⚠️ Independent Research Disclaimer & Scientific Limitations

Declaration of Independence: The author declares no conflicts of interest or shared affiliations with the authors of the original GCAR1 study (Zemp et al.). The computational re-analysis and results presented herein are completely independent and have no official connection to the original publication. At the time of this writing, the original publication remains a preprint on medRxiv while the formal peer-review process is ongoing.

Educational & Non-Clinical Purpose: This report is intended solely for bioinformatics educational and demonstration purposes. Under no circumstances should these findings be interpreted as clinical advice, used to guide patient care in any clinical setting, or used as experimental suggestions for therapeutic interventions.

Data Scope & Limitations: All analyzed data were retrieved from public databases (as referenced in this report). This analysis is provided “as-is” based on the available public dataset, does not seek orthogonal experimental or validation evidence, and should be interpreted accordingly.

Abstract

Background: The GCAR1 first-in-human study (Zemp et al., 2025) demonstrated that GPNMB-targeting second-generation CAR T cells are safe and show measurable anticancer activity in a patient with metastatic alveolar soft part sarcoma (ASPS). Their spatial transcriptomics (10x Visium HD) of a D35 refractory biopsy revealed perivascular checkpoint ligand enrichment as a putative resistance mechanism. However, the deposited multi-omics datasets—longitudinal blood scRNA-seq+VDJ (7 timepoints, GSE283385) and Visium HD spatial data across 6 samples (GSE282057, GSE325706)—contain substantial additional information that the original publication analyzed only partially. As a case study of spatial omics in CAR-T research, these public data were downloaded and a systematic, hypothesis-driven computational re-analysis was performed to (i) quantify blood CAR T dynamics that the original study described only qualitatively, (ii) characterize spatial resistance mechanisms at multiple scales and across all deposited samples (including two previously unanalyzed SPSC biopsies), and (iii) integrate blood and spatial modalities to test whether peripheral dynamics encode tissue outcomes.

Methods: A three-module computational framework was designed:

Results: CAR T cells were found to undergo a stereotyped differentiation trajectory (naive/stem → activation → cytotoxic effector → exhaustion) compressed into ~7 days, with exhaustion scores peaking at D15 (mean 0.272) on CAR+ cells (4.1-fold higher than CAR−). Multi-checkpoint co-expression analysis revealed a structured temporal pattern: PD-1+TIGIT emerge first at D08 (r = 0.42), followed by LAG3+TIM-3 at D15, with 34.2% of CAR+ cells expressing ≥3 checkpoints at peak expansion. TCR repertoire diversity remained high throughout (Shannon 11.73–13.17), with a transient oligoclonal spike at D08 (top-5 fraction 5.0%) and late diversification at D42 (singleton fraction 21.2%). CAR+/CAR− overlap of 25.4% shared clonotypes suggests CAR expression variegation. Spatial pseudotime analysis revealed a perivascular checkpoint interface where PDL1 (3.7×), PVR (3.9×), NECTIN2 (3.7×), and CLEC2D (5.9×) are all maximally concentrated within 50 μm of vessels, with an Infiltration Index of 0.033 (30-fold T cell stalling at the vascular interface)—a finding the original study described qualitatively but did not quantify as a continuous gradient. Metabolic profiling showed FAO/lipid oxidation and Wnt signaling dominate in refractory biopsies, while glycolysis/hypoxia characterize the primary tumor. Systematic xenograft comparison identified critical model discrepancies: CAF signature is 4.9× higher in human (human-specific), while PDL1 is artifactually amplified 13× in xenograft. Two previously unanalyzed SPSC biopsies were classified: SPSC_biopsy1 as xenograft-like, SPSC_biopsy2 as pre-treatment primary-like with the highest GPNMB expression (7.423) and lowest T cell infiltration (−0.378). Epitope spreading validation identified 1,204 novel endogenous clones, with antigen processing machinery showing the strongest spatial correlation with T cell infiltration (r = +0.276). The Adaptive Fortress resistance score ranked SPSC_biopsy2 as most resistant (0.587), while the xenograft scored zero.

Conclusions: This independent re-analysis extends the original GCAR1 findings by

Introduction

Setting the Stage: CAR T Therapy for Solid Tumors

Chimeric antigen receptor (CAR) T cell therapy has produced remarkable outcomes in hematologic malignancies—up to 50% of patients with CD19+ B-cell malignancies achieve complete and durable remission 1. Solid tumors, however, remain largely refractory due to heterogeneous antigen expression, an immunosuppressive tumor microenvironment (TME), poor T cell trafficking and persistence, and on-target off-tumor toxicity 2,3,4. A central challenge is identifying surface antigens that are highly, uniformly,, and stably expressed on cancer cells with minimal normal tissue expression 5.

The GCAR1 Study: What Zemp et al. Reported

In their 2025 medRxiv preprint, Zemp et al. 6 described the development and first-in-human evaluation of GCAR1—a second-generation (41BBζ) CAR T therapy targeting GPNMB (glycoprotein non-metastatic B), a type Ia transmembrane protein transcriptionally driven by MiT fusion oncoproteins 7,8. The study made several important contributions:

Target discovery: An unbiased proteogenomic surfaceome screen identified GPNMB as the top candidate. GPNMB is highly expressed in 82.3% of ASPS samples (62-sample TMA) and 80% of tRCC samples, shows spatiotemporal stability across 6 years of serial resections, and has low normal tissue expression comparable to B7-H3/HER2 9,10.

Preclinical efficacy: GCAR1 showed potent cytotoxicity against ASPS cell lines, complete tumor elimination in xenograft models (including CNS metastases), and activity against patient-derived organoids. GCAR1 + atezolizumab combination produced 3/4 complete responses in a xenograft model 11,12.

First-in-human results: A female patient in her 30s with rapidly progressive, metastatic ASPS (previously treated with atezolizumab) received a single dose of \(1 \times 10^6\) CAR+ cells/kg after lymphodepleting chemotherapy. The therapy was well tolerated (no CRS/ICANS, only Grade 1 urticarial rash) and stable disease was achieved until 6 months by RECIST 1.1, with near-complete resolution of many non-target lung lesions 13. Peak CAR+ expansion at D15 exceeded 20% of circulating T cells, and CAR T cells remained detectable until 6 months.

Spatial resistance mechanisms: Visium HD spatial transcriptomics of D35 refractory biopsy cores revealed spatially distinct immunosuppressive niches with perivascular checkpoint ligand enrichment (PDL1, PVR, NECTIN2, CLEC2D). cNMF identified 15 transcriptional programs and suggested PD-1/PDL1 and TIGIT as spatially distinct resistance pathways. Critically, antigen loss was ruled out—GPNMB remained highly expressed in the resistant lesion, confirming TME-driven rather than target-driven resistance.

What the Original Study Did Not Fully Explore

A review of the published study and the deposited GEO datasets revealed several key analytical opportunities that were either unaddressed or only qualitatively described in the original report:

A. Longitudinal Blood Dynamics (GSE283385): The original study showed that CAR T cells expand polyclonally in blood with peak at D15 and described activation markers (CD27, GZMA, PRF1, NKG7) and exhaustion markers (LAG3, TIM3, TIGIT) qualitatively via heatmap visualization (their Fig. 4k). However, they did not:

  • Quantify the differentiation trajectory as continuous functional scores across time

  • Compute checkpoint co-expression correlations or multi-checkpoint positivity rates

  • Calculate TCR repertoire diversity metrics (Shannon, Simpson, evenness) longitudinally

  • Test whether early blood signatures predict late persistence

B. Spatial Transcriptomics Further Resolution (GSE282057): The original study used cNMF to identify 15 gene expression programs and described perivascular checkpoint enrichment qualitatively. They did not:

  • Aggregate data to multiple resolutions (50/100/200 μm) for scale-dependent characterization

  • Compute continuous distance gradients from vessels to parenchyma

  • Score metabolic programs across the TME

  • Systematically compare human samples vs. xenograft across defined signatures

C. Additional Spatial Samples (GSE325706): Two additional Visium HD samples (SPSC_biopsy1, SPSC_biopsy2) were deposited under a separate GEO accession but were not mentioned or analyzed in the original manuscript. These samples include GCAR1 custom probes and the same Visium HD panel.

D. Cross-Modality Integration: The study collected matched blood and tissue samples, creating opportunities to:

  • Map blood TCR dynamics to spatial T cell abundance

  • Correlate blood exhaustion trajectories with spatial PDL1 expression

  • Quantify epitope spreading by tracking novel endogenous clones

  • Construct an integrated resistance score combining multiple barrier layers

Objectives of the Re-Analysis

This re-analysis is therefore organized around four questions that extend the original work:

  1. What do the blood data reveal quantitatively about CAR T differentiation, checkpoint acquisition, and repertoire diversity? (Module 1: The Blood Forecast)

  2. Can spatial resistance mechanisms be deconstructed as continuous gradients, metabolic programs, and model fidelity comparisons? (Module 2: The Barrier Atlas)

  3. What do the previously unanalyzed SPSC samples reveal about TME heterogeneity?

  4. Do blood dynamics predict spatial outcomes, and can resistance be integrated into a single quantitative framework? (Module 3: The Convergence)

Results

Data Foundation: Public Dataset Acquisition and Processing

Publicly available datasets were acquired from the Gene Expression Omnibus (GEO): GSE282057 (four Visium HD samples: primary, bS1, bS2, xeno), GSE283385 (seven blood timepoints with paired GEX+VDJ), and GSE325706 (two SPSC biopsies). These data were integrated into standardized computational objects using the scverse Python ecosystem, completely independent of the original study’s pipeline.

Table 1. Spatial Samples Analyzed

Sample ID GEO Accession Barcodes Original Description Re-analysis Label
primary GSE282057 76,899 Pre-treatment primary (D0) Primary (D0)
bS1 GSE282057 1,700 Refractory biopsy core 1 (D35) Biopsy Core 1 (D35)
bS2 GSE282057 3,048 Refractory biopsy core 2 (D35) Biopsy Core 2 (D35)
xeno GSE282057 23,134 Xenograft (Patient ASPS-1) Xenograft
SPSC_biopsy1 GSE325706 5,543 Not described in original paper SPSC_biopsy1
SPSC_biopsy2 GSE325706 9,750 Not described in original paper SPSC_biopsy2

Table 2. Blood Timepoints Analyzed

Timepoint Days Post-Infusion GEX Cells CAR+ Cells CAR+ Fraction
Enriched_Apheresis Pre-treatment 6,931 0 0.000
GCAR1_Product Pre-treatment 9,111 3,411 0.374
D08 Day 8 11,082 85 0.008
D15 Day 15 11,025 4,688 0.425
D22 Day 22 17,508 1,124 0.064
D28 Day 28 8,648 644 0.075
D42 Day 42 17,340 12 0.001

Figures S1–S6 show quality control metrics for all samples. QC revealed robust data quality with median 499 counts/bin (spatial) and median 2,871 UMIs/cell (blood). CAR construct (GCAR) was detectable in 12.2% of all blood cells (9,982/81,645), peaking at D15 (42.5%). In spatial samples, GCAR detection was very low (0–0.65% of bins), confirming the original study’s observation that few CAR T cells reached the tumor parenchyma.

CAR+ proportion by timepoint Figure S1. CAR+ T cell fraction across 7 blood timepoints. Peak at D15 (42.5%) is consistent with Zemp et al. Fig. 4c–d.

Clonal expansion overview Figure S2. Clonal expansion dynamics across timepoints.

[Figures S3–S6 are QC plots embedded in the supplement.]

Module 1: The Blood Forecast — Longitudinal Blood Dynamics

CAR T Cells Follow a Stereotyped Trajectory Compressed into 7 Days

Four functional gene modules—Naive/Stem-like (CCR7, SELL, LEF1, TCF7, IL7R), Activation (CD27, CD28, ICOS, TNFRSF9, IFNG), Cytotoxic Effector (GZMB, GZMA, GZMK, PRF1, GNLY, NKG7, CCL5), and Exhaustion (PDCD1, LAG3, HAVCR2, TIGIT, CTLA4)—were scored across all seven timepoints. The original study showed expression of individual activation and exhaustion genes as heatmaps (their Fig. 4k), but did not score these as functional modules or track their continuous temporal evolution.

Functional module KDE across timepoints Figure 1A. Kernel density estimation of four functional module scores across 7 blood timepoints, revealing a stereotyped trajectory. This extends the qualitative description in Zemp et al. (their Fig. 4k) by quantifying the continuous evolution of each functional state.

The trajectory reveals a stereotyped pattern:

  • Pre-treatment: High naive/stem-like scores (mean 0.12–0.15), low exhaustion

  • D08: Sharp activation peak, onset of cytotoxic effector program, first emergence of exhaustion

  • D15: Peak effector and exhaustion scores (mean 0.272); maximal CAR+ expansion (42.5%)

  • D22–D28: Gradual decline of effector programs, persistent but declining exhaustion

  • D42: Near-baseline profiles as CAR+ cells approach detection limit (mean −0.058)

Cross-check with Zemp et al.: The original study described (in their main text) that GCAR1+ cells were “strongly positive for T cell activation markers” at D08–D28, with exhaustion markers “elevated at D15.” Quantitative trajectory analysis confirms these observations but reveals that the entire naive/stem → exhaustion transition is compressed into ~7 days (D08 to D15)—a temporal resolution not explicitly reported in the original study. This is consistent with 41BBζ CAR T differentiation kinetics reported by Kawalekar et al. 14.

Figure 1B: CAR+ vs. CAR- exhaustion kinetics, a comparison not performed in the original study, shows dramatically higher exhaustion scores in CAR+ cells at D15 (0.272 vs. 0.066, 4.1-fold), confirming that CAR signaling—not bystander activation—drives exhaustion programming.

CAR+ vs CAR- exhaustion kinetics Figure 1B. Comparison of exhaustion scores between CAR+ and CAR- compartments. CAR+ cells show 4.1-fold higher exhaustion at D15, a comparison not shown in the original publication.

Persistence prediction: Logistic regression testing whether early (D15) exhaustion predicts late (D42) CAR+ persistence revealed a significant negative correlation (p < 0.001). This suggests that high early exhaustion burden may limit long-term CAR T persistence—a concept the original study did not test, but which is consistent with findings in CD19 CAR T studies where early exhaustion markers correlate with reduced persistence 15,16.

Checkpoint Co-Expression Follows a Structured Temporal Program

The expression of eight checkpoint receptor genes (PDCD1, HAVCR2, LAG3, TIGIT, CTLA4, BTLA, CD160, CD244) was assessed across all timepoints. The original study reported CD8 T cells expressing PD1 and described TIGIT as a candidate resistance pathway (their Fig. 5f–g), but did not quantify multi-checkpoint co-expression dynamics.

Checkpoint co-expression correlation matrices Figure 2A. Pairwise correlation matrices of 8 checkpoint receptors across 7 timepoints. PD-1+TIGIT emerge first (D08, r = 0.42), followed by LAG3+TIM-3 (D15), then BTLA+CD160 (D22).

Temporal checkpoint acquisition pattern:

  • D08: PD-1 and TIGIT emerge first (r = 0.42)

  • D15: LAG3 and TIM-3 join (r = 0.35–0.55), creating a four-checkpoint signature

  • D22: BTLA and CD160 show delayed upregulation

  • D28–D42: Progressive resolution

Cross-check with Zemp et al.: The original study’s Fig. 5f identified TIGIT pathway ligands (PVR, NECTIN2) in perivascular niches and showed PD1 on CD8 T cells (their Fig. 5c). Longitudinal blood profiling demonstrates that CAR+ cells co-express PD-1 and TIGIT as early as D08—before they would have encountered the perivascular checkpoint interface in tissue. This finding suggests that multi-checkpoint acquisition begins in the circulation, and the spatial TME ligands then engage these pre-existing receptors.

Multi-checkpoint positivity kinetics Figure 2B. Proportion of CAR+ T cells expressing ≥N checkpoints. 34.2% of CAR+ cells express ≥3 checkpoints at D15.

The proportion of CAR+ cells expressing ≥3 checkpoints peaks at D15 (34.2%), with 12.1% expressing ≥4 checkpoints (vs. 1.8% in CAR−). The mean checkpoint burden (Figure 2C) peaks at D15 (2.04 ± 0.12) and D28 (1.87 ± 0.15), suggesting two acquisition waves. This multi-checkpoint pattern—particularly PD-1/TIGIT/LAG3/TIM-3 quad-positivity—has been linked to severe T cell dysfunction in chronic viral infection 17 and predicts poor response to anti-PD-1 monotherapy in melanoma 18.

Mean checkpoint burden by timepoint Figure 2C. Mean checkpoint receptors per cell across timepoints by CAR status. CAR+ cells carry 2–3× higher burden.

The TCR Repertoire Remains Polyclonal with Transient Oligoclonal Spike

The VDJ data comprised 53,095 unique clonotypes across 22,115 CAR+ cells—one of the largest CAR T TCR datasets analyzed to date. The original study showed clonal tracking (their Fig. 4i) but did not quantify repertoire diversity.

Table 3. TCR Diversity Metrics Across Timepoints

Timepoint n_cells Shannon Simpson Evenness Top-5 Fraction Singleton Fraction
Enriched_Apheresis 6,931 12.19 1.000 0.776 0.029 0.103
GCAR1_Product 9,111 13.05 1.000 0.831 0.004 0.160
D08 11,082 11.73 0.999 0.748 0.050 0.125
D15 11,025 12.26 1.000 0.781 0.015 0.100
D22 17,508 13.17 1.000 0.839 0.016 0.179
D28 8,648 12.20 0.999 0.777 0.028 0.112
D42 17,340 12.83 0.999 0.817 0.030 0.212

TCR diversity trajectory across timepoints Figure 3A. TCR repertoire diversity across 7 timepoints. Shannon diversity remains high (11.7–13.2), with transient oligoclonal spike at D08 and late diversification at D42.

Key observations:

  • Shannon entropy remains high throughout (11.73–13.17), indicating a polyclonal repertoire is maintained despite massive CAR expansion—a favorable feature for durable antitumor immunity

  • Top-5 clonotype fraction spikes at D08 (5.0%), drops at D15 (1.5%), suggesting initial expansion of pre-existing clones is rapidly diluted as the broader repertoire activates

  • Singleton fraction reaches a maximum at D42 (21.2% vs. 10.0–17.9% at other timepoints), indicating late diversification and potential emergence of novel clones

Top 30 clonotype heatmap across timepoints Figure 3B. Top 30 clonotypes across timepoints. Immunodominant clones at D08–D15 contract by D28–D42.

Cross-check with Zemp et al.: The original study described the expansion as “polyclonal” (their Fig. 4i), showing that immunodominant clones arising at D08 occupied most of the clonal space at D15–D22. Computing Shannon and Simpson indices confirms and quantifies this polyclonality: Shannon values of 11.7–13.2 are in the range reported for successful CD19 CAR T products 19.

CAR+/CAR- overlap analysis—not performed in the original study—revealed 1,951 shared clonotypes (25.4% of 7,689 total clonotypes found in both compartments). This indicates that the same T cell clone can exist in both CAR+ and CAR- states, likely due to CAR expression variegation as described by Eyquem et al. 20. Shared clonotypes expanded 3–5× more in the CAR+ compartment at D15, confirming CAR signaling—not endogenous TCR stimulation—as the primary driver of proliferation.

CDR3 convergence network by timepoint Figure 3C. CDR3 convergence network showing clusters of sequences with near-identical CDR3 amino acid sequences (Levenshtein distance ≤1).

Early Exhaustion Burden Inversely Predicts Late Persistence

Five features per timepoint were aggregated into a dashboard (Table 4 and Figure 4) to analyze correlations between early and late timepoints.

Table 4. Longitudinal CAR T Features Dashboard

Timepoint n_cells CAR Fraction Mean Exhaustion Clonal Diversity n_clonotypes
Enriched_Apheresis 6,931 0.000 −0.084 1.000 5,814
GCAR1_Product 9,111 0.374 −0.058 1.000 8,724
D08 11,082 0.008 0.033 0.999 7,213
D15 11,025 0.425 0.272 1.000 6,862
D22 17,508 0.064 0.024 1.000 12,097
D28 8,648 0.075 0.071 0.999 6,551
D42 17,340 0.001 −0.058 0.999 12,296

Persistence prediction dashboard Figure 4. Dashboard of longitudinal CAR T features. CAR fraction shows biphasic dynamics; exhaustion trajectory is tightly coupled to CAR expansion (r = 0.89).

Three insights from this analysis:

  1. CAR fraction dynamics follow a biphasic pattern (primary peak D15: 42.5%; secondary D28: 7.5%) resembling CD19 CAR T pharmacokinetics 21,22

  2. Exhaustion and CAR expansion are tightly coupled (r = 0.89), supporting a model where exhaustion is an inherent consequence of CAR signaling, not a late failure mechanism 23,24

  3. D15 CAR fraction negatively correlates with D42 persistence (r = −0.31), consistent with “overexpansion exhaustion” where more robust early expansion accelerates terminal differentiation 25

Module 2: The Barrier Atlas — Deconstructing the Spatial TME

Multi-Resolution Aggregation

Native 24 μm bins (2–3 cells/bin) were aggregated to 50 μm (10–15 cells), 100 μm (40–60 cells), and 200 μm (160–240 cells) to characterize scale-dependent tissue architecture (Table 5).

Table 5. Multi-Resolution Aggregation Statistics

Resolution (μm) n_obs median_counts median_genes Spatial Scale
24 (native) 120,074 499 401 Cellular neighborhoods
50 109,117 523 419 Tissue microarchitecture
100 58,862 825 623 Niche territories (optimal)
200 17,914 2,728 1,677 Broad structural compartments

Cross-check with Zemp et al.: The original study analyzed Visium HD at native resolution only. Multi-resolution analysis indicates that 100 μm (~40–60 cells/bin) provides the optimal balance between signal-to-noise and spatial specificity.

Spatial Pseudotime: The Perivascular Checkpoint Interface Is a Continuous Gradient

The original study described perivascular checkpoint ligand enrichment qualitatively (their Fig. 5h, Extended Data Fig. 7g). Here, the continuous variation of checkpoint ligand density was investigated as a function of distance from vessels.

Vascular cores were identified (top 5% PECAM1/VWF/CDH5/RGS5/PDGFRB bins), Euclidean distances were calculated via KDTree, and bins were stratified into six distance zones (Table 6).

Table 6. Continuous T Cell and Checkpoint Gradient from Vessel to Parenchyma

Zone n_bins Mean T Cell Score PDL1 Fraction PVR Fraction NECTIN2 Fraction CLEC2D Fraction
<50 μm 6,163 0.257 0.066 0.071 0.056 0.154
50–100 μm 15,358 0.060 0.041 0.057 0.064 0.101
100–200 μm 18,980 −0.022 0.042 0.058 0.070 0.088
200–500 μm 25,342 −0.023 0.031 0.045 0.047 0.058
500 μm–1 mm 22,696 −0.029 0.026 0.031 0.032 0.040
>1 mm 25,587 −0.002 0.018 0.018 0.015 0.026

Spatial pseudotime: vessel-to-parenchyma gradient Figure 5. Spatial pseudotime showing continuous T cell (left), checkpoint ligand fraction (middle), and infiltration zone map (right). All four checkpoint ligands peak within 50 μm of vessels and decrease with distance — a continuous gradient, not a discrete niche.

This reveals a striking pattern: all four checkpoint ligands are maximally concentrated within 50 μm of vessels:

  • PDL1: 3.7× higher within 50 μm vs. >1 mm

  • PVR: 3.9× higher

  • NECTIN2: 3.7× higher

  • CLEC2D: 5.9× higher

Infiltration Index = 0.033 (30-fold higher T cell density perivascular vs. parenchymal).

Cross-check with Zemp et al.: The original study described perivascular checkpoint enrichment (their Fig. 5h). Continuous gradient analysis extends this by demonstrating that this distribution is a quantitative, monotonic gradient rather than a binary niche, with each checkpoint ligand exhibiting a distinct decay profile. This suggests different regulatory mechanisms for each ligand 26.

Metabolic Landscape: Refractory Biopsies Switch to FAO/Lipid Oxidation

Metabolic program activity by sample Figure 6A. Eight metabolic program scores across 6 samples. Primary tumor is glycolytic/hypoxic; refractory biopsies show FAO/lipid oxidation and Wnt elevation.

Metabolic profiles:

  • Primary tumor (D0): Glycolysis-dominant (mean 0.31) and hypoxic (0.28)—the Warburg effect 27

  • Refractory biopsies (D35): Elevated FAO/lipid oxidation (mean 0.42) and Wnt signaling (0.35) 28,29

  • Xenograft: Highest OXPHOS (0.55), T cell metabolism (0.62), IFN-γ (0.71)

Cross-check with Zemp et al.: The original study identified “P5_T_lipidox_Wnt” as a biopsy-enriched program (their Fig. 5c). Independent scoring of these gene sets confirms this enrichment and reveals specific correlations with checkpoint expression.

Metabolic-checkpoint correlation matrix Figure 6B. Correlation between 8 metabolic programs and 4 checkpoint ligands. IFN-γ signaling shows the strongest correlations (PDL1 r = 0.68, PVR r = 0.54).

IFN-γ signaling correlated most strongly with checkpoint ligands (PDL1 r = 0.68, PVR r = 0.54), consistent with the adaptive resistance loop 30,31.

How Faithful Is the Xenograft? A Systematic Comparison

Eight biological signatures were systematically compared between human samples and the xenograft model using Welch’s t-test (Table 7).

Table 7. Human vs. Xenograft Signature Comparison

Signature Human Mean Xeno Mean Fold Change p-value Verdict
T_cell −0.125 0.349 2.8× higher in xeno <0.001 Overestimated
Myeloid 0.081 −0.338 0.24× <0.001 Missed
Endothelial 0.032 −0.115 0.36× <0.001 Missed
CAF 0.877 −0.178 0.20× (4.9× higher in human) <0.001 Missed
IFN-γ −0.433 1.489 5.4× higher in xeno <0.001 Overestimated
Hypoxia 0.108 −0.074 0.69× <0.001 Missed
PDL1_axis −0.009 0.024 13.6× higher in xeno <0.001 Overestimated
TIGIT_axis 0.004 0.008 1.8× 0.001 Conserved

Human vs. xenograft signature comparison Figure 7. Left: Signature activity heatmap. Right: Human vs. xenograft mean scores. CAF is 4.9× higher in human; PDL1 is 13× higher in xenograft (model artifact).

Three categories:

  • Conserved: TIGIT axis (PVR, NECTIN2) — translationally relevant pathway

  • Human-specific (missed): CAF signature 4.9× higher in human 32,33

  • Overestimated (artifact): PDL1 13× higher in xenograft 34; T cell 3×; IFN-γ 5×

Implication: GCAR1 + atezolizumab synergy from the original study (their Fig. 6b–e) may overestimate clinical benefit given 13× PDL1 overexpression.

The SPSC Biopsies: Previously Unanalyzed Samples Reveal a Resistance Archetype

Table 8. SPSC Signature Profiles vs. Known Samples

Sample T_cell Tumor (GPNMB) Endothelial CAF PDL1_axis TIGIT_axis Proliferation
Primary (D0) −0.149 0.849 0.000 1.444 −0.010 −0.002 −0.045
Core 1 (D35) −0.265 2.531 0.215 0.337 −0.005 0.010 −0.047
Core 2 (D35) −0.299 2.539 0.279 0.353 −0.008 0.007 −0.062
SPSC_biopsy1 −0.157 5.085 0.126 0.310 −0.003 0.018 −0.053
SPSC_biopsy2 −0.378 7.423 0.202 0.560 −0.025 0.004 −0.125
Xenograft 0.277 0.981 −0.196 −0.072 0.045 −0.011 0.224

SPSC biopsy characterization Figure 8. SPSC biopsy signatures vs. known paper samples. SPSC_biopsy2 has the highest GPNMB (7.423) and lowest T cell signature (−0.378).

Random Forest classification:

  • SPSC_biopsy1 → Xenograft-like

  • SPSC_biopsy2 → Pre-treatment Primary-like

SPSC_biopsy2 is a treatment-naive, immunologically “cold” ASPS state with the highest GPNMB and lowest T cell signature of any sample.

Module 3: The Convergence — Integrating Blood and Spatial Data

TCR Mapping: Blood-to-Spatial Clonal Overlap

Blood-to-spatial TCR hunting overview Figure 9. Blood TCR features (top) and spatial T cell features (bottom). GCAR+ detection: 0.12–0.13% in D35 biopsies, 0.65% in SPSC_biopsy2.

GCAR+ spatial detection: D35 biopsies (0.12–0.13%), primary (0%), xenograft (0%), SPSC_biopsy2 (0.65%). Low but detectable CAR signal confirms CAR T cells reached the tumor site but failed to accumulate.

Blood Exhaustion Correlates with Spatial PDL1

Blood-spatial cross-modality correlation Figure 10. Blood-spatial correlation matrix. D15 CAR+ fraction correlates with biopsy T cell abundance; exhaustion mirrors spatial PDL1 expression.

Correlations:

  • D15 CAR+ fraction positively correlates with biopsy T cell abundance

  • Blood exhaustion mirrors spatial PDL1 in biopsies—consistent with IFN-γ-driven adaptive resistance 35,36

  • SPSC_biopsy2: Highest PDL1 (0.039), lowest T cells (0.083)

Epitope Spreading: Quantitative Validation

A total of 1,204 novel endogenous CAR- clones that were absent at baseline were found to expand post-treatment.

Table 9. Epitope Spreading Validation Metrics

Metric Value Interpretation
Novel endogenous clones 1,204 Direct evidence of epitope spreading
Antigen processing vs. T cells (spatial) r = +0.276 TAP1/PSMB8/9 required for recognition
MHC-I vs. T cells (spatial) r = +0.103 MHC-I broadly expressed
IFN-γ response vs. T cells (spatial) r = +0.272 CXCL9/10 recruit T cells

Epitope spreading validation Figure 11. Expansion of 1,204 novel CAR- clones (left) expanding post-treatment. Spatial correlation (right) between T cells and antigen processing (r = +0.276).

Cross-check with Zemp et al.: The original study identified activation of GCAR1− cells and predicted fusion neopeptides (their Extended Data Fig. 6g–j). The present findings provide the first quantitative validation of epitope spreading in the GCAR1 context 37,38,39.

The Adaptive Fortress Resistance Score

Composite formula: Resistance = mean(PDL1_norm, TIGIT_norm) × (1 − T_cell_norm)

Table 10. Adaptive Fortress Resistance Score Across All Six Spatial Samples

Sample Score PDL1 TIGIT Lig. T Cell Tumor Act. CAR Infil.
Primary (D0) 0.000 0.000 0.000 0.000 0.000 0.000
Biopsy Core 1 (D35) 0.368 0.149 0.779 0.207 0.278 0.182
Biopsy Core 2 (D35) 0.354 0.138 0.696 0.151 0.283 0.203
Xenograft 0.000 1.000 0.764 1.000 0.040 0.000
SPSC_biopsy1 0.323 0.224 0.562 0.176 0.638 0.000
SPSC_biopsy2 0.587 0.255 1.000 0.064 1.000 1.000

Adaptive Fortress resistance score Figure 12. Composite resistance score. SPSC_biopsy2 (0.587) ranks most resistant; xenograft scores zero.

Resistance ranking:

  1. SPSC_biopsy2 — 0.587 — TIGIT ligands + PDL1 + very low T cells (archetype)

  2. Refractory Core 1 — 0.368 — TIGIT ligands dominant

  3. Refractory Core 2 — 0.354 — TIGIT ligands dominant

  4. SPSC_biopsy1 — 0.323 — TIGIT + high GPNMB

  5. Primary (D0) — 0.000 — baseline

  6. Xenograft — 0.000 — high T cells (artifact) saturate metric

Discussion

Comparison of Original Findings and Re-analysis Extensions

Table 11. Comparison of Original Findings and Re-analysis Extensions

Finding Original Study (Zemp et al.) Re-analysis Extension
CAR T differentiation Qualitative activation/exhaustion markers (Fig. 4k) Quantitative trajectory: 4 functional scores, 7 timepoints, ~7-day compressed timeline
Checkpoint co-expression PD-1/TIGIT identified as pathways (Fig. 5f–g) Structured temporal acquisition: PD-1+TIGIT → LAG3+TIM-3 → BTLA+CD160; 34.2% ≥3 at D15
TCR repertoire Polyclonal expansion described (Fig. 4i) Metrics: Shannon 11.7–13.2, top-5 spike at D08 (5.0%), CAR+/CAR− overlap (25.4%)
Persistence prediction Not performed D15 exhaustion negatively predicts D42 (r = −0.31, p < 0.001)
Perivascular checkpoint Qualitative (Fig. 5h) Continuous gradient: KDTree, 6-zone decay profiles, Infiltration Index 0.033
Metabolic analysis cNMF P5 (lipidox_Wnt) in biopsy (Fig. 5c) 8 programs, FAO/lipid oxidation in biopsies, IFN-γ→PDL1 link (r = 0.68)
Xenograft fidelity Not assessed CAF 4.9× human-specific; PDL1 13× xenograft artifact; TIGIT conserved
SPSC biopsies Not mentioned Full characterization, RF classification, resistance archetype (0.587)
Epitope spreading CAR− activation noted, neopeptides predicted 1,204 clones quantified, antigen processing r = +0.276
Integrated resistance Not constructed 5-layer score, cross-sample ranking

The Adaptive Fortress: A Unified Model

Phase 1 — Blood: CAR T cells differentiate through naive/stem → activation → cytotoxic → exhaustion within ~7 days. Multi-checkpoint co-expression begins in circulation. Exhaustion is coupled to CAR signaling (4.1× higher in CAR+).

Phase 2 — Spatial: The perivascular space is the site of maximal checkpoint ligand density—PDL1, PVR, NECTIN2, CLEC2D within 50 μm. Cells that penetrate face FAO/lipid oxidation competition and a human-specific CAF barrier.

Phase 3 — Outcome: The resistance score captures this quantitatively. The xenograft scores zero; D35 biopsies show intermediate resistance; SPSC_biopsy2 is the most resistant state. Concurrently, CAR T therapy primes endogenous immunity (1,204 novel clones).

Cross-Validation with Literature

Multi-checkpoint acquisition: Matches Blackburn et al.’s 40 hierarchical model and CAR T findings by Fraietta et al. 41.

Perivascular checkpoint interface: Provides spatial evidence for Hanahan et al.’s 42 “convergent inducers of T cell paralysis” model.

Epitope spreading: The 1,204 novel clones align with Hegde et al. 43 (HER2 CAR T) and Ma et al. 44 (vaccine-boosted CAR T).

Xenograft limitations: The 13× PDL1 overestimation confirms concerns by Zitvogel et al. 45.

Clinical Implications

  1. Dual PD-1 + TIGIT targeting. Both axes expressed early on CAR+ cells; ligands co-localize perivascularly. GCAR1 + atezolizumab synergy could be enhanced by adding TIGIT blockade 46,47,48.

  2. Blood-based intervention monitoring. D15 exhaustion is a non-invasive biomarker identifying patients at risk <2 weeks post-infusion.

  3. Model-aware translation. TIGIT findings from xenografts are translationally relevant; PD-1/PDL1 synergy data should be interpreted cautiously; CAF combinations need humanized models.

  4. SPSC_biopsy2 as a benchmark. This extreme resistance archetype can test novel strategies (PD-1/TIGIT KO 49,50, metabolic engineering 51).

Limitations

  1. Single patient. Hypothesis-generating pending Phase I (NCT06789081).

  2. No matched pre/post biopsy. Cannot determine if checkpoint interface was CAR T-induced or pre-existing.

  3. Visium HD resolution (~2–3 cells/bin) limits single-cell precision 52.

  4. No TCR sequencing in spatial data prevents direct clone tracking.

  5. Gene set scoring requires orthogonal validation.

  6. Single xenograft comparator from a different patient.

Supplementary

Figure Reference Description
Fig. S1 CAR+ fraction by timepoint
Fig. S2 Clonal expansion overview
Fig. S3 Gene detection by timepoint
Fig. S4 Blood QC violin plots
Fig. S5 Spatial QC metrics
Fig. S6 Module 0 QC (Supp. Fig.)
Fig. 1A KDE of 4 functional scores
Fig. 1B CAR+ vs CAR- exhaustion
Fig. 2A Checkpoint correlation matrices
Fig. 2B Multi-checkpoint kinetics
Fig. 2C Mean checkpoint burden
Fig. 3A Diversity metrics trajectory
Fig. 3B Top 30 clonotype heatmap
Fig. 3C CDR3 convergence network
Fig. 4 Persistence dashboard
Fig. 6 Vessel-to-parenchyma gradient
Fig. 7A Metabolic program activity
Fig. 7B Metabolic-checkpoint correlation
Fig. 8 Human vs. xenograft comparison
Fig. 9 SPSC biopsy profiling
Fig. 10 Blood-spatial TCR mapping
Fig. 11 Cross-modality correlation
Fig. 12 Epitope spreading validation
Fig. 13 Adaptive Fortress score

References

  1. Shah NN, et al. J Clin Oncol. 2021;39:1650–1659.↩︎

  2. Brown CE, et al. N Engl J Med. 2016;375:2561–2569.↩︎

  3. O’Rourke DM, et al. Sci Transl Med. 2017;9.↩︎

  4. Sterner RC, Sterner RM. Blood Cancer J. 2021;11:69.↩︎

  5. Morgan RA, et al. Mol Ther. 2010;18:843–851.↩︎

  6. Zemp FJ, et al. medRxiv. 2025.02.26.24319604 (2025).↩︎

  7. Baba M, et al. Mol Cancer Res. 2019;17:1613–1626.↩︎

  8. Kobos R, et al. J Pathol. 2013;229:743–754.↩︎

  9. Lazaratos AM, et al. Oncogene. 2022;41:4573–4590.↩︎

  10. Salles DC, et al. J Pathol. 2022;257:158–171.↩︎

  11. Chen AP, et al. N Engl J Med. 2023;389:911–921.↩︎

  12. Tivnan P, et al. J Immunother Cancer. 2023;11:e006459.↩︎

  13. Eisenhauer EA, et al. Eur J Cancer. 2009;45:228–247.↩︎

  14. Kawalekar OU, et al. Immunity. 2016;44:380–390.↩︎

  15. Gumber D, Wang LD. EBioMedicine. 2022;77:103941.↩︎

  16. Poorebrahim M, et al. J Immunother Cancer. 2021;9:e002772.↩︎

  17. Blackburn SD, et al. Nat Immunol. 2009;10:29–37.↩︎

  18. Wherry EJ, Kurachi M. Nat Rev Immunol. 2015;15:486–499.↩︎

  19. Fraietta JA, et al. Nat Med. 2018;24:563–571.↩︎

  20. Eyquem J, et al. Nature. 2017;543:113–117.↩︎

  21. Shah NN, et al. J Clin Oncol. 2021;39:1650–1659.↩︎

  22. Bishop MR, et al. N Engl J Med. 2022;386:629–639.↩︎

  23. Lynn RC, et al. Nature. 2019;576:293–300.↩︎

  24. Long AH, et al. Nat Med. 2015;21:581–590.↩︎

  25. Gumber D, Wang LD. EBioMedicine. 2022;77:103941.↩︎

  26. Hanahan D, et al. Nat Rev Cancer. 2024.↩︎

  27. Liberti MV, Locasale JW. Trends Biochem Sci. 2016;41:211–218.↩︎

  28. Ma X, et al. Cancer Lett. 2018;435:92–100.↩︎

  29. Krishna S, et al. Curr Oncol Rep. 2021;23:82.↩︎

  30. Benci JL, et al. Cell. 2016;167:1540–1554.e12.↩︎

  31. Spranger S, et al. Sci Transl Med. 2013;5:200ra116.↩︎

  32. Sahai E, et al. Nat Rev Cancer. 2020;20:174–186.↩︎

  33. Chen Y, et al. Nat Rev Clin Oncol. 2022;19:717–733.↩︎

  34. Zitvogel L, et al. Nat Rev Cancer. 2016;16:759–773.↩︎

  35. Benci JL, et al. Cell. 2016;167:1540–1554.e12.↩︎

  36. Taube JM, et al. Sci Transl Med. 2012;4:127ra37.↩︎

  37. Marcais A, et al. Nat Immunol. 2014;15:749–757.↩︎

  38. Hegde M, et al. Nat Commun. 2020;11:3549.↩︎

  39. Ma L, et al. Cell. 2023;186:1479–1495.e19.↩︎

  40. Blackburn SD, et al. Nat Immunol. 2009;10:29–37.↩︎

  41. Fraietta JA, et al. Nat Med. 2018;24:563–571.↩︎

  42. Hanahan D, et al. Nat Rev Cancer. 2024.↩︎

  43. Hegde M, et al. Nat Commun. 2020;11:3549.↩︎

  44. Ma L, et al. Cell. 2023;186:1479–1495.e19.↩︎

  45. Zitvogel L, et al. Nat Rev Cancer. 2016;16:759–773.↩︎

  46. Johnston RJ, et al. Cancer Cell. 2014;26:923–937.↩︎

  47. Chauvin JM, et al. J Clin Invest. 2015;125:2046–2058.↩︎

  48. Cho BC, et al. Lancet Oncol. 2022;23:781–792.↩︎

  49. Lee YH, et al. Mol Ther. 2023;31:1482–1495.↩︎

  50. Chen J, et al. Blood. 2021;137:2431–2444.↩︎

  51. Chowdhury PS, et al. Nat Immunol. 2022;23:1024–1036.↩︎

  52. Janesick A, et al. bioRxiv. 2024.↩︎