Recent years have seen the development of in silico 'virtual screening' of very large libraries of molecules as an integral part of the drug development process. An initial library might contain millions of compounds that are potentially available, either in-house or from vendors' catalogues. Virtual screening has the twin goals of finding molecules with both favourable ADMET properties and suitable bioactivity. The first goal involves searching for molecules with favourable values of relevant properties such as solubility, polarity, logP, possible toxicity, absorption and likely routes of metabolic breakdown, hence guiding the medicinal chemist towards molecules of good bioavailability and low toxicity. While this is a very important aspect of virtual screening, hereafter we shall concentrate on the second goal of finding molecules with suitable bioactivity.

In the favourable case where the three dimensional structure of the target, usually a protein, is known, it is possible to computationally dock thousands of molecules into the active site, looking for those that will have suitable spatial and chemical complementarity and hence bind strongly 1. The simplest case is rigid body docking, where we take given fixed conformations of the protein and ligand and find where in the protein the ligand will bind, and how it will be oriented, in order to obtain the (spatially and chemically) best fit. Even with the assumption of rigid bodies, the search space is six dimensional.

However, in reality the problem is harder than this. Both protein and ligand are liable to undergo conformational change upon docking. This means that the method must allow conformational variation, ideally in both molecules. The search space acquires a high dimensionality and the flexible docking problem is difficult. In the present work, we carry out semi-flexible docking.

One of the most successful strategies for docking is to use a genetic algorithm, as in the program GOLD 2. Such an algorithm mirrors Darwinian evolution, representing the solution as a 'chromosome'. Genetic algorithms allow a population of solutions to exist, and in each 'generation' these can evolve by processes such 'breeding' and 'mutation'. Poor solutions are killed off, while good ones leave their offspring in future generations. Such algorithms may typically reach an excellent solution in a few tens of generations.

Scoring functions, either identical to or different from those utilised as measures of fitness within docking programs, are used to assign predicted binding affinities and rank ligands relative to one another, with a view to selecting and testing experimentally a small subset for biological activity. The development of suitable scoring functions for ranking the solutions produced by docking methods, and especially for accurate prediction of protein-ligand binding affinities, remains a considerable challenge. The scoring function must accurately measure both intramolecular conformational strain energy and intermolecular interaction energy. Several contrasting kinds of scoring function have had some success, including some based on molecular mechanics force fields (Coulomb + van der Waals + hydrogen bonding + bond stretching & bending + torsions) 3 and others centred on modelling each of the relevant terms of a 'Master Equation' describing the free energy of interaction. An alternative is provided by knowledge-based scoring functions, such as BLEEP 4 and PMF 5, where the objective is to use the co-ordinates of hundreds of three dimensional protein-ligand complex structures as a knowledge base. Using this knowledge, a putative protein-ligand interaction geometry can be assessed on the basis of how similar its features are to those of the ensemble of known structures. The features used are the distributions of atom-atom distances between protein and ligand in the complex. Commonly observed features, such as donor/acceptor type nitrogen/oxygen distances at typical hydrogen bonding distances around 3Å, score favourably. Less frequently observed interactions, such as close polar/non-polar contacts, score unfavourably. When the contributions are summed over all pairs of atoms in the complex, the resulting score indicates how much the putative structure 'looks like' a real protein-ligand complex.

When the binding affinity of a series of homologous inhibitors into a particular site is known, it is possible to generate 'customised' scoring functions to fit the data 6. Ideally, the combination of the search algorithm and the scoring function should result in a single solution close to the experimental ligand position 7. General-purpose scoring functions, in contrast, are designed to be applicable to a wide variety of protein-ligand complexes, and are therefore parameterised using a diverse set of protein ligand complexes. This work concentrates on five general-purpose scoring functions.

The application of virtual screening techniques in parallel with High-Throughput Screening (HTS) technology, coupled with structural biology 8, can extend the scope of screening to external databases. This allows more diverse chemical entities to be identified as hits, and as a consequence can help to reduce the assay-to-lead attrition rate observed from HTS 9.

There are many questions, however, associated with the tools employed for docking-based virtual screening. A number of approximations are often employed for the docking/scoring searches (e.g., neglect of protein flexibility in rigid docking, lack of a rigorous treatment of solvation, and the choice of one particular protonation state) in order for the virtual screen to be completed within an acceptable time limit, as well as other unavoidable approximations such as the limitations of X-ray crystal structures. Despite the above, virtual screening can be improved by taking into consideration additional information about the receptor of interest and using this information advantageously in docking/scoring applications 10111213. Recent advances in virtual screening include various physics-based methods 141516 and consensus scoring 1718.

In our study, we concentrate on heat shock protein 90 (Hsp90), which is a chaperone and a target for anti-cancer therapeutics 19. Prior to screening, the binding site was prepared by using the SYBYL^® 7.0 software of Tripos 20. The docking program GOLD 2.2 2 was used to perform docking with and without the presence of a tether. The ligands docked were taken from 'active' and 'inactive' datasets 21; we also used a set of 'decoys' retrieved from the CIPSLINE cancer database 22. Post-dock scoring was calculated using multiple scoring functions: GOLD 2, ChemScore 23, DOCK 24, PMF 5, BLEEP 425262728, and a Consensus generated from the preceding five. We used two ranking methodologies: best GOLD_rank and BestScore_rank (see Methods).

We analysed crystal structures (PDB Codes: 1YC1/1YC3/1YC429, 1BYQ30) containing the ligands 4BC (Figure 1), 43P (Figure 1), and ADP (Figure 2) bound to the N-terminal ATP binding domain of human Hsp90α, as described in detail in the Methods section.

We considered the conformation of the ADP bound Hsp90 (PDB Code: 1BYQ), representing a ligand bound structure, as a suitable starting point for virtual screening. Thr184 was used as a tether, since it generated a low RMSD (root mean square deviation) from its corresponding crystallographic conformations (< 1Å, Table 1) and, in preliminary work, strongly outperformed the alternative Asp93 tether. Each library compound was docked to the binding site (PDB Code: 1BYQ), firstly with no tether and secondly with the tether Thr184. We separately used the GOLD and ChemScore functions for on-dock scoring. The RMSD values between the docked ligands and their crystallographic conformations (1YC1/1YC3/1YC4) are given in Table 1.

A larger set of 'active' (261) and 'inactive' (54) compounds similar to 4BC and 43P and a 'decoy' (~3600) set of known drugs, with similarly druglike physicochemical properties (molecular weight, logP, numbers of hydrogen bond donors, hydrogen bond acceptors and rotatable bonds, see Methods), but assumed inactive against this target, were docked to the binding site, with and without the Thr184 hydrogen bond tether, using GOLD 2.2 and the GOLD scoring function. Post-dock scoring used the SYBYL^®-CScore™ module, BLEEP as stand alone software, and Consensus (see Methods).

By ranking all ligands via their score values, the enrichments were calculated for each scoring function to establish how many decoys had to be picked in order to find all the original actives, based on poses chosen with either the GOLD_rank or BestScore_rank methodologies. GOLD_rank is based on ranks generated using GOLD. Each scoring function is applied to the pose ranked number one by GOLD for that ligand. BestScore_rank chooses, in each case, the best pose as ranked by the particular scoring function in question, rather than always using the pose ranked first by GOLD. The virtual screening was carried out both with and without the tether.

Scores normalised for molecular weight 31 were obtained by dividing the raw score by the number of heavy atoms to the power of 1/3. This is designed to reduce the inherent bias towards larger molecules that arises from the additive nature of scoring functions. Normalisation also reduces the prevalence of high molecular weight molecules amongst the hits, which is likely to be beneficial from a lead optimisation perspective.

We give the results in Table 2 (no tether) and Table 3 (Thr184 tether). Some of the same data are shown as Receiver Operating Characteristic (ROC) curves in Figure 3. The ROC curves are presented as plots of the proportion of all actives recovered versus the proportion of all inactives recovered as one proceeds from the top to the bottom of the ranked list. The areas under these ROC curves are a convenient measure of performance, and are included in Tables 2 and 3. An ideal case would recover all actives before recovering any inactives and hence have an area of unity. The apparent contradiction between the retrievals of actives for GOLD between Tables 2, 3 (e.g., 28% of actives for GOLD_rank and 25% of actives for BestScore_rank in the top 10% in Table 2) is due to the different implementations of the GOLD algorithm in GOLD 2.2 and in the SYBYL^® -CScore^™ module.

Some clear trends are apparent from these results, based on analysis of the data in Tables 2 and 3, and especially the areas under the relevant ROC curves.

(1) The relative performance of the scoring functions is typically given by

GOLD ≈ Consensus > DOCK ≈ ChemScore > BLEEP >> PMF.

The Consensus score used in this work is a simple sum of the Z-scaled scores from five scoring functions. As such, it is less sophisticated than other consensus strategies considered elsewhere 1727. Nonetheless, it is generally a robust method, comparable in performance to the best individual scoring function. The good performance of the Consensus scoring method result is to some extent in agreement with recent virtual screening studies where Consensus scoring improves the enrichment of true hits 323334 in various systems. However, the improvement given by using the Consensus method is small, and on occasions Consensus fails to outperform the best individual function.

The performance of PMF here is usually worse than random and PMF is consistently the poorest performer in all applied protocols. We used the implementation of PMF in SYBYL^® 7.0; our previous use of the SYBYL implementation of PMF also gave disappointing results 27, though the present ones are certainly poorer. PMF gave much better results in its authors' own in-house implementation 535.

(2) The tethered results are in all cases better than the corresponding untethered ones. This effect is particularly strong when the BestScore_rank protocol is used.

The utilisation of tethering during docking requires prior knowledge of ligand-protein X-ray structures, which is not always available. Inspection of the structures shows that in general the tether is satisfied as expected in the better scoring structures and structures unable to satisfy the tether appear further down the ranked list.

(3) When the tether is used, the BestScore_rank protocol always gives better results than a corresponding calculation using the GOLD_rank protocol. For untethered docking, there is little difference in the performance of the two protocols.

The GOLD_rank protocol tends to be biased towards to the GOLD function in relation to the other scoring functions; BestScore_rank proved to be an unbiased method selecting the best score for each scoring function independently.

(4) In most cases, normalisation has little effect on performance, and any such effect is often deleterious. However, for the particular combination of tethered docking and the BestScore_rank protocol, normalisation gives a significant improvement (though not for ChemScore).

(5) This combination of tethered docking, the BestScore_rank protocol and normalisation by dividing the raw score by the number of heavy atoms to the power of 1/3 gives the best results found in this study; this is true for every scoring function except ChemScore.

We consider that this optimal combination gives a good virtual screening performance (other than with PMF), with the percentages of actives found in the first 10% of the ranked library being 90%, 79%, 69% and 52% for the four best individual scoring functions and 89% for Consensus. The ROC curves for this combination are shown in Figure 3.

(6) The performance ranking of the scoring functions for this optimal combination of tether, BestScore_rank and normalisation is somewhat atypical of those found in our other calculations and is given by

DOCK ≈ Consensus > GOLD > BLEEP > ChemScore >> PMF.

This work has demonstrated the successful development of a virtual screening methodology, as has been achieved by other groups for different therapeutically relevant targets 3637. A library of ~3600 compounds was docked semi-flexibly into the active site of Hsp90. Five scoring functions, including BLEEP, were used to discriminate active from inactive compounds. The present work offers alternative protocols for virtual screening of chemical libraries with an emphasis on the effect of using multiple ligand poses for scoring with some of the most common scoring functions and also tethered and un-tethered docking.

For tethered docking, we find that consideration of multiple poses for each ligand in our BestScore_rank protocol is superior to relying on the best scoring pose generated by a single scoring function. The different scoring functions are thus judged on the basis of their own top-scoring poses, which may be different from one another.

Though normalisation has little effect on enrichment elsewhere in this work, in the case where the Thr184 tether is combined with the BestScore_rank protocol, normalisation generates a significant improvement in enrichment. This combination of tethered docking, the BestScore_rank protocol and normalisation gives the best results found in this work. Normalisation also reduces the prevalence of high molecular weight molecules amongst the hits, which is likely to be beneficial from a lead optimisation perspective.

Although we use only a very simple implementation of Consensus scoring, we find it to be a generally robust methodology. It performs similarly to the best individual scoring function in each virtual screening run.

Overall, we demonstrate the validity of virtual screening as a method for identifying new leads from a pool of ligands with similar physicochemical properties and we believe that the outcome of this study provides useful insight into the setting up of a suitable docking/scoring protocol, resulting in enrichment of 'target active' compounds.

CKK carried out the work under the supervision of JAL (industrial supervisor) and JBOM (academic supervisor). All authors cooperated in the analysis of the results and the writing of the manuscript.