[论文解读] Predicting Hydroxyl Mediated Nucleophilic Degradation and Molecular Stability of RNA Sequences through the Application of Deep Learning Methods
本研究提出并评估了三种深度学习模型——LSTM、GRU 和图卷积网络(GCN)——利用斯坦福 OpenVaccine 数据集预测羟基介导的亲核降解反应及 mRNA 序列的化学稳定性。GRU 模型在降解风险预测方面表现最佳,准确率达 76%,均方根误差(RMSE)最低(0.266),证明了基于计算方法筛选稳定 mRNA 治疗制剂的可行性。
Synthesis and efficient implementation mRNA strands has been shown to have wide utility, especially recently in the development of COVID vaccines. However, the intrinsic chemical stability of mRNA poses a challenge due to the presence of 2'-hydroxyl groups in ribose sugars. The -OH group in the backbone structure enables a base-catalyzed nucleophilic attack by the deprotonated hydroxyl on the adjacent phosphorous and consequent self-hydrolysis of the phosphodiester bond. As expected for in-line hydrolytic cleavage reactions, the chemical stability of mRNA strands is highly dependent on external environmental factors, e.g. pH, temperature, oxidizers, etc. Predicting this chemical instability using a computational model will reduce the number of sequences synthesized and tested through identifying the most promising candidates, aiding the development of mRNA related therapies. This paper proposes and evaluates three deep learning models (Long Short Term Memory, Gated Recurrent Unit, and Graph Convolutional Networks) as methods to predict the reactivity and risk of degradation of mRNA sequences. The Stanford Open Vaccine dataset of 6034 mRNA sequences was used in this study. The training set consisted of 3029 of these sequences (length of 107 nucleotide bases) while the testing dataset consisted of 3005 sequences (length of 130 nucleotide bases), in structured (Lowest Entropy Base Pair Probability Matrix) and unstructured (Nodes and Edges) forms. The stability of mRNA strands was accurately generated, with the Graph Convolutional Network being the best predictor of reactivity ($RMSE = 0.249$) while the Gated Recurrent Unit Network was the best at predicting risks of degradation ($RMSE = 0.266$). Combining all target variables, the GRU performed the best with 76% accuracy. Results suggest these models can be applied to understand and predict the chemical stability of mRNA in the near future.
研究动机与目标
- 解决因 2'-羟基基团介导的磷酸二酯键自水解导致的内在 mRNA 不稳定性问题。
- 通过实现序列稳定性的计算预测,减轻 mRNA 药物开发中的实验负担。
- 评估长短期记忆网络(LSTM)、门控循环单元(GRU)和图卷积网络(GCN)在不同环境条件下预测反应活性与降解风险的性能。
- 识别在不同 pH 值、温度和金属离子条件下预测 mRNA 稳定性的最有效深度学习架构。
- 提出一种二分类扩展方法,以最小化在合成前筛选不稳定序列时的假阴性结果。
提出的方法
- 在斯坦福 OpenVaccine 数据集的 6034 条 mRNA 序列上训练长短期记忆网络(LSTM)、门控循环单元(GRU)和图卷积网络(GCN)模型。
- 将结构化数据(最低熵碱基配对概率矩阵)和非结构化数据(作为 GCN 的节点与边)作为输入。
- 采用 k 折交叉验证,并在 107 个核苷酸(训练集)和 130 个核苷酸(测试集)的序列上进行训练。
- 在多种环境条件下优化模型:pH 10、pH 10 加 Mg2+、50°C 和 50°C 加 Mg2+,使用 RMSE 和 MAE 作为评估指标。
- 整合多目标预测结果,并通过所有条件下均值 RMSE 和 MAE 评估整体性能。
- 提出一种二分类扩展方法,以最小化稳定性筛选中的假阴性,重点识别高风险序列。
实验结果
研究问题
- RQ1深度学习模型能否在不同环境条件下准确预测 mRNA 序列的反应活性与降解风险?
- RQ2LSTM、GRU 和 GCN 架构在预测 mRNA 稳定性与降解倾向方面表现如何比较?
- RQ3在多种降解条件下,哪种深度学习模型的预测误差(RMSE 与 MAE)最低?
- RQ4这些模型在多大程度上可减少 mRNA 药物开发中对实验测试的依赖?
- RQ5二分类扩展是否能提高对不稳定 mRNA 序列筛选的可靠性?
主要发现
- 图卷积网络(GCN)在预测反应活性方面表现最佳,RMSE 为 0.249。
- 门控循环单元(GRU)模型在预测降解风险方面表现最佳,RMSE 为 0.266。
- 在综合所有目标变量时,GRU 模型实现了最高的整体准确率(76%)。
- GRU 模型在多种条件下也表现出最低的 MAE,表明其预测误差较小。
- GCN 模型在多个条件下表现出较低的 MAE,说明其个体误差更小,不易出现大幅偏差。
- 尽管性能良好,但模型因预测值范围较大(0.5–4),仍存在 24% 的平均百分比误差,表明存在不可忽视的残余误差。
更好的研究,从现在开始
从阅读论文到最终审阅,大幅缩短您的研究时间。
无需绑定信用卡
本解读由 AI 生成,并经人工编辑审核。