What are the atom features?

[ekarais]

Figure out what the atom features are and where they come from.

[ekarais]

The atom features seem to be the following:

• first 5 features are a one-hot encoding of the atom type (H, C, O, N, F)
• The second 5 features are again the one-hot encoding, but they are each multipled by Z_tilde. Z_tilde is just the atomic number of the atom divided by 9 (because 9 is the largest atomic number (it belongs to Fluoride) in the dataset, so Z_tilde is normalized).
• The last 5 features are again the one-hot encoding, but they are multiplied by Z_tilde ** 2

And that is it.

I strongly suspect that these features come from the so-called cormorant paper (https://github.com/risilab/cormorant) because the EGNN paper says they use cormorant's preprocessing and EMPSN tries to succeed EGNN. I need to look into it more carefully though.

[ekarais]

After some more digging, I found out that these atom features are used in EMPSN, EGNN and Cormorant. This feature set originates from the Cormorant paper (see Appendix, p. 1).