Target Identification of Compounds from a Cell Viability Phenotypic Screen Using a Bead/Lysate-Based Affinity Capture Platform

Abstract

The pharmaceutical industry has been continually challenged by dwindling target diversity. To obviate this trend, phenotypic screens have been adopted, complementing target-centric screening approaches. Phenotypic screens identify drug leads using clinically relevant and translatable mechanisms, remaining agnostic to targets. While target anonymity is advantageous early in the drug discovery process, it poses challenges to hit progression, including the development of backup series, retaining desired pharmacology during optimization, discovery of markers, and understanding mechanism-driven toxicity. Consequently, significant effort has been expended to elaborate the targets and mechanisms at work for promising screening hits. Affinity capture is commonly leveraged, where the compounds are linked to beads and targets are abstracted from cell homogenates. This technique has proven effective for identifying targets of kinase, PARP, and HDAC inhibitors, and examples of new targets have been reported. Herein, a three-pronged approach to target deconvolution by affinity capture is described, including the implementation of a uniqueness index that helps discriminate between bona fide targets and background. The effectiveness of this approach is demonstrated using characterized compounds that act on known and noncanonical target classes. The platform is subsequently applied to phenotypic screening hits, identifying candidate targets. The success rate of bead-based affinity capture is discussed.

Introduction

Drug development organizations are facing reductions in drug approvals accompanied by rising costs of research and development.^1–4 There are multiple culprits facilitating this trend, including poor clinical trial design and an incomplete understanding of target biology. A culprit that has gained significant attention is low target diversity in drug development pipelines.^5–7 This has led many drug developers to implement “nontraditional” techniques to augment their portfolios with fresh targets. One practice is phenotypic screening, which may identify drug candidates that modify disease pathology in unique ways.^8–10

Phenotypic screening has multiple advantages over target-based screens. When properly designed, phenotypic screens may produce hits that more often respond positively in animal models and the clinic since a live model was screened instead of an isolated recombinant protein. Additionally, remaining agnostic to the target liberates researchers from preconceived notions of “druggability,” permitting the relevant pathways to self-select. These advantages progress to liabilities when advancing the compound to later stages of the drug discovery process. If a general phenotypic endpoint is used, lead optimization campaigns become unreliable since structural alterations may result in divergent pharmacology while retaining phenotype. This convolutes marker identification, patient selection, and liability assessment. Consequently, the deconvolution of phenotype-driving targets is essential in progressing these drug candidates.^11–14

A suite of technologies may be leveraged to deconvolute targets, including hybrid screens,^15,16 display technologies,^17,18 and protein arrays. ¹⁹ A widely used technique is affinity capture (AC), which provides direct evidence of drug–target interaction. ²⁰ AC relies on the generation of affinity probes from the hit compounds using phenotypic structure–activity relationships. After the probes are incubated in lysates, targets bound to the probe are abstracted from the biological milieu and identified by mass spectrometry (MS). AC-based target identification has been successful in profiling kinases,^21–23 PARPs, ²⁴ and HDACs. ²⁵ Recently, examples have emerged where AC successfully identified targets of compounds from phenotypic screens.^26,27

Despite these successes, AC is hampered by nonspecifically associated proteins, which interfere with differentiation between bona fide targets and background.^28,29 To address this liability, we have developed a three-pronged strategy that combines target displacement, reproducibility, and uniqueness to increase the probability of identifying phenotype-driving targets. We further apply this technique to phenotypic screening hits, identifying new target candidates.

Materials and Methods

Materials

2-(2-(2-Aminoethoxy)ethoxy)ethanol was obtained from Anichem (North Brunswick, NJ). Di-tert-butyl dicarbonate was obtained from Chem-Impex International (Wood Dale, IL). Trifluoroacetic acid was obtained from Oakwood Products (West Columbia, SC). 2,2-Dimethyl-4-oxo-3,8,11-trioxa-5-azatridecan-13-oic acid was obtained from Ark Pharm (Libertyville, IL). Affi-Gel resin and Bradford assay were obtained from Bio-Rad (Hercules, CA). Extract Clean SPE reservoirs, caps, and stopcocks were obtained from Grace Davison (Deerfield, IL). Millex filters were obtained from Millipore (Darmstadt, Germany). Trypsin was obtained from Promega (Madison, WI). C18AQ magic reversed-phase resin was obtained from Michrom Bioresources (Auburn, CA). EDTA-free protease inhibitor cocktail was obtained from Roche Diagnostics (Indianapolis, IN). All other reagents were obtained from Sigma-Aldrich (Milwaukee, WI).

Preparation of the Linked Compounds

Synthetic procedures to obtain compound precursors are in the supplemental information. Briefly, the phenolic compound (1 equiv.), dissolved in anhydrous dimethylformamide, was added to 2,2-dimethyl-4-oxo-3,8,11-trioxa-5-azatridecan-13-yl methanesulfonate (1.25 equiv.) and potassium carbonate (3 equiv.). The reaction was stirred at 60–80 °C and monitored by mass spectrometry. The reaction was filtered and the isolated solids washed with 9:1 DMSO/water. Samples were dissolved in 9:1 DMSO/water and purified using a Gilson HPLC with a Waters DeltaPak C18 radial compression column (200 × 25 mm, 15 μm). The products were lyophilized to dryness.

Preparation of Affinity Resins

The linked compounds were dissolved in trifluoroacetic acid (>1 mL/100 mg) and dried under nitrogen. Affi-Gel 10 resin (12.5 mL, 0.188 mmol) was placed in a 25 mL Extract Clean reservoir and drained before washing 5 × 12.5 mL in isopropanol and 2 × 12.5 mL DMSO. The compound (6.25 μmol, 0.033 equiv.) was dissolved in 10 mL of DMSO containing diisopropylethylamine (0.073 g, 3 equiv.) and added to the resin, gently shaking overnight. Loading was confirmed by verifying loss of compound in the supernatant by mass spectrometry and/or high-performance liquid chromatography (HPLC). The resin was washed with 12.5 mL 1 M ethanolamine in DMSO, followed by 2 × 12.5 mL DMSO and 5 × 12.5 mL isopropanol.

Cell and Tissue Lysate Preparation

One tablet of EDTA-free protease inhibitor cocktail was dissolved in 50 mL of lysis buffer (50 mM HEPES [pH 7.5]/150 mM NaCl/0.5% Triton X-100/1 mM EDTA/1 mM EGTA) and chilled to 4 °C. The volume of lysis buffer added to the cells/tissue was four times its mass. The preparations were gently agitated at 4 °C for 30–60 min, then centrifuged (~14 °C) at 20,000g for 30 min. The pooled supernatants were filtered through a Millex 0.45 µm filter and clarified by ultracentrifugation (100,000g, 60 min, ~12 °C). The lysate was divided into 1 mL aliquots prior to freezing. Total protein concentration was determined using Bradford assay.

Affinity Capture–Mass Spectrometry

Sixty microliters of 50% resin slurry was diluted in lysis buffer, mixed, and centrifuged to pellet the resin and remove the solution. The wash was repeated three times before adding 3–5 mg of lysate in 1 mL of lysis buffer. The sample was mixed for 3 h (4 °C) before centrifuging and removing the supernatant. The resin was washed three times in lysis buffer and five times in 50 mM ammonium bicarbonate, pH 7.2. The resin was suspended in 100 μL of ammonium bicarbonate with 0.01 mg/mL trypsin and permitted to digest for 12 h (37 °C) before centrifugation and collection of the supernatant. When conducting competition reactions, 10 μM parent compound was added to the lysate for 12 h (4 °C) prior to treatment with resin.

LC-MS/MS Analysis

The samples were dried in a SpeedVac and reconstituted in 10 µL of water in 0.1% formic acid; 9.5 µL were injected onto a 15 cm capillary column (ID 75 µm) packed with C18AQ magic reversed-phase resin (5 µ particle size, 200 Å pore size). The peptides were eluted using a Waters NanoAcquity in a 35 min linear gradient (15%–35% acetonitrile in 0.1% formic acid) at 0.25 µL/min. The eluent was directed into a LTQ-Velos Orbitrap Pro mass spectrometer. Data-dependent scans were collected in the Orbitrap with tandem mass spectra collected in the ion trap.

Data Reduction and Analysis

LC-MS/MS data files were searched using Mascot Daemon v2.2 (Matrix Science) against the NCBInr mammalian database (August 7, 2012). Search parameters included 2⁺ to 3⁺ charge states, two missed cleavages, oxidized methionine variable modification, and mass errors of ±20 ppm for intact spectra and ±0.8 Da for tandem mass spectra. Search results were compiled in Scaffold v2.0 (Proteome Software) and exported to Microsoft Excel. Gene identifications were assigned using in-house software with ambiguous protein entries subjected to BLAST searching against the SWISS-PROT database. All keratin, trypsin, and immunoglobulin proteins were eliminated from the datasets as exogenously introduced contaminants.

Results

Identification of Candidate Compounds from a Cell Viability Phenotypic Screen

The compounds used in this study were derived from a phenotypic screen using the Cell Titer Glo assay against nine non-small-cell lung carcinoma cell lines. The screening results are in preparation for publication and a brief treatment of compound selection follows. The cell lines included Calu-1, Calu-6, NCI-H2009, NCI-H358, NCI-H441, NCI-H1975, NCI-H446, NCI-H522, and NCI-H661 and were chosen based on driver gene status (i.e., Kras mutation, c-Myc amplification). Compounds (~116,000) were screened, and those inducing ≥70% growth inhibition at 5 μM were shortlisted. EC₅₀ values were generated, and hits that reproduced were profiled in several in vitro panels (i.e., kinome). Compounds active in the in vitro panels were set aside in favor of hits driven by unknown targets. Compound analogues were phenotypically profiled to establish structure–activity relationships. Those that contained a site that tolerated chemical expansion with minimal phenotypic liability were transformed into affinity capture reagents. All affinity capture reagents were screened in the cellular panel to verify retention of activity.

An Affinity Capture Resin Matrix Defines the Background Profile in K562 Cell Lysate

When discriminating targets from background, it is useful to understand how often proteins are observed in the biological matrix of choice. This may be accomplished by profiling several affinity capture reagents and cataloging the observational frequency of the background proteins. These data may be compiled into a “uniqueness index,” where the enrichment propensity of proteins is expressed as a percent hit rate (%EPP). The construction of a uniqueness index constitutes the first prong of our approach.

Affinity capture resins typically consist of the compound linked to Affi-Gel 10 resin by an ethylene glycol linker. This linker configuration was found to be generalizable through the testing of many compounds (data not shown); however, background profiles may vary as the linker composition is altered. Nonreacted sections of the resin are capped with ethanolamide. While ethanolamide-capped Affi-Gel 10 resin will nonspecifically associate with protein on its own, its polar nature reduces interactions to nominal levels. The protein background observed in affinity capture experiments is largely governed by the physicochemical properties and linkage vector of the loaded compound in addition to the nature of its target profile. Therefore, to better approximate nonspecific interactions, five compounds of divergent chemotype identified from the cell viability phenotypic screen described above were made into affinity capture resins. Four of the compounds were coupled to resin using multiple linkage vectors, resulting in the production of 15 capture reagents. These reagents are referred to as 1a–1e, 2a–2c, 3a–3c, 4a–4c, and 5 in subsequent figures and supplemental information. No obvious target candidates were initially identified with these reagents when conducting affinity capture. The cellular activities of these compounds were later confirmed to be driven by mitotic arrest precipitated by disruption of microtubule formation (data not shown). Since polymerized microtubules are not commonly amenable to affinity capture methods, these compounds represent excellent tools to assess background since their primary targets will not be specifically enriched and mistaken as frequently observed background proteins.

Background binding was determined by mixing the resins in K562 cell lysate followed by resin isolation, washing, and proteolysis. Experiments were conducted in triplicate. After data compilation, proteins that were introduced contaminants (i.e., keratin) were removed, resulting in the identification of 414 proteins across all 15 resins.

Protein signal was measured in total spectral counts. Spectral counts are defined as the number of times peptides belonging to a protein were detected, isolated, and fragmented by the mass spectrometer. Since the number of observable peptides scales by molecular weight, larger proteins produce higher spectral count levels, rendering intra- and intersample comparisons between proteins of differing molecular weight unreliable. This phenomenon also produces a bias against lower-molecular-weight proteins since they proteolyze into fewer peptides. Since spectral counts also scale by protein abundance, counts may be reproducibly compared to the same protein across experiments within an error window. Examination of the spectral counts produced by the 414 proteins in the dataset revealed many proteins failing to display reproducibility between replicates. Since most of these proteins possessed low spectral counts and were predicted to be medium to high molecular weight, it is unlikely that the stochastic signal is due to a molecular weight signal bias. It is more likely that the lack of reproducibility is driven by the poor enrichment and/or low endogenous levels of the proteins. Consequently, a spectral count cutoff was implemented in order to focus on high-quality, reproducible hits, producing a more accurate uniqueness index. The cutoff was obtained by plotting the percent standard deviation of all proteins enriched from the 15 affinity capture resins against the average spectral count. The plot was funnel shaped (Suppl. Fig. S1a), where the error decreased as spectral count increased. The error plateaued at spectral count levels of 6 and greater, tapering as the counts increased (Suppl. Fig. S1b, Suppl. Table S1). Consequently, all proteins enriched with an average of 5 or less total spectral counts were removed from consideration, leaving 147 proteins as background candidates.

Examination of the dataset yielded several common background themes across the 15 affinity capture reagents (Suppl. Table S2), including some proteins that were previously recognized as contaminants in other bead-based affinity capture approaches.^28,29 The chaperonins containing TCP1 (CCT), flotillin (FLOT), heat shock (HSP), heterogeneous nuclear ribonucleoprotein (HNRNP), hydroxyacyl-CoA dehydrogenase (HADH), and ribosomal (RPS/L) protein families are enriched across most resins. Single proteins, such as actin gamma 1 (ACTG1), eukaryotic translation elongation factor 1 alpha 1 (EEF1A1), and Y-box binding protein 1 (YBX1), represent specific protein family members that are also commonly observed. These species are enriched independent of compound linkage vector, although levels of enrichment and specific members of large protein families vary. In order to prescribe a metric to the observational frequency of these proteins, the %EPP was calculated by dividing the number of times a protein was identified at a total spectral count of ≥6 by the number of capture reagents tested. Proteins that are commonly observed will possess a high %EPP, while those that are uniquely observed will possess a low or zero %EPP. The identities of the background proteins, their observational frequency, and affinity capture resin distribution are found in Supplemental Table S2.

Examination of the dataset also revealed proteins whose enrichment profiles suggest preferred compound and/or linkage vector associations. These proteins are shown in Figure 1 . Notable among these proteins are the NAD(P)H quinone dehydrogenases (NQO1/2). The NQO family of dehydrogenases have been previously identified as off-targets for a variety of compounds, including the antioxidant resveratrol ³⁰ and dasatinib. ³¹ It is possible that vector-specific enrichment of the NQO family, as well as the other proteins in Figure 1 , is indicative of actual binding events against promiscuous and lower-affinity off-targets. Further investigation of the binding mode for these proteins is required to verify their specificity.

Figure 1.

Commonly enriched proteins from K562, Calu-6, and human liver cell lysates that demonstrate compound and/or linkage vector binding specificity. Note that the specificity is largely conserved across cell types.

Background Profiles between Lysate Sources Are Similar

K562 cells are a myeloid leukemia cell line often used as an affinity capture medium since they can be bulked at scale and possess increased protein expression due to their oncogenic nature. Despite these logistical advantages, background profiles, when using other cell types, may not mirror the background observed with K562 cells. To explore this, the non-small-cell lung carcinoma cell line Calu-6 was subjected to affinity capture experiments identical to those described above. The list of protein identifications for the Calu-6 uniqueness index may be found in Supplemental Table S3. Additionally, homogenates of human liver were subjected to the same affinity capture workflow, and the results are presented in Supplemental Table S4. Comparison of the profiles in Supplemental Tables S2–S4 reveals that certain species are commonly observed across the cell types investigated. Ribosomal, heat shock, and additional structural proteins are most notable. The background profiles also display notable divergence ( Fig. 2 ). The chaperonin-containing TCP1 (CCT) protein family appears to be exclusively enriched from the K562 lysate background, while members of the myosin protein family are specific to the Calu-6 cell lysate. The resins enriched for alcohol and aldehyde dehydrogenases from the liver tissue background that were not identified in the other biological mediums. Further, the proteins in Figure 1 that displayed preferential binding to one or more affinity capture resins recapitulated their binding profiles in the other cell lysates, reinforcing the hypothesis that their binding is specific. These data suggest that application of a uniqueness index created from one cellular source against experiments conducted using a different cellular source will likely provide general trends that help discriminate binding specificity. However, given the aforementioned unique background aspects, it is best to re-create the index for each new medium utilized.

Figure 2.

Background proteins broadly associated across all 15 affinity capture resins that are unique to a single cellular source.

Implementation of the Affinity Capture Workflow Deconvolutes Targets across Protein Classes

While the uniqueness index represents the first prong of the target deconvolution strategy, additional metrics reinforce target hypotheses. Affinity capture experiments may be set up in a competitive format, adding parent compound to the cell lysate before the capture reagent to block enrichment (Suppl. Fig. S2). Comparing profiles obtained from affinity capture of a sample pretreated in parent compound versus an untreated sample, target candidates may be discerned by the modulation of their signal. The magnitude and reproducibility of target displacement constitute the second and third prongs of our target capture approach.

To examine the ability of displacement experiments to narrow datasets to the principal targets, three compounds of known target profile were examined: a broad-spectrum kinase inhibitor (6), a histone deacetylase inhibitor (7), and a Bcl-2 inhibitor (8) ( Fig. 3a ). Each compound was linked to Affi-Gel 10 resin based on established structure–activity relationships with their targets, and their syntheses are described in the supplemental information. The compounds possessed nanomolar potencies against their primary targets, which was not affected by linker conjugation (data not shown). Capture resins were incubated in K562 lysate that was pretreated with 0.1% DMSO or 10 μM parent compound for 12 h. The resins were exposed to the lysate for 3 h prior to their isolation, washing, and proteolysis with trypsin. Visualization of the captured proteins was accomplished by plotting the difference in displacement between the duplicates against the average magnitude of displacement ( Fig. 4 ). Proteins demonstrating >50% displacement upon parent compound pretreatment with >67% similarity in the displacement values are considered prime target candidates. The %EPP values previously calculated from the K562 uniqueness index were applied to the hits and assigned a color based on their magnitude (0% indicating a unique protein and 100% indicating a probable nonspecific binder). Additionally, the size of the marker was scaled as a function of total spectral counts.

Figure 3.

(A) Compounds of known target used to validate the affinity capture workflow. Compound 6 is a broad-spectrum kinase inhibitor, compound 7 is a histone deacetylase inhibitor, and compound 8 is a Bcl-2 and Bcl-XL inhibitor. (B) Compounds derived from a cell viability phenotypic screen of unknown target. Colored circles represent sites of linkage to affinity capture resins.

Figure 4.

Proteins enriched using compounds 6–8 from K562 cell lysate. The most relevant target candidates are highlighted in the dashed red box. Size of the markers is indicative of the number of total spectral counts, and color of the marker represents the %EPP, which was calculated from the K562 uniqueness index. (A) Data from affinity capture using compound 6. Sixty-six candidate targets were found in the region of interest ranging from 1 to 165 spectral counts. (B) Data from affinity capture using compound 7. Fifty-four candidate targets were found in the region of interest ranging from 1 to 133 spectral counts. (C) Data from affinity capture using compound 8. Thirty-six candidate targets were found in the region of interest ranging from 1 to 178 spectral counts.

For all affinity capture reagents, the principal targets were successfully identified as highly and reproducibly displaced by parent compound pretreatment, in addition to a low %EPP in the K562 uniqueness index (Suppl. Table S5). Compound 6 was chosen for an initial proof-of-concept study since kinases have been shown throughout the literature to be highly amenable to affinity capture approaches. Since kinase expression and activation may vary, a perfect correlation between in vitro kinase panel and affinity capture data is not expected. Nevertheless, the affinity capture reagent enriched for many protein kinases, including five members of the cyclin-dependent kinase family, for which the compound was originally designed ( Fig. 4a ). Further, the bulk of the nonkinase proteins that demonstrated uniqueness, as well as high and reproducible competition, are known binding partners to kinases, such as the mediator complex and various cyclins. Compound 7 is a known histone deacetylase inhibitor and chosen for the proof-of-concept study for the same reason as compound 6. The principal targets were, again, easily identified in addition to many binding partners known to form larger epigenetic complexes ( Fig. 4b ). Compound 8 is a Bcl-2 and Bcl-XL inhibitor that was chosen to challenge the ability of the affinity capture workflow to successfully identify targets. The lipophilic nature of this compound lends itself to the nonspecific enrichment of a large number of proteins, as shown throughout the plot in Figure 4c . Despite this obstacle, examination of the %EPP from the K562 uniqueness index, in conjunction with the analysis of the competitive displacement characteristics, allowed the principal targets to be identified alongside a smaller complement of background proteins. It should be noted that many of the proteins that failed to be adequately and/or reproducibly displaced by parent compound pretreatment for compounds 6 and 7 generally also possessed a higher %EPP, suggesting they are nonspecific associations. For compound 8, a large number of the nonspecifically enriched proteins that both passed and failed the competitive displacement criteria were found to be highly unique. It is possible that the uniqueness indexes need to be expanded to include more lipophilic compounds to provide a better prediction of targets for these cases.

Identification of Unique Targets from Phenotypic Screening Hits

Given the successful proof-of-concept studies, the method was applied to two compounds derived from the aforementioned cell viability phenotypic screen ( Fig. 3b ). The sites of linkage were chosen by screening close analogues of the compounds in the cell viability assays and identifying modifications that failed to significantly alter activity. Upon appending a linker at these sites, the probes were rescreened in the cellular assays and shown to retain their activity prior to affinity capture resin generation (Suppl. Table S6). The results from the affinity capture experiments are shown graphically in Figure 5 , and the target candidates are described in Table 1 . In both cases, a small number of target hypotheses met the displacement, reproducibility, and uniqueness criteria. In particular, the iron binding nuclear protein pirin (PIR) enriched by compound 9 and the serine hydrolase prolyl-endopeptidase L (PREPL) enriched by compound 10 are presently under further investigation as target candidates.

Figure 5.

Proteins enriched using compounds 9–10 from K562 cell lysate. The most relevant target candidates are highlighted in the dashed red box. Size of the markers is indicative of the number of total spectral counts, and color of the marker represents the %EPP, which was calculated from the K562 uniqueness index. (A) Data from affinity capture using compound 9. Four candidate targets were found in the region of interest. (B) Data from affinity capture using compound 10. Four candidate targets were found in the region of interest.

Table 1.

Target Candidates for Hit Compounds from the Cell Viability Screen.

	Gene Name	Displacement Reproducibility (%)	Displacement Magnitude (%)	Spectral Count	%EPP
Compound 9	ADK	100	100	2	0
	CMBL	100	100	4	0
	EFTUD2	100	100	2.5	0
	HSP90AA1	100	50	4	100
	LDHA	100	100	2	0
	PIR	93.6	68.2	41	0
Compound 10	FLAD1	71.3	54.4	13	0
	GAK	68.6	55.7	6	0
	HRNR	66.7	50	3	13.3
	PREPL	100	100	7.5	0
	TUBA1A	100	100	4	0

The %EPP is the observational frequency of the protein from the background uniqueness index. Larger %EPP values suggest the protein may be nonspecifically associated with the capture reagent.

Discussion

Phenotypic screening has become increasingly commonplace in modern drug discovery workflows, with the hope of impinging on unique therapeutic targets and mechanisms. The key bottleneck to hit progression in phenotypic screening is an adequate understanding of the molecular targets underpinning the desired phenotype. Affinity capture technologies have historically been relied upon to mitigate the problem; however, capture techniques suffer from several limitations. Before affinity capture experiments may be considered, phenotypic screening hits must be vetted to exclude compounds that work by generalized mechanisms such as metal ion chelation, membrane perforation, and DNA alkylation. Since these mechanisms do not involve the engagement of a target in the traditional sense, affinity capture experiments will fail at the expense of significant resource investment. In addition to vetting the mechanism, the compounds themselves require careful examination. A key consideration is if they display phenotypic structure–activity relationships, suggesting they can be modified into capture reagents without significant loss of activity. A further consideration is whether the compound acts as a prodrug, being cellularly metabolized into an active component. In these cases, affinity capture experiments conducted on beads in lysates may fail to produce the same transformation and, consequently, will not bind the target. Finally, not all target classes and structures are inherently amenable to affinity capture experiments. In particular, multiple-pass integral plasma membrane proteins such as ion channels and G protein–coupled receptors have proven difficult to enrich owing to their propensity to aggregate and precipitate upon cell lysis.

Despite the litany of challenges associated with affinity capture, a large number of targets and compounds remain amenable to the process provided due diligence is observed. In these cases, the key challenge to target identification resides in the discrimination of bona fide targets from background upon collection of the dataset. In some cases, such analyses are simple, requiring perhaps a single displacement experiment (Suppl. Fig. S2) or a comparative analysis between two capture reagents of similar structure but divergent function. In most cases, however, the profiles become too complex to easily interpret and a more holistic approach is required. This is especially true when considering compounds that bind to targets such as those listed in Figure 1 . When conducting affinity capture experiments in isolation, it is tempting to consider these proteins as targets that drive phenotype. It is only upon expanding the scope of the experiments and determining what is commonly enriched across chemotype that it becomes evident that these proteins are promiscuous binders. Conversely, the binding of these proteins should not be written off as false positives either. Several of the proteins listed in Figure 1 possess broad roles in managing cellular stress and detoxification. If the compounds of interest are inhibiting the functions of these elements, liabilities may manifest in later stages of drug development as idiosyncratic toxicities.

Examining a large scope of affinity capture data for protein uniqueness is highly useful, but not without fault. As shown in Figure 4c , despite extensive analysis of the background commonly observed in K562 lysate, compound 8 enriched a large number of proteins considered unique that are not the targets or binding partners to the targets of interest. This is likely due to the lipophilic nature of the compound attached to the resin. While it may be argued that augmenting the uniqueness indexes with compounds containing higher lipophilicity would solve this issue, it does not take into account the propensity of associated proteins attracting additional nonspecifically associated proteins to create a background corona that may not be easily predicted or measured. It is for this reason that uniqueness is combined with the magnitude and fidelity of target displacement to obtain a more relevant picture of binding specificity. This workflow was successfully applied to two compounds of unknown target derived from a cell viability phenotypic screen ( Fig. 3b ). In both cases, the candidate lists could be distilled to a small subset of targets that are presently being explored for validation ( Fig. 5 , Table 1 ). However, these experiments represent only a snapshot of the data derived from a phenotypic screen and beg the question of how successful bead/lysate-based affinity capture is on a larger scale. As shown in Supplemental Figure S3, the screen was divided into four sections, including a pilot screen and three phases of the primary screen. When considering the pilot screen and phases 1 and 2 of the primary screen, a total of 23 compounds were advanced to probe design, having demonstrated no obvious target when profiled against in vitro panels and for generalized mechanisms of action. These compounds also possessed clear attachment points where modification did not significantly affect the activity profile. From these compounds, 41 affinity capture reagents were generated, taking advantage of multiple linkage vectors and changes in linker composition to improve the probability of success. To date, eight of these compounds have produced target hypotheses that have entered validation studies. Consequently, from the perspective of providing a unique target for validation efforts, the technique has a 35% success rate on the basis of these data. It should be noted, however, that some of these target hypotheses are anticipated to fail in the validation process, which will inevitably drive the success rate lower. When considering this high level of attrition, several factors emerge as culprits, such as the relevance of the cellular source and potential changes in target binding precipitated by an alteration in the drug’s presentation upon resin attachment. More likely is the fact that many proteins rely on a concentrated, sequestered cellular environment containing auxiliary binding partners to retain their native fold and compound engagement properties. The dilution from cell lysis inevitably causes the dissociation of many stabilizing protein complexes, exposure of the proteins to deactivating factors, or outright denaturation, rendering lysate-based affinity capture unproductive. To adequately address the increasing demand for target and mechanism identification technologies arising from phenotypic screening efforts, affinity capture platforms will need to be revised to incorporate bio-orthogonal techniques and photochemical cross-linking to facilitate the engagement of the targets in situ, followed by their stable extraction for identification purposes.