When the tumor, not the donor, sets the immune tone

What the work claims

This is a primary benchmarking preprint from a Toronto group.¹ Its claim is deliberately narrow and, for that reason, credible: in humanized mice carrying human tumors, the composition of the immune infiltrate is set mainly by the tumor rather than the stem-cell donor, and, more modestly, the resulting infiltrate often resembles the one in the patient tumor it came from. The comparison that carries the argument is a contrast of two variables that are usually confounded, the identity of the implanted tumor and the identity of the stem-cell donor used to reconstitute the immune system.

Why this is worth a careful read: humanized models are widely suspected of producing immune infiltrates that reflect the donor lottery rather than tumor biology. If that were true, they would be near useless for studying how a specific tumor evades immunity. The paper sets out to test that suspicion head-on with high-dimensional single-cell measurement rather than assertion.

How it works

The platform is the NOG-EXL mouse, which carries human IL-3 and GM-CSF transgenes to support human myeloid cells, a lineage that older humanized strains struggle to generate. The authors humanized 88 such mice using cord-blood CD34-positive hematopoietic stem cells from eight donors, then engrafted 82 of them with fragments of ovarian, HPV-negative head and neck, or renal cell tumors that had previously been grown as xenografts in immunodeficient NSG mice. Tumors were implanted a mean of 23 weeks after humanization. Sixty-six percent of the biology here rests on one measurement platform: cytometry by time of flight (CyTOF), a mass-spectrometry-based flow cytometry that reads dozens of markers per cell, applied to tumor, spleen and bone marrow, with human cells gated on human CD45.

Tumor implantation produced 59 evaluable tumors, a 72 percent engraftment rate, leaving 11 matched primary-tumor-to-PDX pairs for direct comparison. The decisive numbers are similarity correlations. PDX tumors derived from the same primary tumor had immune compositions far more alike than tumors from different primaries (median Pearson correlation 0.909 versus 0.537, p less than 0.0001). PDX tumors that merely shared a stem-cell donor were no more alike than tumors from different donors (0.560 versus 0.558, p = 0.93). In other words, the tumor decided, and the donor did not. Blood chimerism did vary by donor but did not predict how heavily a tumor was infiltrated (correlation minus 0.217). Those correlations, though, measure reproducibility across PDX replicates, not fidelity to the patient. Compared directly against the parental clinical samples, the resemblance was real but modest: primary tumors were more like their matched PDX than unmatched ones (median correlation 0.628 versus 0.520, p = 0.012), yet only seven of eleven primaries were closest to their own match, and average immune infiltration ran lower in the PDX than in the patient tumors (15.9 versus 37.8 percent, a non-significant trend). The gaps concentrated in myeloid cells: in two models, monocytes fell from roughly 60 and 80 percent of the patient immune compartment to about 10 percent in the PDX, so even this myeloid-supportive strain under-builds the macrophage compartment it was engineered to rescue.

The pattern held below the level of cell-type counts. CD8 T-cell states, from cytotoxic (granzyme B and K) to exhausted (PD-1 and TIM-3), clustered by tumor of origin: ovarian OV-3 tumors looked exhausted, head and neck HNSCC-4 tumors looked cytotoxic, across different donor backgrounds. Macrophage states defined by FOLR2, TREM2 and PD-L1 did the same. An orthogonal spatial method, imaging mass cytometry over 69 regions from 14 renal tumors, reproduced the tumor-driven organization (same-tumor similarity 0.863 versus 0.802, p less than 0.0001), and xenografts from a patient-derived cell line resembled those from the matched primary tumor. Systemic compartments behaved differently: in spleen and bone marrow, both donor and tumor left a mark.

Where a skeptic should push

The single most load-bearing move is the slide from "reproduces immune composition" to "preserves the tumour immune microenvironment," because the human immune system here is allogeneic. The T cells come from an unrelated cord-blood donor and are not HLA-matched to the tumor, so they cannot mount the ordinary HLA-restricted, tumor-antigen-specific response the model appears to stand in for. Much of the observed reactivity, including the exhaustion markers that look so tumor-specific, could reflect chronic recognition of mismatched HLA rather than genuine antitumor immunity. The paper convincingly shows that the tumor governs which immune cells arrive and what surface states they adopt; it does not show that those cells are fighting the tumor the way a patient's own lymphocytes would. For predicting checkpoint-inhibitor efficacy, that gap is the whole game. The same mismatch reaches back into the headline itself: because each tumor carries its own distinct load of mismatched HLA, the tumor-specific clustering could partly reflect allo-recognition of that tumor's HLA, or tumor-secreted cytokines shaping recruitment, rather than a faithful human immune contexture, and the experiment that would separate them, an HLA-matched or autologous graft set against a mismatched one, is absent. The similarity metric flatters the result too, because a correlation across immune-cell proportions is dominated by the most abundant lineage, so the GM-CSF-boosted macrophages this strain favors can lift a number like 0.909 without fine agreement in the rarer subsets that often decide a therapy.

Two structural caveats reinforce this. The implanted tumors were passaged through NSG mice before humanization, so the patient's original stroma and tumor-infiltrating lymphocytes were long gone; what is modeled is a tumor cell population recruiting a fresh, foreign immune system, not the persistence of the original microenvironment. And the NOG-EXL strain supplies IL-3 and GM-CSF but not IL-15, a survival factor for human natural killer cells; the study does detect NK cells and even finds their fraction fairly concordant between patient and PDX, so this is a caveat inferred from the strain genotype rather than a deficiency the paper reports, but it still bounds confidence in NK-dependent readouts. The headline claim is also hedged in the paper itself, "generally" resembled, "in most models," with poorly infiltrated tumors alongside richly infiltrated ones, and the donor-effect test in the spatial subset was underpowered. This is a strong composition result, not a validated efficacy model.

Where organoids cede ground to humanized mice

The paper never mentions organoid screening, but it draws the boundary of the organoid's validity envelope precisely. Its own framing concedes the organoid's weak point: cultures and ex vivo fragments keep patient immune cells only briefly and cannot capture an antitumor response inside a living organism. The immune compartment is exactly where organoids decay. What this study adds is the reason a living humanized model can be trusted for immune questions at all: the intratumoral infiltrate is driven by the tumor (same-tumor similarity 0.909 versus 0.537; donor effect not significant at p = 0.93), so the signal a screen cares about is not drowned by the donor lottery people feared. That is a claim about reproducibility, not fidelity: a model can be consistently the same across donors and still be an imperfect mirror of the patient, as the modest primary-to-PDX resemblance here shows. Donor-independence is necessary for a usable model, not sufficient for a faithful one.

That reframes model choice as a per-question decision rather than a ladder. For questions about how a tumor shapes its own immune surroundings, which myeloid cells it recruits, whether its macrophages take on a FOLR2 or TREM2 state, how much checkpoint ligand it displays, the humanized PDX is the better-posed of the two models, allogeneic caveat attached, and a tumor-only or short-lived-immune organoid is outside its competence. Running a checkpoint or macrophage-targeting drug through an organoid that lacks a durable T-cell and macrophage compartment produces a mechanistically empty result, the immuno-oncology version of a screen that cannot see its own target.

The threat cuts both ways, which is the honest part. The allogeneic mismatch means the humanized PDX is not a free upgrade either: it models tumor-driven immune recruitment and composition well, but antigen-specific antitumor immunity and checkpoint efficacy poorly. So the constructive reading is a division of labor grounded in mechanism, not a winner. Organoids keep the edge on throughput, patient-matched tumor-cell-intrinsic targets, and short-term autologous immune co-culture where donor and tumor come from the same person. Humanized PDX earns the durable, in-vivo, tumor-driven immune-composition questions, with the allo caveat stapled to every checkpoint claim. The finding that cell-line xenografts resemble matched PDX tumors even hints at a scalable feeder for the in-vivo arm, which is where the two model classes could be stitched into one pipeline.

The bottom line

Established: across 59 humanized PDX tumors and an independent spatial dataset, intratumoral immune composition and CD8 and macrophage cell states cluster by the implanted tumor, not by the stem-cell donor (0.909 versus 0.537 for tumor; 0.93 not significant for donor). The resemblance to the actual patient tumor is real but modest (0.628 versus 0.520), and weakest in the myeloid compartment. Not established: that these infiltrates constitute a faithful model of a patient's antitumor immunity. Because the immune system is non-HLA-matched cord blood and the tumors were pre-passaged in NSG mice, this is a model of tumor-driven immune recruitment and composition, not of autologous, antigen-specific response. What would confirm the stronger reading is an autologous or HLA-matched humanized model showing the same tumor-intrinsic dominance and predicting real checkpoint responses; what would break it is evidence that the composition and exhaustion patterns are driven by allo-reactivity or fail to track any therapeutic outcome.