A Deep Learning Framework for Multi-Objective Codon Optimization Integrating Protein Expression and mRNA Stability = 단백질 발현과 mRNA 안정성 향상을 위한 다중 목표 코돈 최적화 딥러닝 프레임워크

다국어 초록 (Multilingual Abstract)

Codon optimization plays a central role in improving gene expression across diverse host organisms, yet many existing approaches treat it as a single-objective problem, focusing primarily on codon frequency adaptation or mRNA structural stability in isolation. Here, we propose a multi-objective deep learning framework that jointly optimizes translational efficiency and mRNA structural stability while strictly preserving the encoded amino acid sequence. The model employs a hybrid 1D convolutional neural network (CNN) and Transformer encoder to capture both local codon–amino acid relationships and long-range contextual dependencies. To incorporate biologically meaningful structural constraints, the framework is trained using structure-aware reference sequences generated by LinearDesign. Optimization is guided by a composite objective that integrates supervised codon prediction, contrastive representation learning, and a joint loss based on the Codon Adaptation Index (CAI) and Minimum Free Energy (MFE), enabling balanced consideration of translational efficiency and folding stability. Evaluation across Homo sapiens, Escherichia coli, and Saccharomyces cerevisiae demonstrates that the proposed framework consistently outperforms rule-based tools and single-objective deep learning baselines. The model achieves higher CAI, lower (more favorable) MFE, and maintains GC content within biologically plausible ranges, indicating robust and biologically coherent sequence design. These results highlight the importance of biologically grounded multi-objective learning in codon optimization and establish a generalizable framework for rational mRNA sequence design in therapeutic and synthetic biology applications.

국문 초록 (Abstract)

코돈 최적화는 다양한 숙주에서의 단백질 발현을 향상시키기 위해 필수적이지만, 기존 접근법은 주로 코돈 사용 빈도나 mRNA 구조 안정성 등 단일 목표에 집중하는 한계를 갖고 있다. 본 연구에서는 번역 효율성과 mRNA 구조 안정성을 동시에 고려하는 다중 목적 기반 딥러닝 프레임워크를 제안한다. 제안된 모델은 1D-CNN과 트랜스포머를 결합한 인코더 구조를 사용하여 코돈–아미노산 간의 국소적 패턴과 서열 전반의 장거리 문맥 정보를 함께 학습하며, LinearDesign으로 생성된 구조 최적화 서열을 학습 데이터로 활용하여 생물학적으로 의미 있는 접힘 제약을 반영한다. 또한 감독 학습, 대조 학습, CAI–MFE 다중 목적 손실을 통합하여 번역 효율성과 구조 안정성 간의 균형을 달성한다. Homo sapiens, Escherichia coli, Saccharomyces cerevisiae에 대한 평가 결과, 제안된 모델은 규칙 기반 및 단일 목적 딥러닝 모델보다 높은 CAI, 더 낮은 MFE, 안정적인 GC 조성을 달성하며 우수한 최적화 성능을 보였다. 이는 생물학적 제약을 고려한 다중 목적 학습이 코돈 최적화에서 매우 중요함을 보여주며, 향후 치료제 개발과 합성생물학에 활용될 차세대 mRNA 설계를 위한 일반화된 기반을 제공한다.

목차 (Table of Contents)

• 1. Introduction 1
• 2. Previous studies 5
• 3. Proposed method 8
• 3.A Sequence embedding construction 9
• 3.A.1 Codon and amino acid tokenization 9
• 3.A.2 Embedding construction 10
• 3.B Contextual feature extraction 11
• 3.B.1 Local feature extraction with 1D Convolution 11
• 3.B.2 Global context modeling with Transformer encoder 12
• 3.C Multi-objective learning framework 13
• 3.C.1 Contrastive loss for embedding coherence 13
• 3.C.2 Supervised loss for codon prediction 15
• 3.C.3 Joint loss for CAI-MFE optimization 16
• 4. Dataset and preprocessing 18
• 4.A Dataset construction 18
• 4.B Label generation 19
• 4.C Data preprocessing 20
• 5. Evaluation and results 22
• 5.A Evaluation metrics 22
• 5.A.1 Codon Adaptation Index 23
• 5.A.2 Minimum Free Energy 23
• 5.A.3 GC content 24
• 5.B Experimental setup 25
• 5.C Quantitative results 27
• 5.C.1 Machine learning-based evaluation 27
• 5.C.2 Statistical evaluation 29
• 5.C.3 Biologically grounded evaluation 30
• 5.C.4 Comparative evaluation with LinearDesign 33
• 6. Ablation study 35
• 7. Conclusion 37
• Reference 38
• Korean Abstract 42