Beyond Protein Binding: Recent Advances In Screening DNA-Encoded Libraries

Abstract

DNA-encoded library (DEL) screening has emerged as an important method for early stage drug and probe molecule discovery. The vast majority of screens using DELs have been relatively simple binding assays. The library is incubated with a target molecule, which is almost always a protein, and the DNAs that remain associated with the target after thorough washing are amplified and deep sequenced to reveal the chemical structures of the ligands they encode. Recently however, a number of different screening formats have been introduced that demand more than simple binding. These include a format that demands hits exhibit high selectivity for target vs. off-targets, a protocol to screen for enzyme inhibitors and another to identify organocatalysts in a DEL. These and other novel assay formats are reviewed in this article. We also consider some of the most significant remaining challenges in DEL assay development.

DNA-encoded library (DEL) technology has emerged as a powerful method for the discovery of protein ligands.^1–7 DELs are comprised of hundreds of thousands to billions of unique small molecule-DNA conjugates. DELs are created by split and pool synthesis⁸ where both a chemical step and the ligation of an encoding DNA fragment are carried out prior to pooling. The most common format for DEL creation is solution-phase synthesis, which is made possible by the fact that the DNA-small molecule conjugate can be precipitated whenever desired by addition of ethanol to an aqueous solution ³ or binding to a positively charged support.^9–12 While the nature of the small molecule-encoding tag linkage can vary substantially depending on how the DEL is created, the unifying features of all of these libraries created by solution-phase synthesis are that: 1) every molecule has an appended DNA tag (Fig. 1A)., 2) the amount of any particular molecule in the library is vanishingly small (such that in a binding assay the concentration is often fM or even aM), and 3) the molecules are present as an inseparable mixture in a single tube.

Fig. 1. Traditional and OBOC DELs.

A. Composition of traditional DELs in which every molecule has an encoding DNA attached to it. The structure reflects a library reported by Clark, et al.³. B. Composition of an OBOC DEL in which <1% of the molecules have an encoding DNA attached to it. The grey spheres represent 10 μm TentaGel beads. C. A typical screen for traditional DELs. D. Graphic representation a FACS-based screen of an OBOC DEL.

Though much less common, one can also create DELs by solid-phase synthesis, resulting in a one bead one compound (OBOC) DEL. ^{7, 13} These libraries are fundamentally different than those created by solution-phase synthesis in that all of the molecules are not connected to the encoding tag. In the original work of Gallop and co-workers,⁷ the DNA tag and the library molecule (a peptide in this case) were built off of different sites on the bead. In the more recent work of MacConnell, et al. ¹³ less than one percent of the displayed molecules are linked to a DNA tag (Fig. 1B), each bead carries a substantial amount of compound, (≈ 0.1 picomoles on a single 10 μm TentaGel bead), and it is possible to physically segregate the beads. As will be discussed below, the different characteristics of DELs created by solid-or solution-phase synthesis provide different opportunities and limitations in how they can be screened.

The vast majority of screens reported to date using DELs simply score binding to a protein target. For DELs created by solution-phase synthesis, the protocol is quite similar that developed originally for phage display peptide libraries. An immobilized target protein is incubated with the library. After washing, the encoding tags that remain associated with the immobilized protein are amplified and deep sequenced to reveal the predicted structures of the attached ligands (Fig. 1C).³ For OBOC DELs, the beads are mixed with a fluorescently labeled target protein in the presence of a large excess of unlabeled, diverse, competitor proteins. After washing, beads that retain a level of the fluorescent tag are separated from the remainder of the library using a fluorescence-activated cell sorter (FACS) and the encoding tags on these beads are amplified and deep sequenced (Fig. 1D). ^{7, 14, 15}

In both formats, screening noise (defined as the presence of a molecule in the hit pool that proves to be a poor ligand upon attempted validation) is a formidable problem and one must take great care to account for it.¹⁶ A detailed discussion of this topic is beyond the scope of this article. Briefly however, in screening traditional DELs, 2–3 rounds of binding/release, are generally carried out to increase enrichment of bona fide ligands prior to deep sequencing. Some effort has also been made to examine the possible advantages of cross-linking ligands to target proteins to increase enrichment.¹⁷ It is also helpful to include in the DNA tag a UMI (unique molecular identifier)¹⁸ that allows counting of each individual DNA that is isolated to correct for artifacts due to PCR bias. Finally, it is much easier to distinguish noise from true hits if the theoretical diversity of the library is significantly lower than the total number of reads in the deep sequencing step.^{19, 20} OBOC libraries, whether DNA-encoded or not, also have issues with high false positive rates. Many of these stem from the heterogeneity of the density of molecules displayed on different beads in the population.²¹ A small fraction are quite dense and can be thought of as a “molecular kelp forest”. Proteins that enter this forest can easily be trapped in microdomains of extremely high small molecule concentration even if that molecule is a terrible ligand. Several fixes to this issue have been explored.²² One relatively simple solution is to screen several copies of the library and distinguish compounds that score as hits on multiple independent beads (redundant hits) from so-called “singletons”.²¹ Since the kelp forest beads are relatively rare in the bead population, redundant hits are likely to be bona fide ligands whereas the singletons are almost always false positives. To implement this solution in the context of OBOC DELs, each bead must carry a specific barcode as part of the encoding tag.¹⁵

With the appropriate precautions, these simple binding screens are effective for the discovery of novel ligands for many different proteins. Nonetheless, there is interest in expanding the scope of the types of screens that can be done with DELs. A modest extension of the ligand discovery methodology would be to move beyond purified proteins and attempt to engage RNAs, multi-protein complexes in cell extracts, intact cells or other more complex targets. More adventuresome would be to develop screens that demand function, rather than simple binding, in order for a compound to score as a hit. In this article we review recent efforts to develop screens for DELs that demand more than binding to a single target protein. We also attempt to predict how these efforts will develop in the near future given current trends.

Protein-binding screens demanding high selectivity.

The ideal probe molecule is both potent and highly selective for the intended target.²³ Screens of traditional DELs created by solution-phase synthesis can, in theory, be manipulated to demand higher or lower potency (or in this case affinity) by manipulating the concentration of the protein target, though this is complicated by the fact that the target protein is not free in solution, but generally bound to a support (see refs. ^{24, 25} for a discussion). But it is less common to find reports of DEL screens that also manipulate conditions to demand a high level of binding selectivity so that resources are not wasted post-screening on the synthesis and attempted validation of promiscuous ligands.

There are two levels of selectivity that drug/probe molecule developers must consider. The first is the gross, proteome-wide selectivity of a ligand for the target protein over dissimilar proteins. Some organic molecules are truly promiscuous. These are often greasy, flat molecules, sometimes with electrophilic units, that engage many different proteins. These compounds are called PAINS (pan-assay interference compounds²⁶) in traditional high-throughput screening because they arise frequently as hits in various functional screens against many different targets. In a binding screen, these undesirable compounds can be avoided relatively easily by including a large excess of diverse, untagged proteins (for example an E. coli extract) in the screen to compete their binding to the tagged target protein.²⁷ Another level of selectivity has to do with how well a ligand distinguishes one member of a protein family from another. This has been a major problem in the development of inhibitors of kinases, histone deacetylases (HDACs), matrix metalloproteinases (MMPs) and a number of other enzymes that belong to large families. Strategies to address this level of selectivity are evolving. In most cases, a screen is conducted against a single family member, then the hits are re-synthesized and tested against others. Re-synthesis and validation are time-consuming and expensive endeavors, so it would be desirable to eliminate non-selective ligands prior to this step. An improved strategy in this regard is to carry out parallel screens against several different family members, then identify selective ligands informatically by comparing the hit pools. This has been effective in the identification of selective ligands for rat, bovine and human serum albumins²⁸ and compounds that can distinguish between the closely related DDR1 and DDR2 proteins.²⁹ Parallel screening has also been used to advantage in the development of selective drug leads against a variety of bacterial targets.³⁰

An alternative to parallel screening is to include both targets and off-targets in the same screen that incorporates a mechanism to collect ligands that bind the target but not the off-targets. Recently, a two-color screening strategy of this type was described for DELs created by solid-phase synthesis.¹⁵ This screen focused on the identification of small molecules that would bind to the antigen-binding sites of antibodies rich in the serum of patients with active tuberculosis (TB), but not to antibodies present in patients with latent TB (Fig. 2). Since no such antibodies were known, this screen was unusual in that the identity of the target was unknown. It was simply assumed that IgG antibodies were likely to exist that distinguish these two patient populations. For the purposes of this discussion however, we can consider the target to be antibodies unique to patients with active TB and the off-targets to be all other IgG antibodies. Therefore, to bind selectively to the target, a ligand would presumably have to engage the antigen-binding site of the active TB-specific antibodies. To accomplish this, all of the IgG antibodies in the serum of latent TB patients were colored green using an Alexafluor 488 (A488)-labeled anti-IgG Fab and all of the IgG antibodies in the serum of active TB patients were colored red using an Alexafluor 647 (A647)-labeled anti-IgG Fab. The two samples were then mixed together and incubated with about five copies of an OBOC DEL containing 500,000 unique compounds. After washing, beads that retained a high level of red, but not green, fluorescence on their surface were isolated using FACS (Fig. 2). The encoding tags on these beads were amplified and deep sequenced. Compounds found on multiple beads were then re-synthesized (again on-resin) and shown to indeed be able to distinguish serum samples derived from patients with latent or active TB. Finally, competition binding assays using the secretome and cell lysates from Mtb were employed to identify the native antigens recognized by these diagnostically interesting antibodies, revealing that they bind to a post-translationally modified forms of the secreted bacterial proteins Ag85A and Ag85B.

Fig. 2. A method to screen OBOC DELs for highly selective ligands.

The protein target(s) are labeled red and the off-target(s) green. The proteins are then mixed together and exposed to an OBOC DEL. The FACS is gated to collect beads that retain a high level of red, but a low level of green, protein. The encoding tags on these beads are deep sequenced. This method has been employed to identify ligands for antibodies that are abundant in the serum of patients with active, but not latent, tuberculosis (TB).

This screening platform, which took advantage of the ease of carrying out multi-color experiments on a FACS, could presumably be adapted to the simpler task of identifying ligands that distinguish between different members of a given enzyme family (kinases, HDACs, etc.). In this mode, the desired target and one or more off-targets would be given different tags so that the target could be labeled red and the off-target(s) green. The same screen could then be carried out to identify beads that display ligands that bind the target to the exclusion of the off-target(s).

A similar method could presumably be worked out for DELs derived from solution-phase synthesis. In this case, the differentially-tagged target and off-targets would both be present in the solution, then pulled down separately. The encoding DNAs associated with each precipitate would be amplified, deep sequenced and compared to identify compounds unique to the target-binding pool. It remains to be determined if this would yield result similar, or superior to, the parallel screening strategy discussed above.

We suspect that some type of selectivity filter, whether it be at the experimental or informatic level, will be more commonly applied to DEL screens in the future. It will be especially important when using focused libraries that are biased to have a known binding element for the target protein family of interest, such as an ATP-analogue when targeting kinases or an NAD+ mimic when targeting enzymes that utilize this co-factor.³¹

Targeting cell surface proteins in their native environment.

G protein coupled receptors (GPCRs), receptor tyrosine kinases (RTKs), ion channels and other integral membrane proteins are important classes of drug targets. Ideally, one would like to screen against these proteins in their native environment, which would necessitate screening against intact cells in a way that biases the results to report ligands against the receptor of interest.

This has been accomplished by a group at GSK using a traditional DEL containing 4.1 × 10⁷ unique compounds built around a triazine core scaffold (Fig. 3A).³² They targeted tachykinin receptor neurokinin-3 (NK3), a GPCR, which was expressed in HEK293 cells at high levels (≈ 500,000 receptors per cell) after infecting the cells with a BacMam (baculovirus gene transfer into mammalian cells) virus containing a NK3 receptor expression vector. The library was panned against immobilized, NK3-overexpressing cells three times to enrich binders. A parallel screen was done using HEK293 cells infected with a BacMam virus lacking an NK3 gene and compounds that bound to both cells were eliminated from consideration bioinformatically.

Fig. 3. Screening DELs against targets on the surface of living cells.

A. Schematic representation of a screen of a traditional DEL against cell overexpressing the NK3 receptor (blue ovals). B. A two-color screen to identify ligands to the VEGFR2 receptor (blue ovals). The red lines represent peptoids derived from a split and pool synthesis C. A phenotypic screen to identify peptide ligands that drive a cell death pathway via interaction with a cell surface receptor. Dead cells are shown in gray. The blue lines represent an anchoring ligand. The red lines represent peptides derived from a split and pool synthesis.

The results of the screen were impressive. The compounds that were most commonly identified in deep sequencing the hit pool proved to be antagonists of NK3 with IC50s in the low nM range. Moreover, even though the screen did not demand this, they proved to be more than 1,000-fold selective for NK3 over the related NK1 and NK2 receptors. This study clearly demonstrates the potential promise of using DELs for the discovery of ligands to integral membrane proteins.

To the best of our knowledge, this is the only such report in the literature to date. It will be interesting to determine the scope of this approach, especially with respect to the level of target protein expression required to achieve sufficient enrichment to distinguish bona fide ligands from noise. In the process of establishing their screening system, the authors used DNA-tagged chimeras of known, low nM NK3 ligands to establish the degree of enrichment, which proved to be about 15-fold per round of panning. Thus, the three rounds of panning provided an enrichment of about 3,000-fold relative to the level of the ligand in the naïve pool. This degree of enrichment would presumably be reduced if the target protein was present at a lower level and almost surely would be much lower for a ligand that binds less well. This may be a good thing, if the goal is to only find the very best ligands, but it also suggests that it may be problematic to identify leads for more difficult targets where a primary screening hit with low nM affinity may not be present in the library.

Cell-based screens have yet to be attempted with OBOC DELs. One might imagine that it would be easier to pick up lower affinity ligand-receptor interactions because of avidity effects between multiple bead-displayed ligands and integral membrane receptors (Fig. 3B). Indeed, there have been several reports of screening non-encoded OBOC libraries on larger beads against living cells.^33–39 One such study³⁷ employed a design similar to that employed in the NK3 screen discussed above, except that it employed a two-color strategy. HEK293 cells that did or did not express high levels of the Vascular Endothelial Growth Factor Receptor 2 (VEGFR2) were labeled with red or green dyes, respectively, and an OBOC library of peptoids was panned against a mixture of these cells. Beads that bound only red, not green, cells, were picked under a low power fluorescent microscope and the molecules displayed on the beads sequenced by tandem mass spectrometry. The best ligands identified in this were indeed only of low micromolar affinity when assayed as monomers for binding to the soluble VEGFR2 ectodomain. Note that this study is an example of several we will touch on in this review that describes studies done with non-encoded OBOC libraries. While this might be perceived to be outside the scope of this review, we suspect that assay formats similar to those developed for non-encoded OBOC libraries will soon be adapted to OBOC DELs, thus we have chosen to discuss them here.

Another noteworthy screen in this arena was an effort to identify ligands capable of distinguishing cancerous adenocarcinoma cells (HCC4017) from normal bronchial epithelial cells (HBEC30KT) derived from the same tissue of the same patient³⁹ using the same two-color screening strategy employed in the VEGFR2 study. A peptide-peptoid hybrid molecule was identified that was able to clearly distinguish these cell lines. Again, when its affinity for the adenocarcinoma line was measured in a format where avidity effects were not possible, the “K_D” was found to be 5 μM. Interestingly, an effort to identify the target of the ligand demonstrated that it was phosphatidylserine (PS), rather than a protein.⁴⁰ PS is known to be abundant in the membranes of cancer cells, but is present at much lower levels in the membranes of non-malignant cells.

Another non-encoded OBOC library screen against cells that merits mention sought to identify not simply ligands for a cell surface receptor, but compounds that would activate a cell death pathway in the bound cells (Fig. 3).³⁶ In this case, all of the beads displayed an adhesion ligand that allowed the cells to stick to all of the beads. In addition, each bead also displayed a hexameric peptide derived from a split and pool synthesis. After allowing the cell bead complexes to incubate for five days, they were stained with propidium iodide to identify the beads that had dead cells adhered to them. Two peptides were identified in this screen that, when added to lymphoid cancer cells, were able to induce cell death, though neither was very potent. This clever study is interesting in that it represents a type of phenotypic assay using an OBOC library. Cells can easily be engineered to report phenotypes other than death via a fluorescent readout, so this approach likely has broad applicability for identifying molecules that trigger interesting signaling pathways via interaction with a cell surface receptor.

Based on these reports, we predict that living cells will grow in importance as targets for DELs, both of the traditional and OBOC type. As mentioned above, it will be interesting to see whether this approach will be mostly limited to artificial situations in which the target receptor must be massively overexpressed, or whether further technical innovations and improved screening designs will be able to push this approach solidly into the realm where native cells can be employed as targets routinely.

Targeting nucleic acids.

One of the great leaps forward in biology over the last 2–3 decades has been the realization that RNAs play many important roles in cells beyond simply acting as a conduit for information to flow from DNA to proteins. Some have become appealing drug targets. To the best of our knowledge, DELs have yet to be deployed for the discovery of selective RNA ligands. Does this reflect a fundamental problem with the screening format? For example, one can imagine that some of the attributes that would make a molecule a good RNA ligand could also contribute to DNA binding. If so, there might be competition between those molecules binding to the RNA target and to DNA encoding tags. However, we speculate that this will not be a major obstacle and that DELs have a bright future in targeting RNA, though libraries may have to be designed with non-selective nucleic binding as a problem to be avoided, for example excluding the possibility of multiple positive charges, or interchelating moieties from the library.

In particular, it is now become clear that folded RNAs can be engaged by drug-like small molecules⁴¹ (whereas unfolded RNAs are probably best targeted by other nucleic acids). These molecules present ligandable pockets just as do folded proteins. There are, in general, two approaches one can take to targeting folded RNAs. One is somewhat akin to fragment-based drug discovery against proteins.⁴² Collections of small molecules immobilized on a surface or trapped within a gel were screened against a library of all possible sequence variants of a small RNA loop structure, each of which is labeled with a tag. Molecules that bound the tagged RNA were noted, then the captured nucleic acid was sequenced to determine the identity of the bound nucleic acid(s). Each RNA loop ligand was assigned a fitness score, which roughly corresponds to its selectivity for a particular loop over all other loops. With this fragment collection in hand, large RNAs containing several such structures can be targeted by linking together the binding fragments with an appropriate linker.⁴³ Many different RNAs have been targeted successfully using this general strategy, yet there remains a need for more such ligands since not all loops are currently targetable. It may also be useful to revisit even the loops that can already be liganded, since the chemical matter in hand may or may not be suitable for the development of effective drugs. One can imagine that DELs of low MW molecules could be extremely useful in this regard.

The second approach is to treat a large, folded RNA exactly as one would treat a protein in a standard DEL screen. While DELs have not been employed so far in this role, non-encoded OBOC libraries have been shown to be reasonably effective in this format. The earliest such screen of which we are aware was against trans-activation-responsive (TAR) RNA, which plays an important role in the life cycle of Human Immunodeficiency Virus (HIV-1).⁴⁴ An OBOC library of almost 40,000 peptoids was pre-blocked with bovine serum albumin and M1 RNA, then incubated with fluorescently labeled TAR RNA. Beads displaying ligands that retained the labeled RNA were collected manually and the structures of the peptoids determined by mass spectrometry following release from the bead. The screen yielded several sub-micromolar TAR RNA ligands, one of which was demonstrated to have modest cellular activity. A similar screen of a library of over 11,000 peptides yielded micromolar ligands with good selectivity for the CUG secondary structure motif found in the repeat structure of DM1 RNA, which is an interesting target in muscular dystrophy.⁴⁵ In this approach, the library was incubated with Cy3-labeled CUG repeat RNA, which enabled the manual selection of hit beads by fluorescence microscope.

More recently, an OBOC library of macrocyclic peptidomimetic structures called γ-AA peptides was screened successfully against GAS5, a long non-coding RNA (lncRNA) implicated in type two diabetes mellitus (T2DM).⁴⁶ Reduced GAS5 levels result in insulin resistance, so stabilization or upregulation of GAS5 could be of therapeutic benefit. GAS5 levels are mainly regulated by its degradation through nonsense-mediated decay (NMD), whereby the degradation factor UPF1 binds to a premature stop codon on GAS5 upstream of the poly(A) tail. With the goal of disrupting this RNA-protein interaction the screen employed the UPF1-interacting portion of GAS5 as the fluorescently-labeled target in the presence of a large excess of unlabeled, random sequence nucleic acid. Sub-micro molar GAS5 ligands were obtained and shown to have the ability to increase GAS5 levels in cells. The same type of library has also been screened successfully against pre-miR-155, demonstrating the generality of this approach.

Finally, it should be noted that some DNAs fold into unique three-dimensional structures other than the canonical double helix. An interesting example is the G quadraplex found in the promoter of the c-Myc gene,⁴⁷ which encodes a desirable target for cancer chemotherapeutics that has so far proven “undruggable”. Very recently, Litovchick, et al. reported a successful screen against this nucleic acid using 33 different DNA-encoded libraries as the starting point.⁴⁸ They identified several ligands with affinities in the mid-nM range and demonstrated activity in cells.

Taken together, these early studies in the application of DELs to the discovery of nucleic acid ligands suggest that this area of investigation will be highly fruitful.

Functional screens using DELs: Biochemical assays.

The work discussed above involves using DELs in binding screens against targets other than purified proteins. More adventurous would be to use DELs in screens that demand the molecule perform some sort of function to be scored as a hit. This is typical in traditional HTS where the molecule in each well is scored for its ability to agonize or antagonize an enzyme, turn on or off a gene or raise or lower the level of a protein in a cell, etc. Traditional DELs cannot be used in this way because all of the molecules are in a single tube with no way to segregate them. Moreover, the concentration of any one molecule in the mixture is well below the IC₅₀ that could be reasonably expected in a biochemical assay.

OBOC DELs, on the other hand, are fundamentally different (Fig. 1B). Each bead carries a significant amount of compound, the vast majority of which are not tethered to an encoding DNA. If the beads could somehow be spatially segregated, then each compound could be tested in an activity-based assay after release from the bead. This has been accomplished using a droplet-based microfluidics system.⁴⁹ An OBOC DEL comprised of 67,100 molecules, comprised of amino acids capped with a carboxylic acid, was constructed on 10 μm TentaGel beads with a photocleavable linker placed between the bead and the library compound (Fig. 4). A microfluidics device was employed to incorporate a bead, the enzyme autotaxin (ATX; a phosphodiesterase) and a fluorogenic ATX substrate into picoliter-sized droplets. The device shuttles the beads past a laser that releases a given amount of the bead-displayed compound through cleavage of the photolabile linker. After time in an incubation chamber, each droplet passes a sorting electrode that is activated if the fluorescence intensity in that droplet falls within a certain range. The encoding tags of the beads so collected are then amplified and deep sequenced to reveal the predicted structures of the hits. This novel screening platform identified several μM inhibitors of ATX.⁵⁰

Fig. 4. A microfluidic device that allows OBOC DELs to be screened for function.

A. Schematic of the architecture of the device. An enzyme (the phosphodiesterase autotaxin (ATX)), a fluorogenic substrate, an OBOC DEL and oil are combined to create tiny droplets, most of which contain a single bead. The beads are passed by a light source that cleaves a photolabile linker connecting the compound to the bead. After time in an incubation chamber, droplets are sorted depending on their level of fluorescence. B. Blow up of the two chemical processes that occur in the droplets.

This work represents a significant advance as the first function-demanding biochemical screen using a DEL. In its current state, this technology is capable of handling libraries of a few hundred thousand compounds, comparable to what would normally be done in a traditional HTS campaign, but at lower cost.

A fair question to ask is whether this approach is more efficient than first screening a much larger DEL using the standard binding screen and then resynthesizing and testing the putative ligands that arise from that screen for the desired activity. Time will tell. However, whether or not this technique supplants or significantly complements affinity-based selections for primary library screening, it seems likely that it will be extremely useful for secondary screens of focused libraries that more deeply explore “chemical space” around the structure of a primary screening hit.⁵¹ For this application, the capacity of the device to screen a few hundred thousand compounds is more than acceptable and the ability to obtain dose-response data (by varying the amount of photocleavage)⁴⁹ in solution is a significant advantage. Assuming the platform can eventually be made widely available, it should have a significant impact on the “hit to lead” phase of drug development.

Functional screens using DELs: Phenotypic assays.

While there is room for argument regarding the relative merits of biochemical assays as an alternative to affinity selection for DEL screening, there is no question that the ability to use DELs in cell-based phenotypic screens would represent a major advance. Phenotypic screens are fundamentally different in that they are target agnostic and simply score the ability of a compound to drive a particular process in a living cell or organism. The molecular target(s) of active compounds must be worked out later (often a considerable challenge). Traditional DELs cannot be used in phenotypic screens for many of the same reasons they cannot be employed for biochemical assays. Each molecule carries a DNA encoding tag, making it completely cell impermeable. Moreover, there is no way to segregate different molecules in the library and each is present at levels too low to drive a cellular response. So, the goal of screening hundreds of millions or even billions of molecules in a cell-based assay will be impossible for the foreseeable future. But again, OBOC DELs may be applicable to this problem, albeit for screening more modest numbers of compounds. The microfluidic approach depicted in Fig. 4, for example, could potentially be adapted to this purpose if cells were incorporated into the microdroplets and could function normally in this environment.

Alternatively, there is a long history of screening non-encoded OBOC libraries embedded in a matrix, such as agar, that limits diffusion of the compound once it is released from the bead.^{36, 52–56} In an illustrative application, beads and bacteria are mixed with top agar and the liquid matrix is poured onto Petri plates. Once the top agar solidifies, the compounds are released from the beads by chemical-or light-mediated of a linker, forming a “cloud” of compound around the bead to which it was bound originally bound. After some time, the bacteria will grow to form a lawn in the top agar, but beads that harbor a cytotoxic or cytostatic compound will have a clear zone around them. These beads are picked manually from the agar and the identity of the compound remaining on the bead (less than 100% is released in these assays) is determined. Of course, visual inspection and manual recovery of beads would waste many of the advantages of a DEL. Assay formats far more efficient than looking for clear zones around beads will have to be devised. But the main point is that there is clear precedence for phenotypic screening of OBOC libraries once the beads are physically dispersed, suggesting that some clever adaptation of this technology may allow OBOC DELs to be used for this purpose in the near future.

While we have avoided any discussion of the chemical content of DELs in this article (see ref. ⁵⁷ for a recent article from our laboratory focused on that issue), one such comment is justified here. As mentioned above, a major limitation of phenotypic screening approaches to probing biology is that it can often be difficult to identify the target(s) of a compound discovered in such a screen. Therefore, we would suggest that when DELs are adapted for this purpose, it would be wise to use libraries containing moieties that allow covalent engagement of the ligand with the target, for example an acrylamide group with a cysteine sulfhydryl. Impressive advances in chemical proteomics techniques have made it relatively routine to identify the proteins engaged by a covalent ligand in a proteome-wide fashion.⁵⁸ This would make the search for the physiologically relevant target of a bioactive compound much easier. While one might have some initial concerns about making DELs with electrophilic units, this is not likely to be a problem. ^{17, 59, 60}

Functional screens using DELs: Discovery of organocatalysts.

The discovery of ligands for biomolecules is not the only possible application of DELs. There is also interest in their use as a source of novel organocatalysts. One of the earliest examples of a synthetically useful organocatalytic reaction was the proline-catalyzed aldol addition of enolizable ketones to aldehydes (Fig. 5A). ^61–63 For some substrates, the aldol or Mannich products (Fig. 5A) are formed with excellent diastereoselectivity and enantioselectivity, though for others the stereoselectivity is far from perfect. A number of proline analogues have been examined over the years, providing superior results for certain substrate combinations. However, these advances were the result of testing newly synthesized catalysts one by one, a tedious process. Even now, there are many substrate combinations that provide mediocre results in organocatalytic aldol and Mannich reactions. Furthermore, the catalysts are not terribly efficient. Reactions take hours even with 5–10 mol% catalyst.

Fig. 5. Discovery of an organocatalysts from a combinatorial library.

A. The proline-catalyzed aldol (left) and Mannich (right) reactions, which are often used as model reactions for organocatalysts discovery. Ab. B. Proof of concept experiment in which a known diproline catalyst for the aldol reaction was amplified from a naïve library. This experiment conjugated the catalyst and the encoding DNA to a large PEG chain to improve the solubility of the molecule in organic solvents. C. Schematic depiction of the OBOC library screen conducted to identify small peptides with the ability to catalyze aldol reactions.

Since then, the field of organocatalysis has exploded, with catalysts now available for a wealth of different reactions.⁶⁴ However, it is still the case that catalyst discovery and improvement is usually done by synthesizing and testing candidate catalysts one at a time. It seems as if this should be an area of science where a combinatorial approach could be applied to great effect. A vast DEL of possible catalysts could be synthesized and screened using an assay that allows the enrichment of the most active (or most stereoselective) catalysts in a vast library of candidate molecules.

Building on previous work done with non-encoded OBOC libraries (discussed below), Hili and colleagues have demonstrated the feasibility of such a scheme.⁶⁵ They co-localized a known aldol catalyst (di-proline), a substrate ketone, and an encoding DNA in a single molecule that was attached to a long polyethylene glycol (PEG) chain. The PEG chain was included to improve the solubility of the DNA-containing conjugate in organic solvents (Fig. 5B). Without it, the DNA conjugates precipitate form organic solvents. The di-proline-containing molecule was then incubated in the presence of a biotinylated benzaldehyde substrate and the extent of the aldol reaction was monitored by following the conjugation of the biotin tag to the ketone-containing molecule by gel electrophoresis and staining with streptavidin. This demonstrated that the diproline was able to catalyze the addition of the soluble, biotinylated aldehyde to the PEGylated molecule. To simulate a library screen, the authors spiked the diproline-displaying molecule into a solution containing a 2000-fold excess of molecules outfitted with encoding DNA but lacking a tethered catalyst. After incubation with the soluble, biotinylated aldehyde and pulling down with streptavidin, a greater than 1200-fold enrichment of the encoding tag on the diproline-containing molecule was observed, suggesting that an active catalyst could be pulled out of a complex library using this approach. To the best of our knowledge, a novel organocatalysts has yet to be mined from a DEL, but this study provides clear proof of principle that it should be possible to do so.

This work was clearly inspired by experiments done using non-encoded OBOC libraries. Indeed, the design of the proof of principle aldol study described immediately above was similar to a system developed by Wennemers and colleagues for screening non-encoded OBOC libraries of peptides for novel organocatalysts (Fig. 5C).⁶⁶ They first developed the substrate-catalyst co-immobilization design and also employed a dye-tagged tagged, soluble benzaldehyde to visualize beads that harbor an active catalyst. Using this method, peptide catalysts for acylation reactions and aldol condensations have been discovered from OBOC libraries.⁶⁷ Finally, this work was preceded by pioneering efforts of Miller and colleagues that employed different formats to retrieve catalytically active peptides from complex libraries.⁶⁸

This wealth of precedent strongly suggests that it should be possible to use OBOC DELs for catalyst discovery, though this has yet to be reported. Combined with the single proof of principle of using a traditional DEL for the rediscovery of an aldol catalyst, we predict that organocatalyst developers will use DEL-driven approaches heavily in the future. In particular, once a catalyst with some activity is identified, it is likely that library-driven approaches will be particularly useful in accelerating the search for more active catalysts, since modest differences in the relative stereochemical orientation and electronic properties of the active functional groups in the molecule can have enormous effects on catalytic efficiency that are not easily predicted. It would be much better to screen hundreds of thousands or millions of catalyst candidates in a single experiment than to examine rationally designed derivatives of a primary catalyst one by one.

Summary and conclusions

DELs have already proven themselves to be an important source of protein ligands. In some ways, DELs represent a “best of both worlds” scenario with respect to the alternative approaches of traditional HTS and screening phage⁶⁹ or mRNA display⁷⁰ libraries against a target of interest. Like the latter technique, DELs allow huge numbers of molecules to be screened quickly and cheaply in a batch format. But like traditional HTS, DELs allow the use of libraries populated with compounds that display traditional drug-like properties. One is not limited to peptides as is the case for the biological display technologies. The utility of DELs will only increase as the scope of chemistry compatible with this platform is expanded and library designs improve.

As discussed here, innovations in assay design can also contribute to an increase in the utility of DELs. The incorporation of stringent selectivity filters in a primary screen will have a significant impact in the use of screening hits as probe molecules for interrogating biology, where off-target effects can greatly complicate the interpretation of “chemical genetics” experiments.⁷¹ Expansion of the types of targets addressable with DELs to RNAs, whole cells, etc. will also be useful. But we believe that the most impactful advances in this area will be the establishment of methods that allow DELs to be employed in phenotypic screening campaigns. This will be challenging, but can almost certainly be accomplished, particularly with OBOC DELs. As mentioned above, if combined with careful library design to facilitate more facile identification of the target(s) of the screening hits, a phenotypic screening platform suitable for DELs will have a major impact on discovery biology and drug development. Finally, we also believe that DELs have the potential to revolutionize the field of organocatalysis by accelerating catalyst discovery and optimization by orders of magnitude.