Novelty Is Not Free | AI Learning Guide

Look at the three pressures in the comparison. Novelty wants a structure unlike anything known — that is how you escape existing patents and known problems. But unfamiliar structures are exactly the ones chemists struggle to build, so synthesizability pulls the other way, toward familiar chemistry. And safety pulls a third way: features that make a molecule bind tightly can also make it hit unintended proteins. A model told only to maximize binding will happily hand you a molecule that cannot be made or should not be given to anyone. The real target is a molecule that is acceptable on all three at once.

0:00 / 0:00

Auto-play

Novelty, synthesizability, and safety are not independent goals. They compete, and a candidate is only useful if it clears all three at once.

What each pressure rewards and punishes

Novelty

Rewards: structures unlike known compounds, escaping existing patents and known liabilities
Punishes: unfamiliar scaffolds are harder to make and less covered by prior safety data

Synthesizability

Rewards: chemistry close to known reactions, short routes, good yields
Punishes: pushing toward familiar structures, which reduces novelty

Safety

Rewards: features that avoid off-target binding and instability
Punishes: tight, unusual binding features that may also hit unintended proteins

A model that optimizes predicted binding alone will propose molecules that cannot be made or should not be given to anyone. The design goal is a molecule acceptable on all three axes, not maximal on one.