Seeing Macrophages Flip: A Closed-Loop Theranostic Readout Borrowed from Radio Astronomy

Solid tumors stay cold partly by converting the immune system's own frontline cells into guards: tumor-associated macrophages, which normally kill and present tumor antigens, are pushed into a wound-healing, T-cell-suppressing state. A promising class of therapies flips them back — but nobody can watch the flip happen in a living animal or patient. Every demonstration of macrophage "reprogramming" verifies it by killing the tissue and staining or sequencing it afterwards. Theranostic agents — single constructs that carry both an imaging beacon and a therapy, so that where the drug goes is visible on a scanner — exist, but current ones report only particle location or tumor shrinkage, not the immune state the therapy is supposed to change. The result is an open loop: clinicians deliver a reprogramming drug blind and learn weeks later, destructively, whether it worked.

The other three inputs to this proposal supply, respectively, the actuator, the obstacle, and — unexpectedly — the readout. The actuator comes from theranostic chemistry: ferrous-iron platforms that image on MRI while catalyzing ferroptosis (Xie et al.) and manganese-dioxide particles that enhance MRI contrast while relieving hypoxia (Wang et al.) prove that one construct can report and act in the same microenvironment. The obstacle is the immunosuppressive stroma itself: dense hyaluronic-acid matrix and macrophage-dominated immune exclusion (Zhao et al.; Pan et al.). And the readout comes from an unlikely source: astronomers studying fast radio bursts — millisecond radio flashes from magnetars — faced an exactly analogous measurement problem. Their signal is a coherent emission smeared by the plasma it traverses, and they built an unsupervised encoder that maps these smeared signals onto a continuous, interpretable coordinate axis (Representation learning for fast radio burst dynamic spectra). No physical connection between radio transients and cancer is claimed or needed; the transfer is entirely methodological, and this proposal says precisely where it applies.

Core Claim

The claim is that a macrophage-targeted theranostic particle, combined with an unsupervised embedding model of the kind built for FRB dynamic spectra, can give a quantitative, non-invasive, longitudinal readout of the macrophage phenotype — a single scalar $\phi(t) \in [0,1]$, read as a linear map of the embedding $z(t)$, tracking position on the M1-to-M2 continuum — and that this readout is accurate and early enough to close the therapeutic loop, redosing only when the tumor's cytokine pressure drives reprogrammed macrophages back toward the suppressive state. This targets the field's most concrete measurement gap: every existing reprogramming study confirms phenotypic conversion only at endpoint histology — the M1-extracellular-vesicle work in glioblastoma (Wang et al.) and the metal-organic-framework pazopanib study in breast cancer (Xu et al.) both verify M2-to-M1 shifts this way — while theorists of closed-loop theranostics (Sengar; Xiao et al.) have specified the need without a mechanism for reading immune state. The nearest existing platforms — the ferroptosis-inducing iron theranostics of Xie et al. and the ferroptosis-plus-checkpoint-blockade delivery systems reviewed in Boosting cancer immunotherapy — image particle accumulation and treat tumor cells, but do not measure macrophage phenotype at all. We add exactly that measurement, plus a propagation model that corrects it.

Two transfers from the FRB literature do real work here. First, the architecture: the convolutional autoencoder with an information-ordered bottleneck (IOB-CAE) was shown on 140 real CHIME/FRB bursts and 5,000 synthetic bursts to place FRBs on a continuous morphological spectrum rather than in discrete clusters — no quantitative performance number is reported on Lacuna for that paper, but the qualitative finding is the one we need. Macrophage phenotypes are likewise a continuum, not two switches (Immunotherapy for Glioblastoma), and FRB population modeling independently supports treating apparent binary classes as one continuous distribution (Synthesising the repeating FRB population using frbpoppy); a supervised classifier trained on "M1 vs M2" labels would impose the discreteness the biology doesn't have. Second, the propagation model: the work on interpreting FRB observable distributions shows how an unknown medium smears an underlying signal (Interpreting the distributions of FRB observables). We model the dense stroma as exactly such a medium for the particle's MRI signal:

$$S_{\mathrm{obs}}(\mathbf{x},t) \;=\; \big(h_{\mathrm{stroma}} \otimes S_{\mathrm{true}}\big)(\mathbf{x},t) \;+\; \varepsilon,$$

where $S_{\mathrm{true}}$ is the true per-voxel particle accumulation, $h_{\mathrm{stroma}}$ is a depth-dependent smearing kernel set by matrix density and interstitial pressure, and $\varepsilon$ is noise. Just as CHIME/FRB astronomers learned that the apparent cutoff in observed burst properties could be an artifact of propagation and instrumentation rather than astrophysics, the observed distribution of a nanoparticle in tissue is the convolution of true delivery with a stromal transfer function — and inverting it is what lets the embedding track biology rather than barrier geometry.

Research Program

The construct assembles known parts rather than inventing chemistry: a metal-organic-framework nanoparticle whose metal nodes are MRI-visible (iron for $T_2$ contrast and Fenton activity, per Xie et al.; manganese for $T_1$-weighted signal and oxygen generation, per Wang et al. and the three-in-one radiosensitizing platform of Engineered multifunctional nanoparticles), encapsulating a repolarizing cargo in the manner of the pazopanib MOF (Xu et al.), with a dansylated albumin-binding moiety to enforce long circulation and high target uptake, following the design logic that raised tumor uptake from 17.44 %ID/g to 127.36 %ID/g in prostate xenografts (Dansylated Amino Acid–Modified Long-Acting PSMA Derivatives). The engineering novelty is deliberately modest; the scientific novelty is the readable loop.

The readout is built in three stages. First, an IOB-CAE is pretrained on the dynamic MRI sequences of the Cancer-Net PCa-MultiSeg benchmark — 200 patients with diffusion-weighted imaging, apparent diffusion coefficient, and synthetic correlated diffusion imaging, with public code (Jdewbury/Cancer-Net-PCa-MultiSeg, Cancer-Net PCa-MultiSeg) — to learn a low-dimensional embedding of tissue microstructure without labels. Second, the embedding's macrophage axis is calibrated: the anchor is the public single-cell atlas of the oncolytic-virus OH2 study, 61,713 single cells with matched bulk and spatially resolved treatment response (A Multi-Omics Dataset of Prostate Cancer Response to Oncolytic Virus OH2 Treatment), which lets us map where macrophage states sit on the embedding by pairing regions whose ex vivo composition is known with their in vivo imaging. Third, the stromal transfer function $h_{\mathrm{stroma}$ is fitted in animals by varying hyaluronic-acid density — with and without hyaluronidase co-treatment in the spirit of the sPH20-secreting CAR-T design of Zhao et al. — and measuring how the observed particle signal blurs with depth, so that $\phi(t)$ can be corrected for barrier geometry. The clinical target population is chosen by the biomarker the field already trusts: advanced salivary gland cancer, where M2 macrophages make up roughly 41% of the immune compartment and outnumber T cells nearly three-to-one, and where the T-cell-to-macrophage ratio crossing 1 marked who responded to immunotherapy (A macrophage-predominant immunosuppressive microenvironment) — a threshold our readout aims to estimate non-invasively.

The loop itself is then tested: animals receive the particle, and the therapy arm is split into a fixed schedule versus an adaptive schedule that redoses only when $\phi(t)$ shows reversion toward the suppressive state. This addresses the known failure mode that tumors exert ongoing metabolic pressure that reverts reprogrammed macrophages — pressure that can be biochemical, as with the MTAP-deletion-driven methylthioadenosine accumulation that directly polarizes macrophages in glioblastoma (Hansen et al.) — and it mirrors the logic that led FRB astronomers to time-resolved monitoring in the first place: CHIME/FRB polarimetry found rotation-measure swings of hundreds of radians per square meter and even sign reversals over months, proving the local environment changes and must be tracked, not sampled once (Polarization Properties of 28 Repeating Fast Radio Burst Sources with CHIME/FRB).

Baselines

The do-nothing reference point is the field's current practice: endpoint-only verification by immunohistochemistry, flow cytometry, or single-cell sequencing after the animal is sacrificed, exactly as practiced in the M1-vesicle glioblastoma study (Wang et al.) and the MOF-pazopanib breast cancer study (Xu et al.). It is the strongest competitor on cost and regulatory maturity, and it is the ground truth our readout must agree with. Beating it means demonstrating, in paired animals, that the non-invasive trajectory $\phi(t)$ predicts the terminal flow-cytometric M1/M2 ratio at matched time points with a pre-registered Spearman correlation of at least 0.8, and detects phenotypic reversion at least several days before any change in tumor volume — the two things endpoint analysis cannot do at all.

The strongest existing method on the imaging side is supervised multimodal segmentation with radiomic feature extraction: the best configuration on the very benchmark we pretrain on, Swin UNETR with triple-modality input (synthetic correlated diffusion, diffusion-weighted imaging, and apparent diffusion coefficient maps), achieved an absolute Dice score of 34.30, with Attention U-Net showing a 72.5% relative Dice improvement (14.82 to 25.57) when synthetic correlated diffusion was added to the diffusion baseline (Cancer-Net PCa-MultiSeg). We will run this network on the identical longitudinal scans, with the same data, compute, and tuning budget, derive radiomic features from its segmentations, and fit the same linear map to the ex vivo phenotype labels. Beating it means a higher correlation with terminal macrophage fractions at equal label budget — a fair fight, since both get the same calibration animals. A second, simpler competing representation, principal component analysis — the method the FRB paper itself used as its simpler baseline (Representation learning for fast radio burst dynamic spectra) — will be run identically.

The key ablation isolates whether the signal is phenotype or mere accumulation: an otherwise identical particle with the MRI-visible metal nodes but no repolarizing cargo. If $\phi(t)$ moves the same way without the cargo, the readout is tracking particle uptake, not immune conversion, and the core claim fails. A second ablation replaces the information-ordered bottleneck with a randomly ordered one of matched capacity, to isolate whether the ordering — the specific ingredient imported from the FRB work — is what makes the latent axis interpretable as a phenotype coordinate.

Why not just…?

Why not just co-administer an approved imaging agent and the reprogramming drug separately? This is the clinically simplest option, and for late-stage deployment it may win on regulatory grounds. The evidence against it is quantitative: separate molecules have discordant circulation, permeation, and clearance (Xiao et al.; Yang et al.), and the cost of that discordance is measured in the same benchmark we use: the standard radioligand achieved 17.44 %ID/g tumor uptake at 4 hours while the pharmacokinetically unified construct achieved 127.36 %ID/g, an 8-fold radiation area-under-the-curve gain, and equivalent tumor inhibition at half the dose (Dansylated Amino Acid–Modified Long-Acting PSMA Derivatives). Still, our own transfer-function experiment partially settles this: if a co-administered tracer's observed distribution, after stromal deconvolution, predicts the drug's true accumulation as well as the unified construct's signal does, the single-construct argument weakens to a pharmacokinetic one, and we will report that honestly.

Why not just use the supervised segmentation network everyone already has? This is a real threat, and we do not argue around it: Swin UNETR is mature, and on lesion delineation it wins by default. But its reported absolute Dice scores on this benchmark sit in the 0.25–0.35 range, with large-cohort baselines as low as 0.32, precisely because microenvironmental boundaries are ill-defined (Cancer-Net PCa-MultiSeg) — and, more fundamentally, a phenotype trajectory is not a segmentation task: there are no voxel-level M1/M2 labels to supervise on, which is the same reason the FRB community went unsupervised. The linear-probe comparison above is designed to settle this empirically rather than by assertion, and if supervised features correlate better with terminal macrophage fractions, the finding will be published either way.

Why not just deplete the macrophages instead of reprogramming them? Depletion with CSF-1R inhibitors or clodronate is simpler and clinically further along. The counter-evidence is that depletion removes the innate antigen-presenting cells needed to recruit and cross-prime cytotoxic T cells, leaving a cellular vacuum, whereas repolarization converts the tumor's stromal shield into an active immune engine (Harnessing the innate immune system; Tumor-Associated Macrophages: Recent Insights and Therapies). This, too, we test rather than assert: a depletion arm in the murine study will compare T-cell-to-macrophage ratio recovery and tumor control head-to-head with the reprogramming arm. If depletion matches reprogramming on both endpoints in this model, the reprogramming premise of the whole program is weakened, and the readout would then be worth pursuing mainly as a diagnostic for the immunosuppressive state itself — still valuable in tumors like adenoid cystic carcinoma, where checkpoint targets are absent and macrophage dominance is the defining feature (A macrophage-predominant immunosuppressive microenvironment).

Decisive Checks

Three results decide the proposal. First, calibration: across serially sacrificed animals, $\phi(t)$ must track the flow-cytometric M1/M2 ratio at $\rho \geq 0.8$ and beat the Swin UNETR-radiomics and PCA competitors under the matched-budget protocol; the literature reports no comparable non-invasive number on Lacuna, so this comparison defines the first result of its kind. Second, mechanism isolation: the no-cargo particle must leave $\phi(t)$ flat while accumulating signal, confirming the readout measures immune state and not delivery; and the stromal deconvolution must recover depth profiles matching biodistribution endpoints in the 127.36-versus-17.44 %ID/g style of assay. Third, the loop must pay off: in the reversion-prone setting — where tumor-derived pressure of the kind documented for MTAP-loss glioblastoma (Hansen et al.) drives macrophages back toward suppression — the adaptive arm, redosing on $\phi(t)$ alone, must hold a higher T-cell-to-macrophage ratio and better tumor control than the fixed-schedule arm at equal total dose. If all three hold, the field gains something it currently lacks: the ability to watch a tumor's immune defenses change in a living subject, and to act on what it sees.

References

  1. Carbon nanoparticles-Fe(II) complex for efficient theranostics of xenografted colonic tumor
  2. Dimeric Her2-specific affibody mediated cisplatin-loaded nanoparticles for tumor enhanced chemo-radiotherapy
  3. Human Hyaluronidase PH20 Potentiates the Antitumor Activities of Mesothelin-Specific CAR-T Cells Against Gastric Cancer
  4. The MYBL2–CCL2 axis promotes tumor progression and resistance to anti-PD-1 therapy in ovarian cancer by inducing immunosuppressive macrophages
  5. Representation learning for fast radio burst dynamic spectra
  6. Exploration and functionalization of M1-macrophage extracellular vesicles for effective accumulation in glioblastoma and strong synergistic therapeutic effects
  7. Determining M2 macrophages content for the anti-tumor effects of metal-organic framework-encapsulated pazopanib nanoparticles in breast cancer
  8. Precision in Practice: Nanotechnology and Targeted Therapies for Personalized Care
  9. Gold Nanorods Conjugated with Doxorubicin and cRGD for Combined Anticancer Drug Delivery and PET Imaging
  10. Boosting cancer immunotherapy: drug delivery systems leveraging ferroptosis and immune checkpoint blockade
  11. Immunotherapy for Glioblastoma: Current State, Challenges, and Future Perspectives
  12. Synthesising the repeating FRB population using frbpoppy
  13. Interpreting the distributions of FRB observables
  14. Engineered multifunctional nanoparticles for enhanced radiation therapy: three-in-one approach for cancer treatment
  15. Dansylated Amino Acid–Modified Long-Acting PSMA Derivatives68Ga/177Lu-LNC1011 as Prostate Cancer Theranostics
  16. Cancer-Net PCa-MultiSeg: Multimodal Enhancement of Prostate Cancer Lesion Segmentation Using Synthetic Correlated Diffusion Imaging
  17. A Multi-Omics Dataset of Prostate Cancer Response to Oncolytic Virus OH2 Treatment
  18. A macrophage-predominant immunosuppressive microenvironment and therapeutic vulnerabilities in advanced salivary gland cancer
  19. MTAP loss correlates with an immunosuppressive profile in GBM and its substrate MTA stimulates alternative macrophage polarization
  20. Polarization Properties of 28 Repeating Fast Radio Burst Sources with CHIME/FRB
  21. Harnessing the innate immune system and local immunological microenvironment to treat colorectal cancer
  22. Tumor-Associated Macrophages: Recent Insights and Therapies