Some Questions about the MPNN Implementation

[zengkaipeng]

Hi, I have some questions regarding the implementation details in the paper. Below are the specific points I’m unclear about:

1. How is the molecular graph constructed in the paper? Is it fully connected? According to the description in Section 5.2.2, when the distance between two atoms $r > r_c$, they are considered not connected by an edge. However, in the structures R, P, and TS, the distances between a pair of atoms $i$ and $j$ may vary across these structures. This could lead to situations where $i \in \mathcal{N}(j)$ in R but $i \notin \mathcal{N}(j)$ in P. How does this affect the definition of $m(i, j)$?

2. Following up on the first question, in Section 5.2.4, the intramolecular MPNN is mentioned. How is the neighborhood $\mathcal{N}(i)$ defined here? Is it based on atoms within the same molecule whose distance to $i$ is less than $r_c$, or are atoms considered neighbors if their distance is less than $r_c$ in any of the three structures (R, P, or TS)?

3. In Section 5.2.4, for intermolecular message passing, how is $\mathcal{N}(i)$ defined? Specifically, how do we determine whether two atoms $i$ and $j$ from different molecules are neighbors?