🧪 Molecule Screening Engine — Similarity Search Live

Tanimoto similarity (also called the Jaccard coefficient in this binary form) measures overlap between two sets as |A ∩ B| / |A ∪ B| — the number of shared members divided by the number of members in either set. Applied to molecular fingerprints, where each bit flags the presence of a substructure or feature, it gives a score from 0 (no shared features) to 1 (identical fingerprints). It is the de-facto standard similarity metric in cheminformatics because it is simple, fast to compute on bit vectors, and correlates well with shared biological activity between compounds.

A molecular fingerprint encodes a compound's structure as a fixed-length bit vector, where each bit (or hashed bit position) represents the presence or absence of a particular substructure, functional group, or circular atom environment. Real cheminformatics tools compute fingerprints like ECFP or MACCS keys from a molecule's actual graph structure. This simulation uses simplified synthetic 128-bit vectors to demonstrate the same downstream mathematics — the similarity search, ranking, and clustering logic — without requiring a real chemistry parser.

Virtual screening ranks a large compound library by similarity to a known active molecule (the target), on the assumption that structurally similar molecules tend to share biological activity. In practice this lets researchers prioritise a small number of promising candidates from a library of millions for expensive laboratory testing, rather than testing everything blindly. This simulation reproduces that exact ranking step: every candidate's real Tanimoto score against the target determines its position in the ranked list.

The synthetic library is generated from a handful of scaffold fingerprints, each built by mutating the target fingerprint by a different amount, then each candidate is a further small mutation of one scaffold. That mirrors how real compound libraries contain chemical series (families of related analogues) rather than pure random noise. Because node position is derived directly from each candidate's Tanimoto score, molecules from the same scaffold family naturally land near each other, and the whole cluster moves closer to the target sphere as similarity increases.

The similarity mathematics is real: every score shown is the exact |A∩B|/|A∪B| Tanimoto coefficient computed over live 128-bit vectors, recalculated whenever the target, library size, scaffold diversity, or threshold changes. What is synthetic is the chemistry itself — the fingerprints are procedurally generated bit patterns rather than derived from real molecular graphs, since there is no chemistry parser running in the browser. The ranking, thresholding, and score distribution you see are genuine outputs of that computation, not scripted animation.