Figure 14: Multi-graph tensor representation of RNA conformational ensembles,

and the associated symmetry groups acting on each axis. We process a set of $k$ RNA backbone conformations with $n$ nodes each into a tensor representation. Each multi-state GNN layer updates the tensor while being equivariant to the underlying symmetries; pseudocode is available in Listing 1. Here, we show a tensor of 3D vector-type features with shape $n\times k\times 3$. As depicted in the equivariance diagram, the updated tensor must be equivariant to permutation $S_n$ of $n$ nodes for axis 1, permutation $S_k$ of $k$ conformers for axis 2, and rotation $SO(3)/O(3)$ of the 3D features for axis 3.