Oral IRAK-4 Inhibitor CA-4948 Is Blood-Brain Barrier Penetrant and Has Single-Agent Activity against CNS Lymphoma and Melanoma Brain Metastases

The advent of immunotherapy in the form of immune checkpoint inhibition (ICI) and the development and refinement of small molecule inhibitors, namely tyrosine kinase inhibitors, have revolutionized oncologic practice. Significant shortfalls in current therapy remain for diseases that develop in or metastasize to the brain (1, 2). Primary brain malignancies and brain metastases represent a significant driver of cancer mortality, due in large part to a lack of drugs with specific design or activity in the brain space (3–5). The restrictive nature of the blood-brain barrier (BBB) creates a physical barrier to drug entry; thus, only 2% of all FDA-approved anticancer agents are capable of reaching significant activity in the brain (6). Further, the unique characteristics of the brain tumor microenvironment (TME) reduce the efficacy of drugs that reach their target through the activation of reactive astrogliosis and astrocyte drug efflux pumps coopted by tumors (7–9). In addition, brain metastases express altered molecular mutations and characteristics compared with their parent visceral tumors, creating a heterogeneous tumor landscape further hindering treatment (10, 11). Cancers that are very responsive to treatment when present in circulation or visceral tissue, but become more recalcitrant to treatment when sequestered beyond the BBB include lymphoma, melanoma, and breast cancer (12–15). Lymphoma specifically carries a dichotomous prognosis and survival when comparing stage IV diffuse large B-cell lymphoma (DLBCL) and primary central nervous system lymphoma (PCNSL), even though these tumors share many characteristics and mutations.

PCNSL most commonly represents a non-germinal B cell (non-GCB) lymphoma, an aggressive subtype of DLBCL. Although peripheral non-GCB lymphomas are also aggressive, the overall survival, relapse rate, and rate of progression-free survival are worse for PCNSL (16, 17). The standard of care for PCNSL involves induction chemotherapy using a backbone of high-dose methotrexate (HD-MTX) in a polychemotherapy regimen in patients who are young or with good performance status (18). Consolidation with additional chemotherapy, autologous stem cell transplant, or whole-brain radiotherapy is then employed to improve the rate of survival (18, 19). This need for multi-agent chemotherapy, which requires hospitalization and carries significant toxicity, limits many patients from being candidates for treatment, particularly the elderly (16, 18, 20). A recent discovery in DLBCL is the constitutive activation and upregulation of nuclear factor kappa B (NF-κB) signaling through driver mutations in myeloid differentiation primary response protein 88 (MyD88) and its downstream chaperone interleukin-1 receptor-associated kinase-1 and 4 (IRAK-1 and 4; ref. 21).

MyD88 is part of the inflammasome complex acting as an intracellular signaling molecule downstream of Toll-like receptor and/or interleukin-1 receptor (IL-1R) activation (22, 23). When these receptors bind to their target ligands, MyD88 co-localizes with chaperone molecules IRAK-1 and IRAK-4 (23). This complex facilitates the activation of multiple cellular kinases and transcription factors [c-Jun N-terminal kinase (JNK), activator protein-1 (AP-1), mitogen-activated protein kinase (MAPK), and NF-kB], inducing the production of inflammatory chemokines and cytokines (24, 25). The commonly shared point mutation MYD88 L256P is seen across the spectrum of B-cell malignancies, including roughly 60% of PCNSL (21). Focusing on the role of MyD88 across inflammatory disorders, recent reports indicate that chronic activation of this cascade results in defective immunologic responses and can perpetuate a wide range of inflammation-associated diseases (26). Constitutive activation of known oncogenic signaling cascades including MAPK and NF-κB is a direct consequence of aberrant MyD88 and IRAK-4, creating an ideal environment for tumor initiation and progression across numerous malignancies beyond lymphoma (27–30). Studies in human melanoma have shown that IRAK-4 is upregulated and activated in most cutaneous tumor samples (31, 32).

In contrast to diseases like DLBCL that carry activating mutations in MYD88, phospho-IRAK-4 activation in melanoma occurs through response to the inflammatory milieu of the TME (33). The resulting inflammatory cascade leads to an autocrine feedback loop co-opted by tumor to propagate tumor cell proliferation, survival, and metastasis (32, 33). Similar to PCNSL, melanoma brain metastases (MBM) carries a worse prognosis and response to therapy than systemic melanoma (25, 34). In this manuscript, we explore the role of MyD88/IRAK-4 in central nervous system (CNS) diseases, contrasting mutationally driven PCNSL with inflammation-driven MBM. Our results implicate the MyD88/IRAK-4 signalosome, or myddosome, as a novel target of therapy in PCNSL and MBM where the novel oral IRAK-4 inhibitor CA-4948 is a compelling agent for treatment in this space.

A20 and B16F10 cells were purchased from ATCC. OCI-LY3 and SU-DHL-4 were generously shared by Dr. Han Tun, with genomic profiling performed November 2022 (GENOMIC). A20 cells were cultured in RPMI (Fisher-Scientific) supplemented with 10% FBS (VWR) and 1% Penn-Strep (Life Technologies). OCI-LY3 and SU-DHL-4 cells were cultured in RPMI (Fisher-Scientific) supplemented with 20% FBS (VWR) and 1% Penn-Strep (Life Technologies). B16F10 cells were cultured in DMEM (Fisher-Scientific) supplemented with 10% FBS (VWR) and 1% Penn-Strep (Life Technologies). All cell lines were maintained at 37°C in humidified conditions with 5% CO₂. At the beginning of the study, cells were expanded, stocks made, and thawed vials were maintained in culture for no more than 3 weeks.

BALB/c, C57BL/6J, and athymic nude mice were purchased from Jackson Laboratory. Protocols were reviewed and approved by the University of Florida Institutional Animal Care and Use Committee. Five×10⁴ tumor cells suspended in 50% methylcellulose and 50% saline (Fisher-Scientific) were stereotaxically injected into murine brain at a depth of 3 mm, 2 mm lateral to bregma, at a volume of 2 μL in 8- to 12-week-old animals.

CA-4948 provided by Curis, Inc. CA-4948 was prepared at 10 mg/mL in vehicle containing 0.25% w/v hydroxyethyl-cellulose (Sigma-Aldrich) and 0.5% w/v Tween-20 (Sigma-Aldrich), administered by oral gavage. For in vitro studies, CA-4948 was solubilized in DMSO at a stock concentration of 10 mmol/L.

De-identified patient tissues were procured by the Florida Center for Brain Tumor Research and 1Florida ADRC Brain and Biospecimen Repository under the University of Florida Institutional Review Board protocols 201300482 & 202002826.

Tissues were extracted from snap-frozen patient specimens on dry ice. NF-κB proteomic profiling was done using Proteome Profiler Array Kit (R&D Systems, #ARY029) according to the manufacturer's protocol. Protein concentration determined via NanoDrop, with 200 μg of lysate used for each preparation. Images captured using Bio-Rad ChemiDoc MP Imaging System with ImageLab 6.1 software over a series of exposure times. Mean voxel intensity per capture antibody was calculated using Imaris x64 v9.7.0, and protein signal was normalized against internal reference controls.

Crystal structure of CA-4948 bound to human IRAK-4 (7C2V; refs. 35, 36) was downloaded using the Protein Preparation Wizard in the Schrödinger suite (37, 38). Homology modelling was conducted using the Homology Modelling program within Schrödinger (37, 39–41). The homology model was constructed using mus musculus IRAK-4 Fasta sequence Q8R4K2 obtained from Uniprot (42), with 7C2V used as the template. Energy-based modelling was conducted. For docking, the ligand CA-4948 had already been prepared during the protein preparation of 7C2V. Docking was performed using GlideXP (37, 43), with the grid box centered around the crystal structure ligand centroid (Inner Box: 10 × 10 × 10; Outer Box: 34 × 34 × 34). The ligand was docked to both human IRAK-4 (7C2V) and the generated mouse homolog, with the results analyzed on the basis of docking score and root-mean-square deviation (RMSD) to the input ligand. Generated poses were not restricted to the input structure conformation in any manner. The top poses (based on docking energy) were exported as pdb files, along with their respective receptors, and submitted to DeepAtom (44–46) for machine learning–based rescoring.

A total of 10⁴ tumor cells were plated per well in 12-well plates in 1 mL of complete media. After 12 hours, CA-4948 or DMSO was added at a dilution factor of 1:1,000. Media and fresh drug were applied on day 3 of the experiment. Cell count and viability were determined on day 5 using a Vi-CELL XR Analyzer (Beckman-Coulter). In ‘chronic’ proliferation studies, media and drug were reapplied on day 3, whereas only media was refreshed in ‘acute’ samples.

Analysis in tumor-bearing mice was performed 2 weeks postimplantation. Animals were euthanized at 0.25, 0.5, 1, 4, and 8 hours following oral CA-4948 treatment at 150 mg/kg⁻¹ in 8- to 10-week-old BALB/c mice. Plasma was separated by centrifugation at 850 × g at 4°C ×10 minutes and stored at −80°C following cardiac puncture. Whole brain was rinsed with ice-cold saline to remove extraneous blood, and stored at −80°C. Brains were homogenized with 3 volumes of 5% BSA. Cerebrospinal fluid (CSF) was extracted from the cisterna magna and stored at −80°C. Ultraperformance LC/MS-MS (UPLC/MS-MS) was done using a Waters Acquity Class I Plus UPLC with a Waters Xevo TQ-S Micro triple quadrupole mass spectrometer. Chromatographic separation was achieved using Acquity UPLC BEH C18 column (2.1 mm × 50 mm, 1.7 μm) and the mobile phase consisted of 0.1% formic acid (A)–acetonitrile (B) with a gradient program of 90% A held for 0.1 minutes, then decreased A to 50% reaching 2.5 minutes and held at 50% until 3.0 minutes, then sharply decreased back to the initial conditions by 3.1 minutes and maintained until 3.5 minutes to equilibrate the column. The column and autosampler temperatures were kept at 40°C and 4°C, respectively. The mobile phase was delivered at a flow rate of 0.35 mL/min and injection volume was set to 2 μL. MassLynx software v4.2 was used for instrument control and TargetLynx for data analysis. The mass spectrometer was operated in positive ion mode and detection of the ions was performed in MRM mode, monitoring transition of m/z 492.16 precursor ion [M+H]⁺ to the m/z 286.18 product ion for CA-4948, m/z 180.12 precursor ion [M+H]⁺ to the m/z 110.03 product ion for internal standard (phenacetin). A simple protein precipitation method was used for the extraction of CA-4948 and removal of endogenous impurities from all samples prior to analysis. Plasma/CSF samples were precipitated with acetonitrile containing 0.1% formic acid and 25 ng/mL of phenacetin as internal standard at a ratio of 1:3 and vortex-mixed for 5 minutes followed by filtration through 0.45-μm filter plates. Brain samples were homogenized in water at a ratio of 1:2 and processed using the same method as plasma samples. Linearity in the range of 5 to 1,000 ng/mL or ng/g was prepared for CA-4948.

Five days following intracranial tumor implantation, animals were treated with CA-4948 (100 mg/kg, orally) for a total of 5 days. Brain tissue was collected 1 hour after final treatment following cardiac perfusion with cold saline followed by PBS supplemented with 4% acrylamide (Sigma-Aldrich), 0.05% N,N’-methylenebisacrylamide (Sigma-Aldrich), 4% paraformaldehyde and 0.25% VA-044 (TCI America), and 0.5 mg/mL fixable Fluorescein-labeled Dextran, 70,000 MW (Thermo-Fisher). Tissues were stored at 4°C for 3 days to allow hydrogel permeation of tissues. Following hydrogel polymerization at 37°C × 3 hours, whole brain was sectioned to 2 mm and passively cleared over 3 to 7 days with PBS containing 200 mmol/L boric acid (Sigma-Aldrich) and 4% SDS (Fisher-Scientific), pH 8.5 at 50°C. After clearing, samples were washed in PBS with 0.1% Triton X-100 for 2 days, and immunostained at 4°C for 2 days each with: phospho-P38 (Thermo-Fisher, catalog no. MA5–15177), phospho-ERK1/2 (Cell Signaling Technology, catalog no. 4370), goat anti-rabbit Alexa Fluor 647 (Thermo-Fisher, catalog no. A-21244), and DAPI (Sigma-Aldrich). Samples were whole-mounted onto slides using 62% 2,2′-Thiodiethanol (Sigma-Aldrich). Images were acquired using a Nikon A1RMP confocal microscope and analyzed using Imaris x64 v9.7.0 software.

Five days following intracranial tumor implantation, animals were treated with CA-4948 (100 mg/kg, orally) for a total of 5 days. One hour following the final dose, tumor tissue was resected and protein lysis, antibody hybridization, and chemiluminescent imaging was performed as previously described (47). Primary antibodies include: (Cell Signaling Technology) phospho-P38 (catalog no. 4511), P38 (catalog no. 9212), phospho-ERK1/2 (catalog no. 4370), ERK1/2 (catalog no. 4695), phospho-NF-κB P65 (catalog no. 3033), NF-κB P65 (catalog no. 8242), and β-actin (catalog no. 4970), followed by species-specific horseradish peroxidase IgG (catalog no. 7074). Images were taken using a Bio-Rad Gel Doc XR+ Imaging System. Signal intensity of individual bands was measured using Imaris x64 v9.7.0 imaging software, and is directly proportional to protein concentration. Target expression levels are normalized against β-actin.

Frozen brain tissue was cryosectioned at 12 μm, mounted onto TruBond 380 slides (Fisher-Scientific), fixed in 10% buffered formalin (Fisher-Scientific) for 10 minutes, and washed 3x in PBS for 5 minutes. Slides were blocked using SuperBlock (Thermo-Fisher) containing 0.3% Triton X-100 for 1 hour at room temperature. Samples were immunostained with primary antibodies overnight at 4°C: (Cell Signaling Technology) phospho-NF-κB P65 (catalog no. 3033T), phospho-ERK1/2 (catalog no. 4370S); (Thermo-Fisher) CD45 (catalog no. 14–0451–82 and catalog no. MA5–17687), phospho-P38 (catalog no. MA5–15177), KI67 (catalog no. PA5–19462), phospho-IRAK-1 (catalog no. PA5–105752), phospho-IRAK-4 (catalog no. BS-10208R), GFAP (catalog no. PA1–10004), and CD19 (catalog no. 14–0193–82). Slides were washed 3x in PBS containing 0.1% Tween-20 (PBS-T, Sigma-Aldrich) for 5 minutes. Fluorescently-labeled secondary antibodies were added for 1 hour at room temperature: (Thermo-Fisher) goat anti-rabbit AlexaFluor 647 (catalog no. A-21244), goat anti-chicken AlexaFluor 488 (catalog no. A-11039), goat anti-rat AlexaFluor 568 (catalog no. A-11077), and DAPI (Sigma-Aldrich). Slides were washed 3x in 1 x PBS-T for 5 minutes each and coverslips mounted using ProLong Diamond (Thermo-Fisher). Images were acquired using a Nikon A1RMP confocal microscope and analyzed using Imaris x64 v9.7.0 imaging software.

Statistical analyses were performed using GraphPad Prism 9 as described in figure legends. Significance was determined as P < 0.05. Cell counts on IHC images were established using Imaris x64 v9.7.0, where cell counts were based on target channel fluorescence from ‘spots’ identified as a diameter of ≥ 6 μm and a set minimum voxel intensity, with background subtraction applied. For all IHC, individual specimen data includes 2 to 3 tissue sections separated by ≥ 50 μm, where mean values are used for quantitative comparisons between groups. For survival studies, animals were randomized prior to treatment. Human tissue was limited by biospecimen availability from each respective tissue bank.

The data that support the findings of this study are available from the corresponding author upon reasonable request.

To determine if MyD88 is aberrantly activated in CNS lesions of PCNSL and MBM, we performed multiparameter IHC analyses of downstream proteins phospho-IRAK-1, phospho-IRAK-4, and phospho-NF-κB P65 in clinical tissue specimens. The expression of all three markers was found in all tumor samples tested, and observed in both tumor cells and cells that comprise the TME including immune cells and tumor-reactive astrocytes, which are identified as CD45 and GFAP positive, respectively (Fig. 1A, C and E). Protein expression levels in both PCNSL and MBM tumors are significantly increased as compared with unmatched normal human cortex (Fig. 1B, D and F; Supplementary Fig. S1), supporting that MyD88 activation is a specific event in those tumors.

NF-κB and MAPK pathway activation in human PCNSL and MBM tumors was further validated through proteomic analysis, and compared with levels found in unmatched normal human cortex (Fig. 1G). The relative protein expression of IRAK-1 and MyD88 was found to be elevated in both PCNSL and MBM, along with increased expression of interferon regulatory factor 5 (IRF5), a mediator of inflammasome activity directly downstream of MyD88 (Fig. 1H; ref. 48). NF-κB pathway activation in tumor was evidenced through increased protein expression of IκB kinase (IKK1, IKK2), NF-κB1, NF-κB2, RelA/P65, and c-Rel (Fig. 1H). MAPK signaling activation was also observed in both tumor types, with elevated protein expression of JNK1/2 and JNK2 (Fig. 1H). Increased expression of reflexive inhibitors of chronic inflammation were also observed, including IκBα, IκBε, and suppressor of cytokine signaling 6 (SOCS6; Fig. 1H). Together, these data support aberrant activation of the MyD88 inflammasome in both PCNSL and MBM, highlighting a new avenue for therapeutic intervention.

CA-4948 is an oral first-in-class small molecule inhibitor of IRAK-4 (Supplementary Fig. S2A), with desirable ADME and pharmacokinetics profile (35). Current clinical studies of CA-4948 demonstrate preliminary anticancer activity with acceptable tolerability and safety (NCT04278768, NCT03328078). Given that this study involves in vivo analysis of the activity of CA-4948 in mice, it was prudent to computationally assess the ability of CA-4948 to bind to murine IRAK-4.

CA-4948 has been crystallized with human IRAK-4 (PDBID: 7C2V; refs. 35, 36). The binding mode can be visualized in Fig. 2A and Supplementary Fig. S2B–S2C. From this complex, a number of key interactions are evident. The ligand forms a hydrogen bond between its amide carbonyl oxygen and the backbone N-H of MET265, and it forms a number of water-mediated hydrogen bonds between its oxazolopyridine 4-nitrogen and ARG273, ASP272, and SER269, as well as between its secondary alcohol and SER269. For the secondary alcohol, it is likely that this water-mediated hydrogen bonding is the driving force behind the improved selectivity of the R enantiomer over the S enantiomer (35). This binding pocket also has a number of hydrophobic residues which provide favorable interactions with this largely hydrophobic ligand, as well as TYR262, which forms favorable π–π interactions with the ligand pyridine ring. There could also be lone-pair cation interactions between the pyridine nitrogen and LYS213. While this murine homolog has been sequenced, its crystal structure has yet to be solved. Therefore, homology modelling was conducted using Schrödinger's Homology Modelling functionality (37, 39, 40). Mouse IRAK-4 shows an 87% sequence identity with human IRAK-4 (Mus musculus, Uniprot code Q8R4K2; ref. 42; Supplementary Fig. S2D). An energy-based homology model was then developed (Supplementary Fig. S2E). There is only one residue within the binding pocket that is altered in mouse IRAK-4 (residue 263; alanine in mouse and valine in human), and this residue does not play a role in ligand binding outside of contributing to binding pocket hydrophobicity. To further validate the binding of CA-4948, this ligand was re-docked to human and mouse IRAK-4 using GlideXP (refs. 37, 43; Fig. 2B and C). The top binding poses were rescored using DeepAtom (44–46), a deep learning model for predicting ligand-receptor binding affinity, as a complementary means of assessing binding pose quality. Fig. 2D shows the docking data for the top poses. The GlideXP docking score differs for these two compounds (−6.5 versus −8.9 kcal/mol for docking CA-4948 to hIRAK-4 and mouse homolog, respectively), but this difference is likely driven by the dual hydrogen bonding captured by the secondary hydroxyl and SER269 and ASP272 in the mouse homolog docking results (Fig. 2C). It should be reiterated that in the crystal structure, the hydrogen bonding between this secondary hydroxyl and the receptor is water-mediated rather than directly facilitated. In addition, the RMSD of each docking pose to the original crystal structure ligand pose is similar (0.743 and 0.870 Å, respectively) and the DeepAtom score is identical for both (−11.1 kcal/mol). Therefore, it is clear that the differences in sequence between human and mouse IRAK-4 are not sufficient to negatively impact ligand binding to this domain.

We next assessed the potency of CA-4948 against human PCNSL harboring a MYD88 activating mutation (OCI-LY3, L265P) and wild-type MYD88 (SU-DHL-4), where the former may carry increased sensitivity to targeted IRAK-4 inhibition through oncogenic dependency on MyD88. As predicted, the proliferative IC₅₀ value for OCI-LY3 was lower in comparison to SU-DHL-4, 1.21 μmol/L and > 10 μmol/L, respectively (Fig. 3A and B). In addition, we leveraged the use of two distinct and highly aggressive murine preclinical CNS tumor models: A20 B-cell lymphoma and B16F10 melanoma. Although these preclinical models do not carry activating mutations in the MyD88 signalosome, both A20 (Supplementary Fig. S3A–S3B) and B16F10 (Supplementary Fig. S3C–S3D) intracerebral tumors exhibit high protein expression of active IRAK-4, with lower expression in distal normal tissue (Supplementary Fig. S3B and S3D). These data indicate that elevated IRAK-4 is a tumor-specific feature that may confer sensitivity to targeted inhibition. Growth inhibition studies confirm sensitivity of A20 and B16F10 to CA-4948 treatment, with respective proliferative IC₅₀’s of 2.45 μmol/L and 0.55 μmol/L (Fig. 3C and D). CA-4948 treatment did not significantly impact tumor cell viability (Supplementary Fig. S4), suggesting that direct antitumor activity of this agent is likely cytostatic.

UPLC/MS-MS analysis of brain tissue and CSF was performed to determine if this agent is capable of crossing the BBB and reaching therapeutic dose levels. Drug concentration was measured across fixed time points following a single dose of CA-4948. Time to maximum concentration (T_max) was similar for both brain parenchyma (0.5 hours) and CSF (0.25 hours) in naïve mice as compared with plasma (0.38 ± 0.14 hours), with a slight delay in brain parenchymal accumulation in mice harboring A20 PCNSL tumors (0.83 ± 0.29 hours; Fig. 3E and F; Supplementary Fig. S5). Plasma half-life (T_1/2) was 2.73 hours, approximately double the T_1/2 found in naïve brain parenchyma (1.39 hours), naïve CSF (1.33 hours), and tumor-bearing brain parenchyma (1.19 hours; Fig. 3E and F), suggesting that CA-4948 is more rapidly cleared from the CNS. Although brain concentrations of CA-4948 was lower as compared with plasma (1.53%, 4.26%, and 4.95% for naïve CSF, naïve brain parenchyma, and tumor-bearing brain parenchyma, respectively; Fig. 3E and F), drug concentrations in all CNS compartments exceed the proliferative IC₅₀ values established for each OCI-LY3, A20, and B16F10 tumor models (Fig. 3A–D) indicating CA-4948 is capable of achieving therapeutic dose levels in the brain space. Mean CA-4948 concentrations in each brain compartment over time is summarized in Fig. 3G.

To assess the biologic activity of CA-4948, MAPK, and NF-κB protein biomarker expression downstream of IRAK-4 was evaluated in total tumor lysates prepared from vehicle and CA-4948 treated intracranial A20 tumors following a 5-day treatment. CA-4948 was effective at reducing phospho-P38^{Thr180/Tyr182} (pP38) in all specimens (Fig. 4A and B). Cumulatively, CA-4948 also modestly reduced phospho-NF-κB P65^Ser536 (pNF-κB; Fig. 4A and B). Phospho-ERK1/2^{Thr202/Tyr204} (pERK1/2) was reduced in a few CA-4948 treated specimens, but cumulatively was not statistically significant (Fig. 4A and B). To resolve cellular localization of biomarker signal in A20 intracranial tumors, protein expression of active MAPK and NF-κB was measured via fluorescence confocal microscopy. Three-dimensional imaging for pERK1/2 and pP38 was done using the CLARITY tissue clearing method (49), which included perfusion with a vascular dye to illuminate tumor-associated blood vessels (BV), to evaluate the geospatial distribution of signal. The pNF-κB antibody used in these studies was not amenable to the CLARITY technique, and therefore standard IHC methods were applied. B16F10 tumors were also evaluated using standard IHC methods given their high melanin content and subsequent opacity. Protein expression levels were measured as the number of biomarker-positive cells over the total number of cells based on DAPI nuclear stain. In the A20 model, CA-4948 treatment cumulatively reduced pERK1/2 expression from 8.9% to 5.3% (Fig. 4C and D). A trend for reduced pERK1/2-positive cells in CA-4948–treated B16F10 tumors was observed, however these values were not significant (Supplementary Fig. S6A–S6B). In A20 tumors, baseline expression of pP38 was notably higher, with CA-4948 treatment significantly reducing expression from 44.3% to 11.8% (Fig. 4E and F). Similar results were detected in the B16F10 model, with CA-4948 reducing pP38 expression from 38.1% to 11.4% (Supplementary Fig. S6C–S6D). In vehicle-treated A20 tumors the expression of pNF-κB P65 did not show significant co-localization with CD19 (tumor), and was mostly observed in the TME bordering the tumor (Fig. 4G). CA-4948 treatment was successful in reducing the number of pNF-κB–positive cells from 17% to 7.1% (Fig. 4H). As with A20, we found pNF-κB expression predominantly in the TME bordering B16F10 intracranial tumors, where CA-4948 treatment reduced pNF-κB expression from 9.9% to 4.3% (Supplementary Fig. S6E–S6F). Collectively, these data support that CA-4948 treatment suppresses signaling pathways downstream of IRAK-4 in intracranial tumors, signifying biologic activity for this agent within the CNS space.

Given the discrepancy in cellular localization for each MAPK (tumor) and NF-κB (TME), voxel-based co-localization analyses were performed to dissect which cellular subsets within the TME express pNF-κB. Within B16F10 tumors, approximately 70% of pNF-κB–positive cells express GFAP, a marker for astrocytes, ∼15% express CD45 (immune-specific), and the remaining 15% were undefined (Supplementary Fig. S6G–S6H). Co-localization analyses were also performed in patient PCNSL and MBM specimens to determine the composition of pNF-κB expression in human tumors. In PCNSL, approximately 35% of pNF-κB–positive cells were identified as GFAP-positive astrocytes, 30% as CD45+ immune (likely tumor) cells, and 35% of expression was undefined (Supplementary Fig. S7A–S7C). In MBM, approximately 30% of pNF-κB expression derived from GFAP-positive astrocytes, 40% from CD45-positive immune cells, and the residual ∼30% remaining undefined (Supplementary Fig. S7D–S7F) which may include tumor cells or other TME subpopulations. These data suggest an important role for the TME, including astrocytes and immune cells, in mediating NF-κB inflammatory responses to tumor presence within the brain parenchyma, where CA-4948 may stifle these reflexive interactions.

We next tested if CA-4948 monotherapy could mitigate tumor growth in preclinical models of PCNSL and MBM. Five days following stereotaxic tumor implantation, animals were treated with vehicle or CA-4948 for 2 weeks (Fig. 5A). Due to toxicity concerns as a result of long-term continuous treatment, dosing of CA-4948 was reduced to 100 mg/kg and 50 mg/kg. In athymic nude mice harboring the MYD88 L265P mutant human tumor OCI-LY3, 100 mg/kg CA-4948 treatment improved median survival by 68%, with 3 of 8 mice experiencing durable survival outcomes (Fig. 5B). In syngeneic BALB/c mice harboring A20 PCNSL tumors, 100 mg/kg CA-4948 treatment improved median survival by 61% (Fig. 5C). High-dose CA-4948 treatment of mice harboring B16F10 CNS melanoma tumors also improved the mean overall survival by 35% (Fig. 5D), with markedly reduced tumor burden observed in brain tissue isolated on day 7 of treatment (study day 12; Fig. 5E). We next assessed long-term (chronic) versus short-term (acute) CA-4948 treatment of PCNSL to determine if antitumor benefit would be enhanced in the former setting. In in vitro studies of A20 proliferation where CA-4948 was transiently applied and then removed, the proliferative capacity of these cells was quickly restored, slightly exceeding those of vehicle control, whereas continuous treatment consistently subdued proliferation (Fig. 5F). To determine if chronic CA-4948 administration would enhance the survival benefit initially observed in the A20 PCNSL model, we compared survival between groups receiving 100 mg/kg CA-4948 for 2 and 4 weeks. Unlike the in vitro findings, chronic (4 weeks) CA-4948 treatment resulted in a similar survival advantage as compared with acute (2 weeks) treatment (Fig. 5G). We also assessed the proliferative index of control and high-dose CA-4948–treated tumors excised on day 12 of study (day 7 of treatment) via protein expression of Ki67. Ki67 expression was significantly reduced in both CA-4948–treated A20 PCNSL (Fig. 5H and I) and B16F10 melanoma tumors (Supplementary Fig. S8A–S8B), indicating that CA-4948 is capable of direct cytostatic antitumor activity in the CNS.

Aberrant activation of myddosome signaling by virtue of mutated MYD88 is a known driver of lymphomagenesis, leading to the recent clinical investigation of CA-4948 in patients with systemic hematologic malignancies (NCT03328078, NCT04278768). Early clinical data shows that CA-4948 has an acceptable safety and tolerability profile, even with long-term treatment (> 6 months), and shows preliminary antitumor activity in heavily pretreated patients (50). In the present study, we confirmed that myddosome signaling is strongly activated in both PCNSL and MBM, as demonstrated by protein detection of phospho-IRAK-4, phospho-IRAK-1, and heightened NF-κB signaling both in cancer cells and the TME, validating this pathway as a compelling therapeutic target. Even in the presence of an intact BBB we demonstrate that CA-4948 can achieve therapeutic dose levels within the CNS. Furthermore, CA-4948 treatment subdues critical signaling mechanisms downstream of IRAK-4, supporting on-target inhibition of this kinase within the brain. Importantly, single-agent CA-4948 prolongs survival in preclinical models, supporting the clinical investigation of this agent in patients with PCNSL and MBM.

Interestingly, we found that a large proportion of active NF-κB expression stems from the TME in both PCNSL and MBM preclinical models, including tumor-reactive astrocytes and immune cells. This paradigm was also seen in patient tissues, identifying this as a clinically relevant phenomenon. This study corroborates the findings of others (25, 51) with MyD88 signaling detected in patient melanoma TME, shown here for the first time in brain metastases, although neither MBM nor primary melanoma presents with activating mutations in MYD88. While anticipated for MBM, this was an unexpected finding in PCNSL where myddosome activation is thought to be tumor-specific (21, 52). These data suggest that the brain tumor TME activates myddosome signaling as a compulsory response to persistent inflammatory signals stemming from cancerous lesions, perhaps potentiating a pathologic environment that supports tumor expansion possibly agnostic of brain tumor origin. This notion of chronic pro-tumor inflammation is further evidenced by concomitant overexpression of negative regulators of NF-κB and inflammation, including IκBα, IκBε, and SOCS6.

Recent studies indicate that MyD88 is perversely activated in several solid cancers, particularly within the myeloid cells (25, 53). Here, MyD88 emerges as an adaptor molecule that can distinguish pro-tumorigenic tumor-associated macrophages (TAM) from other myeloid subsets (25, 53). Previous work demonstrates MyD88 as responsible for regulating PD-1/PD-L1 immune checkpoint expression on TAMs in melanoma, and promoting tumor resistance to ICI (25). In metastatic urothelial cancer, interleukin-1 alpha (IL1α) and IL1β present as the most differentially expressed ligands inferred to regulate myeloid cell mediated resistance to ICI, where MyD88 is required to relay signaling downstream of IL-1R activation (53). These data combined with ours highlight a unique therapeutic opportunity where selective inhibition of pro-tumorigenic cells within the TME (myeloid cells, tumor-reactive astrocytes) via targeted inhibition of IRAK-4 may resensitize tumors to ICI.

Further work is required to discern the contributions of astrocytes to the inflammatory milieu, where NF-κB may serve as a biomarker of aberrant tumor-induced inflammation in these cells. In our study, we found that NF-κB is activated in tumor-reactive astrocytes, and can be therapeutically attenuated with CA-4948 treatment. Through the lens of other neurologic disease and injury, astrocytes are shown to evolve into transcriptionally distinct activation states based on factors such as spatial adaptation and disease origin, and directly inform immune surveillance and inflammation by resident microglia and infiltrating leukocytes (54, 55). It will be important to elucidate if MyD88 performs as an adaptor molecule in reactive astrocytes similar to myeloid cells by mediating cross-talk in the brain TME, and potentially conferring a tumor-protective effect.

In both PCNSL and MBM, we show that CA-4948 is capable of suppressing MAPK signaling in addition to NF-κB, by virtue of decreased pP38 and pERK1/2 activation. Unlike NF-κB, MAPK expression in vehicle-treated preclinical tumors appears to largely derive from the tumor cells themselves, indicating a direct antitumor effect. In melanoma specifically, recovery of MAPK/ERK signaling is a well-described mechanism of resistance to BRAF/MEK inhibitor treatment, one of the mainstays of therapy for patients presenting with BRAFV600E/K-activating mutations (56). In both extracranial and intracranial disease, the initial response to BRAF/MEK inhibition is robust; however, resistance rapidly develops, with intracranial failure being observed at < 9 months versus ∼12 months systemically (57). In both instances, suppression of MAPK via CA-4948 treatment may represent a novel combinatorial strategy alongside targeted BRAF/MEK inhibition that could prolong therapeutic benefit and prevent the development of resistant escape pathways.

In summation, MyD88/IRAK-4 is upregulated in both PCNSL and MBM and has far-reaching interactions within the unique TME of these brain malignancies. We highlight the capacity of CA-4948 to overcome the restrictive BBB and reach a therapeutically meaningful level in brain malignancies in preclinical models of PCNSL and MBM, producing single-agent antitumor responses. On the basis of the mechanisms of interaction with intracellular signaling pathways in both tumors and the TME, there exists the potential for CA-4948 to work synergistically with other agents. These include ICI and targeted therapeutic agents such as BRAF/MEK in MBM and BTK inhibitors in PCNSL, where CA-4948 could contribute to changing the landscape in brain malignancies for which good therapeutic options are scarce.

C.A. Von Roemeling reports grants from Curis, Inc. during the conduct of the study. B.P. Doonan reports grants from Curis, Inc. during the conduct of the study. L. Hoang-Minh reports grants from Curis, Inc. during the conduct of the study. V. Trivedi reports grants from Curis, Inc. during the conduct of the study. H.W. Tun reports grants from Curis, Inc. during the conduct of the study, as well as grants and personal fees from Gossamer Bio and Acrotech Biopharma outside the submitted work. D.A. Mitchell reports grants from Curis Inc. during the conduct of the study. No disclosures were reported by the other authors.

C.A. Von Roemeling: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. B.P. Doonan: Conceptualization, resources, data curation, funding acquisition, validation, writing–original draft, writing–review and editing. K. Klippel: Data curation, formal analysis, writing–review and editing. D. Schultz: Conceptualization, data curation, formal analysis, funding acquisition, visualization, writing–original draft, writing–review and editing. L. Hoang-Minh: Data curation, writing–review and editing. V. Trivedi: Data curation, formal analysis, writing–review and editing. C. Li: Resources, supervision, methodology, writing–review and editing. R.A. Russell: Data curation, formal analysis, writing–review and editing. R.S. Kanumuri: Data curation, formal analysis, validation. A. Sharma: Data curation, formal analysis, supervision, validation. H.W. Tun: Conceptualization, writing–review and editing. D.A. Mitchell: Conceptualization, resources, supervision, funding acquisition, project administration, writing–review and editing.

This work was supported through agent and funding provided by Curis, Inc. (to D.A. Mitchell). This research was conducted in part through funds from the J. David Vandivier Memorial Melanoma Research Fund (to B.P. Doonan, F006025). This work was supported in part by the University of Florida Clinical and Translational Science Institute, which is supported in part by the NIH National Center for Advancing Translational Sciences under award number UL1TR001427. This work was supported in part by the University of Florida Interdisciplinary Center for Biotechnology Research with funding provided by NIH Grant (1S10OD020026). This work was also supported through the University of Florida Graduate School Preeminence Award (to D. Schultz).

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.