Research

How do you measure affinity at high throughput
without losing the quality a model needs?

Protein models have already been trained on nearly every structure ever solved,
yet for most targets that matter there is almost no affinity data.

What does a metric have to capture
before a higher score translates to the bench?

A docked pose has long been judged correct by how close it sits to the crystal pose,
not by whether it clashes with the protein
or keeps the contacts that hold the ligand in place.

How do you tell real improvement
from overfitting to the last result?

An agent can now run an experiment, read the result,
and rewrite its own method before the next one.