A TfR-binding cystine-dense peptide promotes blood-brain barrier penetration of bioactive molecules

Abstract

The impenetrability of the blood-brain barrier (BBB) to most conventional drugs impedes the treatment of central nervous system (CNS) disorders. Interventions for diseases like brain cancer, neurodegeneration, or age-associated inflammatory processes require varied approaches to CNS drug delivery. Of recent interest as drugs or drug-delivery vehicles are cystine-dense peptides (CDPs). Found throughout the phylogenetic tree, often in drug-like roles, their size, stability, and protein interaction capabilities make CDPs an attractive mid-size biologic scaffold to complement conventional antibody-based drugs. Here, we describe the identification, maturation, characterization, and utilization of a CDP that binds to the transferrin receptor (TfR), a native receptor and BBB transporter for the iron chaperone transferrin. We developed variants with varying binding affinities (K_D as low as 216 pM), co-crystallized it with the receptor, and confirmed murine cross-reactivity. It accumulates in the mouse CNS at ~25% of blood levels (CNS blood content is only ~1–6%), and delivers neurotensin, an otherwise non-BBB-penetrant neuropeptide, at levels capable of modulating CREB signaling in the mouse brain. Our work highlights the utility of CDPs as a diverse, easy-to-screen scaffold family worthy of inclusion in modern drug discovery strategies, demonstrated by the discovery of a candidate CNS drug delivery vehicle ready for further optimization and preclinical development.

Introduction

Drug delivery to targets in the central nervous system (CNS) is hampered by the blood-brain barrier, a term for the particularly impassable vascular endothelial cells in CNS capillaries [1,2]. This system exists to keep toxic metabolites and pathogens out of the brain, but also renders diseases of the CNS difficult to treat using conventional medicines [3]. It is largely for this reason that primary CNS tumors (e.g. glioma) and neuroinflammatory / neurodegenerative diseases (e.g. MS and Alzheimer’s Disease) respond poorly to therapeutic modalities that would otherwise be more effective in cancers or inflammatory conditions affecting peripheral tissues.

While the BBB allows blood osmolytes and nutrients into the brain, and CNS astrocytes provide many of the growth and survival signals required by neurons, several larger hormones and proteins still require access to the CNS from blood, including transferrin (an iron chaperone), insulin, and leptin [4]. Transport of these proteins across the BBB is accomplished by receptor-mediated vesicular transcytosis (transport of cargo from apical to basal side, or vice versa, in intracellular vesicles) using variants of their native receptors (TfR, InsR, and ObR, respectively) employed by CNS vascular endothelial cells. In this highly selective way, certain large molecules like transferrin (~75 kDa) can access the brain parenchyma.

This mechanism has been explored for therapeutic purposes, particularly using TfR. Groups have tested delivery of agents to the CNS using either transferrin itself, or an antibody against TfR, as a ferry for cargo with otherwise poor BBB penetration properties. Examples include transferrin-decorated liposomes, peptides that bind transferrin itself, and antibodies that target TfR to deliver TNFα decoy receptors or anti-BACE enzymes [4]. Furthermore, it has been well established that many cancers and cancer cell lines overexpress TfR [5,6], likely to meet the increased metabolic demands of oncogenesis. Similar strategies to achieving BBB transport via the TfR have been employed for tumor targeting. Examples of delivered agents include small drugs, proteins, oligonucleotides, and radionuclides [7].

While antibodies make up the majority of recently approved biologics, we have explored alternate protein therapeutic scaffolds in order to bypass some of the liabilities of the immunoglobulin scaffold, including poor tissue permeability [8–10], immunogenicity [11,12], and long serum half-life that can become problematic if toxicities arise. Miniproteins in the 10–100 amino acid range represent a mid-sized medicine space [13], with many of the binding specificity and affinity characteristics of antibodies but with improved stability, reduced immunogenicity, and simpler manufacturing methods [14–19]. Among miniprotein scaffolds found in nature, cystine-dense peptides (CDPs) are of particular interest. Their intramolecular disulfide architecture provides particularly high stability metrics [15–17], reducing fragmentation and immunogenicity, while their smaller size could improve tissue penetration and facilitate tunable serum half-life. Their use in drug-like roles throughout every evolutionary clade highlights their utility as drug scaffolds. As such, we sought to identify a CDP that binds to TfR in order to provide easier cargo transport into the CNS. Using a library of mutagenized variants of native CDPs and our mammalian surface display screening platform [20], one such CDP was identified that specifically binds to human TfR. It was matured, crystallized, and validated to deliver a neuromodulatory agent into the mouse brain. Derivatives of this molecule have the potential to carry payload therapeutics for oncology, autoimmune disease, acute and chronic neurodegeneration, and pain management.

Results

Production and validation of soluble TfR for mammalian display screening.

For our CDP screen, the full-length human TfR ectodomain (residues 121–760) was chosen for the screening target, as this was found to be the most soluble and stable version of the receptor. The ectodomain was cloned into the Daedalus soluble protein production lentivector [21], and protein was purified from the growth media (Fig. 1a and S1). The same strategy was used to produce and purify human apo-transferrin (residues 23–698), and some of the material was iron-loaded to produce holo-transferrin. Both apo-transferrin and holo-transferrin were tested for binding to the soluble TfR ectodomain via surface plasmon resonance (Fig. 1b), where only holo-transferrin demonstrated an interaction with the captured TfR.

Fig. 1.

Soluble TfR ectodomain retains native ligand binding properties. (a) SDS-PAGE and Coomassie stain of soluble TfR demonstrating >95% purity and expected ~75 kDa molecular weight mobility. (b) Surface plasmon resonance at physiological pH (7.4) showing captured TfR interacts with iron-bound transferrin (HoloTf) but not empty transferrin (ApoTf). (c) Illustration of SDGF vector CDP display scaffold and target engagement for cell labeling. (d) Flow cytometry demonstrating specificity of TfR:ligand interaction for cell labeling. 293F cells transfected with SDGF-MaCV (Machupo virus glycoprotein) are labeled with a combination of biotinlylated TfR and Alexa Fluor 647-labeled streptavidin (Strep-647). Similar staining using either SDGF-MaCV cells and Alexa Fluor 647-labeled elastase, or SDGF-elafin cells and TfR + Strep-647, both fail to label cells. GFP+ cells were selectively gated for analysis.

As further validation that the soluble TfR represents the human, native structure, interaction with the Machupo virus glycoprotein (referred to as MaCV), which uses TfR to determine tropism and mediate cell entry [22], was tested. For this, the MaCV gene was cloned into the mammalian surface display vector SDGF [20] (Fig. 1c), and then transfected into suspension 293 Freestyle (293F) cells. Elafin, a native CDP binder / inhibitor of elastase, served as a control. Transfected cells were stained with 200 nM biotinylated TfR (all TfR used in cell binding assays is biotinylated) and 200 nM Alexa Fluor 647-labeled streptavidin, and then analyzed by flow cytometry (Fig. 1d). Cells transfected with SDGF-MaCV were successfully stained with TfR, while SDGF-Elafin cells were not. Meanwhile, SDGF-MaCV cells incubated with 200 nM fluorescent elastase were not stained. This demonstrates both the specificity of TfR binding to a native ligand, and the utility of SDGF to identify novel TfR binding partners.

Mammalian display screening to identify and mature a novel TfR binding CDP.

To identify a TfR-binding CDP, a library of oligonucleotides encoding 10,000 native CDPs [20] was amplified and mutagenized. The CDPs were 17–50 amino acids in length, with 4, 6, 8, or 10 cysteines. While there was some weighting of the library towards annotated knottins or defensins, the library contained CDPs from every domain / kingdom of life. This library was cloned into SDGF, made into lentivirus, and transduced into suspension 293F cells. The transduced cells were subjected to staining with TfR (200 nM) and co-stain over the course of one round of magnetic cell sorting and three rounds of flow sorting (Fig. 2a, left), each round enriching for cells stained with TfR.

Fig. 2.

Mammalian surface display CDP screening yielded TfR binders. (a) Primary screening and and affinity maturation of TfRB1. Left: 293F cells expressing a library of native-derived CDPs via transduced SDGF vector were screened using one round of magnetic sorting (performed prior to flow cytometry experiments) and three rounds of flow sorting, producing a clonal population of TfR binding cells (arrow in bottom-left plot); the binder was designated TfRB1G1. Staining conditions were 200 nM stain and co-stain (Alexa-Fluor 647 labeled streptavidin), incubated simultaneously to confer tetravalent avidity. Right: TfRB1G1 binding was improved by two rounds of affinity maturation using site-saturation mutagenesis (SSM), producing two improved variants (TfRB1G2 and TfRB1G3) as seen by differential staining in transiently transfected 293F cells. Staining and sorting conditions for the SSM screens differed from the primary screen by using lower concentrations (20 nM and 8 nM for the first and second SSM screens, respectively) and performing staining in a step-wise fashion to eliminate the tetravalent avidity of streptavidin co-staining. (b) Sequences of the original native CDP and the three resulting generations of TfRB1 proteins; colored fonts match enriched mutations novel to a given TfRB1 variant. (c) Soluble TfRB1 variants were produced, showing high purity and proper disulfide-driven folding via SDS-PAGE Coomassie stain and RP-HPLC. (d) Surface plasmon resonance demonstrated improved binding characteristics as affinity maturation progressed. Instrumentation could not resolve a binding affinity for TfRB1G1, while TfRB1G2 was analyzed by a steady-state binding model that does not yield separate on- and off-rates. Full binding curves providing affinity measurements are found in Fig. S5.

A single TfR-binding CDP, designated TfRB1G1 (TfR binder 1 generation 1), was identified by DNA sequencing of the final enriched cell population. It represents a randomly mutated variant of cytochrome BC1 complex subunit 6 from the marine choanoflagellate Monosiga brevicollis (Uniprot ID: A9V0D7), is 49 amino acids in length (six cysteines), and has a predicted molecular mass of 5.6 kDa. TfRB1G1 was then subjected to affinity maturation using site saturation mutagenesis (SSM) (Fig. 2a, right), wherein a library was created containing every possible non-cysteine single amino acid substitution (43 non-Cys amino acids x 18 possible non-Cys substitutions = 775 variants, including TfRB1G1). This library was screened with a modified staining protocol, using a lower concentration of target and co-stain (20 nM) and separate staining steps; the latter further increases stringency by eliminating the tetravalent avidity granted by streptavidin. Three rounds of flow sorting and enrichment were used to identify point mutants with improved TfR binding characteristics; the flow cytometry profiles throughout this maturation screen (Fig. S2) demonstrate progressive shifting of the population towards high TfR staining. Permutation of enriched variants (Fig. S3a–b) identified an optimal mutant (TfRB1G2), and this process was repeated again (Fig. S3c, 8 nM TfR and co-stain; otherwise, identical protocol) to generate TfRB1G3. This twice-matured variant contains 14 point mutations from the original library member: 4 point mutations from the native M. brevicollis sequence are found in TfRB1G1, while G2 and G3 contain 6 and 4 mutations, respectively, from the previous generation (Fig. 2b).

TfRB1 variants were produced as soluble CDPs and validated by reversed-phase HPLC (RP-HPLC), SDS-PAGE, and mass spectrometry (Fig. 2c and S4). Based on their masses of ~5–6 kDa, their slower-than-expected mobility in SDS-PAGE was a demonstration of the interesting electrophoretic mobility characteristics that some CDPs possess [17]. All variants showed markedly different mobility upon DTT reduction (10 mM) in both SDS-PAGE and RP-HPLC, confirming cystine stabilization. Their binding to TfR was verified by surface plasmon resonance (Fig. 2d and S5), also confirming increased affinity of matured variants (TfRB1G3 [K_D = 216 ± 1 pM] > TfRB1G2 [K_D = 8.7 ± 0.4 nM] > TfRB1G1 [K_D not determined]). They were further evaluated for stability under conditions of reduction, protease exposure, and heat (Fig. S6 and S7). All variants demonstrated full or partial resistance to cellular reducing conditions (10 mM glutathione), while the affinity matured variants showed partial resistance to pepsin (though all were vulnerable to trypsin proteolysis). The intact, non-reduced TfRB1G3 protein demonstrated substantially improved heat tolerance to that of the DTT-reduced protein, with no substantial change in circular dichroism characteristics until well above common ambient temperatures (>50°C) and a failure to observe complete unfolding up to 95°C.

TfR:TfRB1G3 co-crystal structure predicts functional characteristics.

Many drug-like CDPs derive their stability from their disulfide bonding architecture [17], so in order to better predict the impact of structural alterations, the crystal structure of TfRB1G3 was solved, both alone and bound to TfR (Fig. 3a and Table 1). As predicted by structural modeling with I-TASSER [23], the CDP is an anti-parallel two-helix bundle, which is seen in other species’ structurally-characterized cytochrome BC1 complexes [24]. It contains three disulfide bonds with a simple hairpin topology (paired C4:C46, C8:C42, and C18:C32) (Fig. 3a, left). Upon complexing with TfR, the N- and C-terminal residues become ordered, making contacts to TfR. However, the helical core of the CDP (between and including the C8:C42 disulfide) is nearly identical between bound and unbound forms, with backbones aligning to an RMSD of 0.331 ± 0.049 Å (N = 8 alignments). In both crystals, the loop connecting the helices is disordered.

Fig. 3.

Structure of TfRB1G3, unbound and bound to TfR. (a) Left: cartoon representations of the I-TASSER [23] predicted model of TfRB1G3 (top, N-terminus in blue, C-terminus in red, visible side chains colored by atom type), the crystal structure of free TfRB1G3 (middle, blue; 2.55 Å resolution, 4-fold superposition), and the crystal structure of TfRB1G3 when complexed with TfR (bottom, red; 1.85 Å resolution, 2-fold superposition). Side-chains with an atom within 4 Å of TfR in the complex structure are shown as sticks. N- and C-termini are labeled. Red arrows mark the most significant structural differences between model, unbound, and bound structures of TfRB1G3 near the N-terminus. Gray crosses mark the location of the disordered inter-helical linker. Middle: zoomed-out view of TfRB1G3 bound to a TfR dimer; a second TfRB1G3 molecule is on the opposite face, confirming 2:2 binding stoichiometry. Disordered inter-helical region (black) is based on the I-TASSER model. Right: zoomed-in view of binding, showing numerous polar and non-polar (the latter are mainly Leu and Met) interactions driving binding. (b) Alignment of the TfR:TfRB1G3 co-crystal with that of the TfR:transferrin complex (PDB 3S9L; TfR hidden) demonstrates an overlap of the binding sites for TfRB1G3 and transferrin. (c) The TfRB1G3 binding site, colored for human / murine homology. “Similar” residues share charge and hydrophobicity characteristics, while “Dissimilar” residues do not. (d) TfRB1 cross-reactivity with murine TfR confirmed in flow cytometry binding assays. SDGF surface expression of the human (Hs) and murine (Mm) TfR ectodomains permitted staining of these cells with 40 nM Alexa Fluor 647-labeled TfRB1 variants. Quantitation of staining on a set of low- or high-GFP cells is shown; the “GFP− Mm” cells approximate staining attributable to the native TfR expression of 293F cells. Confirmation of antibody species identity and raw flow cytometry figures, including gating of GFP+ and GFP− cells, are found in Fig. S9. *: P<0.0001 by Kruskal-Wallis test. #: HsTfR TfRB1G3 staining experienced sample loss during processing, but data variance was comparable to other samples and the K-W test is robust to different sample sizes, so analysis proceeded.

Table 1.

Crystallographic data collection and refinement statistics.

Protein	TfRB1G3	TfR / TfRB1G3 complex
Accession Code	6OKE	6OKD
Space group	P65	C2
Cell dimensions
a, b, c (Å)	37.66, 190.21	102.6, 145.5, 133.5
α, β, γ (°)	90, 120	90, 90.02, 90
Resolution (Å)	50.0 – 2.55 (2.59 – 2.55)	50.0 – 1.85 (1.88 – 1.85)
R merge	0.102 (0.280)	0.086 (0.693)
I/σ(I)	16.7 (2.0)	16.8 (1.8)
Completeness (%)	99.0 (90.5)	99.8 (99.9)
Redundancy	7.1 (2.7)	3.3 (3.3)
Rwork / Rfree	22.4/28.0	0.163 / 0.187
No. atoms
Protein	1,134	10,198
Heterogen	-	140
Waters	24	782
Wilson B (Å2)	24.6	18.2
Average B (isotropic, Å2)	22.0	26.4
R.m.s. deviations
Bond lengths (Å)	0.006	0.006
Bond angles (°)	1.427	1.11
MolProbity score	1.47	1.24

The complex demonstrates 2:2 binding stoichiometry of TfRB1G3:TfR (Fig. 3a, middle). Binding to TfR is mediated by both polar and non-polar interactions; the latter are primarily Met and Leu residues fitting into small hydrophobic pockets at the interface (Fig. 3a, right). The distortion of the N-terminus results from forming three hydrogen bonds to TfR, while adjustments of the C-terminal helix between bound and unbound states accommodates sterics around TfRB1G3-H47. The N-terminal GS stub, a remnant after TEV cleavage from the Siderocalin fusion partner, is also disordered. Overall, the affinity (K_D = 216 ± 1 pM) is quite strong considering the small buried surface area (1001 Ų); compared to the published affinities and buried surface areas of 113 heterodimeric complexes [25], only two of the 86 complexes with smaller buried surface areas demonstrated stronger affinity (coincidentally, one of them is a protease-inhibiting CDP and its target). Enriched mutations from library mutagenesis and subsequent affinity maturation are found in locations both adjacent to and distal to the interface; in particular, R7, E33, E37, and Y41 are all unlikely to be contributing to binding via direct contact. Whatever positive influence they have on affinity is likely via reducing configurational entropy in solution, reducing the entropic penalty of binding.

By comparison with the TfR:transferrin co-crystal structure [26], the TfRB1G3 footprint overlaps with that of transferrin, predicting competitive binding (Fig. 3b). This was confirmed in surface display flow cytometry binding assays in the presence of soluble apo- or holo-transferrin (Fig. S8). While this may be predictive of poor BBB penetration, owing to the high plasma levels of holo-transferrin (reportedly 20 μM or more [27]), the highest affinity TfRB1 variant (TfRB1G3) has a similar off-rate (k_d) but a much faster on-rate (k_a) than that published for the TfR:transferrin interaction [28]. Most of the residues comprising the TfRB1G3 binding site on TfR are conserved between human and murine orthologs (Fig. 3c). This suggested cross-reactivity, which was confirmed in flow cytometry binding assays (Fig. 3d and S9) using surface displayed human or murine TfR (via SDGF transient transfection) and dye-conjugated, soluble TfRB1 variants. This alternate strategy (soluble CDP, surface-bound target) was employed because soluble murine TfR ectodomain of sufficient quality could be neither produced nor commercially obtained, but was well behaved as a surface displayed protein. Cross-species comparative binding quantitation using this method should be interpreted with caution, as there is no way to validate whether cells of equal GFP fluorescence possess equal amounts of binding-capable human vs murine TfR on their surface, but the data provided sufficient support for cross-reactivity to permit in vivo testing with mice expressing only native murine TfR.

TfRB1 variants accumulate in many tissues, including the CNS.

Owing to the human/murine cross-reactivity, we posited that the in vivo tissue distribution properties of TfRB1 variants in mice would be predictive of human tissue distribution. To assess their distribution properties with minimal alteration of the CDPs, TfRB1 variants were radiolabeled with ¹⁴C using reductive methylation of free amines [29]. By this method, a traceable marker is introduced that alters the protein with only 2 novel methyl groups per free amine (lysine side chains and N-terminus). After intravenous delivery to mice (100 nmol), ¹⁴C-TfRB1G3 distribution was measured by collecting blood and tissue homogenates (Fig. 4) up to 48 hours after injection for liquid scintillation counting. TfRB1G3 demonstrated biphasic plasma elimination kinetics, with a first elimination half-life of 15.6 minutes (95.1% fast) and a second elimination half-life of 10.3 hours (4.9% slow). The highest levels were seen in the kidney, liver, and spleen; small plasma solutes and peptides are commonly excreted by the kidney, while both hepatic and splenic tissues have been reported to express high levels of TfR [30,31]. The brain showed smaller but measurable levels, but it was unclear from this assessment whether this represented accumulation beyond that expected for normal CNS blood flow. To further assess distribution of ¹⁴C-labeled TfR binders, TfRB1G1 and TfRB1G3 levels were quantitated in several tissues after 30 minutes and 3 hours using autoradiography of whole body sections (Fig. 5a, S10, S11). Measured tissue values at 30 minutes correlate well with LSC measured values (Pearson’s R² = 0.946), with the highest levels again in the kidney, liver, and spleen. CNS levels were lower than that of other tissues, but still were quantitated to predict levels of ~100–200 nM after 3 hours. This assumes substantial parenchyma access (in which case nmol g⁻¹ tissue is equivalent to μM concentration), but the data support this assumption because brain tissue is reported to consist of only ~1–6% blood, depending on the measurement method [32–34]. Comparing the brain and cardiac blood quantitation shows CNS accumulation at roughly 25% of blood levels after 3 hours (Fig. 5a, right panel, and 5b), well above even the most generous estimates of CNS blood content.

Fig. 4.

Tissue distribution and elimination of ¹⁴C-TfRB1G3. (a) Scintillation counting of serum shows two-phase decay kinetics from mouse blood. Error bars, when visible apart from data points, are SD. R² = 0.962. (b) Scintillation counting of tissue homogenates shows accumulation / elimination of TfRB1G3 from kidney, liver, spleen, muscle, skin, and brain. In all experiments, 100 nmol (~0.5 mg; specific activity 147 Ci mol⁻¹) material was administered via single bolus intravenous injection. Curve fitting for (b) was one-phase decay modeling; statistics for each fit are below each curve. For all tissues including serum, N = 2–3 mice per time point over 8 time points.

Fig. 5.

Whole body autoradiography of ¹⁴C TfRB1 variant distribution. (a) Whole-body autoradiography (WBA) quantitation of TfRB1G1 and TfRB1G3 in several tissues after 30 or 180 minutes. Right inset shows brain quantitation normalized to blood levels. Error is mean ± SD. (b) Representative brain and heart (inset) WBA images of three TfRB1 generations, plus that of a control, non-TfR binding CDP. As CDPs had different specific activities, image intensities at the full body scale were internally normalized and equalized across all images at a given time point for ease of viewing only; no image modification was performed during quantitation. See Fig. S10 for full body images and Fig. S11 for an example of how regions of interest are determined for quantitation.

TfRB1 variants can ferry a bioactive molecule across the BBB.

While CNS accumulation is not inherently predictive of neuronal access [35], owing to the complex kinetics of binder:TfR vs transferrin:TfR interactions at the apical and basal surfaces and within vesicles during transcytosis, seeing CNS accumulation of TfRB1 variants led us to test whether they could be utilized to modulate neuronal activity in mice. TfRB1 variants are not expected to have any pharmacologic activity on their own in this context (single dose, 95% fast-phase serum clearance t_½ = 15.6 min), and we wanted to establish a model that would be amenable to virtually any CDP scaffold, so we chose to demonstrate this property using neurotensin (NT). NT is a neuropeptide used for local signaling in the CNS and digestive tract [36–38]. Its downstream activity involves intracellular Ca²⁺ regulation leading to modulation of cAMP response element (CRE) driven transcriptional programs [39–41] (Fig. 6a). Intriguingly, modulation of this NT-driven signaling pathway has been explored for suppression of chronic pain [42]. Being that NT is a simple 13-mer peptide (sequence ELYENKPRRPYIL) whose receptor-interacting residues are at its C-terminus, TfR binder-NT fusion proteins (Fig. S12 and S13) that engage the neurotensin receptor were generated, with activity demonstrated in a HEK-293 cell line transfected to over-express NTSR1 (Fig. 6b). For a further control, we attempted to generate a NT fusion of the non-TfR binding CDP tested by WBA (Fig. 5b), but it was misfolded and non-homogenous, likely due to a partially occluded C-terminus seen in its modeled structure (Fig. S14). Stability of TfRB1G3 and its NT fusion variant in isolated mouse serum were measured by quantitative RP-HPLC and mass spectrometry, demonstrating half-lives of ~4.2 hours and ~5.6 hours, respectively (Fig. S15). Given the first-pass serum elimination half-life of ~15 minutes shown for TfRB1G3 (Fig. 4a), serum degradation is unlikely to be of significant relative impact to in vivo pharmacodynamic response, so experiments proceeded without further optimization for stability.

Fig. 6.

TfRB1-Neurotensin fusions induce a response in CRE-luciferase (CRE-Luc) mice. (a) Model of the relevant pathways influencing CRE-driven luciferase in the CRE-Luc mice. Non-standard protein or drug abbreviations are as follows. AC: adenylyl cyclase. PDE: cAMP phosphodiesterase. FS: forskolin. Rol: rolipram. (b) In vitro neurotensin (NT) receptor engagement showing IP₁ accumulation only in response to NT or NT fusions in HEK-293 cells expressing NTSR1. N = 3 wells for all except vehicle, which had N = 36. Horizontal bar indicates sample mean. mTF = murine transferrin. Baseline HEK293 = mean assay value for HEK293 cells (N=36 wells) that do not express NTSR1, included as a reference. (c) Left: luminescence (after intraperitoneal luciferin dosage) either before (red circles) or four hours after intravenous administration of TfR binders (NT fusions, blue circles; non-NT fused, grey circles), using matched cohorts. Horizontal bar indicates sample mean. Significance via T-test (unpaired, 2-tailed). *: P < 0.05. **: P < 0.01. #: P < 0.0001. All other intra-cohort comparisions did not reach significance. For both transferrin variants (native and NT fusion), murine transferrin was used. Right: sample images from the TfRB1G3 cohort quantitated in left plot shown with pseudocolored luminescence intensity over B&W brightfield images. Examples are for viewing only, as all images were equally background-subtracted and contrast-enhanced. No image modification was performed during quantitation. See the Methods for censorship criteria. (d) Immunohistochemistry showing subjectively enhanced luciferase expression in the cortex, striatum, thalamus, and hippocampal (Hip) CA1 region of mice 4 hours after NT fusion administration. Scale bar (top left panel) is 100 μm; all panels are the same magnification.

For in vivo activity assessment, mice carrying a cAMP response element (CRE)-driven luciferase transgene [43] were used. CRE-driven luciferase induction was verified by luciferin dosage (3 mg D-Luciferin in 100 μL, intraperitoneal) four hours after administration of forskolin and rolipram (Fig. S16), a potent activator of adenylyl cyclase and inhibitor of cAMP phosphodiesterase, respectively [44,45]. Conversely, no reporter activation was observed by luciferin dosage four hours after administration of a high dose (300 nmol) of free NT peptide (Fig. S17), demonstrating that unmodified NT peptide in the periphery is not able to cross the BBB and activate the CRE-reporter system in the CNS.

CRE-luciferase levels were tested without (unstimulated) or after intravenous administration of TfR binders: 100 nmol TfRB1 or TfRB1-NT variants, or 12 nmol transferrin or transferrin-NT; the latter’s lower dose compensated for the drastic differences in plasma half-life (15 minutes for TfRB1 vs 40 hours for transferrin [46]), allowing roughly equivalent cumulative drug exposure over four hours. Stimulated mice were dosed with luciferin solution four hours after TfR binder dosage, and CNS luminescence levels were measured. Unstimulated mice had a measurable basal level of CRE activity driving luciferase expression in the brain, but the luminescent signal was significantly elevated in mice administered TfRB1G2-NT and TfRB1G3-NT (Fig. 6c). The response was specific to NT-mediated signaling, as none of the parent molecules lacking NT fusions were shown to significantly elevate luminescence (Fig. 6c, left, grey circles). For all four NT fusions, IHC staining of random mice from unstimulated or NT fusion groups show increased luciferase levels in many brain regions in mice treated with NT fusions compared to unstimulated animals (Fig. 6d). Mice were chosen randomly and IHC is highly sensitive, so quantitation cannot be inferred from the IHC imaging, but the cohort-wide increases in luminescence in intact animals seen for TfRB1G2 and TfRB1G3 demonstrated not only CNS accumulation, but BBB penetration and neuronal access of a molecule fused to TfRB1 variants.

Discussion

Our previous work with mammalian CDP display [20] established the utility of the platform for binder screening, using a computationally designed library to identify a binder to the TEAD oncoprotein from in silico engineered scaffolds. Here, we explored the wide structural diversity available in native CDP space to identify a binder to TfR. Even though it does not come from a classical drug-like CDP scaffold family like knottins or defensins [47], it still demonstrated the resistance to heat, protease (pepsin), reduction, and serum degradation that one would hope to see in a CDP drug candidate, in addition to an antibody-like sub-nM affinity for the most mature variant (TfRB1G3). Screening with native-derived CDP libraries allows access to a wide diversity of scaffolds, and therefore surfaces for target-engagement. This permits a higher chance of achieving shape complementarity for varied sites on many target proteins than would be achievable with a library of more limited diversity. To access this diverse structural space, mammalian display has proven to be an effective and efficient screening platform wherein demonstrable expression and folding is seen for hundreds of distinct CDPs [20], each of which can be the basis for further mutation into thousands of variants.

Native, drug-like CDPs fold into complex topologies from which they derive many of their stability characteristics [17]. However, TfRB1 is a simple anti-parallel two-helix bundle. This provides an opportunity for pharmacophore transfer onto other scaffolds. Many protein-based molecules with bioactivity could benefit from enhanced BBB penetration properties, whether native, antibody-based, or designed in silico. Particularly with the latter group, two-helix bundles are common structural elements, and it would be of interest to test engraftment of the required TfR-engaging surface onto such scaffolds. Any protein fusion presents complications for folding and activity maintenance, and it requires empirical confirmation that neither molecule’s activity is impaired by incorporation into the fusion construct. Nevertheless, the simple paired helical motif and small size (~5.5 kDa) are amenable to fusion tagging with minimal disruption to either component’s activity; indeed, considering the surface display binding experiments (N-terminal fusion) and NT fusion (C-terminal), TfRB1G3 retains TfR binding after fusion at either terminus.

While a demonstration of brain-wide cAMP response using TfRB1G3-NT is promising, it is not representative of any specific condition and may be more sensitive than some disease models requiring substantial alteration of neuronal behavior. Each disorder and model will have different minimal requirements for CNS dose delivered and pharmacokinetic/pharmacodynamic liabilities, so empirical demonstrations in each specific model will be necessary to confirm achievement of relevant dosage. Many existing methods for TfR-based CNS delivery have shown great promise [48], and in the context of a given disease model, a direct comparison between TfR-binding CDPs and similar antibody- or peptide-based tools would help parse out their relative strengths. In that context, CDP-based TfR binders are an exciting addition to existing modalities, as they present the opportunity to facilitate BBB penetration with minimal (~6 kDa) molecular weight addition compared to that of transferrin (~75 kDa) or antibodies (~150 kDa for a full IgG, ~25 kDa for a single chain Fv). Meanwhile, a serum degradation half-life of > 4 hours compares favorably with that of the BBB penetrant peptide Angiopep-2 (~20 minutes) [49], which may otherwise be a liability for fusion with partners that require hours-long serum persistence to accomplish sufficient target tissue accumulation.

When comparing CDPs (or any miniprotein more generally) with antibodies for the development of target-protein binders, we note that our molecule binds at a site of high homology between human and murine TfR. Such a surface is unlikely to generate antigenic fragments for conventional monoclonal antibodies, as the region is likely to be recognized as self and fail to be present in the animal’s mature B-cell repertoire. This could be a desirable feature depending on the target; TfRB1 variants could have plausibly failed to demonstrate BBB penetration due to its competition with transferrin binding, while distal regions with less homology and less risk of native ligand competition (e.g. the apical domain) would be seen as more attractive targets for antibody campaigns. On the other hand, human/murine cross-reactivity simplifies pre-clinical testing strategies; the TfRB1 variants, for example, did not require testing in a TfR-humanized mouse model, which would have added to the expense and complexity of the experiments while reducing translational relevance. Furthermore, if the goal of the screening campaign is to identify an inhibitor of a protein-protein interaction, many such surfaces in conserved proteins are likely to have high intra-mammalian homology, as interface-altering variation would require co-evolution of both partners to maintain function. As mentioned above, classical antibody campaigns targeting such an interface would be inherently difficult, so an alternate binder/inhibitor screening strategy where homology is not a variable may have a higher likelihood of success. The vast majority of clinical monoclonal antibodies are designed to disrupt a disease-associated protein-protein interaction, so it is reasonable to propose incorporating surface display CDP screening into modern strategies for identifying novel clinical entities to disrupt such interactions.

TfRB1G1 and TfRB1G3 have similar CNS accumulation properties via quantitative WBA but drastically different TfR affinities. This is unexpected in the context of a simple on-off receptor interaction model. However, receptor-mediated BBB transcytosis is influenced by a number of factors, including (among others) uptake rates; receptor turnover; the differing binding environments of serum, vesicle, and parenchyma; and competing transferrin. For practical purposes, CNS accumulation measurement via low-resolution methods like WBA cannot discriminate whether molecules are accumulating in the CNS but remaining inside (or bound to) vascular endothelial cells vs bona fide parenchyma access. This phenomenon has been explored in other systems in depth [35]. An alternative plausible explanation would be that TfRB1G1 uptake and apparent CNS accumulation is not TfR-dependent, which would limit transcytosis and parenchyma access. While the differential trafficking properties of the TfRB1 variants could be the subject of future work, we prefer a pharmacodynamic measure (here, cAMP response) to properly evaluate BBB transport when clinical relevance is in question, and in this context we see two nearly-identical molecules (TfRB1G1 and TfRB1G3) imposing drastically different responses in the mouse brain upon peripheral administration. The different response can most plausibly be attributed to their differential TfR affinities, and future detailed evaluation of binding and trafficking in vivo could shed further light on this relationship.

In the context of their competitive binding with transferrin, the BBB penetrating capabilities of TfRB1 variants were somewhat surprising. However, the measured and published [28] binding kinetics of TfRB1G3 and transferrin may provide a mechanism. Both molecules have similar off-rates; however, the on-rate of TfRB1G3 is much faster. It could be that, as soon as a vesicle bearing TfR to the apical cell surface fuses and exposes receptor to the bloodstream, TfRB1G3 simply is able to bind with the newly plasma-exposed TfR more rapidly than does transferrin. The half-life of TfRB1G3 release (~6 minutes given its dissociation constant [k_d = 1.85 ± 0.01×10⁻³ s⁻¹]) is longer than the reported endocytosis rate of transferrin (t_½ = 3.5 minutes [50]), and because TfR uptake is not ligand-dependent [51], TfRB1G3 might simply out-compete transferrin for its binding site and occupy it long enough for uptake. CNS accumulation of TfRB1G1 (Fig. 5) conflicts with this model, but if its uptake into vascular endothelial cells is less dependent on TfR, transferrin would be non-competitive with it. Regardless of the mechanism explaining why TfRB1G3 retains function in vivo in spite of competitive binding with transferrin, its function would necessitate a partial disruption of native transferrin:TfR binding. With that in mind, care will need to be taken that any TfRB1-based clinical candidate is engineered with potency at a low enough plasma concentration that iron homeostasis is not appreciably impacted.

TfRB1 variants accumulate in many tissues apart from the CNS. All of the tissues showing high TfRB1 levels (liver, kidney, spleen, skin) are either reported to express high levels of TfR [30,31] or are involved in clearing foreign entities from blood. Nevertheless, TfRB1 variants can achieve levels in the brain capable of inducing a biologically relevant response after a single intravenous dose, even when the CDPs have not been optimized for serum half-life extension or other desirable pharmacokinetic properties. Delivery of molecules with CNS-specific targets can ameliorate the risk of off-target effects due to non-CNS accumulation. For example, NTSR modulation is already used in chronic pain management [42], so TfRB1-NT fusions could plausibly have similar utility. Beyond that, TfR-binding bispecific antibodies containing an anti-BACE arm have been reported to accumulate in the murine brain and specifically inhibit beta secretase activity [35], providing precedent for fusing TfRB1 variants to inhibitors of neurodegenerative or neuroinflammatory processes. Fusion or conjugation to modulators of CNS-restricted ion channels, many of which are themselves CDPs [52,53], is a viable strategy towards this end. In addition to CNS target specificity, addressing chronic neurodegeneration would also require the molecule to be poorly immunogenic, as dosing would likely be repetitive. However, such a stable CDP may turn out to be less likely to produce fragments for display in antigen presenting cells, while precise crystallographic knowledge of the TfR-binding pharmacophore on TfRB1 could permit sequence optimization to further limit MHC binding of CDP fragments. Lastly, CNS tumor targeting is a tantalizing possibility; brain tumors behind an intact BBB are notoriously difficult to treat with chemotherapy, but the high TfR expression found on many tumors [5,6] could provide a mechanism for both BBB passage and tumor accumulation. However it would be used, CDPs like TfRB1 can be produced biologically as fusion proteins or synthetically with chemical conjugation handles, allowing for a number of strategies to incorporate TfRB1 variants’ BBB penetration function into candidate therapeutics for CNS disorders.

In conclusion, we have exploited the structural diversity of native CDP space to identify a novel TfR binding CDP that can deliver a neuromodulatory agent to the brain. TfRB1 variants show potential as drug delivery vehicles to treat otherwise intransigent CNS disorders, and could be adapted to aid in the delivery of both chemical and biologic agents. More generally, this work demonstrates the simplicity and effectiveness of native-derived CDP library screening in mammalian surface display against targets of interest. The addition of native-derived CDP libraries to the current pipeline of target engagement screening can complement existing strategies; each has different strengths and liabilities, but the small size, stability, and diversity of native CDP space merits further exploration.

Methods

Soluble protein methods

Protein production.

TfR, apo-transferrin, and all soluble CDPs were produced and purified as per previous protocols [17,21] at 2 L scale. In brief, protein sequences were cloned into the Daedalus vector, a lentivector driving secretion of proteins of interest. Proteins either contained their native signal peptide or were fused at their N-terminus to siderocalin for stabilization during translation / secretion; siderocalin fusions were cleaved using TEV protease, leaving an N-terminal “GS” behind. (Note: for screened CDP amino acid numbering, we begin after this GS, as it is irrelevant to the surface display scaffold from which the hit arose.) VSV-G pseudotyped lentivirus was produced through standard methods in 293 cells. This virus was used to transduce 1×10⁷ suspension 293 Freestyle (293F, ThermoFisher R79007) cells at an approximate multiplicity of infection (MOI) of 10 in FreeStyle 293 Expression Medium (Invitrogen 12338018). The culture is then grown (suspension culture incubator at 125 RPM, 37°C, 80% humidity, 8% CO₂) and expanded over the course of 10–12 days (2 L final volume), after which media is collected and sterile filtered. The proteins all have a 6xHis tag, so proteins are purified from media by batch Ni-NTA resin binding. CDPs were separated from the siderocalin carrier by TEV protease cleavage, and a final polish with either SEC (ÄKTA Pure, GE Healthcare) or RP-HPLC (Agilent model 1260, C18 column, 214 nm absorbance, with an in-line Agilent 6120 LC/MS) is performed. TfR and transferrin were frozen in PBS + 5% glycerol, while CDPs, purified by RP-HPLC, were aliquoted and lyophilized in vials. CDP quantitation was established by amino acid analysis before use in vivo. Full protein sequences can be found in Table S1. The non-TfR binding CDP (sequence undisclosed) is a mutant variant of a native CDP containing 6 cysteines, and is approximately 4 kDa in mass. Its I-TASSER [23] predicted structural model is shown in Fig. S14.

Quality control.

All proteins were analyzed by SDS-PAGE (NuPAGE 4–12% Bis-Tris, Thermo Fisher) and Coomassie staining; full gels can be found in Fig. S12. CDPs were run both non-reduced and reduced with 10 mM DTT. Proteolysis and reduction resistance testing was performed on the same RP-HPLC instrumentation as used for purification, after treatment of ~10 μg CDP with either 50 U of porcine pepsin (Sigma-Aldrich P7012) in 100 μL simulated gastric fluid [54] or 50 U of porcine trypsin (Sigma-Aldrich 6567) in 100 μL PBS followed by incubation for 30 min at 37°C. Similarly, glutathione or DTT reduction involved treatment of the CDPs with 10 mM glutathione or DTT in PBS for > 5 min prior to RP-HPLC analysis. CD spectra of TfRB1G3 were measured with a Jasco J-720W spectropolarimeter (1 mm path length) in 20 mM phosphate buffer, pH 7.4, with or without 10 mM DTT at a concentration of 15–25 μM. Data are expressed in terms of relative ellipticity [θ] change at 210 nm, reported in mdeg.

Protein labeling.

Holo-transferrin was produced according to the method of Hamilton et al [55]. In brief, a 10 mL solution of 4.5 mM Fe(NTA)₂, containing 4.5 mM FeCl₃ (Sigma Aldrich) and 9 mM nitrolotriacetic acid (NTA, Sigma Aldrich) in water, was titrated to pH 4.0 with HCl and NaOH. Apo-transferrin solution in PBS was loaded with Fe(NTA)₂, 0.4 equivalents at a time every 5 minutes, until 3.2 equivalents had been added. The solution was buffer exchanged into PBS using spin column dialysis (Amicon Ultra-4 centrifugal filter units, 10 kDa molecular weight cutoff) for >1,000-fold free Fe(NTA)₂ clearance.

Alexa Fluor 647 (Thermo Fisher A37573) CDP labeling took place in PBS buffer with a molar ratio of ~1.5 NHS-ester dye : 1 CDP for > 1 hour, after which free dye was removed by spin column dialysis (Amicon Ultra-4 centrifugal filter units, 3 kDa molecular weight cutoff) for >1,000-fold free dye clearance.

Biotinylation of Avitagged TfR used the BirA-500 kit (Avidity) as per manufacturer’s protocol, followed by a final buffer exchange into PBS containing 5% glycerol, and storage in small aliquots at −80°C.

¹⁴C labeling was based on the procedure of Jentoft and Dearborn [29]. Briefly, CDPs (10 mg) were dissolved in 3.2 mL water with 470 μL 10x PBS. 0.72 mCi (12.6 μM) ¹⁴C-formaldehyde (57 Ci mol⁻¹, Pharmaron) and sodium cyanoborohydride (to 100 mmol, in water) were added, followed by vortexing for 15 seconds. The reactions were incubated at RT overnight. 10 μL reaction was set aside in 1 mL water for analysis, after which the ¹⁴C methylated CDP was purified on a Strata-X reversed-phase column (30 mg, Phenomenex, washed with 3 mL methanol and 3 mL water). After loading, the column was washed with 4 mL water and eluted with 4 mL 2% formic acid in methanol. Labeled CDPs were dried in a blow-down evaporator (40°C under nitrogen stream) and stored at 18°C until use. Specific activity was determined by scintillation counting. TfRB1 variant specific activities were as follows. TfRB1G1: 32 Ci mol⁻¹ for 180 minute animals, 187 Ci mol⁻¹ for 30 minute animals. TfRB1G2: 260 Ci mol⁻¹ for 180 minute animals, 179 Ci mol⁻¹ for 30 minute animals. TfRB1G3: 273 Ci mol⁻¹ for 180 minute animals, 147 Ci mol⁻¹ for 30 minute animals.

CDP library construction

CDP constructs were designed as previously described [20]; in brief, the Uniprot [56] database was screened for protein segments of 30–50 amino acids in length and containing 6, 8, or 10 cysteines. This list was winnowed down with an emphasis on annotated knottins and defensins but otherwise with an eye on phylogenetic diversity, supplemented by a limited number of CDPs of interest from other sources, to a final library size of 10,000 CDPs. IP considerations prevent us from disclosing the protein sequence or species composition of the library. Oligonucleotides of the library containing appropriate flanking sequences for amplification and cloning were ordered (Twist Bioscience), amplified (Phusion® High-Fidelity DNA Polymerase, NEB), and mutagenized (GeneMorph II Random Mutagenesis Kit, Agilent; 22 cycles of amplification) prior to Gibson Assembly cloning into BamHI / NotI digested SDGF vector [20] (GenBank MF958494). The library was transformed into Stellar competent cells (Clontech), maxiprepped (QIAgen), and made into lentivirus (VSV-G pseudotyped, produced by standard methods in 293T cells using the envelope plasmid pMD2.G and the packaging plasmid psPAX). Singleton CDP candidates that required individual cloning were cloned from dsDNA (gBlock, Integrated DNA Technologies) into SDGF and tested for binding 2 days after transient transfection (2.5 μg plasmid transfected with 4 μg polyethyleneimine (Polysciences 23966–1) into 2×10⁶ 293F cells in 1 mL FreeStyle media, shaking in 24-well suspension plates as described above. Testing consisted of staining with TfR and dye-labeled streptavidin, either together or sequentially (TfR followed by pelleting and resuspension in streptavidin solution) in Flow Buffer with 1 μg mL⁻¹ DAPI, and then analyzed on an Acea NovoCyte flow cytometer. For transferrin competition experiments, 10 μM apo-transferrin or holo-transferrin were added to the TfR staining solution prior to resuspension of cells in the stain.

Mammalian surface display screening for TfR binders

Our general mammalian surface display CDP screening methods have been published elsewhere [57]. Details specific to this work are described here.

Magnetic cell sorting.

2×10⁸ 293F cells were transduced with the SDGF CDP library (representing roughly 10⁷ CDP genes with intact cystine architecture after mutagenizing the 10,000 member library) at an MOI of ~1, and expanded until 3 days post-transduction. For initial screening, magnetic cell sorting was performed as described [57] using 200 nM biotinylated TfR, 2 mL anti-biotin MicroBeads (UltraPure, Miltenyi 130–105-637), and 21 mL Flow Buffer (PBS + 2 mM EDTA and 0.5% bovine serum albumin) in a final volume of 25 mL. After agitation, dilution, and rinsing, cells were split into four 10 mL aliquots and run through a Miltenyi autoMACS® Pro Separator using the “posseld” protocol and “quick rinses” after each sort. The running and wash buffers were High BSA Flow Buffer and PBS + 2 mM EDTA, respectively. Eluted cells were pooled, pelleted, and their CDP sequences PCR amplified (Terra™ PCR Direct Polymerase Mix (Takara 639271) for 16 cycles followed by Phusion). This sub-library was cloned into SDGF as above, made into lentivirus, and transduced into a new batch of 293F cells (1×10⁷ cells, MOI ~1) for flow sorting.

Flow sorting.

Flow sorting took place as described [57] using 2.4×10⁷ cells stained in 3 mL Flow Buffer with 200 nM TfR, 200 nM streptavidin Alexa Fluor 647 conjugate (ThermoFisher S21374), and 1 μg mL⁻¹ DAPI. Cells were sorted on a FACSAria II System (BD), gating based on FSC-A (medium), SSC-A (medium), DAPI-A (negative), GFP-A (positive), and APC-A (top 7% of GFP+) channels. An example of our standard gating criteria to identify SDGF-expressing 293F cells is shown in Fig. S18. After each flow sort, cells were cultured in FreeStyle Media, expanded, and re-sorted as above. After the third flow sort, cells were expanded and frozen in 1.5×10⁶ cell pellets. Pellets were PCR amplified as above (Terra Direct PCR followed by InFusion), CDP inserts were subcloned into SDGF, transformed (Stellar competent cells), and colonies picked for miniprepping and sequencing of the cloned CDPs. Enriched variants were those that appeared in the sequence analysis more than once in 94 picked colonies; for the TfR binder screen, five enriched variants (representing 68 of 94 colonies) were tested for TfR binding, of which one (representing 44 of 94 colonies) was a bona fide TfR binder. All site-saturation mutagenesis (SSM) affinity maturation screening was done as above, with the following two changes. First staining was sequential; first with TfR, and then with an equimolar amount of dye-labeled streptavidin. Second, TfR / streptavidin concentrations were reduced to 20 nM for the first SSM maturation and to 8 nM for the second SSM maturation screen. After each SSM screen, enriched variants were studied to assemble a compound mutant (TfRB1G2 or TfRB1G3) that showed higher TfR staining than any of the variants containing either 1 or 2 of the individual mutations.

Surface plasmon resonance (SPR) interaction analyses

SPR experiments were performed at 25°C on a Biacore T100 instrument (GE Healthcare) with Series S streptavidin sensor chips. 10 mM HEPES (pH = 7.4), 150 mM NaCl, 3 mM EDTA, and 0.05% v/v surfactant P20 (HBS-EP+) plus 0.1 mg mL⁻¹ bovine serum albumin was used as running buffer unless otherwise noted. Biotinylated TfR at 1–2 μg mL⁻¹ was injected at 10 μL min⁻¹ to capture the RU levels noted below. All experiments used a blank streptavidin flow cell for referencing. Apo-Tf and holo-Tf at 40 nM in 10 mM HEPES (pH= 7.4), 150 mM NaCl, and 0.05% v/v surfactant P20 (HBS-P+) were injected over 56 RUs of captured human TfR for 5 minutes at 50 μL min⁻¹ and allowed to dissociate for 5 minutes. Double-referenced [58] data were analyzed with BiaEvaluation 2.0.4 software (GE Healthcare). A direct comparison of the off-rates of TfR binders was made on a captured human TfR surface with a density of 516 RUs. TfRB1G1, G2, and G3, each preceded by a buffer blank, were injected at 300 nM for 3 minutes at 50 μL min⁻¹ and allowed to dissociate for 5 minutes. An overlay plot of double-referenced data was generated and normalized for off-rate comparison by dividing each curve by its maximum response (2.5, 26.5, and 27 RUs for TfRB1G1, G2 and G3, respectively) with Scrubber v2.0b software (BioLogic Software). TfRB1G2 binding was run on a captured human TfR surface coupled at a density of 1308 RUs. Serial two-fold dilutions of TfRB1G2 (178 nM to 87 pM) and buffer blanks were run in duplicate, injected in a random order, at a flow rate of 50 μL min⁻¹ with 5 minutes of association and 5 minutes of dissociation. Double-referenced data were fit with a steady-state 1:1 interaction model with BiaEvaluation 2.0.4 software. TfRB1G3 binding was measured with human TfR captured on two flow cells at different densities (516 and 1032 RUs) using a kinetic titration method. Three buffer blank cycles were run prior to the TfRB1G3 cycle and averaged for double-referencing. TfRB1G3 was injected sequentially at 0.037, 0.11, 0.33, 1, and 3 nM at 50 μL min⁻¹ for 5 minutes with a final dissociation of 10 minutes. Data from both surfaces were analyzed globally with a single cycle kinetic analysis model [59] with BiaEvaluation 2.0.4 software. Full binding curves can be found in Fig. S5. SPR measurements are presented with error in the text and figures, but note that this error represents a precision estimate based on fitting residuals, rather than an accuracy estimate based on replicate measurements.

Crystallography

Lyophilized TfRB1G3 was reconstituted in 25 mM PIPES (pH = 7.0), 150 mM NaCl, and 1 mM EDTA at a concentration of at 11 mg mL^-1. Preliminary crystallization conditions were determined using the JCSG-plus sparse matrix screen (Molecular Dimensions) in a vapor equilibration format. Diffraction-quality crystals were obtained by vapor equilibration of drops comprising 1 μL protein solution plus 1 μL well solution over wells of 1 mL 9% w/w PEG (MW = 3350), 30 mM (NH₄)₂SO₄, 100 mM Bis-Tris (pH = 5.0) at ambient temperature. Crystals were cryopreserved for data collection by transferring into well solution plus 20% v/v glycerol. Diffraction data (dmin = 2.55 Å) were collected on an in-house Rigaku MicroMax-007HF rotating anode generator with Osmic Varimax optics and a Rigaku Saturn 944+ CCD detector (Table 1). To isolate stoichiometric complexes, purified TfRB1G3 and TfR were mixed at a 1.2:1 molar ratio, incubated, and fractionated by size exclusion chromatography in 25 mM PIPES, 150 mM NaCl, and 1 mM EDTA (pH = 7.0) on a Superose 6 10/300 column (GE Healthcare Life Sciences). The complex was concentrated to 10 mg mL⁻¹ and preliminary crystallization conditions were determined using the JCSG-plus sparse matrix screen (Molecular Dimensions) in a vapor equilibration format. Diffraction-quality crystals were obtained by vapor equilibration of drops comprising 1 μL protein solution plus 1 μL well solution over wells of 1 mL 5% w/w PEG (Mr = 6000), 100 mM MES (pH = 6.5) at ambient temperature. Crystals were cryopreserved for data collection by transferring into well solution plus 15% v/v glycerol. Diffraction data (dmin = 1.85 Å) were collected on beamline 5.0.1 at the Advanced Light Source (Berkeley, CA) at a wavelength of 1.0000 Å. Both data sets were indexed and scaled using the HKL2000 software package [60]. Twin refinement in REFMAC5 [61] was needed to produce interpretable electron density maps for apo-TfRB1G3. Initial phases were determined by molecular replacement using a structure of TfR (3KAS.pdb) for the TfR/TfRB1G3 complex, or the refined structure of TfRB1G3 from the complex structure for the apo structure, as search models with PHASER [62], as implemented in the CCP4 software suite [63,64]. Modeling and additional refinement was performed using Coot and REFMAC5 [61,65], using the I-TASSER [23] computational model of TfRB1G3 as an initial building scaffold. Molecular images were generated with MacPyMOL (Version 2.0, Schrodinger). The cartoon representations of free and TfR-bound TfRB1G3 in Fig. 3a are superpositions of all molecules of the asymmetric unit: four for free TfRB1G3, two for TfR-bound.

HEK293-NTSR1 activity assay

To demonstrate that the neurotensin extension on various proteins was functional, NTSR activity in HEK293 cells, or HEK293 cells transduced with a lentivector delivering human NTSR1 (HEK293-NTSR1), was measured using the IP-One – Gq kit (CisBio 62IPAPEB). Cells were grown in DMEM + 10% fetal bovine serum, removed from the plates with Accutase, pelleted, and suspended in Hanks Buffered Salt Solution at a density of 1.5×10⁶ cells per mL. HTFR reactions were set up in HTFR 96 well low volume plates (CisBio #66PL96025) according to the manufacturer’s instructions. 10,000 cells (7 μL) were used per 25 μL reaction. The plate was incubated for 60 mins at 37°C. Then 3 μL IP1-d2 working solution was then added, followed by 3 μL Anti IP1-Cryptate working solution. After a 1 hour incubation at room temperature, the plate was scanned in a Perkin Elmer 2104 EnVision Multilabel Reader for fluorescence emission after excitement at 665 nm and 620 nm wavelengths. FRET ratio was calculated as 10,000 x (Signal 665 nm / Signal 620 nm).

In vitro serum stability testing

Lyophilized CDPs were reconstituted to 1 mg mL⁻¹ in mouse serum (Invitrogen #10410), divided among three separate tubes, and incubated at 37°C for 8 hours. At each time point (0, 0.5, 1, 2, 4, and 8 hours), a 40 μL aliquot was removed and incubated at 95°C for 10 min. Precipitated serum proteins were pelleted by centrifugation at 17,000 x g for 5 min. Supernatant was collected, and the pellet was then washed twice with 100 μL PBS using pipetting and gentle vortexing to resuspend. Post-wash supernatants were pooled with the initial post-boil supernatant and then spin filtered through a 0.22 μm PVDF membrane (Millipore #UFC30GVNB). Filtrate was run on RP-HPLC as described in the Soluble Protein Methods section. Peaks representing intact CDP (as confirmed by MS) were used for area under the curve recording.

Animals

All rodents were maintained in accordance with the NIH Guide for the Care and Use of Experimental Animals with approval from the Fred Hutchinson Cancer Research Center Institutional Animal Care and Use Committee (IR # 50808 and 50868). Adult female athymic nude mice (Envigo #069) were used for plasma and tissue pharmacokinetic analysis and whole body autoradiography. Adult male and female CRE-luciferase GPCR reporter mouse, line 187 (Taconic, #11520) were used for in vivo CNS luminescence. Males and females were distributed evenly amongst all treatment groups. Mice had unrestricted mobility in their cage with free access to food and water between drug administration and study end point.

Plasma and tissue pharmacokinetic analysis

To quantify accumulation of CDP in tissue and in circulation, mice were administered 100 nmol of ¹⁴C labeled CDP in 100 μL PBS by intravenous tail vein injection. CDP was allowed to circulate for pre-specified time points between 15 minutes and 48 hours. Mice were humanely euthanized by CO₂ asphyxiation in accordance with American Veterinary Medical Association guidelines. Whole blood was harvested by cardiac puncture at euthanasia into K2-EDTA microtainer tubes (Fisher Scientific #02–669-33) and processed by centrifugation into plasma. Tissues were resected and frozen, before being homogenized. 20 μL of plasma or tissue homogenate was analyzed by liquid scintillation counting in Ultima-Gold scintillant (PerkinElmer, #6013329) in a 5 mL glass vial (PerkinElmer #6000167) and counted on a Packard TriCarb scintillation counter (PerkinElmer).

Whole body autoradiography

¹⁴C labeled CDPs were resuspended to 100 nmol in 100 μL PBS for intravenous injection into mice. After 30 minutes or 3 hours, mice were euthanized by being deeply anesthetized with ketamine-xylazine and submerged in a dry-ice chilled hexane bath for 15 minutes. Frozen carcasses were allowed to off gas hexane overnight at −20°C prior to embedding in chilled 2% carboxymethylcellulose (Sigma Aldrich, C5013). Holes were then drilled into each block and radioactive ¹⁴C glycine controls (0.5 μCi mL⁻¹, American Radiolabeled Chemicals) in PBS + 0.5% BSA were pipetted in and allowed to freeze, prior to sectioning (40 μm) on a H/I Bright 8250 Cryostat (Hacker Instruments). Sections were collected onto 4-inch wide tape (Scotch 821, ULINE) at 2–6 depths to sample all of the tissues of interest, and freeze dried in the cryostat for 2–3 days, after which they were mounted on sturdy paper and covered with cellophane. Mounted sections were exposed to storage phosphor plates (Raytest) along with a radioactive standard curve (146S-PL, American Radiolabeled Chemicals) for 7 days. Scanning took place on a Raytest CR-35 imager at the “25 μm sensitive” setting. Quantitation was performed by background-adjusted densitometry using AIDA Image Analyzer v5.1 Whole Body Autoradiography Professional software (Raytest). For quantification, all region of interest measurements from a single tissue within a single section were averaged, adjusting for area. Blood values were assessed in the heart. Reporting values in nmol g⁻¹ tissue required calibration to the commercial standard strips co-exposed with each WBA sample set. The strips were themselves calibrated so that a known amount of radioactive counts per minute (CPM) per gram of tissue, determined by liquid scintillation counting (LSC), corresponded to the measured WBA densitometry at 11 standard strip concentration data points (linear range corresponding to 6.3 nCi g⁻¹ to 3.3 μCi g⁻¹). This calibration, combined with the CDPs’ known concentrations and specific activities (determined by liquid scintillation counting), permitted the conversion to nmol g^-1. The limit of detection for this method is 14 nCi g⁻¹, well below all measured values. For the quantitation shown in Fig. 5, total number of sections analyzed is indicated in the key; not every section contained every tissue, but every tissue quantitation represents 1–3 regions of interest per section over 1–4 sections per mouse from 3–10 mice per treatment (TfRB1G1 or TfRB1G3, 30 or 180 minutes). A sample image from an unrelated radiolabeled CDP is shown in Fig. S11, providing an example of how regions of interest for quantitation are derived.

In vivo CNS luminescence

CRE-luc GPCR reporter mice were administered 1.68 mg rolipram (Sigma Aldrich R6520) and 0.84 mg forskolin (Sigma Aldrich 344270) by intraperitoneal (IP) injection (200 μL in 17% DMSO solution), or TfR binder by intravenous tail vein injection; dosages were 100 nmol TfRB1 variant or 12 nmol transferrin in PBS. Vials of lyophilized CDPs were allowed to reach room temperature before resuspension in PBS to the final injection concentration. 4 hours after administration, mice were administered 100 μL of 30 mg mL⁻¹ D-Luciferin (RediJect D-Luciferin Ultra, PerkinElmer 770505) by IP injection and anesthetized by isoflurane exposure after 10 minutes of unrestricted activity. Luminescence was measured in anesthetized mice on a Xenogen IVIS instrument with a 1 min exposure. Identical regions of interest encompassing the entire brain were draw and quantified using LivingImage software (version 4.0, PerkinElmer). Mice were censored if their luminescence levels were > 4 SD from the mean / SD luminescence of the other mice in the cohort, or if their luciferin injections failed to hit the peritoneal cavity. For the former, this was interpreted as an incidental physiological response unrelated to drug treatment or NT receptor modulation. For the latter, this was identified by failure of a fluorescent dye (pre-formulated in the RediJect D-Luciferin Ultra, read on the ICG filter set over 5 seconds) to distribute evenly throughout the animal’s abdomen; this indicates insufficient perfusion and incomplete exposure of the whole mouse to luciferin, rendering the animal’s brain luminescence quantitation questionable. Mice were treated as matched cohorts, meaning a given group of animals were serially visualized unstimulated, treated with a given TfR-binder NT-fusion, or treated with the relevant non-NT fusion TfR binder control. Mice were given at least 1 week between each treatment (or lack thereof) / visualization; internal controls using forskolin and rolipram demonstrated 1 week to be sufficient time for luciferase reporter values to return to baseline after stimulation. Animal counts for non-NT fusion TfR binders (Fig. 6c) were lower compared to that of NT fusions because 3 NT fusion animals per cohort were euthanized for IHC analysis (Fig. 6d) in order to conserve animals.

Luciferase immunohistochemistry

Formalin-fixed, paraffin-embedded brains were sectioned on a Microm HM 355S microtome at 6 μm and mounted on positively-charged Superfrost Plus slides (Fisher 12–550-15). Staining was automated on a Ventana Discovery Ultra IHC/ISH autostainer using kits and reagents designed for use on this instrument: EZ Prep deparaffinization reagents (Ventana 950–100), Protease 3 antigen retrieval kit (Ventana 760–2020), anti-rabbit HQ (Ventana 760–4815), anti-HQ-HRP (Ventana 760–4815), ChromoMap DAB kit (Ventana 760–159), hematoxylin (Ventana 760–2021), and bluing reagent (Ventana 760–2037), all at manufacturer’s default settings. The primary anti-luciferase antibody was from Abcam (ab21176, concentration 1:350) and was diluted with Ventana antibody diluent with casein (Ventana 760–219). The precise protocol settings for the instrument are available upon request, but in brief, the sequence was as follows. Sections were deparaffinized at 69°C, and then antigen retrieval took place at 35°C for 4 mins. Slides were rinsed and warmed to 37°C prior to primary antibody addition by hand (1:350). Slides were incubated with the primary antibody for 28 mins and rinsed. Subsequent antibody additions (followed by rinsing) were anti-rabbit HQ and then anti-HQ HRP, incubating 16 mins apiece. Slides were then DAB treated and counterstained with hematoxylin and bluing reagent (8 mins apiece). For use in figures, images were mildly contrast-enhanced, with all image modifications performed identically across all images within a set without regard to treatment group. The full images used for cropping in Fig. 6d can be found in Supplemental Data.

Data and reagent availability

Raw data supporting the findings that were not in the manuscript, and samples of the CDPs and proteins used in this study, will be made available upon reasonable request except where restricted by FHCRC IP considerations. The SDGF vector is available upon request, pending a Materials Transfer Agreement with the FHCRC. The diversified native CDP library is not available for distribution, but our group is enthusiastic about collaborations on targets of mutual interest. The control non-TfR binding CDP is not available as of manuscript submission due to IP considerations. Detailed protocols for SDGF mammalian surface display screening of CDP libraries were recently published [57], and the authors will share additional protocols and advice upon request.

Accession numbers

Mammalian surface display vector expression cassette (GenBank accession no. MF958494), TfRB1G3 crystal and co-crystal structures (RCSB PDB IDs 6OKE and 6OKD, respectively).

Highlights.

• Blood-brain barrier penetration is key to treating central nervous system diseases

• A drug-like cystine-dense peptide that binds transferrin receptor was identified

• Brain accumulation and pharmacodynamic effect of peptide cargo were seen

• This peptide represents a small (5 kDa) tag permitting CNS delivery of therapeutics

Glossary

Affinity maturation

systematic mutagenesis of a screened target-binder to identify mutations that improve affinity and/or stability characteristics.

Apical and basal surfaces

in vascular endothelial cells, the apical surface is luminal (exposed to circulating blood), while the basal surface faces the tissue parenchyma.

Cystine-dense peptide

one of a structurally and taxonomically diverse class of miniproteins stabilized by a core of multiple cystine disulfide bridges; includes knottins and defensins.

Ectodomain

for a transmembrane protein (e.g. transferrin receptor), the extracellular portion of the protein expressed without transmembrane or intracellular domains.

Miniproteins

proteins of ~1–10 kDa from diverse native or synthetic sources characterized by rigid structure and versatility in drug-like roles, with capabilities and pharmacology distinct from that of conventional small molecule drugs or larger proteins (e.g. antibodies).

Native

for proteins, unmodified from the sequence found in its source organism.

Parenchyma

an organ’s functional cellular tissue in contrast to connective tissue and vascular components; for the CNS, its neurons and supporting glial cells.

SDGF

Surface Display GFP FasL, the vector for mammalian surface display used in this study made up of an intracellular GFP and a transmembrane and truncated extracellular domain from human Fas ligand used to tether molecules to the surface of mammalian cells (here, HEK-293 or derivatives).

Abbreviations

BBB

blood-brain barrier

CNS

central nervous system

CDP

cystine-dense peptide

TfR

transferrin receptor

SDGF

surface display GFP FasL vector

293F

human embryonic kidney cell line 293 Freestyle

SSM

site saturation mutagenesis

TfRB1G[1, 2, or 3]

TfR binding CDP, generation [1, 2, or 3]

HPLC

high performance liquid chromatography

NT

neurotensin

NTSR1

neurotensin receptor 1

CRE

cyclic AMP response element