henhaobio
Sign in

bioRxivBioinformaticsPreprint

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

Steven M. MuskaliD, Malcolm J. McGregor

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

A three-dimensional pharmacophore fingerprint records the binding features a molecule can present. It is a description of a hand in search of a glove. Because it is defined by presented features instead of two-dimensional structure, it can identify pharmacophoric similarity between structurally distinct compounds, which is what scaffold hopping and non-obvious me-too design require. The descriptor has remained a niche tool because its cost is dominated by conformer generation. In the reference pipeline, generating 100 conformers requires 2.82 s of the 2.86 s needed to fingerprint one screening collection compound; the bit calculation requires 0.039 s. We therefore removed the conformational stage. PharmCast is a feedforward neural network that predicts all 10,549 bits of a PharmPrint ensemble fingerprint directly from a SMILES string. On the same machine, PharmCast generated pharmacophore fingerprints for two molecules and compared them in 0.584 ms, whereas the conventional conformer-based pipeline took 5.71 s. PharmCast version 10 was trained on 5,887,229 molecules drawn from a screening collection, activity-backed ChEMBL compounds from 142 to 1000 Da, and peptide loops excised from crystal structures. We evaluated 155,648 purchasable catalog compounds excluded from every training set, 139,700 activity-backed ChEMBL compounds not present in the version 10 training set, and 13,500 peptide loops reserved for testing. Median fingerprint error, Pearson r, and pairwise ranking accuracy were 0.008, 0.980, and 0.936 for screening collection chemistry; 0.016, 0.984, and 0.952 for loop peptides; and 0.027, 0.936, and 0.889 for activity-backed ChEMBL compounds. The reference calculation reproduces itself at an error of 0.006 and r of 0.995. Ensemble pharmacophore fingerprints can therefore be predicted from two-dimensional structure alone at a cost suitable for large-scale collection screening and virtual library exploration. Keywords: pharmacophore, fingerprint, scaffold hopping, surrogate model, virtual screening, applicability domain

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (8 Sept 2026) · CC BY 4.0

A feedforward network predicts 10,549-bit ensemble pharmacophore fingerprints from SMILES without conformers: 0.584 ms versus 5.71 s per pair comparison, with median MCC 0.860 to 0.914.

Three-dimensional pharmacophore fingerprints are costly because conformer generation takes 2.82 s of the 2.86 s needed per molecule, which confines them to small curated sets. The paper removes the conformational stage by predicting the conformer-ensemble fingerprint directly from 2D structure.

Key results

  1. ShownPredicting the fingerprint instead of computing it removes the conformer bottleneck. A similarity from two SMILES strings takes 0.584 ms against 5.71 s for the conventional route, and a whole-collection screen takes minutes.

    Per molecule in a library batch: 0.288 ms, of which 0.009 ms is the forward pass. Full collection (4,653,831 compounds): 29.3 min at 2,644 compounds/s on six worker processes. The conventional figure of 3,697 core-hours is extrapolated, not measured. Timings are on CPU. · 3.1 Computational performance; Abstract; 1.3 Computational cost; 3.7 Application: screening collection search…

  2. ShownPredicted fingerprints track the reference on three held-out sets. Agreement is weakest on ChEMBL by every measure. Error is lowest on screening-collection compounds, while per-molecule MCC is highest on loop peptides. Screening-collection error is close to the reference's own reproducibility.

    Median error / Pearson r / ranking accuracy: 0.008 / 0.980 / 0.936 (155,648 screening-collection compounds); 0.016 / 0.984 / 0.952 (13,500 peptide loops); 0.027 / 0.936 / 0.889 (139,700 ChEMBL). Reference vs itself: 0.006, r 0.995. Median MCC: 0.881 (screening collection), 0.860 (ChEMBL), 0.914 (loop peptides). · Abstract; 2.4 Test sets and evaluation; 3.2 Fidelity by test population

  3. ShownPharmCast tracks the reference pharmacophore similarity across all 2D-similarity bands (Pearson 0.942 to 0.981), so it is largely independent of 2D similarity, which scaffold hopping needs.

    On 5,000 held-out screening-collection molecules, Pearson agreement across Morgan-similarity bands (lowest to highest) is 0.966, 0.981, 0.973, 0.959 and 0.942. Pairs were balanced across bands (25,165 plotted), since 97.5% of uniformly drawn pairs fall below 0.20. · 3.3 Independence from two-dimensional similarity

  4. ShownOut-of-domain error is governed by training-set composition, not size. Adding more molecules from the same library (filtered at 600 Da) did not help large compounds. Adding chemistry absent from the collection (ChEMBL) did, and agreement flattened over versions 8 to 10.

    Ranking accuracy on 573 ChEMBL compounds above 600 Da: 0.701 (v2), 0.711 (v4, after 51% more training data), 0.746 (v5), 0.811 (v6), 0.832 (v7), 0.826 (v10). Screening-collection median MCC rises from 0.833 at 1.72 million molecules to 0.881 at 5.89 million. · 3.5 Performance by training-set size and…; 3. Results

  5. SuggestedThe surrogate works as a search objective. A screen against orforglipron found a purchasable compound with high pharmacophore similarity but low 2D similarity. A design campaign used the same ranking-then-rescoring workflow.

    Top purchasable hit: reference ensemble Tanimoto 0.786 at 591 Da, Morgan similarity 0.113, from 4,653,831 compounds screened in 29.3 min. Design campaign: 0.810 pharmacophore vs 0.145 2D similarity, with survivors rescored by the reference calculation. A single reference compound, no experimental validation. · 3.7 Application: screening collection search…; 4. Discussion

Methods, data and reproducibility

A feedforward network (2,059 inputs: Morgan radius-2 bits plus 11 descriptors; layers of 1,024 and 512; 8.05 million parameters) predicts the OR-over-conformers three-point pharmacophore fingerprint. It is trained with binary cross entropy on reference fingerprints from 100 RDKit conformers, and evaluated against that reference on held-out sets.

  • Enamine screening collection (June 2026), reference fingerprints for training; 4,774,670 source compounds · n = 4,609,488 reference fingerprints
  • Activity-backed ChEMBL compounds, 142 to 1000 Da, training source · n = 1,214,214 reference fingerprints (of 2,737,190 eligible)
  • Loop peptides (2 to 6 residues) excised from PDB crystal structures (2.5 Å or better, R-free 0.23 or better) · n = 136,494 loops (122,994 train, 13,500 test)
  • Held-out Enamine screening-collection test compounds excluded from all training sets · n = 155,648 scored (157,335 selected)
  • Held-out ChEMBL compounds not in the version 10 training set · n = 139,700
  • Real ChEMBL compounds above 600 Da, excluded from every model's training set (version comparison) · n = 573
Methods & tools
PharmCast, PharmPrint, PharmSim, RDKit ETKDGv3, UFF, Morgan fingerprints, PyTorch 2.9.1, ChEMBL, Protein Data Bank, Enamine screening collection, CHIP
Data & accessions
https://pharmcast.ai/
Software & environment
PyTorch 2.9.1; macOS 26.5.1; Apache-2.0 licence for code and model weights; each checkpoint records its model card (training composition, mass range, applicability domain).
Assay conditions & compute
Single Apple Mac Studio, M3 Ultra, 28 CPU cores, 60-core GPU, 256 GB unified memory; timings on CPU; screening of 4,653,831 compounds in 29.3 min on six workers; 100 training epochs.
Reproducibility gaps
The reference fingerprint generator and its toolchain are not shipped with the package, and the training sets are not redistributed (Enamine terms apply; ChEMBL-derived data under CC BY-SA 3.0).
Limitations (5)
  • The model predicts the OR-ed ensemble fingerprint, not per-conformer geometry; applications needing a specific conformer must use the reference calculation. · stated by the authors
  • Performance is lowest outside screening-collection chemistry, especially on diverse ChEMBL compounds, and declines above 700 Da where the training set is almost entirely peptidic. · stated by the authors
  • Screening-collection test compounds are not fully independent of training: agreement is 0.857 for the 0.35 to 0.50 similarity band vs 0.922 for 0.70 to 1.00. Predicted similarities carry a small positive bias (median signed error +0.002). · stated by the authors
  • The conventional full-collection cost (3,697 core-hours) is extrapolated, and all timings come from one machine; the model is compared only against its own reference pipeline, not against external methods. · noted in this summary
  • The software is developed by the first author's company, and the evaluation and the design-campaign application are run by the developers, with no prospective experimental validation. · noted in this summary

Read this if you screen or design molecules by 3D pharmacophore similarity and want a conformer-free surrogate, with its applicability domain stated by mass and training-set composition.

Citations
1

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISVenkatraman '22Axen '17Huber '26Imrie '21Muskal '26Arvidsson '24
62% semantic matchPharmacology & Toxicology
V. Venkatraman, A. Roy, D. Demekas+3 more·bioRxiv·Sep 20, 2022