The Tumor Microenvironment: Why Pancreatic Cancer and Glioblastoma Resist Treatment

A tumor is more than its malignant cells

Imagine a therapy that kills pancreatic cancer cells in a laboratory dish. In a patient, those cells sit among collagen, fibroblasts, vessels and immune cells. The drug may arrive unevenly, while neighboring cells provide signals that help the cancer survive. Now imagine the same experiment with glioblastoma. The laboratory result omits another decisive feature: malignant cells can infiltrate functioning brain beyond the visible mass, including regions where vessels still restrict drug entry. [1–3]

The tumor microenvironment is the living tissue system in which cancer develops: nonmalignant cells, extracellular matrix, vessels, chemical signals and local conditions such as oxygen availability. Cancer changes this environment, and the altered environment changes the cancer. A malignant cell’s behavior therefore reflects both its internal machinery and its surroundings.

Pancreatic ductal adenocarcinoma, or PDAC, and glioblastoma, or GBM, make a revealing comparison. Both resist treatment through overlapping physical, immune and evolutionary mechanisms. But they exploit different organs, supporting cells and routes of spread. Here, PDAC means ductal pancreatic cancer, not pancreatic neuroendocrine tumors; GBM refers principally to adult IDH-wildtype glioblastoma under modern classification, not every high-grade glioma. [1,3,30]

Three tests every treatment must pass

Can it reach the relevant disease? The meaningful measurement is exposure at viable cancer cells, including resistant niches and microscopic deposits. Blood concentration, cerebrospinal-fluid concentration and a measurement from one resected region answer different questions.

Can it work once it arrives? Adequate exposure must produce a biological effect. A drug can encounter its target yet fail because another pathway substitutes for it. An immune cell can arrive but lack the antigen recognition, metabolic resources or functional state needed to kill.

Can the benefit last? Treatment changes which cells survive. A response becomes durable only if residual populations cannot escape, adapt or repopulate the disease. These are useful tests for evaluating a platform, not a claim that a single universal sequence explains every patient’s cancer.

Architecture determines which cells treatment can reach

PDAC builds a dense and biologically active stroma

PDAC often contains extensive desmoplasia: a fibrotic reaction rich in extracellular matrix and cancer-associated fibroblasts, or CAFs. Collagen organizes and stiffens tissue; hyaluronan binds water; fibroblasts continually remodel the surrounding material. The resulting mechanical environment can compress small vessels and create poorly perfused regions. Perfusion means blood actually flowing through tissue, not simply the presence of vessels on a slide. [1,4]

A drug must leave the circulation and distribute through this uneven tissue. Solid stress, matrix organization and fluid transport all matter, but they are not interchangeable physical quantities. In particular, elevated interstitial fluid pressure should not be used as a catch-all explanation for vessel compression. The balance of solid and fluid forces has been explicitly debated in pancreatic cancer research. [4,5]

The stroma also carries biological information. Cells interpret matrix stiffness, adhesive contacts and soluble signals. Two locations with similar amounts of collagen can support different cellular behavior. A pathology description of “fibrosis” therefore does not identify one uniform therapeutic target. [1,2]

GBM infiltrates a functioning organ

Glioblastoma is not simply PDAC behind a different wall. It grows within a neural environment containing astrocytes, microglia, neurons and specialized vessels. Brain extracellular matrix differs from the collagen-rich desmoplastic architecture typical of PDAC. GBM contains matrix remodeling and mesenchymal or fibroblast-like populations, but these should not be presented as a direct equivalent of pancreatic CAF biology. [3]

Its dense tumor regions may contain abnormal vessels, hypoxic areas and necrosis. Beyond them, dispersed malignant cells intermingle with normal brain and follow vascular and white-matter structures. The tissue that must be preserved for speech, movement or cognition can also contain the disease that must be eradicated. [3,30]

The blood–brain barrier is a regional problem

Normal brain vessels use tightly connected endothelial cells, regulated transport and drug-efflux systems to control exchange with the blood. Pericytes and astrocytic endfeet help maintain this neurovascular system. In GBM, tumor-associated vessels can become abnormal and more permeable, creating a heterogeneous blood–tumor barrier, or BTB. [6]

The crucial qualification is geographic. A contrast-enhancing region can have a disrupted barrier while infiltrative, non-enhancing regions retain a relatively intact blood–brain barrier, or BBB. There is also variation within these regions and over time. MRI contrast leakage is not a map of effective exposure for every drug: molecules differ in size, binding, transport and susceptibility to efflux. [6]

Even within a leaky tumor, abnormal flow and interstitial transport can produce uneven delivery. Conversely, a drug can enter the brain yet fail biologically. The question is whether its active, unbound concentration is sufficient at the relevant cells for long enough. Showing entry into an enhancing core does not establish coverage of infiltrating disease. [6,30]

For PDAC, an analogous sampling error arises when measurements from the primary mass are treated as proof of adequate exposure in every metastatic site. For either cancer, the clinically important unit is the distribution of treatment across disease, not the average concentration in a convenient specimen.

Fibroblasts reveal why removing a defense can backfire

CAF is a functional umbrella, not a single cell type with one purpose. Pancreatic studies distinguish myofibroblastic CAFs, which emphasize contractile and matrix-associated programs, from inflammatory CAFs, which emphasize cytokine secretion. Antigen-presenting CAFs express MHC class II machinery, but that label does not mean they prime protective T-cell immunity as effectively as professional antigen-presenting cells. [7,8]

These categories describe overlapping, context-dependent states. Location, signals from malignant cells and treatment can change a fibroblast’s behavior. An inflammatory state is not automatically beneficial because it is “inflammatory,” and a matrix-producing state is not automatically harmful because it makes collagen. Recent reviews emphasize function and spatial neighborhood over a rigid catalogue of permanently separate subtypes. [2,9]

The warning against indiscriminate depletion comes from experiments, not semantics. Özdemir and colleagues depleted αSMA-positive myofibroblasts in mouse pancreatic cancer models and observed more aggressive disease, greater immune suppression and reduced survival. Rhim and colleagues found that reducing Hedgehog-dependent stromal support could similarly yield less differentiated, more aggressive tumors. Some stromal components were restraining malignancy. [10,11]

These findings do not prove that every stromal intervention harms patients or that all fibroblasts should be preserved. They show why “less stroma” is an inadequate therapeutic objective. The better objective is to modify a defined harmful function while preserving useful restraint. Which function, in which tumor, at which treatment stage remains an empirical question.

Immune cells can be abundant while immunity is ineffective

An immune-cell-rich tumor is not necessarily under effective immune attack. PDAC can recruit macrophages and other suppressive myeloid populations, including cells described functionally as myeloid-derived suppressor cells. Together with regulatory T cells and fibroblast-derived signals, they can impair effector T-cell access and activity. Macrophage–fibroblast interactions also couple immune regulation to matrix remodeling. [12]

GBM has a particularly prominent myeloid compartment. It includes brain-resident microglia and recruited blood-derived macrophages, whose histories differ but whose functional programs can overlap. A 2025 human glioma study identified immunomodulatory programs shared across myeloid identities. This argues against treating origin alone—or a simple “M1 versus M2” classification—as a reliable guide to what a cell is doing. [13]

Suppression can involve inhibitory cytokines, altered antigen presentation, nutrient use and effects on neighboring cells. Persistent antigen exposure can also leave T cells dysfunctional. Blocking PD-1 addresses one inhibitory interaction; it does not by itself generate an adequate repertoire of tumor-reactive T cells or remove every suppressive mechanism. [14,15]

Brain tumors add a difficult therapeutic tradeoff. An effective inflammatory response can threaten neurological function through edema within a confined space. Corticosteroids may be necessary to control symptoms but can oppose immune activation; radiation and chemotherapy can also affect lymphocyte numbers. These treatment-related influences complicate both biology and interpretation of immunotherapy trials. [14,30]

Antigen presentation links recognition to killing

A mutation is not automatically an actionable immune target. It must produce a suitable peptide, be processed and displayed in the patient’s HLA context, and encounter a T cell capable of recognizing it. Conventional CD8 T cells recognize peptide–MHC class I complexes on target cells; antigen-presenting cells also help initiate and organize the response. Losing any link can interrupt immunity.

PDAC offers a concrete example. Yamamoto and colleagues showed that autophagy-dependent lysosomal degradation can reduce surface MHC class I, making cancer cells less visible to CD8 T cells. Restoring presentation improved antitumor immunity in mouse experiments. That establishes a mechanism and a combination rationale; it does not establish clinical benefit from routinely adding an autophagy inhibitor. [16]

Fibroblasts expressing MHC class II introduce a different issue: presenting an antigen without the appropriate activating signals need not generate useful immunity. The presence of antigen-presentation machinery and the quality of the resulting response must be measured separately. [8]

Vaccines and engineered cells intervene at different points. A vaccine tries to expand immune recognition, but its effect still depends on tumor visibility and effector function. CAR T cells can recognize selected surface proteins without conventional peptide–HLA presentation, yet require sufficient target expression and must cope with uneven antigen distribution, suppression and toxicity. Broadening recognition can reduce one escape route without eliminating the others. [17,18]

Hypoxia and metabolism change the rules of survival

Poor oxygen delivery is a biological signal as well as a supply problem. Hypoxia-responsive programs can alter angiogenesis, stress tolerance and cell state. Oxygen deprivation can also reduce the effectiveness of radiation, whose DNA-damaging effects are partly oxygen dependent. Hypoxic niches therefore connect abnormal vasculature to treatment resistance and evolutionary selection. [19]

Nutrient use is equally uneven. Cancer cells and activated immune cells both need metabolic resources, but the local constraints depend on the nutrient, cell type and location. Lactate accumulation, acidity and other metabolites can reshape immunity. It is too simple to claim that cancer consumes all the glucose and therefore switches off all T cells. Metabolic competition, metabolite signaling and cooperation can coexist. [20]

In PDAC, stromal cells can supply alternative fuels. Sousa and colleagues showed that pancreatic stellate cells can release alanine through an autophagy-dependent process, supporting cancer-cell metabolism in experimental systems. This is a useful example of a tumor drawing resources from neighbors rather than relying exclusively on delivery from the blood. Its quantitative importance need not be identical in every human tumor. [21]

In GBM, hypoxic and necrotic regions coexist with better oxygenated or vessel-associated niches. Different malignant and immune populations can adopt different metabolic programs. Targeting one pathway can therefore select for another route, and a metabolic inhibitor may also injure the immune cells one hopes to activate. A persuasive metabolic therapy needs evidence of selective vulnerability, not merely evidence that a pathway is active. [22]

Heterogeneity includes the ability to change state

Tumors differ between patients, between regions and over time. Genetic variation is only one layer. Cells with related genomes can express different transcriptional programs, change differentiation state or respond differently to their surroundings. A biopsy can faithfully describe its sampled region while missing another clinically decisive population.

In PDAC, classical and basal-like programs can coexist, with intermediate states and differences between primary and metastatic environments. Raghavan and colleagues showed that microenvironmental conditions influence malignant cell state and drug response; culture conditions themselves can alter the state being studied. A model can preserve a tumor’s mutations while losing an important aspect of its behavior. [23]

In GBM, Neftel and colleagues described major malignant states resembling neural progenitors, oligodendrocyte progenitors, astrocytes and mesenchymal cells. Their abundance is influenced by genetics and local conditions, and cells can transition between states. “Stem-like” behavior is therefore better understood as a capacity that can be maintained or reacquired than as a perfectly isolated, permanent cell caste. [24]

This creates two routes to recurrence: treatment can select a pre-existing resistant population, or surviving cells can adapt. Both may occur together. A drug that eliminates the dominant state can leave a smaller population able to rebuild the tumor in a different form. Durable benefit requires understanding residual disease, not only profiling the pretreatment bulk.

Invasion creates different problems in pancreas and brain

PDAC combines local invasion, including growth along nerves, with the ability to establish distant disease. A metastatic deposit encounters a new organ environment rather than simply reproducing the primary tumor’s surroundings. This is one reason findings in an operable pancreatic primary cannot be assumed to describe advanced systemic disease. [1,23]

GBM typically causes failure through diffuse intracranial infiltration and recurrence rather than widespread extracranial metastasis. Removing the contrast-enhancing mass cannot guarantee removal of dispersed cells in functioning brain. Wider treatment coverage must be weighed against injury to the organ the patient depends on. [30]

Neural tissue also participates actively. Studies in human specimens and experimental glioma systems demonstrated functional neuron-to-glioma synaptic signaling, including AMPA-receptor-mediated input. Connected malignant-cell networks and neuronal activity can support tumor behavior. This expands the meaning of microenvironment beyond immune cells and connective tissue, although a mechanism demonstrated in models does not yet establish an effective treatment strategy for adult GBM. [25,26]

PDAC and GBM compared

These are recurring patterns, not properties shared uniformly by every region or patient. The comparison summarizes the mechanisms discussed above. [1–26,30]

Therapeutic strategies must solve a defined problem

The implication is not to abandon tumor-cell-directed therapy. It is to ask what prevents that therapy from delivering its full benefit. A microenvironment intervention may work best as an enabling component, with a partner capable of exploiting the change. That hypothesis still needs testing against a relevant comparator.

Improve access without confusing access with efficacy

In PDAC, matrix modification and stromal reprogramming aim to improve tissue conditions or interrupt harmful signaling. Selectivity matters because some stromal functions restrain progression. In GBM, approaches include brain-penetrant drug design, local or convection-enhanced delivery, and ultrasound-mediated barrier opening. Their routes and coverage differ; local administration does not guarantee uniform distribution through infiltrated brain. [2,27]

A phase 1 ultrasound study in recurrent GBM measured increased drug concentrations in sonicated brain tissue, providing direct human evidence that a delivery intervention can change exposure. That is an important achievement at the first test. It does not by itself prove that the drug will control every relevant population or improve survival. [27]

Change immune function and recognition together

Myeloid-directed strategies seek to reduce harmful recruitment or reprogram suppressive states; fibroblast-directed strategies may alter the signals that maintain those states. Vaccines, dendritic-cell approaches, oncolytic platforms and engineered lymphocytes address other steps. A combination is most convincing when each component corrects a measured limitation in the same biological setting. [12,14]

PDAC vaccine data illustrate both promise and restraint. In the 2025 follow-up of a small phase 1 study combining surgery, atezolizumab, autogene cevumeran and chemotherapy, vaccine-induced T-cell responses persisted and were associated with delayed recurrence. The immune-responder versus nonresponder comparison was not a randomized test of vaccine benefit; the study cannot isolate the vaccine’s effect from patient biology or the treatment combination. [17]

Local GBM CAR T studies address access and recognition simultaneously. In a 2025 bivalent EGFR/IL-13Rα2 phase 1 report, 8 of 13 patients with measurable disease had tumor regression, but only one had a confirmed partial response under the study’s response criteria. That distinction matters: any shrinkage is not the same as a confirmed objective response. [18]

A 2026 phase 1 study of intracranial B7-H3 CAR T cells treated 15 selected patients with recurrent GBM. It reported encouraging activity, including one sustained complete response at the reported follow-up, alongside serious grade 3 treatment-related events. With no randomized comparator, these findings support further testing rather than a survival advantage claim. Selection, repeat dosing and neurological safety remain central to interpretation. [28]

Coordinate the sequence of treatment

A therapy that reduces suppressive signaling may create a temporary opportunity for another treatment. The useful schedule may depend on when immune cells arrive, how long a delivery window remains open and whether malignant cells change state. Giving more agents together is not a substitute for demonstrating that the proposed sequence produces the intended effect.

KRAS-pathway inhibition in PDAC is also relevant to the environment: malignant-cell signaling helps organize interactions with stromal cells. Recent synthesis emphasizes feedback between oncogenic signaling and the TME as a potential source of adaptation to targeted therapy. This is a rationale for studying resistance and combinations, not evidence that every KRAS inhibitor needs the same stromal partner. [1]

What translational failures actually teach

HALO 301 showed that a plausible stromal target was insufficient

The randomized phase 3 HALO-301 trial enrolled patients with previously untreated, hyaluronan-high metastatic pancreatic cancer. Adding pegvorhyaluronidase alfa, or PEGPH20, to nab-paclitaxel plus gemcitabine did not improve overall survival: median survival was 11.2 versus 11.5 months; the hazard ratio was 1.00. Response rates were higher, but progression-free survival was not improved. [29]

The trial does not show that stroma never matters. Nor does it establish that stromal depletion caused the aggressive phenotype seen in separate mouse experiments. It shows that this intervention, biomarker and chemotherapy combination did not deliver the anticipated survival benefit. A mechanistic rationale and a biomarker-enriched population were not enough.

Checkpoint blockade did not solve the whole GBM immune problem

In CheckMate 143, nivolumab did not improve overall survival over bevacizumab in recurrent GBM: median survival was 9.8 versus 10.0 months, with a hazard ratio of 1.04. This result tests a specific treatment comparison in a specific setting. It neither disproves tumor immunity nor validates the idea that a single checkpoint controls the disease’s suppressive environment. [31]

A vascular or imaging effect need not be a survival effect

In the phase 3 AVAglio trial, adding bevacizumab to radiotherapy–temozolomide for newly diagnosed GBM prolonged progression-free survival but did not significantly improve overall survival. Anti-VEGF treatment can also reduce vascular leakage and edema, altering contrast enhancement. Imaging therefore needs to be interpreted alongside neurological status, steroid use and survival. Symptom relief can be valuable without proving durable tumor eradication. [32,30]

These examples caution against promoting an intermediate signal into a final outcome. They also show why failure analysis must be specific: insufficient exposure, weak biological activity, resistant residual cells, toxicity and an unsuitable clinical setting require different next experiments.

What would make the next result persuasive

For an access strategy, we would look for human measurements in clinically relevant regions, not only plasma or an accessible tumor core. For an immune strategy, we would ask whether tumor-reactive cells reached malignant cells and remained functional. For a targeted strategy, we would want evidence that the intended pathway was inhibited at tolerable exposure and that escape populations were characterized.

Clinical validation then requires a population and comparator that match the claim. Newly diagnosed and recurrent GBM are not interchangeable. Resected PDAC with microscopic residual risk is not equivalent to bulky metastatic disease. A biomarker should predict the treatment effect, rather than merely identify patients who already have a better prognosis.

For HER4’s therapeutic and commercial assessment, the same distinctions matter. A locally delivered cell therapy may impose surgical access, repeat-administration and monitoring requirements. A personalized vaccine adds tissue and manufacturing constraints. Those burdens may be justified by meaningful benefit, but biological elegance alone does not establish practical value. We should ask what improves, for whom, by how much, and at what cost in toxicity and treatment burden.

Evidence limitations

This is a selective explanatory synthesis of published studies and recent reviews, not a systematic review or a complete treatment guideline. Landmark older experiments are included because they establish mechanisms or document pivotal failures; newer papers update the biological interpretation. The evidence-review date does not imply that every study or ongoing trial available by that date has been captured.

Mouse models establish causal relationships under controlled conditions, but their immune systems, treatment histories and disease distribution differ from those of patients. Organoids improve experimental access while omitting or simplifying vessels, immune recruitment and organ-level interactions. Single-cell and spatial studies reveal associations and candidate communication pathways; proximity or a gene signature alone does not prove a functional causal relationship.

Human sampling is also selective. Resected tumors and safely accessible regions are overrepresented; infiltrative margins and difficult metastatic sites may be underrepresented. Prior treatment, corticosteroids and evolving classification complicate comparison between cohorts. A specimen is a time-and-place sample of an evolving disease.

Early trials can establish feasibility, dosing and biological activity. Small uncontrolled cohorts cannot reliably quantify the incremental survival effect of a therapy, and comparisons with historical outcomes are vulnerable to selection and differences in supportive care. The certainty of a claim should follow the study design, not the novelty of its technology.

Key takeaways

• The tumor microenvironment helps determine delivery, biological response and recurrence; these are separate problems that require separate evidence.

• PDAC’s stroma is heterogeneous and can restrain as well as promote malignancy. Indiscriminate depletion can worsen disease in experimental models.

• GBM’s BBB/BTB varies across regions. A permeable enhancing core does not establish adequate treatment of infiltrating cells elsewhere.

• Myeloid suppression, antigen presentation and metabolism interact. Releasing one immune brake may leave recognition and killing impaired.

• Cellular heterogeneity includes reversible states. Eliminating the dominant population may leave cells able to rebuild the disease.

• The strongest therapeutic claim connects a measured mechanism to durable, clinically meaningful benefit in the intended patient population.

Our assessment

The most useful lesson from PDAC and GBM is that a cancer treatment must be evaluated in the tissue system where it is expected to work. The pancreas and brain impose different constraints, and even one tumor contains multiple environments. A successful intervention may need to preserve some structures, alter others and coordinate treatment with a changing malignant population.

That raises the standard for a promising result. We should be able to explain which obstacle was removed, demonstrate that the relevant cells were affected, and show that patients benefited beyond an intermediate laboratory or imaging signal. The next meaningful advance will be the one that connects those observations convincingly.

Evidence cutoff: November 6, 2026. Selected trials illustrate biological and translational principles; they are not a comprehensive treatment landscape.

← Return to Therapeutics