Extending the Detection Limit in Fragment Screening of Proteins using Reverse Micelle Encapsulation

Abstract

Detection of very weak (K_d > 10 mM) interactions of proteins with small molecules has been elusive. This is particularly important for fragment-based drug discovery where it is suspected that the majority of potentially useful fragments will be invisible to current screening methodologies. We describe an NMR approach that permits detection of protein-fragment interactions in the very low affinity range and extends the current detection limit of ~10 mM up to ~200 mM and beyond. Reverse micelle encapsulation is leveraged to effectively reach very high fragment and protein concentrations, a principle that is validated by binding model fragments to E. coli dihydrofolate reductase. The method is illustrated by target-detected screening of a small polar fragment library against interleukin-1β, which lacks a known ligand-binding pocket. Evaluation of binding by titration and structural context allows for validation of observed hits using rigorous structural and statistical criteria. The 21 curated hit molecules represent a remarkable hit rate of nearly 10% of the library. Analysis shows that fragment binding involves residues comprising two-thirds of the protein’s surface. Current fragment screening methods rely on detection of relatively tight binding to ligand binding pockets. The method presented here illustrates a potential to faithfully discover starting points for development of small molecules that bind to a desired region of the protein, even if the targeted region is defined by a relatively flat surface.

Molecular recognition by proteins is a central feature of biology. How the various contributions to the thermodynamics governing protein-ligand interactions combine to result in specific high affinity interactions remains to be fully understood. In a more pragmatic arena, there remains an on-going effort to develop the capacity to robustly discover and develop new small molecule inhibitors and pharmaceuticals directed at proteins. Fragment based drug discovery (FBDD) is emerging as a more rational complement to empirical screening of small molecule libraries against protein targets.¹

The attraction of FBDD is the potential to limit the size of the screening library, to optimize the chemistry of the starting hit for the subsequent hit-to-lead medicinal chemistry, and to explore chemical space that is otherwise not represented in high-throughput screening libraries. However, the affinity of very small molecules as exemplified by the rule-of-three (Ro3) fragment libraries, for example, and proteins is anticipated to be quite weak,^2,3 even though such molecules may have maximal affinity.⁴ A major motivation for developing very weak binding molecules into stronger binding inhibitors lies in the principles of superadditivity of fragment binding. Reduction of the rotational and translational entropic penalty upon linking fragments together enhances binding well beyond what would be expected from strictly adding the binding free energy of each fragment.⁵ In practice, optimization of fragment linking is a significant challenge, but has found some success.⁶ The power of superadditivity derived from linking suggests that fragments initially binding as weak as the high-mM range may be useful for inhibitor and drug design.⁷ Despite their utility, fragments that bind in this affinity regime are generally out of reach of current methods of detection.

There have been efforts to extend the small molecule binding detection limit using computational methods⁸ crystallography of native^9,10 or heavily cross-linked protein,¹¹ tethering,¹² protein-templated fragment reactions,¹³ and target observed NMR.¹⁴ While some studies have occasionally reached beyond 10 mM affinity, no current technique provides reliable detection and robust determination of binding affinities weaker than K_d ~10 mM.¹⁵ It is important to note that fragments binding at the very edge of the detection limit have nevertheless been advanced to higher affinity inhibitors and drugs.^16–18 And taking the opposite approach, many known high-affinity binders have been broken down to vanishingly weak binding fragments.^19–22 These studies illuminate the potential of extending the fragment detection limit to greatly enhance pools of building blocks for inhibitor design. In this study we have developed a robust and universal method to extend the current detection limit of ~10 mM for fragments binding to proteins to ~200 mM and beyond.

RESULTS AND DISCUSSION

Theory

Here we use a simple idea to enter into the very weak binding regime where dissociation constants on the order of 10s and 100s of mM can be quantitatively evaluated. The key is the encapsulation of single protein molecules within the protective water core of self-assembling reverse micelle (RM) particles prepared in a liquid alkane solvent (Figure 1A).²³ The water core of the RM constitutes approximately 2% of the total sample volume. By simply concentrating the material that is normally used in a solution NMR screen into the RM water core, one can cost-effectively access the ligand concentrations necessary to characterize binding in the K_d 10s and 100s of mM regime. Importantly, the debilitating artefacts often associated with protein NMR spectroscopy of bulk aqueous samples containing large concentrations of small molecules are eliminated.^24,25

Figure 1.

Enhanced binding due to increased effective concentration arising from confinement. (A) Schematic of a protein encapsulated within a RM. The protein is depicted in orange, water in blue, surfactants in green, and the alkane solvent in gray. Approximate dimensions of the water core and the protein are given. (B) Illustration of the enhanced effective concentration due to confinement of protein and fragment within the water core of the RM. Colors as in part A with fragment in yellow. (C) Theoretical ligand binding curves in standard bulk aqueous conditions (solid red line) and confined within the RM (solid blue line). Displayed is the fraction of protein bound (f_PL) as a function of binding dissociation constant (K_d). The same amount of protein and ligand was used in both cases and corresponds to a bulk aqueous sample (0.5 mL) of 100 μM protein and 800 μM of fragment. Confinement to the aqueous core of the RM results in a 50-fold increase in effective concentration. Experimentally determined dissociation constants of AMP, ADP and NADP binding to ecDHFR are shown for bulk aqueous conditions (red squares) and in the RM (blue circles). Error was determined through fitting and error bars are smaller than the symbols. Dashed lines indicate the smallest CSP_sat at which a ligand is detectible for a given K_d for the aqueous sample (dashed red line) and the corresponding RM sample (dashed blue line). This is based on a minimal observable normalized CSP of 0.005 ppm as determined from LMS analysis of the pilot IL-1β screen and as detailed in the Methods section.

Under the conditions used here, reverse micelles are comprised of a shell of surfactants with their polar head groups pointed inwards towards a water core and their hydrophobic tails pointed outwards into a bulk alkane solvent (Figure 1A). Specific sample conditions (surfactant, water, alkane solvent) produce a reverse micelle population that is remarkably uniform in its physical attributes.²⁶ Determination of encapsulation conditions that preserve the structural integrity of the protein is generally straightforward and involves choosing an appropriate RM surfactant system and systematic optimization of co-surfactant, water and buffer content.²⁷ Importantly, at room temperature reverse micelles exchange their contents on times scales much faster 10³ per second and are effectively in equilibrium.^28–30

Using absolute amounts that would correspond to 0.8 mM of ligand in a standard bulk aqueous NMR sample, effective concentrations of ligand at 40 mM can be directly accessed due to the exclusion of ligand from the alkane phase (Figure 1B). Similarly, the amount of protein required for a 100 μM sample results in an effective concentration of 5 mM when confined to the RM water core. NMR target-based fragment screening commonly enables a minimum detectable fraction of protein-ligand complex (f_PL) of ~0.1 to 0.2.³¹ The enhanced ligand concentration within RMs suggests nearly two orders of magnitude higher K_d values would be detected when using the RM encapsulation strategy with equivalent material as used in a typical bulk aqueous NMR sample (Figure 1C). It is noted, however, that the detection limit depends on the magnitude of the NMR chemical shift perturbations brought about by binding. The chemical shift perturbation at saturation (CSP_sat) varies among fragments. Fragment binding is often characterized by very low CSP_sat values.³² Because of the enhanced concentration, the minimum CSP_sat required for hit detection is reduced in the mM range under RM conditions (Figure 1C) and provides greatly enhanced sensitivity to small chemical shift changes.

Experimental validation

To confirm the enhancement of fractional saturation by reverse micelle encapsulation we examined E. coli dihydrofolate reductase (ecDHFR) as a model system. Though a number of suitable surfactant systems are available, the 10MAG/LDAO/DTAB system²⁷ is the most broadly capable³³ and was used for the current study. Following appropriate optimization, RM encapsulation of ecDHFR gives high-quality NMR spectra (Figure 2). ADP (K_d = 17 ± 1 mM) and AMP (K_d = 26 ± 3 mM) represent weak-binding fragments of NADP⁺ (K_d = 9 ± 2 μM), which is the natural cofactor of ecDHFR.³⁴

Figure 2.

¹⁵N-HSQC of ecDHFR encapsulated in reverse micelles. The reverse micelles are composed of 52.5 mM 1-decanoyl-rac-glycerol (10MAG), 18 mM N,N-dimethyldodecylamine N-oxide (LDAO), 4.5 mM dodecyltrimethylammonium bromide (DTAB), with 10 mM hexanol in pentane at pH = 8.5. 10% D-pentane is present for lock signal, otherwise no other components are deuterated. 100 μM of ecDHFR is present in the total sample volume, 5 mM when considering the water phase. Folate is present at 4:1 folate to protein. Spectrum was collected at 25°C and 750 MHz with 32 scans per FID.

We determined the degree of saturation of the protein when employing the same absolute amounts of protein and ligand as in the bulk experiment (Figure S1). Titration data were of high quality and allowed accurate determination of binding affinities (Figure S1). The results confirm the prediction (Figure 1C). Importantly, the K_d values observed in the RM are within experimental error of the bulk solution values. The more strongly binding NADP⁺ saturated the enzyme in both bulk and RM samples (Figure 1C). The enhanced concentration imposed by RM encapsulation enables both the detection of weaker interactions (higher K_d) and interactions defined by a smaller spectral change (lower CSP_sat).

Pilot screen of interleukin-1β (IL-1β)

To examine the general potential of the RM approach to identify weakly binding small molecule fragments, we undertook a fragment screen of interleukin-1β (IL-1β). IL-1β is a 153 residue monomeric protein that is centrally involved in the inflammatory response and signals through a ternary protein-protein complex (Figure 3A).³⁵ IL-1β is a relatively featureless protein that contains no ligand binding pockets and binds no known small-molecule ligands. Targeting IL-1β for inhibition will involve interference at its protein-protein binding interfaces (Figure 3B).^35,36 As a prerequisite to RM screening the protein must first be satisfactorily encapsulated. Sample variables such as the water to surfactant molar ratio (water loading, W₀), co-surfactant hexanol concentration, and pH were adjusted to obtain spectra of the encapsulated protein that were nearly identical to those obtained in bulk solution and result in high-quality NMR spectra (Figure 3C).^25,27 It is important to monitor pH as the high concentrations of added ligand have the potential to significantly alter the pH. Internal protein markers for pH have been established for IL-1β.²⁷ Importantly, the LDAO surfactant provides a very high buffering capacity through its zwitterionic head group,³⁷ with an effective concentration of 1.5 M under the conditions used here.

Figure 3.

(A) Ternary signaling complex of IL-1β (gray) with its receptor binding partner (IL1-RI, cyan) and the receptor accessory protein (IL-1RAcP, magenta). (B) Surfaces of IL-1β (gray) that are in contact with IL-1RI (cyan) or IL-1RAcP (magenta). Residues that were recently suggested by mutagenesis to be essential in the interaction of IL-1β with integrin are depicted in orange. (C) ¹⁵N-TROSY of IL-1β encapsulated in reverse micelles. The reverse micelles are composed of 48.75 mM 10MAG and 26.25 mM LDAO with 20 mM hexanol in pentane at pH = 5.0. 10% D-pentane is present for lock signal, otherwise no other components are deuterated. 100 μM of IL-1β is present in the total sample volume, 5 mM when considering just the water phase. Spectrum was collected at 25°C and 750 MHz with 16 scans per FID.

A protocol for preparation of RM samples for fragment screening was developed (Figure S2). Because of partitioning effects between the water pool of the RM and the bulk alkane phase, only polar molecules are suitable. Calculations suggest that partitioning into the bulk alkane solvent is tolerable for molecules having partition coefficients (logP) less than −0.25 (Figure S3). It should be noted that alkane-water partition coefficients are generally more negative than the more commonly available octanol-water counterparts (Figure S4).^38,39 This is advantageous as it broadens the pool of molecules that can be successfully employed in the RM approach. Here we used highly water-soluble fragments with clogP(octanol/water) values less than 0.5. These criteria resulted in the selection of 233 molecules from the Maybridge Ro3 2500 library. The general screening strategy then followed a standard approach⁴⁰ where mixtures of members of the fragment library were screened using ¹⁵N-TROSY spectra. No visual or spectroscopic signatures of aggregation or disruption of reverse micelle integrity were observed. ¹H spectra were collected and compared to those under aqueous conditions to ensure that effects such as ligand aggregation, interactions with the RM, or partitioning primarily to alkane phase were absent. Further details are provided in the Methods section.

The initial screen was comprised of 54 mixtures and resulted in 23 mixtures containing hits as identified by a least median squares statistical method.^41,42 The mixtures were subsequently deconvoluted and 31 individual molecules were found to interact with the protein (Table S1). The library of 233 polar fragment molecules screened under standard bulk aqueous conditions (100 μM Il-1β and 800 μM fragment) yielded no hits.

Fragment hit characterization

The 31 fragment hits determined from the deconvolutions were further characterized by direct titration of the RM encapsulated protein. These fragments showed fitted dissociation constants, based on individual residue chemical shift perturbations (CSPs), ranging from 50 mM to over 1 M (Table S2). To demonstrate that the observed binding at elevated concentration in the RM is not due to a confounding effect, spectra were collected of IL-1β in a limited but bulk aqueous volume (10 μL) in a 0.6 mm i.d. capillary at the same effective concentrations used in the final titration point for one fragment hit (i.e. 5 mM IL-1β and 320 mM fragment). Chemical shift perturbations were essentially identical though the narrow diameter of the capillary but resulted in a lower quality spectrum (Figure S5).

Hit fragments caused chemical shift perturbations at several residues. A two-step process was used to define specific binding sites. Spatial clusters of amide nitrogens of residues identified by CSPs were identified using a k-means algorithm.^43,44 The residues thereby tentatively identified as responding to the same binding event were then fitted to a common K_d using stringent statistical criteria for goodness of fit (see Methods section). Examples of clustering are shown in Figure S6. Residues that showed binding in isolation, without participation of additional nearby residues, were concluded to engage in ill-defined interactions and were excluded from further analysis (see Table S2). Of the 31 original fragment hits, three did not have any residues that displayed well-fitting titration curves and seven did not contain any residues that fell within the stringent definition of a specific binding site. Twenty-one of the original 31 fragment hits were found to have well-defined specific binding to at least one site on the surface of the protein. Each discrete binding site and the associated K_d values for all titrated fragments are documented in Table S3. Importantly, all fragments but one identified in this way had multiple discrete binding sites, indicating the particularly primitive interaction chemistry anticipated for weak fragments.

It is important to note, that of the remaining 21 fragments that engaged in specific binding interactions, 13 also showed at least one non-specific interaction. Interestingly, residues that engage in non-specific interactions with one fragment are also involved in specific binding with other fragments, precluding the possibility that global effects are involved (i.e. pH, conformational change, etc.).

Surface coverage

The 21 curated fragments perturb the amide NH resonances of residues that constitute 67% of the molecular surface of the protein. A surface representation (Figure 4A) of the distribution and density of binding sites emphasizes the breadth of coverage and illuminates regions of the protein that represent fragment binding “hotspots.” Most of the hotspots correspond to the receptor-binding interface of IL-1β (see Figure 3B). In addition, these hotspots also reflect surfaces of IL-1β that bind to a variety of cosolvents at affinities in the molar range.²⁵ Though binding is distributed throughout the surface of IL-1β, some common binding sites are apparent. These common sites were identified by finding common residues (see Methods). The binding sites that represented the highest similarity are depicted in Figure 4B. Common binding sites are found in or near the interfaces with the receptor-binding partners. In addition, these sites all bind in an affinity range of ~300 mM and below which suggests that may be productive sites for design of inhibitors to block the receptor interfaces. It should be noted, however, that binding sites that are more unique, but do not involve residues comprising the common sites often have tighter binding as well. The distribution of dissociation constants is broad with a maximum near 200 mM but with considerable representation at much lower affinities (Figure 4C). These results reflect the propensity of interface residues to engage in weak intermolecular interactions predicted by theory.⁴⁵

Figure 4.

(A) Heat map of the number of hits detected per residue by the RM fragment screen of IL-1β. Residues are included only if determined to be a member of a binding cluster (see text and Table S3). The surface is color coded for number of hits ranging from one (blue) to ten or more (red). See color bar inset. Gray surface corresponds to no fragment hits detected (ND). IL-1β structure from PDB ID code 4DEP. No hits were detected in the screen undertaken in standard aqueous conditions. (B) Common binding sites of fragments revealed by similarity analysis (see methods). (C) Distribution of binding dissociation constants of curated specific binding events (see Methods section). Data are tabulated in Table S3.

Secondary screen and enhanced affinity

The reverse micelle encapsulation strategy is designed to capture a broader array of starting points for elaboration into high affinity molecules for pharmaceutical development or for deployment in chemical biology. This has been achieved in perhaps surprising fashion with IL-1β. To illustrate the utility of the fragment hits discovered here for elaboration, we performed a simple structure activity relationship experiment for imidazo[1,2-a]pyridine-6-ylmethanol which bound to multiple sites on the protein surface (Figure 5a). Fourteen commercially available compounds of similar structure were screened and the binding affinities of the three most promising compounds were determined by titration (see the Supporting Information, Methods section). The same residues were used to fit the affinity of each analog to ensure that the reported K_d values reflect fragment binding to the original binding site. The secondary screen compounds displayed the behavior expected: some showed enhanced or diminished affinity and site discrimination is evident (Figure 5B). For example, one secondary compound displayed a 4-fold increase in binding (198 mM to 46 mM) and enhanced specificity through a net increase of only one heavy atom. Thus, standard approaches in fragment elaboration appear to be, as might be expected, directly applicable to the weakly binding fragment hits discovered by the RM-NMR strategy. Clearly a more comprehensive medicinal chemistry effort is warranted and is on-going.

Figure 5.

Summary of the structure activity relationship analysis from a primary fragment hit imidazo[1,2-a]pyridine-6-ylmethanol. a) Binding sites summarized here are colored on the surface of IL-1β for reference. Same orientation as Figure 2. b) Binding affinities of the primary hit and three secondary molecules. Binding sites are listed here according to the color in which they are depicted in part a. Residues were used to define and fit the original hit binding affinities were also used for the analogs. Errors determined from global fitting of the cluster residues are in parenthesis.

Discussion

In summary, reverse micelle encapsulation provides a means to capture very weak interactions between small molecules and proteins. The confined space of the reverse micelle allows for high effective concentrations of ligand to be attained while avoiding inordinate material cost and thus making the search for low affinity binders both practical and reliable. Comparable binding affinities are seen in bulk and within the RM water core. Alternatively, the RM approach may be used to reduce fragment consumption (Figure S7). Importantly, RM encapsulation avoids a number of artefacts that generally interfere with solution NMR spectroscopy of proteins in the presence of high concentrations of small molecules.^24,25 Though used with protein observed ¹⁵N-TROSY spectra here, deuteration of the alkane and surfactants²⁷ as well as solvent suppression techniques such as WET suppression⁴⁶ can be used to simplify ¹H spectra and potentially extend RM-NMR to ligand observed NMR for detection. This provides site-resolved information at the atomic scale about the location of protein-ligand interactions. Quantitative and rigorous structural and statistical analysis clearly defines specific binding sites across the protein surface. This type of coverage is unprecedented and would, if general, greatly reduce the burden of relying on chance when searching for initial hits targeted to a specific area of interest.

The RM method offers many advantages over many currently used fragment detection methods. Only NMR and crystallography provide an atomic scale view of binding interactions. X-ray crystallography has had some success in fragment screening, but is challenged by sensitivity to high concentrations of small molecules when weaker interactions are to be characterized. A promising recently introduced approach employed a set of ‘MiniFrags’, smaller than those used here, and provided surface coverage and hit detection rate to that found here.¹⁰ Unfortunately, however, crystallography is generally unable to determine hit affinities, which limits further evaluation and development of discovered hits into higher affinity molecules.

Due to the effects of partitioning between the water core of the RM and the bulk alkane solvent, the procedure is limited to polar molecules. Nevertheless, a significant percentage of a small library of polar fragments decorated two-thirds of the surface of interleukin-1β, which is often considered “undruggable.” The high hit rate displayed here diminishes the need for a large library though further optimization and expansion of the fragment library for RM-based screening is potentially warranted. The high hit rate of polar molecules is an intriguing complement to the more apolar hits generated by traditional screening. Polar hits are anticipated to aid drug development since apolarity tends to increase during the hit-to-lead process and becomes a significant source of attrition.⁷ The ability of the RM based screening method to identify polar fragments that are invisible to standard methods may allow starting materials that display more favourable physiochemical properties.⁴⁷

Though very weak, the binding affinities of molecules discovered using the RM encapsulation strategy represent several productive interactions between the protein and the fragment,^4,48 when considering the significant rotational-translational entropy (ΔS_r-t) penalty that must be overcome.⁵ In practice, combinatorial chemistry must be optimized to achieve binding energies that are more negative than simple addition of the individual fragment affinities.⁴⁹ Exploitation of this phenomenon has advanced fragment hits that abut the current detection limit (~10 mM) into strongly binding drugs and inhibitors.^16–18

It is evident that the RM NMR strategy is able to illuminate those interactions predicted by theory but generally missed by assay-based or other biophysical techniques. The significantly increased breadth of discovered hits may help reduce the barrier to the generation of high affinity binders at desired locations on a protein’s surface.

METHODS

Protein expression and purification

E. coli DHFR was prepared using a method that removes the co-purifying, endogenous tetrahydrofolate to yield apo-ecDHFR.⁵⁰ Human IL-1β was prepared using a modified version of a previously described protocol.⁵¹ Details of protein expression and purification for ecDHFR and IL-1β are provided in the supporting information.

Aqueous Sample Preparation

ecDHFR samples for aqueous experiments were composed of 100 μM ¹⁵N-ecDHFR in 20 mM Tris, pH = 8.5, 100 mM NaCl, 10 mM DTT, and 0.02% sodium azide with 10% D₂O as the lock solvent. Aqueous IL-1β NMR samples consisted of 100 μM ¹⁵N-IL1β in 50 mM sodium acetate, pH = 5.0, 5 mM DTT, 0.02% sodium azide, with 10% D₂O as the lock solvent.

RM sample preparation

LDAO was pH adjusted in water (8.5 for ecDHFR and 5.0 for IL-1β). All surfactant components were mixed together then lyophilized overnight.³⁷ ecDHFR RM surfactants were composed of 52.5 mM 1-decanoyl-rac-glycerol (10MAG), 18 mM N,N-dimethyldodecylamine N-oxide (LDAO), 4.5 mM dodecyltrimethylammonium bromide (DTAB). The lyophilized surfactant mixture was suspended in pentane through rigorous mixing. Aqueous protein phase containing buffer components, DSS, and 5 mM ecDHFR were added via direct injection^27,37 for a W₀ of 15. Hexanol was titrated in and samples vortexed until sample clarity was achieved, typically 20 mM. IL-1β samples were prepared identically, with the surfactant composition of 45 mM 10MAG and 30 mM LDAO. All IL-1β mixture screening and deconvolutions were made with 10% D-pentane as a lock solvent, all titration and assignment samples were made with 98% D-pentane. All samples were referenced to internal DSS.⁵²

NMR experiments

All NMR experiments were collected at 25° C on 500 MHz, 600 MHz, 750 MHz Bruker Avance III spectrometers or a Bruker NEO 800 MHz NMR spectrometer, all equipped with cryogenically cooled probes. ecDHFR titration ¹⁵N-HSQC experiments were collected with 1024 × 64 complex points and 32 scans. All IL-1β screening and titration data were collected with ¹⁵N-TROSY experiments with 1024 × 64 complex points and 16 scans. Spectra were processed in NMRpipe⁵³ and analyzed with SPARKY.⁵⁴

Fragment Screening of IL-1β

Fragment library was constructed from 233 fragments from the Maybridge Ro3 library. Fragments were chosen based on being highly water-soluble and having a cLogP (octanol/water) of < 0.5 to avoid partitioning into the alkane phase.²⁵ Solid fragments were solubilized in DMSO at 200 mM. Liquid and low-melting point solid (LPS) fragments were solubilized at 100 mM in water. Mixtures were screened with 4 or sometimes 5 fragments at 8:1 fragment to protein. These mixtures contained 3 solid fragments and 1 or 2 liquid/LPS fragments. Mixtures were constructed to avoid similar fragments. Solid fragments were prepared from mother plates into mixtures with an automated liquid handler. Plates containing solid fragment mixtures were dried thoroughly via vacuum centrifugation to remove DMSO, which can adversely affect the behavior of reverse micelles.

For screening in aqueous conditions, 100 μM IL-1β in NMR buffer was added directly to the latch-rack tubes containing solid fragment mixtures and vortexed until fragments were solubilized. Liquid/LPS fragments were the directly added to the sample and spectra collected.

Conditions for RM samples are as described above. 10 μL of 5 mM IL-1β in NMR buffer with liquid fragment were added to pentane suspended surfactants. 10 mM of hexanol was titrated in to partially encapsulate protein, this mixture was added to latch rack tubes containing solid fragments and mixed until fragments were solubilized. If needed, more hexanol was titrated in until visual clarity was achieved, with 10 to 25 mM generally required for efficient and stable encapsulation. Visual clarity and comparison to a reference protein spectrum indicated that destabilization from fragment-RM interactions were not present. For both bulk aqueous and RM screens, ¹H spectra were collected and compared to confirm that fragments fully dissolved under both conditions. ¹⁵N-HSQC/TROSY spectra were collected and compared to a reference for detection of hits. Fifty-four mixture samples were examined for both the bulk aqueous and RM screens. Hits detected in a mixture were then deconvoluted by examination of each fragment individually.

Hits were defined by identification of significant chemical shift perturbations (CSP) using the least median squares (LMS) algorithm,⁴² as described previously.⁴¹ CSPs were calculated by the weighted chemical shift ratio:⁵⁵

$$CSP=\sqrt{{(\mathrm{\Delta }{}^{1}H)}^2+{(\mathrm{\Delta }{}^{15}N/9.8655)}^2}$$

LMS is a robust regression and outlier detection technique that is tolerant to outliers, due to elimination of outliers on the final regression coefficient. Minimization of the median of squared residuals yields a scale estimate (σ), which is analogous to a standard deviation. The resulting difference of CSPs from the minimized median is termed the residual (r) and residues with r ≥ 2σ (2σ = 0.005 PPM) are considered outliers and thus hits. One-dimensional LMS analysis was implemented in SYSTAT (Systat Software, San Jose, CA). Two or more residues meeting the LMS criteria were required for considering the fragment a hit. Residues perturbed by fragment binding as assessed by LMS analysis are reported in Supplementary Table S1.

Definition of specific binding and affinity determination

Titrations were achieved by collecting spectra of the following fragment to protein molar ratios: 0:1, 8:1, 16:1, 32:1, 48:1, and 64:1. For each titration series, a 0:1 sample and a 64:1 sample were prepared and used to mix subsequent titration points after spectral collection. CSPs for all residues were fitted to the exact quadratic form of the binding equilibrium equation, with CSP weighting:

$$CSP=CS{P}_{sat}\left(([F]+[P]+K_d)-\sqrt{{([F]+[P]+K_d)}^2-4[F][P]}\right)/2[P]$$

CSP_sat is the chemical shift perturbation at saturation of protein with fragment, [F] is fragment concentration, and [P] is protein concentration. K_d determined here is the residue-specific affinity of the fragment to the protein. Only residues with CSP_sat > 0.025 ppm, K_d < 1 M, and R² > 0.95 were considered reporters of fragment binding. Individual residue fitting is reported in Supplementary Table S2.

Specific binding sites were determined using spatial clustering using the k-means algorithm⁴⁴ within scikit-learn⁵⁶. Backbone N atoms for residues showing CSPs were grouped into k groups. These were informed by examining the distribution of hit residues on the IL-1β structure and individually fitted K_d values. Centroid coordinates were optimized by minimization of standard squared error (SSE) between all atomic coordinates and centroids until a stable distribution of clusters was achieved:

$$SSE=\sum _{i=1}^k\sum _j{(x_j-{\overline{x}}_i)}^2+{(y_j-{\overline{y}}_i)}^2+{(z_j-{\overline{z}}_i)}^2$$

The structure of IL-1β within the ternary signaling complex (PDB 4DEP) was used for the atomic coordinates.³⁵

A common K_d value was fitted for each cluster to achieve site-specific binding affinities. All cluster members were simultaneously fit to the quadratic equilibrium equation with a single global K_d, but CSP_sat was allowed to float for each residue. Residues with R² > 0.95 were included in the resulting group K_d fits. Results of group k-means cluster fitting and cluster K_d determination are presented in Supplementary Table S3.

Common binding sites

Residues comprising binding sites of each of the twenty-one curated fragments were examined for common residues. Each binding site was scored according to how many residues were shared with other binding sites. The number of the shared residues was weighted by the total number of residues within the binding site. Nine highest-scoring binding sites were determined from this analysis and are depicted in Figure 4B.

Secondary screen of similar fragments

Fourteen compounds that are similar in structure to imidazo[1,2-a]pyridine-6-ylmethanol and commercially available (Sigma-Aldrich) were assessed in a secondary screen. The screen was performed against IL-1β in RM encapsulated conditions. The 14 fragments screened here are identified in the supporting information. The secondary screen was performed with an effective concentration of fragment within the RM core of 80 mM. ¹⁵N-TROSY spectra were assessed for binding and the most promising secondary fragments were titrated: 2-(6-methylimidazo[1,2-a]pyridin-2-yl)ethanamine, 1-(6-methylimidazo[1,2-a]pyridin-2-yl)methanamine, and [(7-methylimidazo[1,2-a]pyridin-2-yl)methyl]amine. Binding was assessed by titration and fit using the cluster K_d fitting protocol as outlined above.