Workflow Example: Conditioned Generation¶
This workflow is for cases where you do not just want plausible molecules, but molecules that satisfy a more specific design objective. In MolCraftDiffusion, the conditioning can be either property-based or structure-based.
Two Main Modes¶
Mode |
Question it answers |
Typical method |
|---|---|---|
Property-conditioned |
“Can I bias generation toward a target value or regime?” |
CFG, GG, or hybrid guidance |
Structure-conditioned |
“Can I preserve one part of the molecule while redesigning or extending another?” |
Inpainting or outpainting |
Conceptual Flow¶
[Base or adapted model]
|
+--------------+--------------+
| |
v v
[Property guidance] [Structure guidance]
| |
v v
[Property-directed generation] [Structure-directed generation]
What Changes with Conditioning¶
This workflow adds a design constraint to generation. Rather than only sampling plausible molecules, you guide the model toward a target property profile or a target structural pattern.
Typical Outcome¶
This is often the right workflow when you want:
molecules around a known scaffold,
candidates with a target property profile,
a more directed search than unconditional generation can provide.