Sage Journals: Discover world-class research

Abstract

A first step in fragment-based drug discovery (FBDD) often entails a fragment-based screen (FBS) to identify fragment “hits.” However, the integration of conflicting results from orthogonal screens remains a challenge. Here we present a meta-analysis of 35 fragment-based campaigns at Novartis, which employed a generic 1400-fragment library against diverse target families using various biophysical and biochemical techniques. By statistically interrogating the multidimensional FBS data, we sought to investigate three questions: (1) What makes a fragment amenable for FBS? (2) How do hits from different fragment screening technologies and target classes compare with each other? (3) What is the best way to pair FBS assay technologies? In doing so, we identified substructures that were privileged for specific target classes, as well as fragments that were privileged for authentic activity against many targets. We also revealed some of the discrepancies between technologies. Finally, we uncovered a simple rule of thumb in screening strategy: when choosing two technologies for a campaign, pairing a biochemical and biophysical screen tends to yield the greatest coverage of authentic hits.

Keywords

fragment fragment-based screen privileged substructure naive Bayesian model

Introduction

Fragment-based drug discovery (FBDD) is a proven method in the arsenal of modern drug discovery that aims to identify therapeutically useful chemical matter from a nearly infinite chemical space. At the heart of the method is the identification of simple chemical building blocks, or fragments, that have an affinity for a target of interest. These simple fragments or features of these fragments are then extended or merged with other fragments or larger molecules, with the aim of accessing compounds with improved affinity. A first step in FBDD often entails a fragment-based screen (FBS) against a target of interest. In many ways, high-throughput screening (HTS) is a progenitor of FBS, and although many lessons learned from HTS can be applied to FBS, there are features of FBS that set it apart and require fresh scrutiny.^1–3

As is the case with HTS, particular screening technologies are more amenable to different targets and are also prone to different false positives or negatives. Since this is the case, orthogonal assays are often employed to corroborate putative hits. Although two different technologies are often paired in workflows, either in parallel or in tandem to avoid the biases in any one technology, the relationships such as complementarity, overlap, and orthogonality of assays are not well understood. As a result, when choosing technologies for an FBS campaign, the strategies employed by teams are often a patchwork of available technologies or work flows, as opposed to a best standard of practice.

Methods of utmost relevance to the current study can be divided into biophysical and biochemical assays. For the purpose of this study, we define biochemical assays as monitoring the inhibition or activation of a biochemical event, such as inhibition of an enzyme, and biophysical assays as measuring the binding of a ligand to a protein. Some technologies can be used for both purposes (such as nuclear magnetic resonance [NMR]), but for this study, technologies were used for only one of these types of assays. In the current study, the techniques used for biophysical assays include NMR,⁴ surface plasmon resonance (SPR),⁵ and protein thermal shift.⁶ Technologies employed for biochemical screens, also termed high-concentration screens (HCS),^7–9 include fluorescence,^9,10 capillary electrophoresis (CE),^10,11 and time-resolved fluorescence energy transfer (TR-FRET), to name a few. The technologies used in the current study are summarized in Supplemental Table S1.

Here we describe an assessment of Novartis’ latest fragment library, which contains ~1400 fragments. It was intended as a generic library, suitable for multiple biophysical and biochemical technologies. Over the past 3 years, it has been screened against dozens of targets, using diverse technologies (Suppl. Table S1), by various teams of Novartis scientists located at several research sites, affording a rich set of data from which to extract FBS lessons. While the design of FBS libraries has received much attention in literature,^12–15 a true meta-analysis of their performance and assessment of their suitability for multiple technologies has yet to be reported. The term meta-analysis was coined in 1976 by Gene Glass to refer to “the statistical analysis of a large collection of analysis results from individual studies for the purpose of integrating the findings.”16 In our case, the individual studies correspond to fragment-based campaigns carried out by individual project teams, against various biological targets. Through the integration of this body of work, we aim to investigate three questions: (1) What makes a fragment amenable for FBS? (2) How do hits from different FBS technologies and target classes compare with each other? (3) What is the best way to pair FBS assay technologies?

Methods

Library Design, Data Collection, and Hit Selection

Data collection and hit selection are described in the supplemental material.

Naive Bayesian Modeling

Naive Bayesian (NB) models were built using Pipeline Pilot 8.0.¹⁷ For model building, we only used substructures composed of one or two subunits (as illustrated in Suppl. Fig. S1). While extended connectivity fingerprints (e.g., ECFP4)¹⁸ could have been used to generate accurate models, we chose to use substructures that made chemically intuitive sense, lending themselves more readily to meaningful interpretation (e.g., aromatic rings and functional groups would not be broken). NB models were trained using either the superset hit selection or technology-specific hit selection to define categories of hits and nonhits. The statistical details of NB modeling have been described elsewhere^19,20 and are included in the supplemental material.

Relating Targets Using Naive Bayes Models

We used the NB models trained on the superset hit selection to correlate targets by the similarity of their active fragments. For this purpose, all informative substructure weights (NP_i) were extracted from 16 superset NB models. Perhaps not surprisingly, most substructures were uninformative across all models. The median of the highest weight for each substructure was then calculated (0.154). Because we did not want activity-irrelevant substructures to influence similarity calculations between models, we used this median value to filter substructures in subsequent calculations. That is, to assess the similarity of a pair of targets, we calculated Pearson correlation coefficients for weight vectors of the two corresponding models but limited the vectors to those substructures that achieved a weight >0.154 for at least one of the two targets. Pearson correlation coefficients were then converted into distances between 0 and 2, and the resulting target distance matrix was used as input for multidimensional scaling (MDS, using Sammon’s Non-Linear Mapping provided through the MASS library in R) to get coordinates for the targets in a 2D space. To assess the goodness of fit, the stress was computed and a final stress value of 0.078 was achieved.

Difference Vector for Technologies and Complementary Score between Technologies

To compare specific technologies with each other, we compared NB models derived for a single technology with NB models trained on the superset hit selection. Models were trained with substructures to increase sampling. We derived a complementary score, which reflects how well two technologies compensate for the technology-specific biases inherent in each technology. For the examples in the main text, we focused on a workflow where all hits from primary screens are progressed to subsequent screens. In this scenario, it is desirable to pair a technology that might miss a particular substructure (and thus a fragment or fragment class) with a technology that would find it. Our complementarity score reflects the ability of two technologies, when combined, to identify as many substructures as possible and ranges from 0 (lacking complementarity) to 1 (high complementarity). A detailed derivation of this score is described in the supplemental material.

X-Ray Crystallography, Differential Scanning Fluorimetry, and SPR

Experimental details for X-ray crystallography of Pim-1 and SPR and differential scanning fluorimetry (DSF) control experiments can be found in the supplemental material.

Results and Discussion

Privileged Substructures and Fragments

We first assessed how many fragments were deemed hits as more and more targets are screened ( Fig. 1A ). For comparison, we simulated the number of compounds that would be hits, if the probability of hitting a target was equal for all fragments, while employing the same hit rates found experimentally (see Supplemental Material - Experimental and Simulated Hit Assessment of Fragment Library). Interest-ingly, compared with simulation, a lower fraction of the library is found to be a hit against any target (37% vs. 54%). This implies that some compounds in the library hit more often than expected by chance, while others hit less often than expected by chance (note, over the 20 targets investigated, 63% of the library was never identified as a hit). To investigate this further, we plotted a histogram of the number of fragments that hit a given number of targets ( Fig. 1B ). This revealed that many fragments have not hit a single target. Furthermore, while only a few fragments (20, or 1.4%) have hit six or more targets, this is higher than expected by chance since the cumulative binomial probability of hitting six targets out of 20 trials (campaigns) with a hit rate of 3.72% (the average hit rate for the 20 campaigns) is p(x ≥ 6)= 6.5 × 10⁻⁵.

Figure 1.

The percentage of fragments that hit at least one target as the total number of screened targets increases (A). Percentage of fragments with various hit rates (B). In A and B, experimental findings (black) are compared with simulations (gray). An example of a privileged fragment 1 that hit seven 20 targets (C). The crystal structure of the privileged fragment 1 bound to Pim-1 (D) and a rendering of the crystal structure displaying the hydrogen bond network between the carboxylic acid moiety and both Lysine 67 and a bound water molecule (E). Bound water molecules are depicted as red spheres. Crystal coordinates of complex structure have been deposited in the RCSB Protein Data Bank under accession code 4MTA.

We next questioned whether there were any salient relationships between physicochemical properties and hit rates of fragments. For this purpose, we divided active fragments into the following categories according to the number of protein targets that they bound to: (1) one target, (2) two targets, (3) three targets, (4) four or five targets, and (5) more than five targets. Cumulative distribution curves (Suppl. Fig. S6) clearly showed that the number of targets hit by a fragment increased with its logP and number of aromatic rings or bonds, as well as decreased with its solubility and number of rotatable bonds (i.e., more flexible fragments are more specific). For example, fragments hitting only one target had a median logP of 1.45, whereas the median for fragments hitting more than five targets was shifted to 2.47 (i.e., a rather drastic difference of 1 log unit was observed). To a lesser extent, promiscuous fragments (more than four targets) also tended to have a higher molecular weight and more atoms than other active fragments.

Although it is tempting to deem the high hit rate fragments “frequent hitters,” which often carries along with it a negative connotation, it should be pointed out that unlike larger compounds, fragments often demonstrate affinity for multiple targets and gain specificity as they are evolved into more complex compounds. An example of this is 7-azaindole, which was evolved into vemurafenib, which targets mutant BRAF as well as AZD5363, which targets PKB.^21–23 Furthermore, all of the fragments that were screened were previously subjected to quality control assessments (see Supplemental Material - Library Design), suggesting that the mass, purity, and solubility are desirable.

To assess the value of the compounds with hit rates that were higher than expected by chance (≥6 of 20 targets hit), we ascertained how many had been crystallized with targets considered in the meta-analysis. Interestingly, 5 of 20 (25%) of the high hit rate fragments have been crystallized with target proteins included in the meta-analysis, a rate much higher than the other active fragments (1-5 targets hit, 4.6%; Suppl. Table S4). Thus, these fragments are valuable assets in the screening library, and fragments that have been carefully quality controlled with high hit rates should not necessarily be termed frequent hitters, as this often has a negative connotation, but rather privileged fragments.

An illustrative example of this is a privileged fragment 1 that hits 7 of 20 targets ( Fig. 1C ). In an FBS campaign against Pim-1 kinase (previous to those included in the current meta-analysis), this fragment was identified as a hit by an X-ray crystallography screen ( Fig. 1D , E ). The fragment binds to the adenosine triphosphate (ATP) site of Pim-1, and the hydrogen bond network established between the carboxylic acid and both a bound water and Lysine 67 ( Fig. 1E ) has previously been reported for both fragments^24,25 and small molecules,^24–27 supporting the notions that this is an authentic interaction and a valuable starting point for a project team.

We questioned whether there were substructures enriched for specific target classes. Interestingly, when targets were related to each other using NB models built on the superset hit selection, targets belonging to the same protein family tended to cluster together in 2D space ( Fig. 2 ). For example, the serine/threonine kinases cluster is close to the tyrosine kinases. For serine proteases and serine/threonine kinases, activity-relevant substructures that were exclusive to these families (i.e., did not obtain high weights for targets of other families) are displayed.

Figure 2.

Comparison of target classes based on the chemical matter selected as hits. Naive Bayesian (NB) models were built for each target, based on the active and inactive fragments. The significant feature weights were then extracted and used to derive Pearson correlations between each target model. These correlations were then subjected to multidimensional scaling (MDS) to compare the targets in 2D space. Targets of the same class tend to cluster near each other, revealing that they interact with similar chemical matter. Substructures that were enriched in the actives of two specific target classes but were not enriched in the actives of other target classes are depicted.

In summary, there are fragments that bind more targets than expected by chance—privileged fragments—that may offer valuable starting points for project teams in lead discovery. Furthermore, particular substructures that are enriched in fragments are active against specific target classes.

Physical Properties of Hits versus Nonhits

Often in FBS, a library is screened using one technology, and hits from this screen are confirmed using an orthogonal technology. For example, a primary biochemical screen might be employed, followed by confirmation in NMR, and finally crystallization trials.⁷ Alternatively, some groups will use multiple technologies in parallel and integrate results from each technology into a final hit list.²⁸ This begs one to question how different technologies compare with each other.

As a simple first step, we assessed trends in the physicochemical properties of hits identified by different technologies against various targets, a subset of which are presented in Figure 3 . This level of resolution allows exploration of discrepancies in properties of hits that may be due to specific technologies or particular targets. It is apparent that while for most technologies, the hit rate increases for higher molecular weight ( Fig. 3A ) and higher AlogP ( Fig. 3B ) fragments, this is not true for every target/technology pair. Also, we see that different techniques have different trends for a given target. For example, NMR tends to identify larger and more hydrophobic compounds for oxidoreductase 2, while hits obtained from TR-FRET are more evenly distributed for both of these properties (a trend that seems to be reversed for hydrolase 1 for these technologies). Different trends for particular technology/target pairs can be seen for other physicochemical properties as well, such as fraction molecular polar surface area (Suppl. Fig. S8). In a related work reported by Vernalis,²⁹ the molecular weight (MW) distribution for hits was quite similar to that of nonhits, while more variation was seen for the AlogP distribution. Furthermore, these distributions seemed to vary with target family class.²⁹

Figure 3.

Physicochemical properties of fragment “hits.” The percentage of fragments in each bin that were identified as hits are represented by red bars, while the gray background indicates the percentage of nonhits.

Hit Rates of Different Technologies

Another way of comparing technologies is by assessing the fraction of screens that a compound was classified as a hit. In Figure 4 , the number of fragments versus the binned fraction of screens that a compound classified as a hit for each technique is plotted (0-1, increments of 0.2, 0 means that the fragment was never a hit for that technology, while 1 would mean the fragment was a hit in every assay for that technology). It should be noted that this is a subset of techniques used in the meta-analysis in which at least eight campaigns were available for analysis. For this analysis, we used primary hit selection to include as many screens for each technology as possible. From the validated or consensus hit selection (superset) that we previously performed, we see that fragments tend to hit in <50% of the screens that they were tested in (only one fragment hit >50% of targets; Fig. 1B ). In contrast to this, we see that for most technologies, dozens of fragments hit in more than 60% of the screens ( Fig. 4 ). It should be noted that this is not a direct comparison since the targets used for the validated and consensus hit calling do not always perfectly overlap with those used for technology-specific primary hit calling.

Figure 4.

Number of fragments versus the fraction of screens that a compound was classified as a hit in different technologies. The number of fragments is plotted for each binned average. The total number of fragments for each technology is 1400. While the fraction of screens that a compound is classified as a hit tends to be similar for most technologies, it is higher for surface plasmon resonance (SPR) and lower for differential scanning fluorimetry (DSF).

Since the fraction of screens that a compound classified as a hit was well above the rates observed in the superset (when validated and consensus hit selection was employed), we see that different technologies are prone to frequent hitters. With this in mind, one might question whether fragments that consistently hit in a specific technology should be excluded from a generic library of fragments. By tracking some of these fragments, we learned that indeed for some specific cases, the fragments have been confirmed with orthogonal screens—that is, “privileged fragments” may also hit more often in specific technologies. Perhaps more potentially problematic are fragments that are frequent hitters in specific technologies but less often in other technologies. To assess such fragments, we tracked how many crystal structures have been obtained from fragments that are frequent hitters for specific technologies but were not often found in the superset (<3 targets hit). Surprisingly, the rates are very similar for these fragments (2.1%) compared with fragments that are not frequent hitters (1.3%) (Suppl. Table S4), confirming that frequent hitters observed by a single technology can be real hits and thus serve as good starting points for optimization.

Furthermore, from Figure 4 , we see that for most technologies, the number of fragments with various average hit rates is similar, with SPR and DSF being the exceptions. SPR tends to have more fragments with high hit rates (>60%) than the other technologies assessed, while DSF has no fragments with hit rates above 40%. We investigate SPR and DSF in more depth in the supplemental material. Through control experiments, we were able to identify fragments that tended to nonspecifically accumulate in the matrix of biacore surfaces, explaining in part the high hit rates observed for SPR (Suppl. Figs. S9–S11). Through control experiments for DSF, we identified that ~1% of the fragment library consistently destabilized the thermal shift of proteins, acting like denaturants (Suppl. Fig. S12). In addition, we argue from a thermodynamics standpoint why DSF might lower hit rates than other technologies (supplemental material).

Substructural Differences in Technology Hits and Nonhits

We next examined the difference between hits identified by specific technologies and hits identified in the superset (validated or consensus for any technology) in a more rigorous manner. We did this by examining biases for each technology on a substructural level for enrichments in hits. To do this, we computed the probability that a substructure is enriched in active compounds for all targets, with respect to hit selection performed using one technology (primary or validated when available; Suppl. Fig. S2B), or hit selection performed using the superset (Suppl. Fig. S2C). By looking at the difference in these probabilities (Equation S2 and Suppl. Fig. S2D), we can understand whether a method tends to identify particular substructures in hits that other technologies miss (D_i > 0; this might be a false positive), as well as identify substructures that the technology of interest tends to miss, although the fragment is enriched in the actives of other technologies (D_i < 0; this might be a false negative). We termed this the difference vector D for each technology, because deviations from zero in this vector reflect differences between a specific technology and the superset over many different targets for specific fragment substructures. We subsequently used this vector in a number of ways to both compare the various screening technologies, as well as to assess how to combine the technologies.

We performed principal components analysis on the difference vector for each technology ( Fig. 5 ) to visualize relationships between the technologies. Interestingly, using this approach, we see that two nonoverlapping clusters are formed composed of the biophysical techniques (red markers) and the biochemical techniques (blue markers). Intuitively, one might roughly separate fragment screening technologies into these two groups, so it is reassuring that our method to identify the distinguishing characteristics of specific technologies captures this intuition, although it did not use this information to generate the difference vectors. Although this method of grouping technologies is visually appealing and makes sense intuitively, we sought a quantitative way to ascertain how complementary technologies are to each other based on the complementarity score, which we describe in the following section.

Figure 5.

Comparison of each technology based on the chemical matter that it identifies as hits versus the chemical matter present in the superset. The first two principal components for the difference vector D of each fragment screening technology are plotted. Biochemical technologies are depicted in blue and biophysical technologies are depicted in red.

Selecting Complementary Assays

We next sought to understand how technologies should be paired. We chose to assess one common strategy of hit selection, where all hits from two (or more) technologies are carried forward, for example, to consider for chemistry, SAR by archive, or crystallization studies. We termed this hit selection strategy inclusive (as opposed to consensus hit selection, which is discussed in the supplemental material). To validate our method, we chose two campaigns that had each used five screening technologies, which was the greatest number of technologies used for a single campaign in our study. In our approach, we assess what fraction of the superset would be obtained if only two of the possible five screening technologies were employed. We then try to predict how to select the optimum screening technologies a priori.

In our analysis, we employed the difference vectors D for each technology (as described in the supplemental material) to describe that technology. This vector stores the differences between a specific technology and the superset for substructures enriched in actives. We sought to use the information in this vector to predict how complementary two technologies are, that is, which ones should be paired. To do this, we use these vectors to calculate (1) the Pearson correlation between two technologies and (2) the complementarity score for inclusive hit selection (see supplemental material).

To validate this method, we left out the target of interest and calculated the Pearson correlation and total inclusive hit selection bias for technology pairs for oxidoreductase 1 and oxidoreductase 2 (Suppl. Fig. S13A,B). We see that for both targets, when the complementarity is low (<0.956) and Pearson correlation is positive (>0), a low fraction of the superset is often obtained (upper left corner in Suppl. Fig. S13A,B). Conversely, when the complementarity is high (>0.956) or the Pearson correlation is negative (<0), a high fraction of the superset tends to be obtained. Furthermore, we can quickly ascertain from these plots that while these metrics largely separate the technology pairs that tend to obtain the greatest fraction of the superset, so does a simple rule of thumb: pairs that include a biophysical and a biochemical assay tend to capture the greatest fraction of the superset. This observation is depicted for oxidoreductase 1 and 2 in Figure 6 . This figure (as well as our analysis) also reveals that pairing two biophysical technologies seems to be less desirable than pairing two biochemical technologies, but there were less biochemical technologies used in these specific campaigns (two vs. three), so this assertion carries less weight. As three or more technologies are assessed, it may be more valuable to use a quantitative assessment as carried out here, but with two technologies, this rule of thumb seems to be adequate.

Figure 6.

The fraction of all hits identified when only two of five technologies are considered for oxidoreductase 1 and 2. Pairs consisting of biochemical (blue), biophysical (red), and both (yellow) technologies are depicted.

To assess the robustness of this argument, we also plotted the Pearson correlation and total inclusive bias for technology pairs when all campaigns were assessed. Once again, we observe that when a biochemical and biophysical technology are paired (green diamonds, Suppl. Fig. S13C), they tend to reside in the region that has a low Pearson correlation or low total bias, suggesting, as in the case of our two test targets, they will yield a large fraction of the superset.

In conclusion, we have presented a meta-analysis of fragment-based screening campaigns that, to our knowledge, is more comprehensive than any other FBS analysis to date. This breadth of data allowed us to draw several conclusions, which would not have been possible to reach otherwise.

From our studies of what makes a fragment amenable for FBS, we learned that there are privileged fragments that hit more targets than expected by chance. These fragments are of high value to a screening library as a higher fraction has been crystallized against targets compared with other hits, and in cases that we examined, they were potential seeds for lead discovery. This implies that specificity screens against off-targets should only be used with caution, as these screens may lead to privileged fragments being removed from a hit list.

From our assessment of technology differences, we learned that every technology is prone to idiosyncrasies, although they may differ in nature. While frequent hitters were identified for all technologies (except DSF), they tended to have similar downstream success in crystallography campaigns for specific targets, compared with fragments that were not frequent hitters, and thus their removal from generic libraries may not be desirable. Perhaps flagging them as frequent hitters for a given technology, so that resources will be invested in them only if they also hit orthogonal technologies, would be a judicious choice.

We investigated two technologies in depth: SPR due to especially high hit rates and DSF due to especially low hit rates. For SPR, we identified fragments that accumulate on the surface of chips over the course of an experiment, which may lead to false positives. For DSF, we identified fragments that consistently destabilize proteins, which may lead to false negatives. These in-depth analyses help explain in part the observation that when multiple technologies are used in parallel, there may be inconsistencies between the fragments selected as hits from each technology.

We also proposed a quantitative framework for comparing screening technologies (bias vectors). This framework supports a simple rule of thumb that was also observed: when using inclusive hit selection with two technologies, a biophysical and biochemical technology should be paired to cast the broadest net on fragment hits.

Footnotes

Acknowledgements

The authors thank John Davies, Dan Erlanson, Stephan Reiling, and Eugen Lounkine for helpful comments, as well as Céline Be, Regis Cebe, Rene Hemmig, and Andrea Saenger for excellent technical assistance. The authors also thank the many talented scientists who generated the data used for this meta-analysis.

Supplementary material for this article is available on the Journal of Biomolecular Screening Web site at .

Author Contributions

The manuscript was written through contributions of all authors. All authors have given approval to the final version of the manuscript.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Novartis Institutes for Biomedical Research. P.S.K. and A.M.W. were funded as Presidential Postdoctoral Fellows by the NIBR Education Office.

References

Erlanson

D. A.

McDowell

R. S.

O’Brien

Fragment-Based Drug Discovery. J. Med. Chem. 2004, 47, 3463–3482.

Jahnke

Erlanson

D. A.

Fragment-Based Approaches in Drug Discovery; Wiley-VCH Verlag GmbH: Weinheim, Germany, 2006.

Zartler

E. R.

Shapiro

M. J.

Fragment-Based Drug Discovery: A Practical Approach; John Wiley: Hoboken, NJ, 2008.

Dalvit

NMR Methods in Fragment Screening: Theory and a Comparison with Other Biophysical Techniques. Drug Discov. Today 2009, 14, 1051–1057.

Navratilova

Hopkins

A. L.

Fragment Screening by Surface Plasmon Resonance. Acs Med. Chem. Lett. 2010, 1, 44–48.

Kranz

J. K.

Schalk-Hihi

Protein Thermal Shifts to Identify Low Molecular Weight Fragments. Method Enzymol. 2011, 493, 277–298.

Albert

J. S.

Blomberg

Breeze

A. L.

. An Integrated Approach to Fragment-Based Lead Generation: Philosophy, Strategy and Case Studies from AstraZeneca’s Drug Discovery Programmes. Curr. Top. Med. Chem. 2007, 7, 1600–1629.

Barker

Courtney

Hesterkamp

. Fragment Screening by Biochemical Assay. Expert Opin. Drug Dis. 2006, 1, 225–236.

Hesterkamp

Barker

Davenport

. Fragment Based Drug Discovery Using Fluorescence Correlation Spectroscopy Techniques: Challenges and Solutions. Curr. Top. Med. Chem. 2007, 7, 1582–1591.

10.

Boettcher

Ruedisser

Erbel

. Fragment-Based Screening by Biochemical Assays: Systematic Feasibility Studies with Trypsin and MMP12. J. Biomol. Screen. 2010, 15, 1029–1041.

11.

Austin

Pettit

S. N.

Magnolo

S. K.

. Fragment Screening Using Capillary Electrophoresis (CEfrag) for Hit Identification of Heat Shock Protein 90 ATPase Inhibitors. J. Biomol. Screen. 2012, 17, 868–876.

12.

Sherman

Dixon

S. L.

Hole Filling and Library Optimization: Application to Commercially Available Fragment Libraries. Bioorg. Med. Chem. 2012, 20, 5379–5387.

13.

Schulz

M. N.

Landstrom

Bright

. Design of a Fragment Library That Maximally Represents Available Chemical Space. J. Comput. Aid. Mol. Des. 2011, 25, 611–620.

14.

Baurin

Aboul-Ela

Barril

. Design and Characterization of Libraries of Molecular Fragments for Use in NMR Screening against Protein Targets. J. Chem. Inf. Comput. Sci. 2004, 44, 2157–2166.

15.

Lau

W. F.

Withka

J. M.

Hepworth

. Design of a Multi-Purpose Fragment Screening Library Using Molecular Complexity and Orthogonal Diversity Metrics. J. Comput. Aid. Mol. Des. 2011, 25, 621–636.

16.

Glass

. Primary, Secondary, and Meta-Analysis of Research. Educational Researcher 1976, 5, 3-8.

17.

Pipeline Pilot, 8.0; Accelrys Software, Inc.: 2010.

18.

Rogers

Hahn

Extended-Connectivity Fingerprints. J. Chem. Inf. Model 2010, 50, 742–754.

19.

Lounkine

Kutchukian

Glick

Chemometric Applications of Naïve Bayesian Models in Drug Discovery—Beyond Compound Ranking. In Chemoinformatics: Case Studies and Pharmaceutical Applications, Bajorath

, Ed.; John Wiley: Hoboken, NJ, 2013.

20.

Nidhi Glick

Davies

J. W.

. Prediction of Biological Targets for Compounds Using Multiple-Category Bayesian Models Trained on Chemogenomics Databases. J. Chem. Inf. Model 2006, 46, 1124–1133.

21.

Tsai

Lee

J. T.

Wang

. Discovery of a Selective Inhibitor of Oncogenic B-Raf Kinase with Potent Antimelanoma Activity. Proc. Natl. Acad. Sci. U. S. A. 2008, 105, 3041–3046.

22.

Bollag

Hirth

Tsai

. Clinical Efficacy of a RAF Inhibitor Needs Broad Target Blockade in BRAF-Mutant Melanoma. Nature 2010, 467, 596–599.

23.

Addie

Ballard

Buttar

. Discovery of 4-Amino-N-[(1S)-1-(4-chlorophenyl)-3-hydroxypropyl]1-(7H-pyrrolo[2,3-d]pyrimidin-4-yl)piperidine-4-carboxamide (AZD5363), an Orally Bioavailable, Potent Inhibitor of Akt Kinases. J. Med. Chem. 2013, 56, 2059–2073.

24.

Xiang

Hirth

Asmussen

. The Discovery of Novel Benzofuran-2-Carboxylic Acids as Potent Pim-1 Inhibitors. Bioorg. Med. Chem. Lett. 2011, 21, 3050–3056.

25.

Schulz

M. N.

Fanghanel

Schafer

. A Crystallographic Fragment Screen Identifies Cinnamic Acid Derivatives as Starting Points for Potent Pim-1 Inhibitors. Acta Crystallogr. D Biol. Crystallogr. 2011, 67(Pt 3), 156–166.

26.

Pierce

A. C.

Jacobs

Stuver-Moody

Docking Study Yields Four Novel Inhibitors of the Protooncogene Pim-1 Kinase. J. Med. Chem. 2008, 51, 1972–1975.

27.

Lopez-Ramos

Prudent

Moucadel

. New Potent Dual Inhibitors of CK2 and Pim Kinases: Discovery and Structural Insights. FASEB J. 2010, 24, 3171–3185.

28.

Williams

The Pyramid Approach to Fragment-Based Biophysical Screening. In Label-Free Technologies for Drug Discovery, Cooper

Mayr

L. M.

; Eds. John Wiley: New York, 2011; pp. 241–253.

29.

Chen

I. J.

Hubbard

R. E.

Lessons for Fragment Library Design: Analysis of Output from Multiple Screening Campaigns. J. Comput. Aid. Mol. Des. 2009, 23, 603–620.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

3.12 MB

0.00 MB