Focus on protease: inhibition helps inhibit viral replication; and it is conserved across most coronaviruses; so good point to start working in drug development.
- Take smaller binding fragments covering the binding site, and combine them together
- Try to combine these fragments together into a molecule that fits well with the binding site
protease inhibition is usually achieved with a covalent peptide bond, but this crowd-sourcing effort showed that
machine-learning rapid library synthesis
- begin with some guess for the model molecule
- then, use ML to perform modifications to the molecule really quickly by scanning though (“ML-prioritized rapid library synthesis”) a bunch of changes to the molecule
- pick and repeat
Molecular Transformer
THROW THE FUCKING REACTION INTO AN LLM, as WORDS
I desire death
So; taking reactants + reagents as input; guess the product.
as input; guess the product.
SABER
- decompose molecule into building blocks
- make biostesters of the building blocks
- change crap
limitations
ML can’t extrapolate into unknown search space and it could come up with bullshit; so to fix:
- using physics to create correct docking structures
- use ML to perform last mile optimization