COSMOsim3D and COSMOsar3D


Based on the fact that ligands similar to an active ligand show an increased probability to be also active, ligand-based drug design does not require a target structure. The art of ligand-based drug design thus consists of finding adequate representations of the intermolecular interaction patterns of ligands defining a meaningful similarity measure. Typically, the similarity between ligand molecules is assessed in terms of two- or three-dimensional structure, shape, polar interactions, and pharmacophoric models.

COSMOsim3D is a unique and very robust method for automatic and unsupervised field-based ligand-ligand alignment and similarity searches. It utilizes local σ-profiles, which have been proved to contain the required information on the relevant molecular interactions.

The molecular surface charge approach of COSMOsim3D naturally enables scaffold hopping. Thus, the COSMOsim3D method is suited for datasets with different chemotypes, allowing users to find alternative chemical scaffolds with similar shape and polar features.

See also our paper: "COSMOsim3D: 3D-Similarity and Alignment Based on COSMO Polarization Charge Densities" or examples on Re-Alignment of Sutherland data sets and Validation study on bioisosteric pairs

Local σ-profiles

Justified by the general COSMO-RS finding that intermolecular interactions can be excellently quantified based on the COSMO surface polarization densities σ, COSMOsim3D introduces local σ-profiles (LSPs) for the alignment and similarity measure of ligands. LSPs are σ-histograms, resulting from projecting the molecular σ-surface onto a 3D-grid, and they provide information about

  • electrostatics
  • hydrogen bonding
  • hydrophobic interactions
  • shape

Using LSPs, COSMOsim3D meanwhile has been proven to be a very accurate and extremely robust method for automatic and unsupervised field-based ligand-ligand alignment and similarity searching.


Superposition of eight ligands and their COSMO cavities as obtained from a COSMOsim3D alignment. Molecules and cavities rendered with PyMOL.
Probe molecules are rotated and translated randomly in the grid. For each position, a 3D similarity is calculated. The position with the maximum 3D similarity between the template and probe molecules is used for the alignment.

Key features:

  • Pairwise alignment and similarity assessment: COSMOsim3D utilizes LSPs instead of chemical structure or pharmacophores to assess a molecule’s similarity to a template molecule. This enables scaffold hopping and allows users to deal with datasets of different chemotypes.
  • Multi-template alignment: It is possible to align multiple template molecules and use their superposition as a virtual template molecule. The virtual template molecule can be used for alignment and similarity assessment of potential ligand molecules.
  • Ligand-Based virtual screening: COSMOsim3D can be used to rank a set of potential ligand molecules according to their similarity to a single template or a virtual template stemming from the superposition of multiple ligands. Thus, it allows users to enrich ligand sets with potential cognate drugs, generating fewer false positives and more true hits.


As an alternative approach to similarity-based ligand assessment, QSAR models are used to correlate computed properties of molecules and their biological activity. COSMOsar3D is an extension of COSMOsim3D which uses the LSPs of aligned ligands as a novel set of molecular interaction fields for 3D-QSAR.

In a recent study, LSP-based COSMOsar3D models instantly had a significantly higher predictivity than standard 3D‑QSAR models. Utilizing smooth spatial histograms introduces an increased robustness of the models against small geometrical shifts of the ligands relative to the grid even at larger grid spacing. In a histogram, a local property hotspot is smoothly partitioned over the neighboring grid points, resulting in a position-independent representation.
See also our paper: "COSMOsar3D: Molecular Field Analysis Based on Local COSMO σ-Profiles" or an example on  Predictive performance of several 3D-QSAR methods

Experimental pKi vs. predicted pKi for the AChE data set from a PLS analysis using LSP descriptors

Key Features

  • Robustness and Predictivity: Instead of using properties directly as input for MFA, COSMOsar3D uses a histogram of a property. This approach leads to very robust performance of the models with respect to grid size, grid position and small misalignments, while retaining increased predictive accuracy.
  • Ionic and Neutral Molecules: The polarization charge densities of neutral and charged species are in the same range, which allows for the inclusion of compounds of varying charge states in the same model.
  • Linear Relationship of logKi and LSPs: LSPs constitute an optimally suited set of descriptors for a linear regression analysis of pKi values, according to the 3D-QSAR paradigm. COSMOsar3D provides a rationale that logKi values should be linear functions of the LSP descriptors. To the best of our knowledge, no other set of molecular fields used so far in MFA can claim such a sound theoretical justification for the expectation of a linear pKi model.
  • Description of Hydrogen Bonding: As shown in a recent paper, the polarization charge densities σ are better suited for the description of hydrogen bonding than the electrostatic potential, which is usually employed in MFA. See also “Polarization charge densities provide a predictive quantification of hydrogen bond energies.”