3D CNN的缺陷：
旋转不等变、很难扩展到大的蛋白质口袋
The model was constructed based on the 3D CNN model which is not equivariant to rotation and hard to scale to large protein pockets.
Vector Feature-based Equivariant Networks
基于向量特征的等变网络
However, they require the input and hidden features for each layer to be equivariant, which does not fit the vector features such as side-chain angles of each amino acid. To address this problem, Deng et al. (2021) proposes vector neurons which extends 1D scalar neurons to 3D vectors and also defines a set of equivariant operations in the vector space. Later, Jing et al. (2021) develops the geometric vector perceptrons to efficiently propagate information between scalar features and vector features. Schütt et al. (2021) introduces equivariant message passing for vector representations using equivariant features. However, they can only linearly combine 3D vectors to satisfy the rotation equivariance, in essence, limiting the model’s geometric expressiveness.
等变神经网络
等变神经网络通常采用基于 GNN 的架构来实现 3D 对象的全局旋转等变。然而，它们要求每层的输入和隐藏特征是等变的，这不适合每个氨基酸的侧链角度等向量特征。为了解决这个问题，邓等人。 (2021) 提出了向量神经元（vector neurons），它将 1D 标量神经元（scalar neurons）扩展到 3D 向量，并在向量空间中定义了一组等变运算。后来，Jing 等人。 (2021)开发了几何矢量感知器（perceptrons）来有效地在标量特征和矢量特征之间传播信息。舒特等人。 (2021) 引入了使用等变特征的向量表示的等变消息传递。然而，它们只能线性组合 3D 向量来满足旋转等方差，本质上限制了模型的几何表达能力。
【ICML 2022】Pocket2Mol + Efficient Molecular Sampling Based on 3D Protein Pockets
)
