POOLING IN HIGH-THROUGHPUT DRUG SCREENING

Abstract

Pooling in high-throughput screening (HTS) refers to the act of testing mixtures of compounds in a primary screen to accurately identify hits for secondary screening. The gain in compression of the number of compounds to be further screened by pooling can also be extended to achieve much-needed error tolerance in HTS. Despite the success of pooling in other biological experiments, pooling in high-throughput drug screening has been a controversial and often marginalized paradigm. At first appearance, pooling appears to promise gains from reduced effort, or possibly could create more problems than solutions. However, this article demonstrates that pooling is a practical and necessary part of HTS: discussions include the rationale for pooling compounds in HTS, a unifying view of pooling design theory, a review of past attempts at pooling and their success, and recent advances in the field.

1. Introduction

Traditional high-throughput screening (HTS) has provided impressive gains to drug discovery. However, advances in genomic biology and combinatorial chemistry have led to an increasing number of targets and compounds to screen. The past trend of reducing assay size is aiding HTS to keep pace with the increasing target lists, but miniaturization alone is insufficient [1]. One of several ways to improve the efficiency of HTS is pooling, where mixes or pools of compounds are tested in each well of a HTS assay rather than individual testing of each compound. Because most compound libraries contain only a small fraction of active compounds, pooling can reduce the number of tests needed to identify them. The central rationale of pooling is that most compounds can be quickly identified as inactive via a negative result of a pooled test. Because of this potential, pooling has received intermittent attention from the screening community, which is evident from the references cited in this article.

This review discusses the complexities and solutions to pooled drug discovery, and focuses on the following five topics: (i) design of pooling schemes; (ii) experimental success of pooling; (iii) mixing constraints and cutoff selection of pooling; (iv) additive, synergistic and antagonistic effects of pooling compounds; and (v) recent advances in pooling. An additional goal of this review is to create a repository for literature collected from a very diverse set of journals and researchers in the hope of facilitating a fruitful collaboration of ideas among the HTS community.

2. Pooling Designs

Currently, most high-throughput screens employ the one compound, one well approach. This approach is simple in terms of both the physical implementation of the assay and the analysis of results [2, 3]. However, given the usually small number of active compounds in a library that is typically comprised of a much large number of compounds, this approach can be wasteful. Also, HTS screens are prone to several forms of experimental error, which produce false positive (inactive compound classified as active) and false negative (active compound classified as inactive) results during analysis [4]. False positive results are less of a problem because these compounds are retested in subsequent stages. In contrast, false negative results represent potential lead compounds that are lost to the development process because replicate testing of a compound library is prohibitively costly. Therefore, there is a need to improve the efficiency and robustness of screening, and pooling of compounds presents a solution to this.

The concept of testing pools of samples in biological experiments was conceived during World War II, when blood samples from cadets were pooled to optimize syphilis testing [5]. Over several decades, the mathematical theory of pooling has undergone tremendous advances and has found applications in diverse areas [5, 6, 7, 8, 9]. An excellent introduction to the subject can be obtained from the book written by Du and Hwang [5]. There have been several applications of pooling in high-throughput screens with varying success. These attempts used a variety of pooling schemes, which are reviewed immediately below while results of these schemes will be discussed in the following section.

Pooling schemes can be broadly divided into two categories: adaptive and nonadaptive. Adaptive schemes involve testing pools of compounds in stages, whereby information gained from one stage is used to design the next stage. The resultant gain in reduction in the number of assays required is, however, offset by the delay in obtaining the results of a screen. In contrast, nonadaptive schemes involve testing all of the pooled compounds in a single stage, typically testing each compound multiple times. Many of these nonadaptive schemes guarantee the accurate identification of a specified number of active compounds per library. Aspects of both of these strategies have been merged to form hybrid methods, details of which are available elsewhere [5, 6, 7]. In the following discussion, four types of screen designs (‘one compound, one well’, adaptive pooling, orthogonal pooling and nonadaptive pooling strategies) using a hypothetical example of a 2400-compound library (25 plates of 96 compounds each) with 24 truly active compounds (1% hit rate) are presented. Furthermore, a detailed analysis of pooling designs for a subset of 25 compounds comprising 2 active compounds is presented to enable the properties of various pooling schemes to be highlighted. The probability of finding 0, 1, or 2 active compounds in a subset of 25 compounds selected from a larger set of 2400 compounds with 24 active compounds is 99.8%, based on the use of the hypergeometric distribution.

2.1. One compound, one well strategy

The one compound, one well strategy tests each compound individually and hence requires 2400 wells for 2400 compounds. Figure 1 shows the original compound library distributed over 25 plates (P1 to P25), which contain 96 compounds each, and magnifies the results for a subset of 25 compounds in position A1 on these plates. If compounds number 4 and number 20 (position A1 on plates P4 and P20) were truly active but one false negative testing error occurred, the results from a standard detection method would indicate only one compound, for example, compound number 20, as active (highlighted in yellow in Figure 1). In this pooling design, the benefit of a simple design and analysis method is obtained at the cost of testing a large number of inactive compounds. Additionally, any errors in testing cannot be detected or corrected in this primary stage of HTS. Therefore, there is a need for more efficient and robust screening methods, such as the three pooling designs described below.

Figure 1. One compound, one well strategy: Distribution of a 2400-compound library into 25 plates.

(A) Distribution of 2400 compounds over 25 plates in the one compound, one well strategy for HTS. (B) The analysis for a subset of 25 compounds (position A1 on plates P1 to P25); compounds number 4 and number 20 (highlighted) are active. A false negative (FN) result in well number 4 results in the incorrect designation of compound number 4 as inactive. Compound number 20 is correctly identified as active (yellow-colored well).

2.2. Adaptive pooling strategy

The adaptive pooling designs exploit the presence of large numbers of inactive compounds by dividing a compound library of size n into pools of size r (ie, r compounds are mixed in each assay well) [7, 10, 11]. In subsequent stages, the pools that exhibited activity are retested to identify which compounds were responsible for the activity. The number of active compounds or hits is typically denoted by d, the abbreviation for defectives which is the term used in fault-testing literature.

As shown in Figure 2, the 25 plates are pooled in groups of five (M1 to M5). For example, plates P1 to P5 in the original library are pooled and tested in a single plate, M1. This translates to the pooling scheme shown in Figure 3 (Stage 1) for a subset of 25 compounds (n = 25) in position A1 of each plate (M1 to M5) tested in pools of five compounds (r = 5). Because compounds number 4 and number 20 (position A1 on plates P4 and P20) are the active compounds, pools M1 and M4 (wells A1 from plates M1 and M4) should produce a positive response and only compounds in these pools are carried over to the next stage. However, a false negative result in pool M1 would cause compound number 4 to be considered as inactive. Several similar stages can be conducted, but in this example only two stages are required to identify compound number 20. In the second stage, each of the five A1 compounds in pool M4 is tested in individual wells (M6 to M10, Stage 2; Figure 3) to identify compound number 20 as active. The total number of wells of compounds required to be tested to identify compound 20 as active is 10 (5 in Stage 1 and 5 in Stage 2), instead of 25 for the one compound, one well setting. Adaptive designs provide valuable savings in resources by minimizing the number of wells used, but the resultant procedure from such designs can be time consuming to conduct. In general, adaptive designs require the testing of a maximum of d log₂ n wells to identify d active compounds out of n compounds, when no limit on the number of compounds mixed in each well is imposed (see Chapter 2 of reference [5]). Simple adaptive designs such as the one adopted for this example are vulnerable to testing errors, where an error in a single well in an early stage can strongly influence the overall assay accuracy, as depicted in Figure 3. Hybrid designs that test pools multiple times in each stage can be used to reduce the sensitivity of an assay to testing errors.

Figure 2. Adaptive pooling strategy: A 2400-compound library pooled into five plates.

The adaptive pooling strategy implemented at the plate level, by pooling groups of five plates of compounds into one pooled plate which is screened for activity.

Figure 3. Adaptive pooling analysis.

The interpretation of adaptive pooling strategy is shown for a subset of 25 compounds present in well A1 of each plate in the original library in Figure 2. The pooling scheme is conducted in two stages. Stage 1 eliminates 20 compounds (via the negative results in M1, M2, M3 and M5). A false negative (FN) results in well M1 causes compound number 4 to be incorrectly designated inactive. Stage 2 tests the compounds from pool M4 individually to identify the active compound (20), via well M10.

2.3. Orthogonal pooling strategy

The orthogonal pooling or self-deconvoluting matrix (SDM) strategy can be considered to be an intermediate form of pooling and has been used in a number of screens that have been subsequently published [12, 13, 14, 15]. In orthogonal pooling, each compound is tested twice, each time in combination with a completely different set of compounds. Therefore, for an n compound library tested in pools of size r, this strategy would require $\frac{2n}{r}$ wells to be tested. When more than two compounds are pooled in each well (r > 2) in orthogonal pooling, the number of wells that require testing is an improvement from using n wells in the one compound, one well strategy. In orthogonal pooling, each compound is required to show activity in both wells in which it is tested in order to be classified as a hit. Although this method does not provide robustness to testing error, it does conserve both resources and time by conducting a reasonably efficient screen in a single stage. However, even in the absence of testing error, false positives occur when an inactive compound is tested both times with a different active compound (illustrated in Figure 5). This failure probability and the inability to correct or detect testing errors are the major drawbacks of the orthogonal pooling scheme.

Figure 5. Orthogonal pooling analysis.

The analysis of the pooled wells is shown for the subset of compounds present in well A1 of plates P1 to P25 in Figure 4. Compounds present in two active wells are marked as active. However, owing to the false negative (FN) testing error in well M9, compound number 4 is incorrectly labeled as inactive. Furthermore, the design contains one false positive (FP) result with compound number 5 because the compound was tested twice in the presence of two different actives (ie, compounds number 4 and 20). Compound number 20 is correctly identified as a hit.

The application of an orthogonal pooling scheme to the 2400-compound library example is shown in Figure 4. In this design, plates are pooled once along each row and once along each column of the original library, allowing each compound to be tested twice but with a different set of compounds each time. The 25-compound subset in position A1 in all pooled plates (M1 to M10) translates to the design shown in Figure 5. Compound number 1 in the original library (in well A1 in plate P1) is tested once in pool M1 with compounds number 2, 3, 4 and 5 (row pools) and tested again in pool M6 with compounds number 6, 11, 16 and 21 (column pool). This method uses 10 wells (5 wells from the row pool, ie, M1 to M5, plus 5 wells from the column pool, ie, M6 to M10) compared to 25 in the one compound, one well strategy. The status of a compound is determined by examining the result from both sets of wells. A compound is considered to be active only when it is active in both sets of wells. As seen from the decoding diagram at the bottom of Figure 5, only those compounds with two active results (yellow circles) are considered as active. Therefore, compound number 20 is correctly identified as a hit (highlighted with a red rectangle), while compound number 4 is missed owing to a false negative result in pool M9. In addition, compound number 5 was pooled twice with active compounds and is wrongly identifies as a hit (false positive). This example illustrates the risk associated with orthogonal pooling. However, this false positive risk is minimized when the fraction of active compounds in a library is small.

Figure 4. Orthogonal pooling strategy: A 2400-compound library pooled into ten plates.

An illustration of an orthogonal pooling scheme with a 2400-compound library. Each compound is tested in two pools, one created by mixing plates along each row of the original library and the other by mixing plates along each column.

2.4. Nonadaptive pooling strategy

The nonadaptive pooling schemes, which the authors of this review regard to be the most suited for HTS, are more difficult to design and implement than the schemes of the three previously described strategies. However, the nonadaptive pooling strategy offers a reduction in the number of assays used and the ability to detect and correct testing errors in a single stage [8, 16, 17, 18]. Nonadaptive pooling schemes test each compound multiple times across a single screen and in combination with a different set of compounds, using a low overall number of assays. The results from the pooled tests are then jointly decoded via an analysis algorithm to identify the active compounds. Testing each compound multiple times enables nonadaptive pooling methods to detect and even correct experimental errors. Such schemes are designed using advanced mathematical algorithms that account for compound library size, expected hit rates, expected error rates, pool size, etc. Furthermore, these algorithms are used to construct mixing schemes that are guaranteed to identify all the active compounds successfully, even in the presence of testing errors. Because nonadaptive designs are mathematically complex, these schemes are typically represented as a binary matrix showing which compounds (represented along the columns of the matrix) are to be pooled in which wells (represented along the rows). Random nonadaptive pooling designs also exist that require fewer pools than constructed nonadaptive pooling designs and which can be created to guarantee a high, but not perfect, probability of success [17, 19].

An illustration of the application of a nonadaptive pooling scheme with the 2400 compound test case is shown in Figure 6. The black squares along a row correspond to the plates (P1 to P25) with compounds that are pooled together and tested on a single plate (M1 to M20). In this example, the plates are pooled according to a strategy called the Shifted Transversal Design (STD) created by Thierry-Mieg [18] and adapted to drug screening by Kainkaryam and Woolf [20]. The details of this drug screening design are discussed in the next section. The 25 compounds, corresponding to position A1 on these plates, are arranged along the columns of the binary matrix shown with black and white squares in Figure 6 (and Figure 7). A black square indicates that the compound along that column is pooled in the wells corresponding to that row. Hence, compound number 1 is pooled in wells M1, M6, M11 and M16, along with other compounds specified by the mixing matrix.

Figure 6. Nonadaptive pooling strategy: A 2400-compound library pooled into 20 plates.

A nonadaptive pooling scheme using the Shifted Transversal Design pooling strategy, where black squares represent the plates (P1 to P25; along that column) to be pooled together to form a pooled plate (M1 to M20; along that row). 20 pooled plates are used to screen 25 plates of compounds, resulting in a 20% reduction in the number of plates to be analyzed.

Figure 7. Nonadaptive pooling analysis.

The analysis of a nonadaptive pooling scheme for a subset of 25 compounds (well position A1 on plates P1 to P25 in Figure 6) is shown. Decoding the activity of compounds from pooled assay results is conducted by labeling those compounds present in at least two active wells in which all other compounds are inactive, as active. Despite the false negative (FN) testing error, compounds number 4 and 20 are correctly identified as active.

In this example, each compound is tested four times using 20 wells, thus reducing the number of wells to be tested and analyzed by 5, and each assay contains a pool of five compounds. The decoding method for such a pooling scheme is more involved (details available in reference [18]) than the three previously described pooling schemes, but a simplified version is shown at the bottom of Figure 7. Any compound present in at least two wells showing inactivity (blue-colored wells) is considered as inactive. Furthermore, any compound present in at least two wells showing activity (yellow-colored wells), where all other compounds within the same wells are inactive, is considered as active. Despite the false negative error in pool M9, the nonadaptive design is able to accurately decode the true active compounds. This scheme has been explicitly designed to guarantee the identification of up to two active compounds and to prevent the occurrence of false positive results, unlike the orthogonal pooling strategy at the cost of the need for ten extra wells. Of note is that despite testing each compound multiple times, the total number of assays needed in a nonadaptive pooling scheme can be lower than the one compound, one well strategy. Typically, nonadaptive schemes require approximately d² log₂ n wells to guarantee the identification of d active compounds in an n-compound library [21].

3. Review of experimental results of pooling

This section reviews cases in which adaptive pooling, orthogonal pooling and nonadaptive pooling have been used in published drug screens. These examples are far from exhaustive, but highlight the reasoning that is used in selecting a pooling design for screening. Pooling has been successfully used for screening solid-phase libraries, combinatorial libraries, natural product extracts, etc [15].

One area where adaptive pooling strategies have been successfully used is in combinatorial library screening. This approach takes advantage of combinatorial synthesis of compounds to produce mixtures of compounds, which typically have some common functional unit. In one such study, 810 compounds synthesized from two-way and three-way combinations of monomer units, taken from a nine-monomer set, were tested in a human type II phospholipase (PLA₂) inhibition assay using an adaptive pooling strategy [22]. At each stage of the adaptive test, pools of compounds having a common monomer unit were tested together, resulting in the testing of 27 compounds per well in the first stage. The pool with the highest inhibition was progressed to the next stage, where the common monomer unit from the previous stage was preserved, while a second monomer position was varied for each pool. This process was repeated for all three monomer positions and the screen finally led to the compound with the three-monomer unit with the best PLA₂ inhibitory activity. Simulated measurement errors of approximately 7% for high-activity and 16% for low-activity pools were reported for this study. At these levels of experimental error, an adaptive strategy that does not use replicate measurements can be successfully used for screening. Similar combinatorial synthesis approaches to pooling have been conducted experimentally and reviewed elsewhere [23].

A number of examples of the reasonable successful use of orthogonal pooling by the screening community have been reported [12, 13, 14, 15]. In a recent study, Motlekar et al tested 64,000 compounds in a human cysteine protease cathepsin B assay, using both the one compound, one well approach and an orthogonal pool of ten-compound mixtures per well [13]. Twenty active compounds were identified via the orthogonal pooling procedure, including three compounds that were missed (false negative) by the one compound, one well strategy. Similarly, Warrior et al conducted a 10-compound per well orthogonal pooling study on three different screening formats (a GPCR functional assay, a GPCR binding assay, and a tyrosine kinase assay) and showed that the performance of pooling was equal to or better than one compound, one well screening in all cases [14]. For cases in which the first stage results were ambiguous, the standard orthogonal pooling strategy was augmented with an additional retesting stage. Despite this additional feature, a cost comparison study showed that orthogonal pooling in this instance was significantly cheaper and faster than the one compound, one well strategy. Furthermore, in a comprehensive study published in 2005, Ferrand et al rigorously tested a library of 26,400 compounds in a CXCR3 scintillation proximity assay by both one compound, one well (repeated twice) and orthogonal pooling (repeated once) approaches [15]. Orthogonal pooling was found to significantly reduced resources and time, as well as false positive errors without increasing false negative errors. Data from all five screens (three in one compound, one well and two in orthogonal pooling) were used to identify 64 true actives, against which the statistical performance of the pooling approaches was evaluated.

Although the theoretical performance of nonadaptive designs are superior to orthogonal designs, nonadaptive pooling has not yet been tested in a drug screen to the best of the knowledge of the authors of this review. Strong theoretical and computational results suggest that nonadaptive pooling has the potential to be successfully applied to HTS. In 2006, as already mentioned, Thierry-Mieg published the general nonadaptive pooling strategy STD using a mathematical construction, which is the best-known algorithm to date [18]. The STD algorithm takes in, as inputs, the compound library size (n), the maximum number of active compounds expected (d) and the maximum number of false positive and negative testing errors expected (E) to design a pooling strategy that guarantees the successful identification, in a single stage, of d active compounds in the presence of E testing errors, while maintaining a lower number of tests than that required by the ‘one compound, per well’ approach. The STD algorithm achieves compression and accuracy by testing each compound multiple times in unique pools to guarantee the identification of all d active compounds, even when testing errors occur. The example in Figure 7 shows an STD strategy for n = 25; d = 2 and E ~ 1, using 20 pooled wells to test 25 compounds. In reference [20], the authors report the extension of this approach to drug screening by implementing a constraint on the maximum number of compounds that can be mixed in a well (r) and by modifying the approach to include false testing error rate (e) as an input instead of the number of errors expected (E), which is difficult to determine without knowing the number of wells used.

In practice, nonadaptive pooling methods may be more difficult to implement than adaptive pooling methods because the mechanics of pooling compounds requires a comparatively complex series of mixing steps to create the pools. However, Kainkaryam and Woolf proposed a block pooling design strategy that can be used overcome this problem [20]. As is illustrated in Figure 6, any nonadaptive design can be used on either individual compounds or on a plate basis. If the pooling is performed on a plate basis, more powerful yet complex nonadaptive designs using existing liquid-handling technology can be adopted. The subset of compounds for which the pooling design would apply is defined by their position on the plates. The other design parameters, such as the hit-rate and error-rate, can be chosen appropriately for this subset based on parameters chosen for the whole library. In contrast, pooling on an individual compound basis is possible, but would require more advanced liquid-handling technologies to be practical. This practical limitation may also explain why orthogonal pooling has been more popular in the screening community, despite its design drawbacks. A positive feature of nonadaptive designs is their ability to complete a screen in a single stage, thereby avoiding errors caused by day-to-day variability. Ferrand et al showed that day-to-day variability has a stronger influence on the screening results of standard procedures compared with any differences due to the use of one compound, one well or pooling strategies [15].

4. Note on pool size and choice of threshold

Most pooling schemes in HTS are constrained by the number of compounds that can be mixed in a well. The constraint arises from factors such as: (i) physical limitations of well size; (ii) effect of ionic strength of pool on compounds; and (iii) side effects of pooling, such as synergistic and antagonistic interactions between compounds (discussed in detail in the next section). Therefore, pooling designs used for drug screening include the parameter r, which as already explained represents the limit on compounds that can be pooled in a well. In most designs, this limit on pool size also defines the upper limit of the possible compression and error correction that can be gained by pooling in HTS.

A second issue is the definition of a cutoff to define an assay result as a positive or negative result. In a pooled screen, due to the presence of multiple compounds, the concentration of each compound in the well may be different than if the assay were run using a one compound, one well design. Generally, for a compound showing inhibition (%I) at a certain concentration ([C]), the IC₅₀ value can be calculated by the following expression (equation 4.1) [13]:

$${\text{IC}}_{50}=[C]\frac{100-\%I}{\%I}$$ (4.1)

Assuming that pooled screens would test compounds at lower concentrations than a one compound, one well screen, for a compound tested in both formats, the following relationship (equation 4.2) exists between the compound concentration in both formats ([C]_S and [C]_M for single and pooled formats, respectively) and the corresponding cutoff inhibitions (%I_S and %I_M for single and pooled formats, respectively) [13]:

$${[C]}_S\frac{100-\%I_S}{\%I_S}={[C]}_M\frac{100-\%I_M}{\%I_M}$$ (4.2)

The expression for the choice of cutoff inhibition in a pooled screen (equation 4.3) can be obtained by rearranging equation 4.2. %I_M corresponds to an equivalent choice of cutoff inhibition in a one compound, one well screen by accounting for the difference in the testing concentration of the two screening formats.

$$\%I_M=\frac{{[C]}_M}{{[C]}_S}\frac{\%I_S}{100-\%I_S(1-\frac{{[C]}_M}{{[C]}_S})}\times 100$$ (4.3)

Equation 4.3 represents the cutoff inhibition above which a pooled well can be considered to display activity corresponding to the presence of at least one active compound tested individually. Using the example from reference [13], if a compound tested individually at 10 μM ([C]_S = 10) is labeled as a hit for producing 33% inhibition (%I_S), then the same compound in a well containing several other inactive compounds tested at 5 μM ([C]_M = 5) each, would have to show at least 20% inhibition to be termed a hit:

$$\%I_M=\frac{5}{10}\frac{33}{100-33(1-\frac{5}{10})}\times 100=20$$

5. Additivity, Synergies and Antagonism

Some of the problems that could arise from pooling of compounds in HTS have featured predominantly in the debate about the use of pooling [24, 25]. The following three issues will be discussed: additivity, synergies and antagonistic effects. Additivity assumes independent action of compounds in pools, while synergy and antagonism refer to interactions between compounds in pools. Recent efforts to handle these three issues are described in the section Advances in pooling.

Pooling several moderately active compounds together in a well, such that the combined activity of the compounds mimics the appearance of an single active compound, leads to the production of an additive effect. From the previous example in the section Pool size and choice of threshold, if two compounds that individually produce only 20% inhibition at 10 μM were pooled together at 5 μM each, the result from the pooled well would be an observed inhibition of greater than 20%, assuming both compounds acted independently. Therefore, such a pooled well would be considered a hit, using the previous criteria for pooled wells, resulting in a false positive. Researchers who have published drug screens that used pooling have not searched systematically for this effect, although several have reported finding false positive results [13, 14, 15]. An alternative and favorable view of such false positive results is that additive collaboration of non-interacting compounds could ultimately lead to potential multi-component therapies.

Pooling of compounds could also produce physiochemical effects, such as a reaction or aggregation. Consequently, such compounds could achieve a positive result (synergy) or negate the activity of each other (antagonism or blocking). Typically, pooling designs treat these interactions as false positive (synergy) and false negative (antagonism) testing errors and attempt to correctly identify them by including retesting steps. However, several research groups have systematically screened multi-compound mixtures to identify useful synergistic effects, with reasonable success [26, 27]. For example, Feng and Shoichet studied synergistic and antagonistic effects caused by aggregation-based behavior of compound mixtures [27]. They screened 764 compounds individually and in pools of 10 compounds and compared the observed pooled behavior with that predicted from individual tests. They found that synergistic effects predominated, although antagonism was also observed.

There is great potential for smart pooling design and analysis strategies that successfully identify individual hits while: (i) minimizing resources; (ii) providing error tolerance; and (iii) retaining the ability to identify synergistic effects. Some of the above goals are conflicting, but recent advances in pooling design theory promise reconciliation among them, as discussed in the following section.

6. Advances in Pooling

This section presents recent (mainly) theoretical advances that attempt to address or reconcile the multiple goals of pooling designs. These advances involve both the design of smart pooling strategies as well as the application of wide ranging methods to analyze the results from pooled screens. Typically, these analysis methods are computation-intensive techniques that shift the emphasis of the pooling from experimentation toward assay design and analysis.

Several approaches have been proposed to prevent interactions among compounds during pooling, such as: (i) reducing the number of compounds pooled; (ii) pooling dissimilar compounds; and (iii) systematically preventing chemical classes such as electrophiles and nucleophiles from mixing [14, 28]. For adaptive pooling designs, computational methods have been designed that use chemical structure information to design optimal pools for maximizing coverage of the chemical space, while minimizing overlap [29]. For orthogonal pooling, simulation techniques have been proposed that predict probabilities for the occurrence of synergy and blocking (antagonism) to help select the most suitable efficient pooling and retesting schemes [30].

Given the increasing complexity of the goals of pooling schemes, nonadaptive pooling designs appear to be the most promising strategy for screening, because such designs typically test each compound several times during a screen. The resultant increased amount of data allows for analysis schemes that explicitly account for possible antagonists and exhaustively search for pair-wise or higher-order synergies (see Chapters 6 and 8 in reference [17] for details). Furthermore, the mathematical guarantees of most nonadaptive designs are provided for worst-case scenarios, while on average these designs can identify more active compounds and correct larger testing errors [18]. A simulation-based approach has been proposed to exploit the abilities of nonadaptive pooling to design better pooling strategies and decoding methods [31]. In keeping with this trend of increasing computational efforts to address the complex goals of pooling, a Bayesian network pool decoder has been proposed that generalizes the analysis problem [32]. This Bayesian network decoder can incorporate prior knowledge of compound activity and various possibilities for measurement error and can potentially be extended to specifically identify synergistic or antagonistic effects. Together, these approaches promise to sustain the success of high-throughput drug screening.

7. Conclusions

The recent decline in success of the application of the one compound, one well strategy to high-throughput drug screening suggests that the strategy could benefit from further optimization. Pooling of compounds in HTS offers several potential gains, not limited to resource savings and error tolerance. Pooling also offers to address some of the critical issues experienced by HTS, such as the rapid increase in compounds and targets to be screened, the ever-present possibility of false negative testing errors, and the search for new classes of multi-compound, multi-target therapeutics. There are several challenges to the success of pooling in HTS, including the creation of optimal pooling design and analysis strategies, physical implementation of these pooling strategies, and management of synergistic or antagonistic behavior of compounds. Some of these challenges are also opportunities to extend the current capabilities of HTS in drug discovery. Collaborations between the currently disjointed theoreticians and practitioners of these methods would greatly enable these challenges to be met and opportunities to be realized.

Abbreviations

HTS

high-throughput screening

STD

shifted transversal design

SDM

self-deconvoluting matrix