Pharmprint Evolution: 10K Bit Triplet Fingerprints

PharmCast · Eidogen-Sertanty

PharmCast: predicting the 3D fingerprint from the flat structure

Molecules do not act through their 2D drawing. They act through the three-dimensional arrangement of chemical features they can present, and capturing that means generating hundreds of 3D conformations per molecule. That is the expensive part: about 2.5 seconds a molecule, of which 98 percent is building the conformers and 2 percent is the actual measurement. PharmCast predicts the finished pharmacophore fingerprint directly from the flat 2D structure and skips conformer generation entirely. It runs in 0.0006 seconds per molecule instead of 2.5, and agrees with the real calculation at a correlation of 0.964. A complete comparison of two molecules from structure alone takes 0.034 milliseconds against 4.96 seconds.

This is not a new way of mapping pharmacophores. The features, the triplets and the 10,560-bit encoding are exactly the method described above. What changes is the cost of asking how similar two molecules are, and that cost is what has always bounded how much chemistry a comparison-driven pipeline could afford to look at.

The practical consequence is the whole point: a multi-million compound catalogue can be screened in minutes instead of months, and the expensive real calculation gets spent only on the few hundred candidates that decide the answer. It does not replace the conventional calculation. It is a filter that lets far more chemistry be examined, and the real calculation still decides.

Sequence does not determine structure

Training the peptide-aware model required a corpus of short loops pulled from experimentally determined protein structures, and that corpus produced a result that runs against a common assumption. Searching 19,108 loop sequences for the largest disagreement between a sequence's own observed conformations returns LGGK, seen in 13 independent Protein Data Bank entries, every one of them determined at 1.8 Angstrom resolution or better with a free R-factor at or below 0.20. Those conformations disagree by 2.12 Angstroms of backbone deviation, and across the whole search backbone deviations reach 3.1 Angstroms and all-atom deviations 6.3.

What makes it more than a curiosity is the shape of the distribution. Across the 78 pairs of observed LGGK conformations the median pharmacophoric similarity is 0.56, with 72 percent falling below 0.70, and the values separate into distinct populations with gaps between them. That is not one presentation with noise around it. It is several. Measured more broadly across 1,481 sequences seen three or more times, the median agreement between two observations of the same sequence is 0.878, with 18 percent of comparisons below 0.70. For short loops, sequence identity does not imply pharmacophoric identity, and the spread is signal rather than error.

The model family

PharmCast-S is trained on the screening collection alone. PharmCast-SP adds those protein loop peptides, ensemble enhanced so each peptide's fingerprint carries both the rigid conformations actually observed in the PDB and 100 computed ones, and is reported separately. The newest composite trained on 2,114,521 molecules, 2,056,482 from the collection plus 58,039 loop peptides, and finished at 0.893 validation agreement, converged rather than truncated: the last ten epochs moved it by 0.0011.

The method underneath is the PharmPrint work described above, published with Malcolm McGregor in J. Chem. Inf. Comput. Sci. 1999, 39, 569-574 and 2000, 40, 117-125. For the library these models are trained and tested on, see the screening collection profile. For a campaign that uses pharmacophoric similarity as its objective, see the ChIP project.