Why the Choice of Representation Decides What the Model Can Do | AI Learning Guide

Compare the three columns and ask one question of each: can I get the molecule back? The graph says yes, because the molecule is stored as its parts, so a model can add an atom or change a bond and still hold a real structure. The string says yes, but with a catch — the model is predicting characters, and nothing in that process knows how many bonds a carbon can form, so it can write something readable and impossible. The fingerprint says no. Once the structure has been reduced to a row of ones and zeros, there is no route back. That single test — is the encoding invertible — tells you immediately whether a method could ever be used to design a molecule, or only to sort one.

0:00 / 0:00

Auto-play

Representation is chosen to match the task. Generation needs an encoding that can be decoded back into a real structure; comparison and ranking need a fixed-length encoding that makes molecules directly comparable.

What each encoding supports

Graph

Keeps connectivity and labels
Editable atom by atom
Supports generation
Awkward as a fixed-size model input

SMILES

Keeps connectivity as a sequence
Reads like text, so sequence models apply
Supports generation, but validity must be checked
One molecule has many strings

Fingerprint

Keeps only feature presence
Fixed length, directly comparable
Supports search and ranking
Cannot be decoded back to a structure

A representation that cannot be inverted cannot be used for generation, no matter how good the model is. This is the single most useful test to apply when reading about a new method.