 14
• 2.2. Results 23
• 2.2.1. Molecular dynamics-based site selection for DHDR mutagenesis 23
• 2.2.2. Construction of YFP-DHDR chimeric enzyme for ratiometric assay design 27
• 2.2.3. Production of reproducible mutant DHDR activity dataset 31
• 2.2.4. Enhanced time-course reaction profile of YFP-DHDR-L214S 37
• 2.2.5. Structural analysis of DHDR-L214S and sDHD complex 44
• 2.3. Discussion 50
• 2.4. Conclusion 56
• 2.5. Materials and Methods 57
• Chapter 3. Machine-learnt Small Model with Fully Explainable Parameters for DHDR Activity 68
• 3.1. Introduction 69
• 3.2. Results and Discussion 76
• 3.2.1. Catalytic and Structural Insights into DHDR Function 76
• 3.2.2. Theoretical background and model architecture 84
• 3.2.3. Explainable parameters governing catalytic efficiency 93
• 3.2.4. Preferred binding of the DHDR-sDHD reaction 104
• 3.2.5. Model-based prediction of previously uncharacterized DHDR variants with enhanced activity 109
• 3.3. Conclusion 113
• 3.4. Materials and Methods 114
• Chapter 4. Development of an HPLC-free Measurement Method for High-Throughput Assay of DHDR Activity 121
• 4.1. Introduction 122
• 4.2. Results and Discussion 125
• 4.2.1. Fluorescence spectral characteristics of racDHD and tTHD 125
• 4.2.2. Fluorescent assay of DHDR activity 133
• 4.2.3. tTHD fluorescence as a high-throughput assay measure 138
• 4.3. Conclusion 145
• 4.4. Materials and Methods 146
• Chapter 5. Overall conclusion and further suggestions 149
• 5.1. Overall conclusion 150
• 5.2. Further suggestion 1: Metabolic fine-tuning for the intracellular redox balance of EQL producing bacteria 152
• 5.3. Further suggestion 2: Systematic selection and pruning of MD collective variables 156
• Appendices 158
• Appendix I. Polyphenol-Hydroxylating Tyrosinase Activity under Acidic pH Enables Efficient Synthesis of Plant Catechols and Gallols 159
• AI.1. Abstract 160
• AI.2. Introduction 162
• AI.3. Materials and methods 168
• AI.4. Results 182
• AI.4.1. Measurement of pH Stability of Quercetin and Myricetin 182
• AI.4.2. Screening Bacterial Tyrosinases for Polyphenol Conversion Under Acidic Condition 185
• AI.4.3. Effect of pH on the Kaempferol Conversion by BmTy 190
• AI.4.4. MS Identification of Tyrosinase Products 193
• AI.4.5. Production of Quercetin and Myricetin from Kaempferol Using BmTy 196
• AI.4.6. Inhibitory Action of Quercetin in Monophenolase Activity of BmTy 198
• AI.5. Discussion 202
• AI.6. Conclusion 206
• Appendix II. Regioselective One-pot Synthesis of Hydroxy-(S)-equols Using Isoflavonoid Reductases and Monooxygenases, and Evaluation of the Hydroxyequol Derivatives as Selective Estrogen Receptor Modulators and Antioxidants 207
• AII.1. Abstract 208
• AII.2. Introduction 210
• AII.3. Results & Discussions 215
• AII.3.1. Screening of (S)-equol-hydroxylating monooxygenases 215
• AII.3.2. Modulation of regioselectivity of HpaB 220
• AII.3.3. Site-directed mutagenesis of HpaB to enhance its regioselectivity 224
• AII.3.4. One-pot synthesis of ortho-hydroxylated (S)-equols from daidzein 227
• AII.3.5. Estrogenic efficacies of (S)-equol and hydroxyequol derivatives 230
• AII.3.6. Antioxidative effect of hydroxyequol derivatives 234
• AII.4. Conclusion 236
• AII.5. Experimental Methods 237
• References 246
• Abstract in Korean 261