Fingerprints: Turning a Structure into a Fixed List of Numbers | AI Learning Guide

The point of a fingerprint is to make every molecule the same size. Instead of listing atoms, you ask a fixed set of questions — is this fragment here, is that pair of atoms here — and write down the answers as a row of ones and zeros. Now every molecule is a point in the same space, and comparing two molecules becomes comparing two rows. The formula on the page counts how many ones they share and divides by how many either one has, giving a number from zero to one. Notice what has been given up: from the row of ones and zeros you cannot draw the molecule back. Many different structures produce the same row. That is why fingerprints are used for searching and ranking, and not for building new molecules.

0:00 / 0:00

Auto-play

A fingerprint converts a variable-sized structure into a fixed-length numeric vector that records which local features are present. It makes molecules directly comparable as points, at the cost of discarding the structure itself.

Tanimoto similarity

For two binary fingerprints, let \(a\) and \(b\) be the number of features set in each and \(c\) the number set in both. Then \[ T = \frac{c}{a + b - c} \] gives a value between 0 and 1. A value near 1 means the two molecules share most of their features; a value near 0 means they share almost none. The measure is a similarity score, not a probability.

Graph or SMILES

Variable length
Full connectivity preserved
Structure can be reconstructed
Used when a model must build or edit a molecule

Fingerprint

Fixed length
Only presence of features recorded
Structure cannot be recovered
Used when a model must compare or rank molecules

References

[1]Chemical similarity — Wikipediaen.wikipedia.org