Two ways to obtain a candidate
Library search
- Input: a fixed, finite collection of existing compounds
- Operation: score every entry, rank, keep the top ones
- Boundary: the answer must already be in the collection
- Output: a ranked subset of known molecules
Generative model
- Input: a learned probability distribution over structures
- Operation: sample a new structure from that distribution
- Boundary: only the encoding's expressive range limits the output
- Output: molecules that may never have been made or catalogued
Why sampling is the right word
The model holds probabilities, not a list. Producing a molecule means making a sequence of choices — which atom or fragment comes next, how it connects — where each choice is drawn according to what the model considers plausible. Because those draws are random, running the model twice on the same input gives two different molecules. This is a feature, not a defect: it is the only reason a single trained model can supply a whole pool of candidates instead of one guess.
The encoding constraint returns
A model can only generate in a language it can write. Fingerprints record which features are present but cannot be turned back into atoms and bonds, so they support scoring and similarity but not generation. Graph- and string-based encodings can be decoded, which is why they are the ones used when the task is to propose a structure rather than rank one.