Small-Molecule Library Subset Screening as an Aid for Accelerating Lead Identification

Abstract

Several small-compound library subsets (14,000 to 56,000) have been established to complement screening of a larger Genentech corporate library (~1,300,000). Two validation sets (~1% of the total library) containing compounds representative of the main library were chosen by selection of plates or individual compounds. Use of these subsets guided selection of assay configuration, validated assay reproducibility, and provided estimates of hit rates expected from our full library. A larger diversity subset representing the scaffold diversity of the full library (3.4% of the total) was designed for screening more challenging targets with limited reagent availability or low-throughput assays. Retrospective analysis of this subset showed hit rates similar to those of the main library while recovering a higher proportion of hit scaffolds. Finally, a property-restricted diversity set called the “in-between library” was established to identify ligand-efficient compounds of molecular size between those typically found in fragment and high-throughput screening libraries. It was screened at fivefold higher concentrations than the main library to facilitate identification of less potent yet ligand-efficient compounds. Taken together, this work underscores the value of generating multiple purpose-focused, diversity-based library subsets that are designed using computational approaches coupled with internal screening data analyses to accelerate the lead discovery process.

Introduction

High-throughput screening (HTS) has become a staple for lead discovery in most small-molecule drug discovery programs.^1,2 Its foundation is the testing of large numbers of compounds in assays designed around target molecules, pathways, or phenotypes believed to be relevant in specific disease conditions. It is very labor intensive and requires significant infrastructure that includes compound libraries, automation, and data analysis in addition to assay development. Based on cost and time requirements, ongoing efforts are aimed at increasing the efficiency and productivity of HTS while maintaining its quality.^2,3

Compound libraries often range in size from hundreds of thousands to millions. Consequently, one approach to increasing screening efficiency is reducing the size of the compound screening set, but this must be done without significant loss of diversity of chemical matter in the resulting hit sets. Approaches to reducing the screening set size have included sequential or iterative-focused screening and generation of diverse sets of compounds. ⁴ Iterative-focused screening requires sequential rounds of generation and testing of small-compound sets. As multiple compound selection and picking efforts are required, the efficiency of this approach has been questioned. ⁵ In addition, the diversity of the final hit set is heavily influenced by the composition of the original focused set, so the likelihood of finding novelty is low. A number of approaches to generation of diverse compound sets for use in HTS have been described that explore methods of compound selection and assessment of the constituent chemical diversity.^2,6–9 Diversity selection was central to the generation of each of the subsets described in this article.

Although reduction in the size of the screening set is a typical driver of library subset generation, it was not the key driver in the building of our initial library subsets. Our general strategy has been to evaluate all compounds in all screens as long as the assay format and reagent availability permit. We want every opportunity to identify potential hits, particularly for challenging targets. We routinely screen more than 1 million compounds per week in small volumes that minimize reagent costs. Rather, a key objective of our subset screening was to enhance the quality and productivity of our lead-finding efforts.

Accordingly, our initial subset-generation efforts were aimed at providing tools for assessment of assay and screen performance, providing project teams with preliminary data in advance of full screens, and identifying hit compounds that may not emerge from the full library when screened under typical conditions. Both a plate-based and a cherry-picked validation set were established with the aim of maximizing diversity while maintaining comparability to our main corporate library. In addition, a 50,000-compound diversity set was generated to address challenging targets with difficult assays and/or limited reagent availability and also to facilitate screening of exploratory targets using a modest-sized compound set. Finally, a diverse set of compounds, designated the in-between library, was assembled with the aim of identifying hits with high ligand efficiency (LE) and molecular size intermediate between fragments (traditionally screened at high concentration with biophysical methods) and typical HTS library compounds (normally screened at lower concentrations in biochemical or cellular assays); this aim has been facilitated by screening at concentrations higher than those used for our main library.

Although library subsets and high-concentration screening are not novel and are pursued in many drug discovery organizations, there are relatively few publications describing their design and use. As such, the aim of this article is to describe the rationale and methods for generation of the four diversity-based library subsets to support HTS at Genentech. Furthermore, their value will be demonstrated through examples of the use of small diverse subsets in preparation for full screens, retrospective analyses of subset hit rates and scaffold recovery, and assessment of the efficiency of hit identification with the property-restricted subset screened at elevated concentrations.

Materials and Methods

Origin of the Genentech Compound Library

The HTS library at Genentech has been assembled over the past decade and now contains more than 1.3 million compounds in discrete wells. Compounds in the library fall into three main categories: (1) individual compounds from commercial sources, (2) groups of commercial compounds synthesized around specific chemical scaffolds (these sets number from tens of compounds to more than a thousand compounds), and (3) compounds synthesized for specific therapeutic projects. Because compounds of type 3 were not systematically included in the generation of all subsets, they were not included in any of the analyses detailed below.

Compounds from commercial sources were selected to increase the chemical diversity of the library at the single compound or scaffold level. Compounds were assigned to a particular scaffold solely on the basis of supplier designation without further extension to structurally similar nonscaffold compounds. Selection of compounds was facilitated by the use of filters to eliminate compounds that are unattractive as starting points for a medicinal chemistry program (see below). Physicochemical property filters have also been used to keep the majority of purchased compounds compliant with Lipinski’s “rule of 5” criteria. ¹⁰

Design of Library Subsets

Electronic files containing the normalized chemical structures (SD file format) were exported from our corporate database and analyzed with Pipeline Pilot (version 7.5; Accelrys Inc., San Diego, CA) or command line tools created in house using the OEChem Toolkit (OpenEye Scientific Software, Santa Fe, NM). Simple ring assemblies or Murcko ring assemblies were generated using Pipeline Pilot. ¹¹ Simple ring assemblies contain all ring atoms that share at least one ring atom or ring bond; there may be several individual rings in a simple ring assembly (fused or spiro-linked) and several simple ring assemblies per compound. Murcko assemblies contain all ring atoms and all atoms on a direct path between each ring; there is a single Murcko assembly per compound. Comparison of molecular structures was performed using either an Atom-Atom Path (AAP) similarity method or Tanimoto similarity calculated with extended-connectivity fingerprints (ECFPs) derived from Pipeline Pilot (ECFP4 or FCFP4) or an in-house developed tool (designated CFP and based on Rogers and Hahn ¹² ). More details of the AAP and CFP methods will be presented elsewhere. Undesirable molecular features were described using SMILES or Mol-file format files.^13,14 Filter components in Pipeline Pilot were used to remove structures containing these features prior to all of the analyses described herein, including defining the content of the main library. In line with other recent literature (for example, see Baell ¹⁵ and references therein), these filters remove compounds with potentially reactive functional groups (e.g., aldehydes, Michael acceptors), features known to give rise to false-positives in biochemical assays (e.g., rhodanines) or are undesirable starting points for drug development (e.g., quinones).

The compounds for the plate-based validation set were selected from the compounds in the main library. Ring systems (either simple ring assemblies with attached alpha atoms or Murcko assemblies including atoms with double bonds to the linker) were calculated for those compounds; unsubstituted terminal rings (morpholine, cPropyl, cButyl, cPentyl, cHexyl, and cHeptyl) were electronically removed from the structures prior to ring system generation so that these features would not contribute to the core ring assemblies. In both cases, exocyclic double bonds were retained as part of the assembly. Unique ring systems associated with each 1536-well plate were noted and then plates were ordered so as to maximize the accumulated number of unique ring systems. Plates 1 to 10 and 11 to 20 from each of these two orderings were analyzed independently for consideration as potential validation subsets.

The compounds for the cherry-picked validation set were chosen from compounds in the main library after removal of compounds with molecular weights greater than 500. Compounds that were no longer commercially available were also removed prior to selection to facilitate repurchase of the validation set should the compounds become depleted. Each compound was assigned a random number between 0.0 and 1.0 using the Pipeline Pilot “rand” function with a four-digit integer seed, and the 14,080 compounds with the lowest number were selected. This was repeated six times with different seed values to generate six potential validation subsets.

A diverse set of 50,000 compounds was selected with the aim of containing representative examples of all of the scaffold-based sets of compounds plus a diverse sampling from the nonscaffold sets. Stringent filtering of the 1.3 million compound set based on physicochemical properties was used prior to selection to improve overall assay data quality (molecular weight <500, cLogP <5; see Suppl. Fig. S1). After filtering, approximately 480,000 compounds associated with ~3300 scaffolds containing 25 or more compounds were designated the “scaffold set.” Approximately 250,000 compounds without scaffold designation were combined with ~20,000 compounds from sparsely populated scaffolds (fewer than 25 compounds) to form the “nonscaffold” set.

Diverse compounds from each of the 3300 defined-scaffold groups were selected for inclusion in the subset using an order-dependent, directed sphere-exclusion algorithm.^16,17 Molecules in each scaffold set were ordered according to commercial availability and increasing molecular weight. The lowest-molecular-weight compound from each scaffold was included in the selected set, and then compounds of higher molecular weight were added to the growing set if their AAP similarity metric was less than a specified value. Thus, higher-molecular-weight compounds were added to the selected set only if they contributed significant diversity. The similarity cutoff was adjusted for each scaffold to select approximately 1 compound in 15, resulting in the final selection of ~27,000 scaffold-based compounds. The 270,000 nonscaffold compounds were filtered to remove compounds similar to those already selected from the scaffold set using the CFP similarity metric. Selections from the remaining nonscaffold set were made using the same order-dependent algorithm acting on the molecular weight–sorted list of nonscaffold compounds, using the CFP method as comparator. The similarity criterion was adjusted to select ~21,000 compounds from this set.

Compounds for the initial version of the in-between library were selected from those available in 2008 from ChemDiv (San Diego, CA), ChemBridge (San Diego, CA), Asinex (Moscow, Russia). and LIFE Chemicals (Kyiv, Ukraine). Compound lists were filtered to remove compounds with molecular weights greater than 350 and cLogP values greater than 3.5. Murcko ring assemblies were generated for each compound, keeping exocyclic double bonds and including wild-card atoms at the substituent attachment points. The lowest-molecular-weight example from each Murcko assembly set was selected for purchase. Other compounds from the Murcko assembly set were considered in order of increasing molecular weight and added to the purchase list if they had a Tanimoto similarity (FCFP4 descriptors) coefficient <0.85 versus all compounds already in the list. If the compound was already present in the HTS library, a 0.65 Tanimoto similarity cutoff was used so as to ensure that the in-between set contained representative compounds from the HTS library but was biased toward novel compounds.

Analysis of Hit Rates and Scaffold Recovery

The utility of our library subsets was initially assessed by conducting an in silico analysis of primary screening data from 39 HTS campaigns (Suppl. Table S1). These screens covered multiple target classes and assay technologies. Assay technologies included time-resolved fluorescence resonance energy transfer (TR-FRET), luminescence, fluorescence-polarization, IMAP, Z′-Lyte, kinetic fluorescence, and fluorescent and luminescent cellular reporters. Data were processed to yield a percentage inhibition for each compound; values less than 0% inhibition indicate apparent enhancement. Compounds that inhibited greater than 50% in an assay were defined as hits unless otherwise indicated. Hit rates for the main library and subsets were calculated as percentages by dividing the number of hits by the total number of compounds assayed and multiplying by 100. Linear regression analyses were used to compare hit rates of the library subsets to respective hit rates for the main library.

The ring-assembly recovery rates for the subsets were calculated as the number of unique Murcko assemblies present in the hits obtained from the subset divided by the number in the hit set from the main library; these values differ from the simple compound hit rates as more than one hit may contain the same Murcko assembly. The recovery rates were also calculated for groups of more than 10, 25, or 50 hit compounds in the main library containing the same Murcko assembly; in this case, the scaffold was considered recovered if at least one compound containing that Murcko assembly was present in the subset hits. A similar analysis was conducted on compounds in supplier-defined scaffold sets to determine the scaffold recovery rates.

Library management, the screening system, the current screening workflow, and use of validation subsets for pilot and validation screens are described in the supplementary material.

Results

This work describes the generation of and uses of two validation subsets and one diversity subset of our main library in addition to a property-restricted collection that partially overlaps with the main library. Table 1 summarizes the purpose and methods for subset generation and provides the numbers of compounds and structural assemblies in each of the resulting library sets relative to the Genentech main library. Supplementary Figure S1 summarizes the molecular weight and cLogP distributions for the main library and each subset.

Table 1.

Summary of Genentech high-throughput screening and subsets thereof.

Name	Main library	Plate-based validation	Cherry-picked validation	50k Diversity	In-between library
Purpose	Hit finding	Assay validation & prediction of screen performance	Assay validation & prediction of screen performance	Partial library screen to identify hits	Hit finding
Selection/ creation details	Not applicable	Plate-based to maximize ring assembly content	Random	Scaffold and cluster-based to represent library diversity	Commercially-available compounds with diverse assemblies
Logistics of picking	Not applicable	1536-well plates from current library	Individual compounds from current library	Individual compounds from current library	Individual compounds from commercial sources
Number of compounds (1536-well plates) a	1.3 million (1005)	12,303 (10)	13,485 (10)	49,037 (36)	55,646 (43) (40,802) b
Number of Unique Simple Ring-Assemblies a	16,263	3,200	3,228	9,050	5,771 (1,414) b
Number of Unique Murcko Assemblies a	244,040	6,582	8,731	34,882	15,356 (6,513) b
Number of supplier-defined scaffolds	5,750	57	2,421	4,154	Not Applicable

a

Counts of unique compounds and ring assemblies made after removal of compounds that possess undesirable molecular features; see the Methods section.

b

Number in parentheses indicates number of compounds and ring assemblies that are unique to the in-between set and not in the main library.

Validation Sets

A plate-based validation set has the advantage of being easy to assemble because entire 1536-well plates can be selected without the need to cherry-pick individual compounds. We reasoned that a set of compounds containing diverse ring systems would be representative of the entire library and therefore selected four sets of plates that maximized the number of unique ring assemblies (simple or Murcko) present in the set. Interestingly, the count of unique ring systems as more plates are included in the set increased very rapidly for the simple rings but much more gradually for the typically larger Murcko rings (see Suppl. Fig. S2). Data from 13 screens were used to compare the hit rates obtained with the four potential subsets to the hit rates obtained from the main library ( Fig. 1A , C ; Suppl. Table S1); the resulting linear regressions differed considerably from each other, with slopes ranging from 0.26 to 1.4 and y-intercepts from −0.29 to 1.5. The hit rates of the main library were best matched for plate set 2 based on the simple ring assemblies, so this set became our plate-based validation set. These 10 plates contain 12,303 compounds with 3200 unique simple ring assemblies and 6582 Murcko ring assemblies; thus, the plate-based validation set represents 20% of the simple ring and 2.7% of the Murcko ring diversity in the main library while containing only 1% of the compounds ( Table 1 ).

Figure 1.

Hit rate evaluations for potential validation sets. Scatterplots compare hit rates of the main library to those of (A) the four potential plate-based validation sets and (B) the six potential cherry-picked validation sets. Individual points represent specific screens. The green lines represent the line of identity, where x = y. (A) Compound sets: Murcko set 1 (blue), Murcko set 2 (green), ring set 1 (yellow), ring set 2 (orange). (B) Compound sets: set 1 (blue), set 2 (green), set 3 (yellow), set 4 (orange), set 5 (red), set 6 (brick red). (C) Table of linear regression parameters for comparisons of subset hit rates to main library hit rates.

Although the plate-based validation set contains a good sampling of the available ring diversity, it may not ideally be suited to predict the hit rate expected from the main library because compounds containing specific ring assemblies are not distributed uniformly across the plates. Although scaffold-based compounds make up a significant proportion of the overall library, they most often occur on plates with other compounds from the same scaffold. Thus, the ring diversity present on these plates will be low, and they will not be selected for the plate-based validation set. To circumvent this, we generated a second validation set by cherry-picking individual compounds from the main library. Six potential compound sets were randomly selected, and the hit rates expected from each set were compared with the hit rates obtained from the main library using data from 20 historic screens (Suppl. Table S1). Hit rates for all six compound sets under consideration showed much stronger correlations with those of the main library than did the candidates for the plate-based validation set ( Fig. 1B , C ). From the regression analysis, all sets showed similar slopes and R² values, but set 4 was chosen because it had the y-intercept nearest to zero and, consequently, should show better predictions for low-hit-rate screens. This set contains 3228 simple rings and 8731 Murcko assemblies, representing 20% and 3.6%, respectively, of each of these assemblies in the main library ( Table 1 ). Notably, Murcko assemblies have greater representation in the cherry-picked validation set than in the comparably sized plate-based validation set, whereas the simple ring assemblies occur in similar numbers. This results from the very nonuniform distribution of simple ring assemblies across compounds with multiple rings present in each molecule; regardless of the selection mechanism, the most populous simple rings will be present in at least one compound.

Applications of Validation Sets

During the assay development process, it is extremely valuable to have a set of compounds available for testing assay performance. This compound set should be small enough that it can be run manually with the aid of automated reagent dispensers as a pilot screen or on the uHTS system in several replicate sets as a validation screen. The pilot screen is intended for initial characterization of an assay in terms of sensitivity to inhibition and susceptibility to compound-related artifacts. ¹⁸ The objective of the validation screen is assessment of the impacts of reagent stability, process automation, and liquid handling on the quality and reproducibility of the data. Although assay controls and DMSO plates can provide some of this information, use of a validation compound set enables assessment of compound reproducibility and inhibition distribution as well as estimation of false-positive and false-negative rates. ¹⁹ For purposes of both pilot and validation screens, it is important that the compounds possess a range of properties and activities. Ideally, such compound sets should be representative of the full library so that results obtained with the subset are predictive of what will be obtained for the entire collection.

The plate-based and cherry-picked validation subsets were routinely used to ensure that optimal screening conditions were established prior to commencing the full library screen. For example, a pilot screen of the plate-based validation set with target 18 showed a high initial hit rate of 2.2%. This hit rate triggered further examination of the assay conditions and prompted elimination of the enzyme/compound preincubation. Upon reassay of the plate-based validation set, a number of compounds that inhibited significantly in the original assay no longer inhibited in the modified assay ( Fig. 2A ). The reduction in hit rate to 0.35% was likely due to a reduction in nonspecific inhibition. ²⁰ In a second case, assay of the cherry-picked validation set promoted selection of an elevated compound concentration for the target 37 primary screen. Initial assay of the validation set at 5 µM led to a hit rate of 0.19% with a relatively low inhibition cutoff set at 35%. As this target was known to be difficult to inhibit specifically, we wanted to increase this hit rate. Subsequently, reassay of the validation set at an increased concentration of 7.5 µM yielded a hit rate of 0.33% with the same inhibition cutoff ( Fig. 2B ). After eliminating false-positive hits arising from detection interference (four compounds at 5 µM and seven compounds at 7.5 µM), the final hit rates of 0.16% and 0.28% for the 5 and 7.5 µM concentrations, respectively, led to a decision to screen the main library at a concentration of 7.5 µM.

Figure 2.

Use of validation sets in pilot screens and validation screens. (A) The plate-based validation set was evaluated in the original target 18 assay and then was reassayed in a modified version of the assay. (B) The cherry-picked validation set was evaluated in the target 37 assay first at 5 µM and then subsequently at 7.5 µM. (C) The cherry-picked validation set was evaluated in two different assay formats for target 26. (A–C) Scatter-plots of inhibition compare the results under each of the two sets of assay conditions. The green line represents the line of identity where x = y. The blue lines represent inhibition cutoffs used for these analyses; they were based on three times the compound standard deviations in each evaluation. (D) The cherry-picked validation set was assayed against target 26 four times over a 16-h period on the uHTS system. The inhibition results of sets 1, 2, and 4 are plotted in three dimensions. (A–D) Legend: Total control, complete reaction with DMSO sample (green); background control, reaction lacking target protein and producing minimal assay response (red); antagonist control, complete reaction with inhibitory control compound (yellow); compounds, validation set compounds (blue).

The cherry-picked validation set was also used for making a selection between two potential assay formats for the target 26 primary screen. Our primary considerations in choosing between TR-FRET and fluorogenic activity assays were sensitivity to inhibition and compound interference with assay detection. Assay of the validation set compounds showed more sensitive detection of inhibitors (32 compounds as compared with 7) and a reduced incidence of negative compound interference (6 compounds with less than −50% inhibition as compared with 67) for the TR-FRET assay over the fluorogenic assay ( Fig. 2C ). Consequently, assay of the validation set enabled a clear choice of the TR-FRET assay for the full library screen. With the same validation set, the performance of the target 26 TR-FRET activity assay was validated on the uHTS system prior to launch of the full screen by running four replicate 10-plate sets spanning a 16-h period. Inhibitory compounds showed good reproducibility across the replicates ( Fig. 2D ). However, greater variability was observed for compounds exhibiting negative inhibition, possibly as the result of poor compound behavior (i.e., low solubility, nonspecific interaction, and/or detection interference). The hit rates for the four sets ranged from 0.036% to 0.057% (five to eight compounds), and there was one false-positive and no false-negatives. ¹⁹ As such, the validation set data supported performing the full screen with the existing TR-FRET assay and automation protocol.

A key purpose for the validation sets is prediction of hit rates; accordingly, we compared the hit rates for these sets to the hit rates for the main library. The screens used to select among several options for each validation set (the qualification set) were considered separately from the remaining screens (test set) in our evaluations. In comparison to the main library, the plate-based validation set tended to underpredict hit rates for screens on the low end and overpredict hit rates for screens on the high end ( Fig. 3A , E ). In contrast, the hit rates for the cherry-picked validation set showed closer agreement with the hit rates of the main library, as indicated by linear regressions for the hit rate comparisons with y-intercepts nearer to 0 and slopes nearer to 1 for both the qualification- and test-set screens ( Fig. 3B , E ). Closer correlation of the hit rates for the cherry-picked validation set with the hit rates of the main library is also indicated by the R² value of 0.98 for its test set as compared with 0.90 for the test set for the plate-based validation set.

Figure 3.

Hit rates of library subsets relative to hit rates of the main library (compounds that inhibited greater than 50% in an assay were defined as hits). Scatterplots compare hit rates for the main library to those of (A) the plate-based validation set, (B) the cherry-picked validation set, (C) the 50,000 diversity set, and (D) the in-between library. The green line is the line of identity, where x = y, and individual points represent specific screens. Green data points in A and B indicate screens used to qualify subsets. (E) Table of linear regression parameters for comparisons of subset hit rates to main library hit rates. Linear fits were conducted separately for the screens that were used to qualify the validation sets (qualification sets) and those that were not (test sets).

Design and Analysis of the Diversity Set

The main library contains ~56% scaffold-based compounds with tens or hundreds of compounds containing a common core. A diverse set of 50,000 compounds was constructed to represent this underlying nature of the library. Approximately 27,000 compounds were selected by choosing several examples from each of the scaffolds and the rest of the set selected via a diversity selection of the nonscaffold compounds. In both cases, the selections were biased toward compounds that were still commercially available (to allow rapid confirmation of hits from repurchased solid material) and toward compounds of lower molecular weight. The latter criterion was introduced to bias the physicochemical properties of the set to be well behaved (solubility, permeability ²¹ ) and may also have a positive impact on hit rate due to the reduced complexity of lower-molecular-weight compounds. ²²

The diversity subset representing 3.8% of the library contains 56% of the simple rings, 14% of the Murcko assemblies, and 72% of the scaffolds ( Table 1 ). The screening hit rates for this set show strong agreement with the hit rates of the main library, comparable to the agreement seen for the cherry-picked validation set ( Fig. 3C , E ). The degeneracy of ring types among single-compound hits gives rise to a greater proportional recovery of Murcko assemblies or scaffold types in the diversity set compared with the main library, with this effect being more pronounced as the number of hits sharing the same ring type increases ( Fig. 4A , B ). Similar trends are seen for the two validation sets, although the reduced representation of ring types in these subsets dampens the effect. Interestingly, the ability to recover scaffold hits is greater than for Murcko hits for both the diversity and cherry-picked validation sets (compare Fig. 4A and B ), no doubt related to the greater coverage of scaffold types in these subsets ( Table 1 ). Because of the disproportionately low coverage of the supplier-defined scaffold compounds in the plate-based validation set, the recovery of these features from this subset is very poor unless larger hit clusters are considered. These significant variations likely result from the different sizes of the subsets in addition to the intentional consideration of the scaffolds in the 50,000 diversity set but not in the other subsets and the nonuniform distribution of scaffold compounds across plates.

Figure 4.

Scaffold recovery. Recoveries of specific assemblies in screen hit sets were calculated for the plate-based validation set (black), the cherry-picked validation set (dark gray), and the 50,000 diversity set (light gray). Recoveries of single compounds are included for comparison. (A) Murcko assemblies. (B) Scaffolds.

Design and Application of the In-Between Library

Early screens of the main library identified as hits a number of fragment-sized compounds (≤18 nonhydrogen atoms). Even if IC₅₀ values obtained for these compounds were relatively weak (~25 µM, the detection limit under standard screening concentrations), they are of high interest for lead development due to their high LE (0.35 for the example cited above). ²³ By extension, hits of this size, with even lower potencies, might be of interest from an LE perspective, but they must be assayed under conditions in which the weaker inhibition can be detected. To exploit this idea, we established the in-between library containing compounds that are larger than those compounds used for biophysical fragment screening but smaller than typical HTS compounds. We restricted the physicochemical properties of these compounds (primarily cLogP) to maximize the chance that they have good solubility and allow screening at a higher concentration than the main library. With the intent to plate the compounds at fivefold higher concentration than the main library and assuming this would allow detection of compounds with IC₅₀ values up to 125 µM, LE values for even the largest hits in this set should be ~0.25 or better. For challenging targets with low hit rates, such compounds would be of high value as potential leads to initiate medicinal chemistry programs. The compounds for inclusion in the in-between library were selected to contain a diverse set of Murcko ring assemblies, with overlap with the main library permitted, but novelty preferred. Of the 55,600 compounds in the in-between library, ~15,000 are also present in the main library.

The hit rates for the in-between library show a much poorer correlation with those of the main library than do the hit rates of the other sublibraries, as indicated by a substantially lower R² value ( Fig. 3D , E ). The in-between library was not intended to reflect the composition of the main library; hence, the differences in its properties and ring/scaffold composition likely contribute to the significant differences in hit rates relative to those of the main library ( Table 1 ). In most but not all cases, the hit rates are higher for the in-between library, and this is largely a function of its fivefold higher screening concentration. This is supported by the observation that, in most screens, the ~125,000 compounds in the main library with properties comparable to the in-between set have hit rates very similar to the entire main library (Suppl. Fig. S3).

Redundancy between the main library and the in-between library enabled the comparison of screening results at standard and fivefold elevated primary screening concentrations for specific compounds. Although high-concentration biochemical screening has been described in the literature,^{24 -26} we are not aware of any direct comparisons of screening results obtained at two different concentrations. Rather than restrict the evaluation to raw hit rates with a generic 50% inhibition cutoff, hit rates were evaluated with the more relevant inhibition cutoffs actually used in each screen (based on multiple factors including compound standard deviations and hit rates; see Suppl. Table S2). In addition, we tracked whether each confirmed hit (as defined by production of an IC₅₀ value in a dose-response evaluation in the primary assay with a frozen aliquot of the HTS compound stock) was identified in the low- and/or high-concentration primary screen. The analysis for the high-hit-rate target 18 screen ( Fig. 5A ) indicates that some confirmed hits were uniquely identified by the low- or high-concentration primary screen or were identified in both. The low-hit-rate target 26 screen lies in contrast to this: there were no confirmed hits uniquely defined from the low-concentration screen ( Fig. 5B ). Thus, for the 15,000 compounds in common between the main library and in-between set, the specific conditions used for the in-between screen, namely, fivefold higher concentration, were able to identify hits that would have been missed had the screen included only the main library.

Figure 5.

Comparison of results at two primary screen concentrations. Compounds that occur in both the main library and the in-between library were screened at two concentrations differing by fivefold as part of the full library screens. The primary screen and dose-response results for 25 targets were extracted in silico and analyzed for hit rates and numbers of confirmed hits. Data points represent the mean values of all replicates assayed at a given concentration. Not all compounds identified as primary screen hits were advanced into dose-response evaluations, and some compounds advanced into dose-response evaluations may not appear to meet the selection criteria because of variation among replicate values. (A, B) Scatterplots of inhibition at the two screening concentrations are shown for target 18 (A) and target 26 (B). The blue lines represent inhibition cutoffs used for the primary screens. Legend: Confirmed hits, compounds tested in dose response and producing IC₅₀ values (green); compounds tested in dose-response evaluations and not producing IC₅₀ values (red); compounds not tested in dose-response evaluations (gray). (C) Scatterplots compare the hit rates at the two screening concentrations for each target. (D) Numbers of confirmed hits identified in common at both primary screen concentrations (blue), only the low concentration (red), or only the high concentration (green) were tallied based on the inhibition cutoffs from the primary screens, and the results are displayed graphically. (E) Box plots depict the ligand efficiencies of confirmed hits from the in-between library (red) and the main library (green). Each line represents the mean (center point) ± 1 standard deviation for a given screen and library. Ligand efficiencies are displayed as discrete points for compounds more than 1 standard deviation from the mean.

When all of the screens are considered, hit rates increased 2.2-fold in moving from the low concentration to the high concentration ( Fig. 5C ; Suppl. Table S2). Although most screens showed increased hit rates at the high concentration, Targets 37, 30, and 38 had comparable hit rates at both concentrations, and target 23 had a significantly lower hit rate at the high concentration. In addition to the likely small contribution of normal assay variation, compounds with lower inhibition at the high concentration may also result from compound interference or precipitation; however, the property restriction of this subset is intended to limit the occurrence of these effects.

Not only were overall hit rates increased at the high concentration, but the number of confirmed hits also increased on average greater than 9.7-fold ( Fig. 5D ; Suppl. Table S2). Exceptions include targets 15, 23, and 36, which had higher numbers of confirmed hits from the low-concentration screen. It is important to note that not all positive compounds identified in the primary screens were evaluated in dose response, and this may lead to underrepresentation of confirmed hits from either library for some targets. In many screens, the increased number of confirmed hits from the high-concentration screen was substantial, for example, target 19 (24×), target 29 (25×), and target 31 (30×). Examining the bigger picture for the full library, the confirmed hits from the in-between library have higher mean and higher minimum LE values across all targets than do the confirmed hits from the main library ( Fig. 5E ). Nevertheless, significant numbers of highly ligand-efficient compounds were also identified in the main library for most targets.

Discussion

The Genentech library is an evolving collection of approximately 1.3 million compounds that includes scaffold-based compound sets. With the assumption that compounds around a common core will often have similar bioactivities, ²⁷ screening such sets of compounds has the advantage of providing hit information even if some compounds in the scaffold are lost due to experimental errors (false-negatives). In addition, this built-in structural redundancy gives confidence that resulting hits are authentic (not false-positives) through provision of some structure-activity relationship information directly from primary screens. These advantages are fundamental to our guiding philosophy to screen all compounds against all targets. In this article, we describe how we have used diversity-based library subsets to complement this “screen-all” approach, as our subsets have largely been used to supplement full library screens rather than to replace them.

Numerous methods have been described for assessing molecular similarity (e.g., see Bender et al. ²⁸ and references therein). Reduced topological representations (Murcko assemblies), ECFPs, and atomic property descriptors (BCUT) have all shown promise in the selection of diverse subsets that retain high levels of hit identification.^7–9 The four library subsets described in this article were each designed with a different approach to attaining diversity. The plate-based validation set was generated with 1536-well plates as the unit of selection, and selections were directed at maximizing the number of unique ring assemblies contained within a set of 10 plates. The cherry-picked validation set was approximately the same size but was assembled through random selection of individual compounds. The 50,000 diversity set was designed through the use of AAP matching to select diverse scaffold compounds coupled with ECFP-based selection of nonscaffold compounds. The AAP method is particularly attractive in this application as compounds with atom-type changes in the core ring structure are recognized as being similar, whereas other fingerprint or topological methods differentiate such molecular pairs. Lastly, the in-between library was built prospectively through the purchase of available compounds considering ring assemblies as the index of diversity.

Researchers at Novartis have successfully used plate-based subsets representing 20% of their compound collection to enhance their overall hit rates while reducing resources required for screening. ⁷ Similarly, the team at Pfizer was able to retrieve approximately half of their hit series from screens of one-eighth of their library, thus increasing the efficiency of their screening efforts. ⁸ The plate-based subsets generated at both of these organizations represented large numbers of plates covering significant proportions of their collection and were based on 384-well plates as the selection unit.

In contrast, the plate-based set described herein was developed from a small number of plates (10) each containing a large number of compounds (~1400). In addition, the chemical diversity within our library is not uniformly distributed across the plates, and although a wide range of ring assemblies are represented (2.7% of Murcko assemblies and 20% of simple ring assemblies), these assemblies do not represent the diversity actually present in the library. As a result, this plate-based validation subset is not a robust predictor of hit rates observed for the main library ( Fig. 3A , E ), and it has limited utility with respect to the recovery of ring assemblies representative of the screening hits ( Fig. 4A , B ). Nevertheless, our plate-based validation subset performed a useful function in validating assays for screening. The cherry-picked validation set that replaced this subset better represents the content and diversity of our library, and consequently, it is well suited to serve the additional functions of predicting screen hit rates and early identification of chemical matter.

Both validation sets have aided our preparations for high-throughput screens, enabling us to ensure that assays and screening systems are optimally configured for robust hit identification through provision of early hit rate estimates and compound inhibition profiles. As examples, an overly high hit rate in a pilot screen might be indicative of an assay that is subject to delivering false-positives, or a prevalence of compounds producing negative inhibition may indicate interference with the detection mechanism. Such data have triggered modification of assay conditions followed by reassessment of the validation set and, consequently, have improved the quality of our assays for a number of screens.

The in-between library has enabled the use of high-concentration screening to identify hits with lower potencies than would have been discovered under our standard screening conditions ( Fig. 5 ). The size of these compounds in this library and the use of high-concentration screening are reminiscent of the Needle screening approach, ²⁹ although target-based in silico selection of compounds for assay is not required when using the in-between library. Other examples of high-concentration biochemical screening have also been described, but detailed comparisons of screening at multiple concentrations were not included in these studies.^24–26 Primary screening at a range of compound concentrations (qHTS) has been conducted as a means of increasing screening robustness by facilitating discrimination of false-positives and false-negatives. ³⁰ In addition, the authors were able to identify inhibitors with potencies lower than discernable in single-concentration primary screens.

Increasing the ability to detect hits is particularly important for challenging targets with low hit rates, and increasing the screening concentration is one way to accomplish this ( Fig. 5B ). Although compound insolubility and interference limit the utility of this approach for the full library, restricting some of the properties of the in-between library intentionally created a compound set for which this approach proved successful. Screening of this library at high concentrations against a number of targets identified potential lead matter that may not have been discovered at typical screening concentrations ( Fig. 5D , green bars). Consequently, as an option for identifying low-potency, ligand-efficient compounds with a modest expenditure of resources, we advocate single-well, high-concentration primary screening of a property-restricted library set followed by dose-titration of positives. Hansson et al. ³¹ have noted that higher hit rates are positively correlated with cLogP. Interestingly, the in-between and diversity sets maintain hit rates higher than or comparable to the main library despite their lower average cLogP values (Suppl. Fig. S1). This suggests that hit rates with less lipophilic compounds can be maximized by screening at higher concentrations or ensuring a high scaffold diversity.

In favorable circumstances, screens of our library subsets in advance of full screens have generated chemical matter that has speeded project progression. For example, a pilot screen of the validation set produced a compound that later served as a control for an orthogonal assay used in triage of hits from the primary screen. In addition, two early screens of the in-between library yielded preliminary hit sets that jump-started our medicinal chemistry programs. Moreover, the chemical matter obtained from these early screens has also been used for testing and optimization of hit follow-up strategies put in place to develop hits into leads.

As a fallback strategy in the event that it is not possible to screen our full library, we generated a 50,000-compound diversity set. The retrospective analyses in Figure 4A and B indicate that the diversity set will have value in identifying a diverse range of lead matter in such cases. In addition, there is interest in using this subset for assessing the potential “ligandability” of new or less well-characterized targets.

Taken together, this work demonstrates the merit of building diverse library subsets using computational approaches coupled with screening data to address specific screening needs. It illustrates the importance of adapting compound selection methods to particular characteristics of the library—in our case, its prominent scaffold composition. Moreover, it underlines the value of conducting subset screens in addition to full library screens to enrich the content and quality of our lead discovery efforts. Lastly, we have shown that high-concentration screening of a property-restricted library subset is a productive approach to identifying ligand-efficient chemical matter that would have otherwise gone undiscovered.