Ressources

Anticipated LOTUS

Features

One possible solution to having features per node types in our model would be to use the textual description of the species and molecules

Species

In the case of species, we could think that the textual description of the species is just description of its taxonomy. $\rightarrow$ we could maybe use SciBERT for that. Here is an example:

• https://w.wiki/8paM

Molecules

For the molecules however, this is more difficult.

• one potential solution would be to use the information in the wikidata page and use it as text.
  • Here is an example : https://w.wiki/8qrK
• One other solution would be to use ChemBERT to contextualize the SMILES however that means that we use a different model that for the molecules.
• But then why couldn't we just use the molecular fingerprints ?

ChemBERT : https://huggingface.co/seyonec/ChemBERTa-zinc-base-v1