Pharmprint Evolution: 10K Bit Triplet Fingerprints

From Combined Bits to Individual Conformer Matching

Based on the original PharmPrint method (McGregor & Muskal, 1999, 2000)

3-Point Pharmacophore Triangles → 10K Bits

Each pharmacophore is a triangle of 3 features with specific distances:

d2 d3 d1 p1 p3 p2

7 Pharmacophore Types: A (acceptor), D (donor), N (negative), P (positive), H (hydrophobic), R (aromatic), X (other)

6 Distance Ranges: 2.0-4.5, 4.5-7.0, 7.0-10.0, 10.0-14.0, 14.0-19.0, 19.0-24.0 Å

~10,549 possible triangular pharmacophores → Each gets 1 bit → 10K bit fingerprint

Original Pharmprint
Many conformers
C1
C2
C3
C4
Individual 10K bit fingerprints
C1:
... (10K bits)
C2:
... (10K bits)
C3:
... (10K bits)
C4:
... (10K bits)
All conformer shapes superposed:
OR ALL BITS
Combined 10K bit fingerprint
... (10K bits)
Problem: Cannot identify which conformer matched
Polypharmprint Method
Hundreds of conformers
C1
C2
C3
C4
Keep each 10K bit fingerprint separate
Individual conformer shapes (kept separate):
C1 C2 C3 C4
Individual 10K bit fingerprints (kept separate)
C1:
... (10K bits)
C2:
... (10K bits)
C3:
... (10K bits)
C4:
... (10K bits)
Advantage: Can trace back to exact conformer that matched
Polypharmprint Matching Process
Query Molecule
C1
...
C2
...
C3
...
C4
...
BEST MATCH
Hit Molecule
C1
...
C2
...
C3
...
C4
...
Key Breakthrough:
We can access the exact starting 3D conformation of the matching conformer pharmprint, enabling dramatically more efficient exploration of the conformational space when aligning query and hit molecules. This transforms molecular alignment from a computationally expensive global search into a targeted, efficient process.

Application note · Eidogen-Sertanty

Predicting the fingerprint: a neural surrogate for PFP

An ensemble PolyPharmPrint fingerprint (PFP), the union of the triplet bits set across a molecule's whole conformer ensemble, requires embedding 100 conformers per molecule and costs about ten seconds per compound, which caps how much chemical space a pharmacophore-driven design search can cover. So we trained a network to predict the full 10,560-bit ensemble fingerprint directly from a canonical SMILES, and validated it on 9,975 withheld molecules from the Enamine screening collection.

0.828median Matthews correlation, predicted against real fingerprint
0.911Pearson r on pairwise pharmacophore similarity
98%correct on which candidate is closer, when the gap exceeds 0.10
30,645×faster than explicit conformer generation
Pipeline: a canonical SMILES becomes a 2,048-bit Morgan fingerprint plus 11 descriptors, passes through hidden layers of 1,024 and 512 units to 128 numbers, and an SVD inverse expands those to the predicted 10,560-bit ensemble fingerprint shown against the truth
What goes in and what comes out, at the median accuracy of the test set. A canonical SMILES becomes a 2,048-bit substructure fingerprint plus 11 descriptors, 2,059 numbers in all. Those pass through hidden layers of 1,024 and 512 units to 128 numbers, which are coordinates in a compressed space learned from real fingerprints by truncated singular value decomposition. Multiplying back through that decomposition expands them to the full 10,560 bits, shown against the truth with every disagreement marked.

It is not covertly re-deriving 2D similarity

A model that had learned to reproduce Morgan similarity under another name would be useless for scaffold hopping, which is the whole point of a pharmacophore search. It is not doing that. Agreement holds flat across the full range of 2D similarity, from pairs sharing under a tenth of their two-dimensional structure to close analogues: Pearson runs 0.906, 0.910, 0.917, 0.919 and 0.900 across those bands.

Hexbin plot showing predicted pharmacophore similarity tracking the real value along the line of exact agreement, Pearson 0.911 over 30,000 unique withheld pairs
Predicted pharmacophore similarity tracks the real value over 30,000 unique withheld pairs, at Pearson 0.911 and Spearman 0.910. Mean absolute error 0.054, with a small systematic bias of +0.044 that a search can absorb because it displaces every candidate equally.
Scatter of predicted against real pharmacophore similarity, coloured by 2D Morgan similarity, showing accuracy independent of 2D similarity
Accuracy is independent of 2D similarity. Green where two molecules share almost no 2D structure, red where they are close analogues.

Reading the axes. Both plots use the same notation. For any two molecules A and R, AFP and RFP are their real ensemble fingerprints as computed by pfpall, while APS and RPS are those same two fingerprints as predicted by the surrogate from structure alone. So the horizontal axis, Tanimoto(AFP, RFP), is the true pharmacophore similarity between the pair, and the vertical axis, Tanimoto(APS, RPS), is the surrogate's estimate of it. Perfect prediction would put every point on the diagonal.

Where it applies, and where it does not

The honest reading is that this is a broad-phase instrument. Asked which of two candidates sits closer to a reference, it is 98% correct when the true gap exceeds 0.10, 88% between 0.05 and 0.10, 73% between 0.02 and 0.05, and 57% below 0.02. Its mean absolute error of 0.054 is about the size of the gap between rank 50 and rank 200 in a typical ranking, which is exactly why high bitwise fidelity coexists with modest performance among close finalists. Use it as a fitness proxy while a design search is still ranging widely, at 3,065 molecules per second in a single process, then confirm the survivors with the real calculation.

Limits. Trained below molecular weight 600, and ranking of larger designs collapses above roughly 640. The 0.922 decoder ceiling is a property of the 128-component decomposition rather than of chemistry, and rises with more components. The deeper limit is that an ensemble PFP is a union over 100 conformers, so it encodes conformational behaviour that 2D topology only partly determines. The trained model itself is not distributed.

This is what makes a wider search practical: the ChIP de novo design engine scores candidate molecules on pharmacophore similarity to a reference drug, and the cost of that scoring is what bounds how much route and building-block space a campaign can afford to explore. See the ChIP project for a worked campaign.

References

Original PharmPrint Method:

  • McGregor, M. J., & Muskal, S. M. (1999). Pharmacophore fingerprinting. 1. Application to QSAR and focused library design. J. Chem. Inf. Comput. Sci., 39(3), 569-574. [PDF]
  • McGregor, M. J., & Muskal, S. M. (2000). Pharmacophore fingerprinting. 2. Application to primary library design. J. Chem. Inf. Comput. Sci., 40(1), 117-125. [PDF]