Introduction to Virtual Screening

Note: The following is a summary about virtual screening from Leach and Gillet [2007], chapter 8.


Virtual screening is the computational analogue of biological screening.  Its aim is to filter a set of molecules by scoring and ranking these structures using computational procedures to help researchers take decisions on the tasks being carried out, such as deciding which compounds to purchase from external sources.

There are various ways to assess structures from a virtual screening experiment.  One method to score, filter, or assess structures is by using a previously derived mathematical model that predicts the biological activity of each structure.  Another method is to run substructure queries to filter out molecules with undesirable functionality.  In scenarios where the target structure is known, one can use docking algorithms to find molecules that are predicted to bind strongly to the active site of a protein.  Virtual screening is applied to very large data sets and it is necessary that the process has high throughput.  Multiple techniques can be used to filter out structures in a succession of screening methods [Charifson and Walters 2000].

Wilton et al. [2003] suggested four main classes of virtual screening methods; three classes under Ligand-based methods (machine learning techniques, pharmacophore-based design, and similarity searching) and one class under Structure-Based methods (protein-ligand docking).  All these methods depend on the amount of structural and bioactivity data available.  Additionally, methods for non-specific targets have been developed.  An example of such technique is the prediction of the likelihood that a molecule has “drug-like” characteristics and possesses desired physiochemical properties.

On the turn of the twenty-first century, separate researchers published studies about the concept of “drug-likeness” [Clark and Pickett 2000; Walters and Murcko 2002].  This concepts overcomes the limitations of HTS (High Throughput Screening) and combinatorial chemistry by attempting to select the features of drug molecules that confer biological activity.  One simple method used to assess “drug-likeness” is by using substructure filters to eliminate molecules known to have problems [Roche et al. 2002].  Others tried to find “drug-likeness” by analysing the values of simple properties like molecular weight, partition coefficient, and the number of rotatable bonds.  Based on this kind of analysis, Lipinski et al. [1997] formulate the “rule of five” describing the molecular properties important for a drug’s pharmacokinetics ( in the human body which is more likely when:

  • Molecular weight > 500
  • logP > 5
  • >  5 H-bond donors (sum of OH and NH groups)
  • > 10 H-bond acceptors (sum of N and O atoms)

Other observations conclude that 70% of “drug-like” molecules have:

  • 0 ≤ H-bond donors ≤ 2
  • 2 ≤ H-bond acceptors ≤ 9
  • 2 ≤ rotatable bonds ≤ 8
  • 1 ≤ rings ≤ 4

Other machine learning techniques such as neural networks, genetic algorithms, and decision trees consider more complex possibilities.  The need to go further in defining what makes a good lead has been recognized with the concept of lead-likeness.  It implies cut-off values in the physico-chemical profile of chemical libraries such that they have reduced complexity (e.g. MW below <400) and other more restricted properties [Hann and Oprea 2004].

As the interest in using detailed structural compound knowledge increased, structure-based virtual screening methods have been developed and improved.  Protein-Ligand Docking aims to predict 3D structures when a molecule “docks” to a protein.  There are essentially two components to the docking problem:

  1. Need a way to explore the space of possible protein-ligand geometries (poses)
  2. Need to score or rank the poses to identify most likely binding mode and assign a priority to the molecules.

The difficulty with protein-ligand docking is that it involves many degrees of freedom (rotation and conformation) and solvent effects.  Modern methods explore the orientational and conformational degrees of freedom at the same time.  Most of these protein-ligand docking methods fall into one of the following three categories:

  • Monte Carlo algorithms (change conformation of the ligand or subject the molecule to a translation or rotation within the binding site
  • Genetic and evolutionary algorithms
  • Incremental construction approaches

It is often useful to distinguish between docking and scoring in structure-based virtual screening experiments.  Docking involves the prediction of the binding mode of individual molecules.  Its goal is to identify the orientation closest in geometry to the observed x-ray structure.  On the other hand, scoring ranks the ligands using some function related to the free energy of association of the two units.  Although there are some good scoring functions such as the DOCK function [Desjarlais et al. 1988] and the piecewise linear potential [Gelhaar et al. 1995], there is as yet no universally applicable scoring function able to reproduce experimental binding similarities within acceptable error.

After in silico screening, it is not sufficient for a molecule to bind tightly to its target in an in vitro test, but it must also reach its site of action in vivo.  Thus the requirements for a drug that passes both virtual screening and lab tests can be summarized as:

  • Must bind tightly to the biological target in vivo
  • Must pass through one or more physiological barriers (cell membrane or blood-brain barrier)
  • Must remain long enough to take effect
  • Must be removed from the body by metabolism, excretion, or other means
  • Must not be toxic

Such aspects of drug discovery are often referred to as ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity).  Several ADMET models are dependent upon the hydrogen bonding capacity of a molecule, ranging from simply counting the numbers of donors and acceptors to sophisticated calculations of the overall propensity to be an acceptor or donor to more complex descriptors  derived from linear regression models of properties such as solubility, octanol/water partition coefficient, and blood-brain barrier permeability.

The polar surface area is another molecular property which has proved its importance in ADMET models, especially for the prediction of oral absorption and brain penetration.  It is defined as the amount of molecular surface arising from polar atoms (N and O plus attached H). Clark [1999] observed that polar surface area greater than 140Å2 has poor absorption.

Cruciani et al. [2000] developed the Volsurf descriptors based on 3D molecular fields.  This family of descriptors have particular utility in the prediction of pharmacokinetic properties. These molecular descriptors quantify the molecule’s overall size and shape, and the balance between hydrophilicity, hydrophobicity, and hydrogen bonding.

Toxicity prediction has proved to be a very difficult problem.  The majority of the studies limit to a single toxicological phenomenon or a single class of compound, example polycyclic aromatic hydrocarbons.  Nevertheless, several attempts have been made to predict more general toxicity. The Deductive Estimation of Risk and Existing Knowledge system, DEREK, is one of the widely used expert-based rule method.  However, the use of DEREK and other similar methods still require care and the involvement of human experts.


Virtual screening is a computational technique of scoring, ranking, and filtering molecules that are likely to bind to a target.  Virtual screening enables chemists to search through portions of the enormous 1060 and more conceivable compounds in the chemical space [Bohacek et al. 1996]  to select a much small and manageable set of compounds.  Thus, virtual screening methods have become central to many cheminformatics problems in design, selection, and analysis stages of the drug discovery process.


Bohacek R S, McMartin C and Guida W C (1996), The art and practice of structure-based drug design. A molecular modeling perspective 16:3-50.

Charifson P S and W P Walters (2000). Filtering Databases and Chemical Libraries. Molecular

Diversity 5:185–197.

Clark D E (1999). Rapid Calculation of Polar Surface Area and Its Application to the Prediction

of Transport Phenomena. 1. Prediction of Intestinal Absorption. Journal of Pharmaceutical

Sciences 88:807–814.

Clark D E and S D Pickett (2000). Computational Methods for the Prediction of “Druglikeness”.

Drug Discovery Today 5:49–58.

Cruciani G J, P Crivori, P-A Carrupt and B Testa (2000). Molecular Fields in Quantitative

Structure–Permeation Relationships: The VolSurf Approach. Journal of Molecular Structure:

Theochem 503:17–30.

Hann M, Oprea I, Pursuing the leadlikeness concept in pharmaceutical research. Current Opinion in Chemical Biology 8:255-263.

Leach A, Gillet V (2007). An Introduction to Chemoinformatics. 159-181.

Lipinski C A, F Lombardo, B W Dominy and P J Feeney (1997). Experimental and

Computational Approaches to Estimate Solubility and Permeability in Drug Discovery and

Development Settings. Advanced Drug Delivery Reviews 23:3–25.

Roche O, P Schneider, J Zuegge, W Guba, M Kansy, A Alanine, K Bleicher, F Danel,

E-M Gutknecht, M Rogers-Evans, W Neidhart, H Stalder, M Dillon, E Sj¨ogren, N Fotouhi,

P Gillespie, R Goodnow, W Harris, P Jones, M Taniguchi, S Tsujii, W von der Saal, G

Zimmermann and G Schneider (2002). Development of a Virtual Screening Method for

Identification of “Frequent Hitters” in Compound Libraries. Journal of Medicinal Chemistry


WaltersWP and M A Murcko (2002). Prediction of “Drug-Likeness”. Advanced Drug Delivery

Reviews 54:255–271.

Wilton D, P Willett, K Lawson and G Mullier (2003). Comparison of Ranking Methods for

Virtual Screening in Lead-Discovery Programmes. Journal of Chemical Information and

Computer Sciences 43:469–474.