PharmCast: rapid generation of three-dimensional pharmacophore fingerprints from two-dimensional structure without conformer generation

Technical Summary: PharmCast – Rapid Generation of 3D Pharmacophore Fingerprints from 2D Structure

Problem Statement
Three-dimensional (3D) pharmacophore fingerprints are powerful descriptors for identifying binding features and enabling scaffold hopping, as they capture the spatial arrangement of chemical features (e.g., donors, acceptors, hydrophobic regions) independent of the molecular skeleton. However, their utility in large-scale virtual screening and optimization loops has been historically limited by computational cost. The standard pipeline requires generating an ensemble of conformers for each molecule to determine which pharmacophoric triangles can be presented. In the reference pipeline described, generating 100 conformers consumes 2.82 seconds per molecule, while the actual fingerprint calculation takes only 0.039 seconds. Consequently, processing a five-million-compound library requires approximately five and a half core-months, rendering 3D pharmacophore similarity impractical as a primary screening objective or for use in iterative optimization.

Methodology
The authors introduce PharmCast, a surrogate model designed to bypass the conformer generation step entirely. Instead of calculating 3D geometries, PharmCast predicts the ensemble pharmacophore fingerprint directly from the two-dimensional (2D) molecular constitution (SMILES string).

• Model Architecture: PharmCast is a feedforward neural network. It takes as input a 2,048-bit circular Morgan fingerprint (radius 2) and 11 whole-molecule descriptors (including molecular weight, cLogP, topological polar surface area, rotatable bonds, and Labute accessible surface area), totaling 2,059 inputs. The network maps these inputs to all 10,549 bits of the PharmPrint ensemble fingerprint through two hidden layers (1,024 and 512 units), resulting in approximately 8.05 million parameters.
• Training Data: PharmCast version 10 was trained on 5,887,229 molecules drawn from three distinct sources to ensure broad chemical and conformational coverage:
  1. Enamine Screening Collection: ~4.6 million compounds (filtered to 600 Da), representing synthesized molecules intended for biological interrogation.
  2. Activity-backed ChEMBL: ~1.2 million compounds (142–1000 Da) with recorded biological activity, ensuring synthetic accessibility and biological relevance.
  3. Protein Loop Peptides: ~123,000 loops (2–6 residues) excised from high-resolution PDB crystal structures. These were included to extend the conformational range, as peptides possess significantly more rotatable bonds (median 16) than typical screening collection compounds (median 5).
• Reference Standard: The "ground truth" for training and evaluation was generated using a conventional pipeline: 100 conformers generated via RDKit ETKDGv3 with UFF minimization, followed by bitwise OR combination to form the ensemble fingerprint.
• Evaluation Strategy: The model was evaluated on three strictly held-out populations not present in the training set: 155,648 purchasable Enamine screening collection compounds, 139,700 activity-backed ChEMBL compounds, and 13,500 peptide loops. Performance was measured against the reproducibility ceiling of the reference calculation itself (re-running the reference with a different seed yields a median error of 0.006 and Pearson r of 0.995).

Key Contributions

1. Elimination of Conformer Generation: The primary contribution is the removal of the conformational stage from the fingerprinting workflow. By predicting the ensemble fingerprint directly from 2D structure, the method addresses the computational bottleneck that has confined pharmacophore similarity to small, curated sets.
2. Surrogate Model with Defined Applicability Domain: The paper provides a rigorous measurement of the surrogate's agreement with the reference calculation, explicitly quantifying failure modes relative to training set composition. It demonstrates that simply increasing the size of the training set within the same chemical space (screening collection chemistry) does not improve performance on out-of-domain data (large compounds); rather, adding missing chemistries (ChEMBL and peptides) is required.
3. Version Series Analysis: The authors present a historical analysis of model versions (2 through 10), showing that while screening collection ranking accuracy plateaus in the high 0.93s, performance on large, diverse compounds (ChEMBL >600 Da) improves significantly only when the training set composition shifts to include non-screening-collection chemistry.

Results

• Computational Speed: PharmCast achieves a massive speedup. Generating and comparing fingerprints for two molecules takes 0.584 ms with PharmCast, compared to 5.71 s with the conventional pipeline. Screening a 4.6-million-compound collection takes 29.3 minutes with PharmCast versus an extrapolated 3,694 core-hours for the conventional method.
• Accuracy:
  • Screening Collection Chemistry: Median fingerprint error of 0.008, Pearson r of 0.980, and pairwise ranking accuracy of 0.936.
  • Loop Peptides: Median error of 0.016, Pearson r of 0.984, and ranking accuracy of 0.952.
  • Activity-backed ChEMBL: Median error of 0.027, Pearson r of 0.936, and ranking accuracy of 0.889.
  • These errors are close to the reference calculation's own reproducibility ceiling (error 0.006, r 0.995).
• Independence from 2D Similarity: The model successfully predicts 3D pharmacophoric similarity even when 2D Morgan similarity is low. For example, saquinavir and indinavir (unrelated scaffolds) have a 2D Morgan similarity of 0.303 but a pharmacophore similarity of ~0.84, which PharmCast accurately predicts.
• Applicability Domain: Performance degrades at the upper end of the molecular weight range (>700 Da) where the training set is sparse and dominated by peptides. The model is most accurate on screening collection chemistry and least accurate on the diverse, activity-backed ChEMBL set, highlighting the dependency on training set coverage rather than molecular weight alone.

Significance and Claims
The paper claims that ensemble pharmacophore fingerprints can now be predicted from molecular constitution alone at a cost suitable for screening entire collections and integration into optimization loops. The authors emphasize that the descriptor records what a molecule can present (an OR over conformations) rather than committing to a specific pose, making it a property of constitution that is tractable for machine learning.

The work supports a workflow where the surrogate model (PharmCast) is used for rapid ranking and prioritization of structurally novel compounds (scaffold hopping), while the reference calculation is reserved for rescoring the top candidates to ensure geometric validity. The authors note that the model's performance is governed by training-set coverage, suggesting that future improvements will depend on expanding the chemical diversity of the training data rather than simply increasing the volume of similar data. The paper concludes that this approach enables the practical application of 3D pharmacophore similarity in scenarios previously restricted to 2D fingerprints due to computational constraints.