Computational identification of a phospholipidosis toxicophore using 13C and 15N NMR-distance based fingerprints

Modified 3D-SDAR fingerprints combining 13C and 15N NMR chemical shifts augmented with inter-atomic distances were used to model the potential of chemicals to induce phospholipidosis (PLD). A curated dataset of 328 compounds (some of which were cationic amphiphilic drugs) was used to generate 3D-QSDAR models based on tessellations of the 3D-SDAR space with grids of different density. Composite PLS models averaging the aggregated predictions from 100 fully randomized individual models were generated. On each of the 100 runs, the activities of an external blind test set comprised of 294 proprietary chemicals were predicted and averaged to provide composite estimates of their PLD-inducing potentials (PLD+ if PLD is observed, otherwise PLD−). The best performing 3D-QSDAR model utilized a grid with a density of 8 ppm × 8 ppm in the C-C region, 8 ppm × 20 ppm in the C-N region and 20 ppm × 20 ppm in the N-N region. The classification predictive performance parameters of this model evaluated on the basis of the external test set were as follows: accuracy = 0.70, sensitivity = 0.73 and specificity = 0.66. A projection of the most frequently occurring bins on the standard coordinate space suggested a toxicophore composed of an aromatic ring with a centroid 3.5-7.5 Å distant from an amino-group. The presence of a second aromatic ring separated by a 4-5 Å spacer from the first ring and at a distance of between 5.5 Å and 7 Å from the amino-group was also associated with a PLD+ effect. These models provide comparable predictive performance to previously reported models for PLD with the added benefit of being based entirely on non-confidential, publicly available training data and with good predictive performance when tested in a rigorous, external validation exercise.