[论文解读] AlphaDesign: A graph protein design method and benchmark on AlphaFoldDB

Zhangyang Gao, Cheng Tan+1 more|arXiv (Cornell University)|Feb 1, 2022

Protein Structure and Dynamics被引用 25

一句话总结

AlphaDesign 引入一个基于 AlphaFold 的结构到序列蛋白质设计的新基准，并提出 ADesign，一种带角度特征的基于图的方法、一个简化的图变换编码器，以及一个具备置信度感知解码器，达到当前最先进的准确性和速度。

ABSTRACT

While DeepMind has tentatively solved protein folding, its inverse problem -- protein design which predicts protein sequences from their 3D structures -- still faces significant challenges. Particularly, the lack of large-scale standardized benchmark and poor accuray hinder the research progress. In order to standardize comparisons and draw more research interest, we use AlphaFold DB, one of the world's largest protein structure databases, to establish a new graph-based benchmark -- AlphaDesign. Based on AlphaDesign, we propose a new method called ADesign to improve accuracy by introducing protein angles as new features, using a simplified graph transformer encoder (SGT), and proposing a confidence-aware protein decoder (CPD). Meanwhile, SGT and CPD also improve model efficiency by simplifying the training and testing procedures. Experiments show that ADesign significantly outperforms previous graph models, e.g., the average accuracy is improved by 8\%, and the inference speed is 40+ times faster than before.

研究动机与目标

• 使用 AlphaFold DB 在跨物种和不同长度上建立一个大规模、标准化的结构到序列蛋白质设计基准。
• 开发一个开源的基于图的设计方法（ADesign），在准确性和效率上优于现有模型。
• 研究新特征（蛋白质角度）、简化的图变换器编码器，以及具备置信度感知解码器对设计性能的影响。
• 评估该方法在长度不受限 vs 受限以及物种感知 vs 联合数据集上的性能。
• 通过标准化训练/验证/测试分割并报告恢复率，提供公平、可重复的比较。

提出的方法

• 将蛋白质表示为一个 k-NN 图，包含 12 个节点特征和 23 个边特征，其中包含新的角度特征（alpha、beta、gamma）和二面角。
• 使用简化的图变换器（SGT），其中注意力权重通过单个多层感知机（MLP）学习，而不是通过分离的 Q/K 投影。
• 用平行、具备置信度感知的蛋白质解码器（CPD）替代自回归解码，并使用一个学习得到的置信分数以实现并行预测。
• 结合一维卷积神经网络以捕捉图派生特征的局部序列依赖，从而实现上下文感知的序列预测。
• 端到端训练，采用交叉熵损失来预测残基类型，并在 SL/SF 与 JL/JF 基准上使用相同的优化设置。
• 基准设置利用 AlphaFold DB，包含长度自由和物种感知配置，以及用于公平比较的标准化数据分割。

实验结果

研究问题

• RQ1 Does AlphaDesign provide a fair, large-scale benchmark for graph-based protein design on AlphaFoldDB across species and lengths?
• RQ2 Can ADesign achieve state-of-the-art accuracy and efficiency compared to prior graph-based design methods?
• RQ3 How do the newly introduced protein angle features, the simplified graph transformer, and the CPD decoder contribute to performance gains?
• RQ4 What is the effect of length restrictions and species-aware vs joint training on model performance?
• RQ5 What factors most drive performance improvements in graph-based protein design?

主要发现

• ADesign achieves higher recovery accuracy than prior graph-based models by about 8% on average under SL/SF settings.
• ADesign is substantially faster, with testing times over 40x faster than comparable methods in SL settings.
• Under SF/JF settings, ADesign maintains strong performance with averaged improvements around 9.4% over baselines.
• Ablation studies show that the newly added angle features contribute most to accuracy, followed by the simplified graph encoder and the CPD decoder.
• The confidence-aware decoder significantly enhances inference speed and provides notable accuracy gains.