liveBook · Manning

Jaccard similarity

As explained in Build AI Drug Discovery Pipelines

Tanimoto similarity is a metric that compares two sets or binary molecular fingerprints as the ratio of shared features to their union.

Jaccard similarity, also called Tanimoto similarity, compares two sets or binary molecular fingerprints by dividing shared nonzero features by all nonzero features present in either fingerprint. Scores range from 0 (no shared features) to 1 (complete feature overlap); a score of 1 does not prove two molecules are identical. section 2.5.1

In ligand-based screening, fingerprints act as coordinates in feature space: nearby molecules are more alike according to the selected representation. Similarity searching uses this score to find compounds resembling a query molecule. section 2.5.1

Interpret a similarity search with these practical cautions:

  • Compare thresholds within the chosen metric: Dice scores are generally greater than or equal to Tanimoto scores, so the same threshold can produce more hits with Dice. section 2.5.2
  • Treat thresholds as representation-dependent: descriptors can make searches overly conservative or retain too many false-positive hits. section 2.5.2
  • Use approximately 0.85 over common 2D fingerprints only as a rule of thumb for likely shared biological activity among drug-like molecules. section 2.5.2

Query-based screening

Generate molecular fingerprints, select one compound as a query, compute bulk Tanimoto or Dice scores against a library, inspect the highest and lowest matches, then retain compounds above a chosen threshold.

Figure 2.16 Mathematical definition of Tanimoto and Dice similarity coefficients with example computation for assessing the similarity of molecule A to molecule B.

Use the mathematical comparison and worked feature-overlap example to distinguish Tanimoto from Dice similarity. section 2.5.1

How Build AI Drug Discovery Pipelines covers it

How deep it goes
Defines the set ratio, contrasts Dice, demonstrates query scoring, and stops short of deeper fingerprint theory.
Examples
Uses Malaria Box compounds, a single query molecule, score extremes, thresholded matches, and saved top hits.
What it assumes
Assumes molecular fingerprints and binary features, while defining similarity, intersections, unions, and thresholds from first principles.
What you can do after
Apply it via fingerprint utilities to compute query-library scores, filter hits, and save top similar compounds.
How it connects
Connects molecular fingerprints, binary vectors, Dice similarity, descriptors, and ligand-based screening.

Also explained in

Compare how 3 books explain Jaccard similarity →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Build AI Drug Discovery Pipelines. The section links point to the parts of the book it draws on.