Fig. 7: Phenotypic screening with learned distance reveals scaffold hopping and drug prospecting opportunities.

Learned phenotypic distance (Twin-NN, y-axis) complements conventional chemical informatic distance based on chemical structure (Tanimoto coefficient on ECFP4 circular fingerprints, x-axis). The scatterplot contains all pairwise combinations of 83 NT-650 compounds (each pair is a dot), calculated from their average MI traces and chemical structures. Where the phenotypic distance is low (< 0.3) but the Tanimoto distance is average or high (> 0.4), molecular structure dissimilarity misses neuroactive similarity. We illustrate each quadrant with examples. Bottom left: low Tanimoto and phenotypic distance (both metrics agree that molecules are similar). Bottom right: low phenotypic distance and high Tanimoto distance (scaffold hopping opportunity). Top-left: low Tanimoto distance, but high phenotypic distance (classic “activity cliff”: disparate activity despite high structural similarity). Top right: high Tanimoto and phenotypic distances (both metrics agree that molecules are unrelated). Dot size reflects the Tanimoto similarity between the target profiles (as binary vectors) of the compound pairs. All time series in this plot are scaled to the minimum and maximum of the dataset (0 and 6750 MI units, respectively), and the y-axis is plotted on this normalized 0 to 1 scale. Source data are provided in the Source Data File.