Definition
Structural diversity of AI-selected compounds as a fraction of the screening library's diversity, using pairwise Tanimoto distances. A falling score means the model is mining familiar scaffolds and narrowing chemical space.
The Measurement Gap
A discovery team maintains a 2.4 million compound screening library spanning 47 scaffold families. Over 3 years, the AI virtual screening platform recommended 1,847 compounds for synthesis across 11 programs. A structural diversity analysis reveals that selections come from just 12 scaffold families, with 68% from the 4 most heavily represented families in the training data.
The AI recommends variants of structures it has the most training data for because those produce the most confident predictions. It avoids novel scaffolds because it cannot predict their behavior reliably, so they rank low. The ranking function rewards confidence, confidence correlates with training data density, and training data density correlates with familiar chemistry.
This is invisible in standard metrics. Hit rate is good because selections come from well-characterized families. VSRVR accuracy is decent because the model operates on familiar territory. But structural novelty is declining. The pipeline progressively narrows. The organization spends millions on AI-guided discovery while systematically excluding novel scaffolds that represent the strongest patent positions and breakthrough potential.
No current metric measures whether AI compound selection spans available chemical diversity or concentrates in well-represented structural families.
Formula
CSEB = D(selected) / D(library)
Where D = structural diversity index based on pairwise Tanimoto distances on Morgan fingerprints
Worked Example:
18 months of AI selections for a kinase inhibitor program:
Full screening library (2.4M compounds): D(library) = 0.73. Moderate structural diversity.
AI-selected compounds (412 over 18 months): D(selected) = 0.41.
CSEB = 0.41 / 0.73 = 0.562.
The AI cuts almost half the structural diversity from its recommendations. Of 47 scaffold families, the AI selected from 14. Within those 14, 73% of selections came from 4 families: aminopyrimidines (183), quinazolines (67), pyrazolopyrimidines (42), and indazoles (19). The 33 excluded families include macrocycles, PROTACs, covalent warhead scaffolds, and allosteric binders representing the frontier of medicinal chemistry.
Trending over quarters: CSEB Q1=0.61, Q2=0.57, Q3=0.53, Q4=0.49. The declining trend shows a positive feedback loop: select familiar chemistry, generate data on familiar chemistry, train on familiar data, select even more familiar chemistry.
Proposed formula, pending validation pilots.