Circular dichroism (CD) spectra task, sensitive to stereochemistry, show stereo models outperform non-stereo.

While including stereochemistry aids in some tasks, the added tokens and complexity expands the chemical space. For tasks that are not as stereoselective, this can hinder exploration.
3/3

For rediscovering stereoisomers, and docking ligands, the stereo aware models can outperform non-stereo aware models.

For larger protein pockets, big strucutral changes matter more than stereochemical changes; in these cases, stereo/non-stereo models are comparable.
2/3

Prior works do not consider stereochemistry of generation tasks; so we do just that!

We look at how string-based generative methods, like genetic algorithms and REINVENT, perform on these tasks when given stereo-awareness through chiral and isomeric string tokens.
1/3