search

hanjq17 / GMN

[ICLR 2022] The implementation for the paper "Equivariant Graph Mechanics Networks with Constraints".
https://arxiv.org/pdf/2203.06442.pdf
MIT License
56 stars 6 forks source link

The distance threshold (_lambda=1.6) for edges in MD17 #2

Closed chocolates closed 1 year ago

chocolates commented 1 year ago

Hi, For the MD17 dataset, if I understand it correctly, you use the _lambda (https://github.com/hanjq17/GMN/blob/main/spatial_graph/md17/dataset.py, line 85 ) as the distance threshold to determine which nodes are connected by edges. According to the paper, this threshold should enable connecting 2-hop neighbors for every atom. May I ask how to determine this threshold, e.g., how do you know to set _lambda=1.6 rather than 1.5, 1.7 to guarantee connecting 2-hop neighbors? is there any reference for this value? Thank you in advance!

hanjq17 commented 1 year ago

Hi,

It was just out of empirical reason. The bond connection for these MD17 molecules happens to be distinguishable by the cutoff 1.6. I double checked the bond connection determined by cutoff=1.6 with that produced by rdkit, and it happened to be the same. However, cutoff like 1.5 or 1.7 will not work (I remember the exact threshold was really close to 1.6, otherwise there will be mis-connection). Overall, the threshold only work for MD17 data, and if you would like to work on other data, it would be better to apply packages like rdkit to generate high-fidelity bond connection info.

Hope these explanations well answer your question.

Best, Jiaqi

chocolates commented 1 year ago

Ok, thanks for clarifying it.