USRCAT: real-time ultrafast shape recognition with pharmacophoric constraints

Adrian Schreyer, Tom L. BlundellJournal of Cheminformatics 2012 · DOI

TL;DR

This paper introduces USRCAT, an enhanced molecular shape recognition algorithm that incorporates pharmacophoric information, overcoming a key limitation of its predecessor, USR. USRCAT significantly improves the accuracy of virtual screening for drug discovery, especially for diverse chemical libraries, while maintaining high computational efficiency.

What Problem Does This Paper Address?

Finding new drug candidates often involves screening vast libraries of chemical compounds to identify molecules that are likely to bind to a specific biological target. Ligand-based virtual screening, which relies on the shape of known active molecules, is a powerful approach because it doesn't require the 3D structure of the target protein. However, existing shape-based methods like Ultrafast Shape Recognition (USR) have a critical flaw: they are "agnostic" to atom types. This means they can identify molecules with similar overall shapes but vastly different chemical properties and functionalities, leading to many false positives. This is particularly problematic when screening large, heterogeneous chemical databases where molecules might share a shape but not the necessary chemical features for binding.

Methods

The core innovation is USRCAT (Ultrafast Shape Recognition with CREDO Atom Types). It builds upon the Ultrafast Shape Recognition (USR) algorithm, which represents molecular shape using statistical moments derived from inter-atomic distances. USRCAT enhances USR by:

  • Incorporating Pharmacophoric Information: It defines specific subsets of atoms (e.g., hydrophobic, aromatic, hydrogen bond donors/acceptors) based on SMARTS patterns.
  • Calculating Subset Moments: It calculates additional statistical moments for these atom subsets.
  • Leveraging Global Reference Points: Crucially, unlike some earlier extensions, USRCAT uses the four reference points derived from all atoms in the molecule to calculate the distributions for subset moments. This ensures that the pharmacophoric information is encoded relative to the overall molecular shape, improving discriminatory power.
  • Compact Representation: The resulting USRCAT vector is 60-dimensional (5x12), with the first 12 dimensions being identical to the original USR moments. This compact representation is key to its speed.
  • Efficient Database Implementation: The paper details a robust implementation using PostgreSQL, leveraging its cube data type and GIST indexing for fast querying of millions of conformers.

Datasets

  • Protein Data Bank (PDB): Used as a source for chemical components.
  • Directory of Useful Decoys, Enhanced (DUD-E): A benchmark dataset containing diverse protein targets and associated active ligands and decoys. This dataset was specifically chosen to test the performance of shape-based methods against a challenging, heterogeneous set.
  • ChEMBL: Used for retrieving active ligands for the DUD-E targets.
  • ZINC: Used for sourcing decoys for the DUD-E benchmark.

Key Insights and Ideas

  • Shape is Necessary, but Not Sufficient: The study strongly emphasizes that while molecular shape is a critical factor for molecular recognition, it's not the only one. Incorporating atom-type information (pharmacophores) is essential for accurate virtual screening, especially in diverse chemical spaces.
  • The DUD-E Benchmark's Nuance: The DUD-E dataset, while valuable, presents a unique challenge for shape-based methods. Its design, which prioritizes topological dissimilarity between actives and decoys, inherently favors topological similarity methods over pure shape-based ones. This explains why topological methods performed exceptionally well on this specific benchmark.
  • Scaffold Hopping is Achievable: USRCAT demonstrates a significant ability to retrieve active compounds that do not share topological similarity with the query molecule. This "scaffold hopping" capability is highly desirable in drug discovery, as it can lead to novel chemical structures and circumvent existing patents.
  • Probe Radius Optimization is Key: The performance of USRCAT is sensitive to the "probe radius" parameter, which defines the extent of the bounding box used for similarity calculations. Finding the right balance is crucial for optimizing both speed and accuracy, especially when dealing with databases of varying molecular sizes.

Key Results

  • Improved Enrichment Factors: USRCAT consistently outperformed the traditional USR method in retrospective virtual screening benchmarks, achieving nearly double the enrichment factor at the crucial 0.25% level on the DUD-E dataset.
  • Competitive with ElectroShape: USRCAT performed competitively with ElectroShape, another pharmacophore-aware method, and in some cases, outperformed it.
  • High-Speed Screening: The PostgreSQL implementation, utilizing GIST indexing, can screen millions of conformers in milliseconds, making it suitable for web services and large-scale screening. For example, a linear scan of 1.4 million circular fingerprints took less than 900 ms.
  • Scaffold Hopping Success: Table 2 shows that USRCAT retrieved a substantial number of active hits that were missed by topological similarity methods across various targets. For instance, for the AA2AR target, USRCAT retrieved an average of 5.92 hits not found by topological similarity.

Limitations and Future Work

  • DUD-E Benchmark Limitations: The authors acknowledge that the DUD-E dataset is not ideal for benchmarking pure shape-based algorithms due to its design favoring topological similarity. This might have understated USRCAT's performance in certain scenarios.
  • Query Molecule Size: The performance of USRCAT (and USR) can be affected by the size of the query molecule relative to the molecules in the screening library. Future work could involve better handling of size variations.
  • Parameter Optimization: While the paper discusses the impact of the probe radius, further optimization of pharmacophore weights and probe radius for specific targets and libraries could enhance performance.
  • Generalizability: The authors note that the benchmark results are specific to the DUD-E dataset and the chosen conformer generation tool (Omega TK). The generalizability of these findings to other datasets and tools might need further investigation.
  • Future Directions: The authors suggest that USRCAT's modular design allows for easy adaptation to specific needs, such as defining different atom types for subsets or integrating with other virtual screening methods. The source code is freely available, encouraging further development and application.

Summary updated

Abstract

USRCAT provides a solution to the lack of atom type information in the USR algorithm. Researchers, particularly those with only limited resources, who wish to use ligand-based virtual screening in order to discover new hits, will benefit the most. Online chemical databases that offer a shape-based similarity method might also find advantage in using USRCAT due to its accuracy and performance. The source code is freely available and can easily be modified to fit specific needs.
BibTeX
@article{schreyer2012usrcat,
  author = {Adrian Schreyer and Tom L. Blundell},
  title = {{USRCAT: real-time ultrafast shape recognition with pharmacophoric constraints}},
  journal = {Journal of Cheminformatics},
  year = {2012},
  doi = {10.1186/1758-2946-4-27},
}
Figure