Motixafortide, cemiplimab, gemcitabine and nab-paclitaxel in metastatic pancreatic cancer: a single-arm phase 2 study with single-cell correlatives

Abstract

The C-X-C motif chemokine receptor 4 (CXCR4)/C-X-C motif chemokine ligand 12 (CXCL12) axis drives immune exclusion in pancreatic ductal adenocarcinoma (PDA). Building on our preclinical data, we conducted the first stage of an open-label, single-arm phase 2 clinical trial combining CXCR4 inhibition (motixafortide), PD-1 blockade (cemiplimab), and chemotherapy (gemcitabine/nab-paclitaxel [GN]; MCGN) in 11 treatment-naïve patients with metastatic PDA (NCT04543071). The primary endpoint, objective response rate, was 64% partial response and 55% confirmed partial response per Response Evaluation Criteria in Solid Tumors. Secondary endpoints include median progression-free survival of 9.7 months (95% confidence interval [CI]: 5.9-not reached [NR]), overall survival of 10.1 months (95% CI: 9.3-NR), duration of response of 8.2 months, and 91% disease control rate. One patient achieved a pathological complete response after pancreatoduodenectomy and hepatectomy, remaining disease-free for 18 months without adjuvant therapy. Exploratory single-nucleus RNA-sequencing of serial biopsies showed reduced tumor heterogeneity, depletion of epithelial-to-mesenchymal transitional states, and enrichment of CXCL12+ cancer-associated fibroblasts in responders relative to resistant patients. MCGN also induced tumor inflammation as confirmed by tissue staining and T cell rescue. A multicenter randomized phase 2 trial comparing MCGN to GN in patients with treatment-naïve metastatic PDA is ongoing.

The CXCR4-CXCL12 axis drives immune exclusion and immunosuppression in pancreatic cancer. Here, the authors report a single-arm phase 2 study of metastatic pancreatic cancer evaluating motixafortide (CXCR4 inhibitor), cemiplimab (PD-1 blockade) and chemotherapy, reporting a 64% response rate and 91% disease control, with CXCL12+ fibroblasts and T-cell changes linked to response.

Introduction

Pancreatic ductal adenocarcinoma (PDA) is the third most common cause of cancer-related death in the United States¹. Despite improvements in systemic treatments, long-term survival in patients with metastatic disease remains poor, with 3% alive at 5 years². Immune checkpoint blockade (ICB) therapies have transformed outcomes across many cancers and result in long-term survival in subsets of patients, but to date, they have had a limited impact in patients with PDA, except for those with mismatch repair deficiency (~1% of patients)^3,4. This is likely due to the highly immunosuppressive, cytotoxic CD8⁺ T cell-depleted PDA tumor microenvironment (TME)⁵.

The CXCR4/CXCL12 axis is a well-established contributor to the immunosuppressive and immune-excluded TME in PDA. CXCL12 is secreted by cancer-associated fibroblasts (CAFs)⁶, contributing to the formation of the desmoplastic TME typical of PDA, and increased CXCL12 expression is strongly associated with poor prognosis^7–11. CXCL12 binding to CXCR4 on T cells results in CXCR4 crosslinking and phosphorylation of protein tyrosine kinase 2 and increased T cell adhesion to fibronectin, leading to T cell exclusion from the TME^12,13. Preclinical studies have also demonstrated that transglutaminase 2-mediated crosslinking of CXCL12 and keratin-19 leads to the formation of a filamentous barrier around tumor cells, further suppressing T cell-mediated cytotoxicity¹⁴. Finally, studies using the autochthonous Kras^LSL.G12D/+; Trp53^LSL.R172H/+; tg^Pdx1-Cre (KPC) PDA mouse model revealed that combining a CXCR4 antagonist, AMD3100, with ICB reduces tumor growth and increases infiltration of CD3⁺ T cells in the first week of treatment¹².

Despite these strong preclinical rationale, clinical trials targeting the CXCL12/CXCR4 axis with combined CXCR4 inhibition and immune checkpoint blockade have demonstrated limited benefit. A study evaluating continuous 7-day infusion of AMD3100 in 26 patients and a trial testing the combination of ulocuplumab (monoclonal α-CXCR4) and nivolumab (α-PD1) in 27 patients failed to demonstrate radiographic responses^13,15. In two additional clinical trials combining CXCR4i with ICB, none of the 15 patients treated with motixafortide with atezolizumab (NCT03193190)^16,17, and only one of 29 patients treated with motixafortide with pembrolizumab experienced a radiographic partial response¹⁸. Finally, the COMBAT/KEYNOTE-202 trial tested the combination of 5-FU and liposomal irinotecan (5-FU/nal-I) with motixafortide and pembrolizumab and reported a confirmed overall response rate of 13.2% and a median progression-free survival (PFS) of 3.8 months in patients who had progressed on prior gemcitabine-based chemotherapy¹⁹. These modest outcomes may, in part, reflect other immunosuppressive mechanisms such as persistent infiltration of immunosuppressive regulatory T cells (Tregs), which are known to blunt anti-tumor immunity. Hence, strategies that can reduce Treg-mediated immunosuppression may enhance the therapeutic potential of CXCR4 inhibition with ICB. Gemcitabine, a component of standard chemotherapy for PDA, selectively impairs Treg function and tumor infiltration, and augments anti-PD1 activity^20,21.

With this rationale, we evaluated a treatment regimen composed of AMD3100, mDX400 (anti-PD1), and gemcitabine in the preclinical KPC PDA mouse model. We found that this combination increased intra-tumoral CD8⁺ T cell-to-Treg ratios and prolonged mouse survival, supporting its use as an immunomodulatory chemotherapeutic backbone. Guided by these findings, we designed and completed the first stage of a single-arm phase 2 study using a Simon two-stage design to evaluate motixafortide, a selective CXCR4 antagonist given as subcutaneous injection, with cemiplimab (anti-PD1) and a chemotherapeutic backbone consisting of gemcitabine and nab-paclitaxel (MCGN) in treatment-naïve patients with metastatic PDA. MCGN had a favorable safety profile and resulted in a 64% objective response rate (ORR) with a 91% disease control rate (DCR). One patient achieved an exceptional partial radiographic response, and underwent pancreaticoduodenectomy and resection of a residual hepatic metastasis, revealing no pathological evidence of disease. We performed single-nucleus RNA-seq and multiplexed immunostaining of serial biopsies from all patients, revealing cancer-cell-intrinsic, immune and stromal predictors of response and resistance. Notably, we found that an increase in prevalence of proaxonogenic CXCL12-expressing CAFs associated with treatment response. Here we report the results of the first stage of the single-arm phase 2 cohort, completed under the original Simon two-stage design, after which the protocol was amended to an actively accruing randomized multi-institutional phase 2 trial comparing MCGN with GN in patients with metastatic PDA (NCT04543071).

Results

Preclinical rationale for combining gemcitabine with CXCR4 and PD-1 blockade

In the autochthonous KPC mouse model, we compared the therapeutic efficacy of gemcitabine, AMD3100, and anti-PD1 mAb as monotherapies, double-, or triple-combination therapies, relative to a contemporaneous cohort of control KPC mice treated under identical conditions for another study²² (Supplementary Fig. 1a). As expected, mice treated with gemcitabine or AMD3100 demonstrated decreased or elevated peripheral white blood cell counts, respectively, demonstrating drug-specific pharmacodynamic responses (Supplementary Fig. 1b). Across these seven treatment arms, the triple therapy arm resulted in the best tumor responses, and it significantly extended median overall survival when compared to treatment with gemcitabine monotherapy (mOS = 31 vs 20 days, HR = 0.4; p = 0.015) (Supplementary Fig. 1c, d). Once animals met endpoint criteria (see Methods), mice were euthanized and necropsied; immunohistochemical (IHC) analyses found that tumors from KPC mice treated with the triple combination had elevated rates of apoptosis (activated cleaved caspase3, CC3+ cells) (Supplementary Fig. 1e) relative to gemcitabine-treated controls. The triple regimen was well tolerated based on longitudinal animal weight measurements and there was no change in the proximal cause of death between treatment groups (Supplementary Fig. 1f). To assess the early responses of KPC pancreatic tumors to treatment with these regimens, we performed a short-term intervention study in which pre-treatment biopsies were collected²³ and compared to matched necropsy specimens collected after 6 days of treatment (Supplementary Fig. 1g–j). We observed that only the triple-combination treatment significantly increased intra-tumoral CD8⁺ T cell to Treg ratio (Supplementary Fig. 1g). By contrast, nearest-neighbor tumor-to-CD8 distance in this short-term murine study was variable across treatment groups (Supplementary Fig. 1h) limiting mechanistic interpretation of that metric in isolation.

First stage of an open-label clinical trial testing MCGN in previously untreated patients with metastatic PDA

Based on these promising pre-clinical findings, we designed and completed the first stage of an open-label, multicenter, investigator-initiated, single arm clinical trial evaluating motixafortide, cemiplimab, gemcitabine, and nab-paclitaxel (MCGN) in treatment-naïve metastatic PDA (Fig. 1a). Due to its favorable pharmacokinetics, we chose motixafortide as the favored CXCR4 inhibitor as AMD3100 requires continuous infusion and would not be a feasible option for long-term administration in patients. Twenty-one patients were screened, of whom eleven (one over-enrolled) participated in the study across Columbia University and Brown University (Supplementary Fig. 2). The median participant age at enrollment was 58 years, and the study included nine male (82%) and two female patients (18%) and three (27%) African American patients (Table 1). All participants had microsatellite-stable disease (Supplementary Table 1). All eleven patients received at least one dose of each study drug and were eligible for analysis of both safety and preliminary efficacy. Endpoints (see Methods) included Objective Response Rate (ORR) as the primary endpoint as well as overall survival (OS), progression-free survival (PFS), disease control rate (DCR), and safety as secondary endpoints.

Fig. 1. Clinical outcomes of patients treated with gemcitabine, nab-paclitaxel, cemiplimab (anti-PD1) and motixafortide (CXCR4 inhibitor).

a Biopsy and dose administration schedule of patients treated on the study protocol. b Waterfall plot of tumor best response to the study combination by RECIST criteria and best response classified as confirmed partial response, unconfirmed partial response, stable disease, and progressive disease. c Spider plot of tumor response and duration by RECIST. d Swimmer’s plot displaying survival and major events incurred by patients on the study protocol, including death, treatment discontinuation, progressive disease, definitive radiation therapy, and pancreaticoduodenectomy. e Spider plot of change in CA 19-9 levels in patients enrolled in the study. Kaplan–Meier curves displaying progression-free (f) and overall survival (g) for study patients. Kaplan–Meier curves showing post-hoc analysis of progression-free (h) and overall (i) survival for participants stratified by CXCL12+ pCAF prevalence. Source data are provided in the source data file.

Table 1.

Demographics

Overall (n = 11)	Number (%)
Median age at consent, in years [Min and Max]	58 (46–69)
Baseline ECOG Performance Status
0	5 (45)
1	6 (55)
Sex
Male	9 (82)
Female	2 (18)
Race
Black	3 (27)
White	8 (73)
Ethnicity
Hispanic	2 (18)
Site of Primarya
Head	2 (18)
Neck/Uncinate/Body	5 (46)
Tail	4 (36)
Metastatic Sites at Baselinea
Single	4 (36)
Multiple	7 (64)
Site of Metastasis at Baselinea
Liver	10 (91)
Peritoneum	2 (18)
Lung	1 (9)
Osseous	1 (9)

^aSource data reviewed and confirmed ad hoc by investigators.

Seven out of eleven patients (64%) experienced a partial response (PR) by Response Evaluation Criteria in Solid Tumors (RECIST) criteria, and six of these had a confirmed PR (cPR) (55%) (Fig. 1b). Three patients had stable disease (SD) of which two were confirmed on subsequent imaging (minimum RECIST change -14%, -18% and -22%) and one patient had primary progressive disease (PD), resulting in a disease control rate (DCR) of 91% (Fig. 1b, c). One patient experienced complete resolution of all RECIST target lesions. Of the ten patients who had hepatic metastasis, two experienced radiologic resolution of all liver lesions. The median time to response was 3.5 months and the median duration of response was 8.2 months. An independent, blinded RECIST review of images demonstrated that 5 of 11 patients experienced a PR (45%) with a DCR of 91%.

As of 08/31/25 and a median follow-up of 36.9 months, 2 patients were alive and in follow-up. Causes for end-of-treatment for the 11 patients included radiologic progression (N = 8), clinical progression (N = 1), investigator discretion (N = 1), and curative-intent surgical resection (N = 1) (Fig. 1d, Supplementary Fig. 2). All patients with elevated CA 19-9 levels at baseline experienced a reduction in CA 19-9 levels, with 2/11 (18%) experiencing normalization (≤35 U/mL) (Fig. 1e). The median PFS and OS were 9.7 months (95% confidence interval [CI]: 5.88-not reached [NR]) and 10.1 months (95% CI: 9.34-NR), respectively (Fig. 1f, g). In an exploratory post-hoc analysis stratified by prevalence of CXCL12⁺ pCAFs, a fibroblast subset we identified via single-nucleus profiling of tumor samples (see Single nucleus profiling of paired biopsies below), we assessed PFS and OS separately in high- versus low-pCAF subgroups. Compared to participants with low CXCL12⁺ pCAF prevalence, individuals with high CXCL12⁺ pCAF prevalence had numerically longer median PFS (4.8 versus 12.0 months; HR 0.44, 95% CI:0.11–1.68) and median OS (8.6 versus 18.6 months; HR 0.39, 95% CI:0.10–1.54); however, these differences were not statistically significant (Fig. 1h, i). The most common adverse events of MCGN, regardless of grade, included skin hyperpigmentation (11/11 patients), injection site reaction (9/11 patients) attributed to motixafortide, and alopecia (10/11) attributed to nab-paclitaxel. There were no grade 4 or 5 treatment-related adverse events (TRAEs) and the most common grade 3 TRAEs were anemia (5/11) and rash (3/11) (Table 2, Supplementary Fig. 3). Post-progression treatments are described in Supplementary Table 2.

Table 2.

Most common adverse events from MCGN

	Any adverse event (%)	Grade 3 adverse event (%)
Skin hyperpigmentation	11 (100)	0
Alopecia	10 (91)	0
Injection site reaction	9 (82)	0
Edema	8 (73)	1 (9)
Fatigue	8 (73)	0
Flushing	8 (73)	0
Peripheral sensory neuropathy	8 (73)	0
Pruritus	8 (73)	0
Nausea	7 (64)	0
Anemia	6 (55)	5 (45)
Diarrhea	5 (45)	0
Fever	5 (45)	1 (9)
Paresthesia	5 (45)	0
Rash	5 (45)	3 (27)
Dysgeusia	4 (36)	0
Vomiting	4 (36)	0
Anorexia	3 (27)	0
Chills	3 (27)	0
Dyspnea	3 (27)	0
Hypokalemia	3 (27)	0
Hypotension	3 (27)	1 (9)
Nail discoloration	3 (27)	0
Platelet count decreased	3 (27)	2 (18)
Urticaria	3 (27)	2 (18)
Weight loss	3 (27)	0
Bone pain	2 (18)	1 (9)
Cough	2 (18)	0
Dry skin	2 (18)	0
Facial edema	2 (18)	0
Generalized muscle weakness	2 (18)	0
Hot flashes	2 (18)	0
Hyperhidrosis	2 (18)	0
Hypertension	2 (18)	2 (18)
Hypothyroidism	2 (18)	0
Infusion related reaction	2 (18)	0

Case study of an exceptional responder

Patient 001-12 is a 59-year-old male who underwent an aborted pancreatoduodenectomy due to a metastatic tumor deposit identified on the liver surface (confirmed by intra-operative pathological evaluation) during routine laparoscopy prior to planned resection (Fig. 2a–e). Prior to initiating treatment, the primary tumor measured 3.2 cm in largest dimension, and by cycle 12 (day 332 after initiating therapy), it had decreased by 53% to 1.5 cm. By day 443 of treatment, contrast-enhanced computed tomography (CT) revealed stability of the 1.5 cm tumor mass and no radiographic evidence for other metastatic sites (compare Fig. 2f to Fig. 2b). On day 457, he underwent a diagnostic laparoscopy which identified a single atypical lesion on the liver surface which was biopsied and demonstrated no evidence of carcinoma (Fig. 2g, h). He proceeded with pancreatoduodenectomy, which on pathological review identified therapy-related changes and low- to intermediate-grade pancreatic intraductal neoplasia within the resected pancreas tissue, and no residual carcinoma (Fig. 2i, j). None of the 20 resected lymph nodes harbored carcinoma cells. Given the pathologic complete response in the primary tumor and resolution of a previously identified liver surface metastasis, he declined adjuvant chemotherapy. As of 08/05/2025, he has remained alive and disease-free for 1,023- and 567-days following initiation of therapy on study and curative-intent surgery, respectively; demonstrating that MCGN may induce potent, prolonged disease control in a subset of patients.

Fig. 2. Case of exceptional responder.

a Timeline of diagnosis, treatment, and curative-intent surgery for an exceptional responder to study therapy, including tumor marker changes over time. b Computed tomography (CT) scan at diagnosis. c, Pancreas fine needle aspiration (FNA) stained with hematoxylin and eosin (H&E). Surgical liver biopsy at time of aborted pancreaticoduodenectomy demonstrating metastatic pancreatic adenocarcinoma stained, with H&E (d) and cytokeratin (CK) (e). f CT scan demonstrating best radiologic response, with tumor decrease from 3.2 to 1.5 cm in greatest diameter. Surgical specimens from wedge resection of site of prior liver metastasis (H&E (g) and CK (h)) and pancreaticoduodenectomy (H&E (i) and CK (j)) demonstrating no residual tumor consistent with a complete response to study therapy. Image magnification of pathologic images: c 0.55 mm (40x), d, e, g, h, j 2.2 mm (10x), i 11 mm (2x). Scale bars of pathologic images: c 50 μm, d, e, g, h, j 200 μm, i 1 mm. Representative images for b–f were taken from a single responder from a population of n = 11 patients. Representative images for g–j were taken from a single responder.

MCGN-induced changes of the tumor microenvironment

To understand cellular and spatial changes induced by MCGN on the TME, pre- and on-treatment (on cycle 1, day 18), biopsy specimens were obtained for all 11 enrolled participants and analyzed using quantitative multiplex immunofluorescence (qmIF) (Fig. 3a). In addition, optional on-progression biopsies were also obtained from three patients. Compared to pre-treatment specimens, we observed an increased abundance of CD8⁺ T cells in all on-treatment specimens and a subsequent reduction of CD8⁺ T cells in the three on-progression specimens available (p = 0.007; paired T-test) (Fig. 3b). There were numerically higher although not statistically significant CD8⁺ T cell to FOXP3⁺ Treg ratios in tissue obtained from patients who experienced a PR or SD compared with tissue from patients who experienced disease progression (Fig. 3c). We next interrogated whether MCGN treatment resulted in increased proximity of T cells to malignant cells by analyzing the spatial distribution of the immune cells within 40 microns of the tumor cells (Fig. 3d). We compared on-treatment or on-progression biopsies to baseline tissue from patients who experienced a PR (n = 7), SD (n = 3) and PD (n = 4, including primary progression and post-progression specimens). For the one patient with primary resistance, we compared on-treatment to pre-treatment specimens, and for the three patients who experienced secondary resistance, we compared post-progression samples to their respective pre-treatment specimens. There was a significant increase in peritumoral CD8⁺ T cells in patients with a PR (p = 0.0012), a numerically higher although not statistically significant increase in the SD group (p = 0.1), and no change in the PD group (Fig. 3e). Peritumoral CD20 and CD56 population levels were not significantly changed with treatment (Supplementary Fig. 4a, b). Collectively, these data indicate that MCGN treatment led to increased CD8⁺ T cell infiltration and closer proximity to malignant cells in patients who achieved a PR.

Fig. 3. Quantitative multiplex immunofluorescence performed on pre-treatment, on-treatment, and on-progression biopsy samples from study patients.

a Representative (from population of n = 11 patients, k = 10 image fields per patient) quantitative multiplex immunofluorescence panel staining on pre- and on-treatment biopsy samples from a patient. Scale bars represent 50 μm. b Change in cytotoxic T-cell (CD8 + ), T-regulatory cell (FOXP3 + ), and B-cell (CD20 + ) densities in each patient biopsy, from on-treatment to pre-treatment and on-progression (where applicable, N = 3). Pre- to on-treatment cell proportion was significant for CD8 + T-cells (p = 0.007), but not for T-regulatory cells (p = 0.95) or B-cells (p = 0.09). c Ratio of total CD8+ to FoxP3+ cells in baseline and on-treatment/progression biopsies. d Diagram of methodology for identifying cell proportions within 40 microns of each tumor cell. Created in BioRender. Li, L. (2026) https://BioRender.com/2heehfp. e Change in tumor infiltration with CD8 + T-cells by patient radiographic response by RECIST. Responses are categorized as partial response (p = 0.0012), stable disease (p = 0.1) and progressive disease (p = 0.79). The progressive disease category displays change from baseline to on-progression in one patient with primary progression and three patients with secondary progression and on-progression biopsies. All comparisons were performed by paired Wilcoxon signed-rank tests (two-sided) without multiple hypothesis correction for response subgroups, which were independent, pre-specified, and mutually exclusive. Box plots indicate 25th percentile, 50th percentile, and 75th percentile, with whiskers extending to furthest point within 1.5 times the interquartile range and flier markers for points outside this range. Source data are provided in the source data file.

Single-nucleus profiling of paired biopsies

To gain comprehensive insight into MCGN-induced changes in the tumor ecosystem of these participants, we performed single-nucleus RNA-sequencing of eight paired pre-treatment and on-treatment biopsies and three matched on-treatment and post-progression samples, predominantly from liver metastases (Fig. 4a). After quality control^24–27 (see Methods), we recovered a total of 91,427 high-quality transcriptomes (Fig. 4b, c and Supplementary Fig. 5a) that partitioned into eight major cell type clusters, including malignant cells (k = 20,124), T cells (k = 7384), B cells (k = 3954), myeloid cells (k = 11,435), hepatocytes (k = 22,792), fibroblasts (k = 5513), epithelial cells (k = 16,922), and endothelial cells (k = 3053) (Fig. 4b, c and Supplementary Fig. 5b). Hepatocytes, expected in biopsies from hepatic metastases, formed a transcriptionally distinct compartment with minimal overlap with malignant-cell programs, and downstream analyses were performed within annotated cellular compartments. We validated our annotations by differential expression analysis, finding marker genes specific for each cluster. We observed IGHG1 and FCRL1 in B cells; THEMIS, PTPRC, and CD96 in T cells; CD163 and ZEB2 in myeloid cells; COL3A1 and COL5A1 in fibroblasts; APOB and ALB in hepatocytes; PTPRB and FLT1 (encoding VEGFR1) in endothelial cells; PKHD1 and DCDC2 in epithelial cells; and ADAM9, FLNB, and MIR31HG in malignant tumor cells (Supplementary Fig. 5c).

Fig. 4. Single-cell transcriptomic profiling of patient biopsies.

a Experimental design for transcriptomic analysis. Created in BioRender. Izar, B. (2026) https://BioRender.com/u50zkml. Biopsies were taken from liver (n = 18), lung (n = 1), and peripancreatic lymph node (n = 1) metastatic sites as well as from primary pancreatic sites (n = 2); these biopsies represent pre-treatment, on-treatment, and on-progression time points. b, c UMAP and PCA representations of 91,427 cells integrated across all samples. Individual cell compartments include malignant cells (k = 20,374), T cells (k = 7384), B cells (k = 3954), myeloid cells (k = 11,435), hepatocytes (k = 22,792), fibroblasts (k = 5513), epithelial cells (k = 16,922), and endothelial cells (k = 3053). d CIN70 signature expression strength is highest in malignant metaprograms MP2 and MP6, the programs with the highest representation of biopsies from non-responders. e, f Differential abundance of various cell types between non-responders (NR) and partial-responders (PR) at both the pre-treatment and on-treatment time points. Separate analyses were performed in the immune and fibroblast compartments, respectively, and are depicted in full in Supplementary Fig. 5. g Expression of inhibitory checkpoint-related genes within the immune compartment.

Tumor heterogeneity and cell states associated with treatment response

Our snRNAseq data enabled interrogation of tumor heterogeneity, which is typically challenging given the low tumor purity of PDA, and its relationship to radiographic treatment responses. Using previously established methods (see Methods)²⁸, we found that transcriptional heterogeneity within tumor cells was comparable in pre-treatment specimens across patients. Individuals who experienced a PR had significantly lower tumor transcriptional heterogeneity after treatment than those who experienced SD or PD. Patients who had an initial PR but subsequently progressed maintained a reduced level of heterogeneity compared to non-PR patients, suggesting that MCGN treatment resulted in deletion of some clones, but that proliferation of other clones occurred in those without a partial response (Supplementary Fig. 5d, e). To identify molecular functions associated with treatment outcomes, we performed non-negative matrix factorization (NMF) of malignant cells and identified six meta-programs (MPs) (Supplementary Fig. 6a, b).

Among these, MP2 and MP6 represented conserved epithelial-to-mesenchymal transition (EMT) and chromatin regulation programs, with outsized contribution from patients with SD/PD and underrepresented contribution from patients with a PR (odds ratio = 4.75, p = 0.0038 by Fisher’s exact test for MP2 + MP6) (Supplementary Fig. 6a, c). Gene Set Enrichment Analysis (GSEA) against established PDAC subtype signatures showed that MP2 aligned most strongly with a classical program, whereas MP6 aligned more closely with a squamous/QM/basal-like program, indicating that these resistant metaprograms do not map cleanly onto a single canonical subtype. Particularly, comparing the overlap between each MP and published EMT gene sets from Gavish et al. and MSigDb, we observe a conserved set of FN1, TIMP1, VIM, LAMC2, FLNA, PMEPA1, ITGB1, TGM2, and DST occurring in MP2 but not in other MPs (Supplementary Fig. 6c)^29–31.

Consistently, MP2/MP6 associated with poor survival in the PDA cohort from The Cancer Genome Atlas (TCGA) analyses (Supplementary Fig. 7a, b). Irrespective of clinical response, most patients experienced reduced fraction of cells expressing MP2/MP6 signatures after treatment (p = 0.032, likelihood ratio test of mixed-effects models) (Supplementary Fig. 7c, d), suggesting that MCGN selectively affected these tumor cell populations. MP2 and MP6, but not other MPs, strongly correlated with an established signature (CIN70) of chromosomal instability (CIN), a genomic hallmark of cancer that strongly associates with immune evasion, tumor heterogeneity and acquisition of an EMT state (Fig. 4d). To further investigate transcriptional changes due to treatment, we examined the expression signature of each metaprogram at different treatment stages. We observe that among partial responders, every metaprogram except for MP1 decreased after treatment, whereas in patients with SD/PD, MP2 decreased after treatment while all other MPs increased in expression (Supplementary Fig. 7e). Together, our results suggest that MCGN leads to elimination or reduction of cells states associated with aggressive disease behavior, and that disease progression is associated with a proliferation of cell states that are less aggressive but more heterogenous.

Biomarkers of response and resistance in the stromal and immune compartments

We identified eight subsets of cancer associated fibroblasts (CAFs) which included the presence of a proaxonogenic CXCL12-expressing CAF (CXCL12+ pCAF) subset that also expressed neurogenic genes such as RELN, NEURL1B, and SOX6 (Supplementary Fig. 8a, b). This CXCL12+ pCAF population showed partial overlap with, but was not synonymous with, the canonical iCAF state, and is therefore best considered a distinct fibroblast program enriched for CXCL12 expression in our dataset. We observed that the presence of CXCL12+ pCAFs at both the pre-treatment and on-treatment stages, but not other CAF subsets, was strongly enriched in patients who experienced a PR, suggesting that the presence of these pCAFs may predict successful response to MCGN therapy (Supplementary Fig. 8c, d). Querying TCGA patient data with an expression signature derived from this CAF subset, we found that higher expression of this signature did not predict improved survival (Supplementary Fig. 8e), suggesting that the presence of CXCL12+ pCAFs was non-prognostic outside of the context of MCGN therapy (Fig. 1h, i). Using ContactTracing³² to identify ligand-receptor interactions, we identified the activity of CXCL12-CXCR4 interactions in response to MCGN as potentially contributing to this factor. As part of this analysis, we also examined how chemokine (C-X-C motif) ligand 14 (CXCL14), a cytokine that binds to CXCR4 and influences immune cell trafficking, is altered in response to treatment. CXCL12 and CXCL14 were identified as top ligands for CXCR4 in T cells, and treatment with MCGN disrupted these interactions in patients with a PR but not in those with PD. Importantly, there was restoration of the CXCL12-CXCR4 interaction and potential increased CXCL14-CXCR4 interaction in patients with an initial PR who subsequently experienced PD (Supplementary Fig. 8f). Together, these data suggest that MCGN disrupts a signaling axis between CXCL12-expressing pCAFs and CXCR4-expressing T cells, that presence of CXCL12-expressing pCAFs may predict response to MCGN, and that restoration of this axis with CXCL14 may confer resistance to this treatment regimen.

A pre-existing inflamed TME (in the pre-treatment biopsy), characterized by tumor infiltration with memory B cells, monocytes, memory CD4⁺ T cells, and, also terminally differentiated CD8⁺ T cells (Tex) was associated with PR (Fig. 4e and Supplementary Fig. 9a), while increased myeloid infiltration and Tex associated with lack of response in on-treatment biopsies (Fig. 4f and Supplementary Fig. 9b). We reasoned that the variable association of Tex and treatment response may be related to the depth of dysfunction within the immune compartment, as indicated by the strength of expression in checkpoint and exhaustion-related genes. To explore this, we performed unbiased differential expression analysis to identify genes in the T cell compartment associated with “effective treatment stage”, in which on-treatment biopsies from patients with PD are effectively treated as on-progression biopsies. We observed that in the on-treatment samples without progression, the most differentially expressed genes were in robust T cell response pathways such as IL-2/STAT5 signaling (PLEC, ITGAE, BCL2, ABCB1), IFN-γ signaling (SAMD9L, RNF213, STAT1, XAF1), and TGF-β response (SMURF2, SKI, NCOR2) (Supplementary Fig. 9c). In contrast, the on-treatment samples with progression showed markers of reversion to a more stem-like state (BACH2, TCF7, LEF1, SELL, CCR7), suggesting that T cell dysfunction and exhaustion may contribute to tumor progression after initial treatment (Supplementary Fig. 9c). To this end, we interrogated the expression of several canonical co-stimulatory and -inhibitory immune checkpoints. Co-expression of multiple inhibitory checkpoint marker genes in both myeloid and T cell compartments in on-treatment samples associated with disease progression (Supplementary Fig. 9d). MCGN treatment reduced broad expression of TIM-3 (encoded by HAVCR2) in the myeloid compartment, but after progression, there was stronger expression of PD-1 (encoded by PDCD1), TIGIT, and LAG-3 in T cells, as well as increased expression of BTLA in B cells and more widespread expression of VISTA (encoded by VSIR) across myeloid and T cell subtypes (Fig. 4g). Together these analyses suggest that MCGN alters immune composition and immune cell function which associate with treatment response, and that tumor progression occurs downstream of dysfunction and exhaustion in the T cell compartment.

Discussion

Outcomes for patients with metastatic PDAC remain poor, with standard of care chemotherapy resulting in overall response rates ranging between 23 and 31% and progression-free survival (PFS) duration between 5.5 and 7.4 months^33–35. Despite their promise, immune checkpoint inhibitors have yielded negligible benefits when combined with chemotherapy³⁶. The CXCL12/CXCR4 axis represents a major barrier to effective immunotherapy as CXCL12-producing CAFs confer exclusion of T cells in the desmoplastic PDAC stroma.

Unlike 5-FU, gemcitabine has been reported to reduce circulating peripheral Tregs in PDA patients and to inhibit ex vivo IL-2–mediated Treg induction^20,37–39. Consistent with this, preclinical studies have suggested that gemcitabine can modulate intratumoral Treg biology, suppress tumor growth in immunocompetent settings, and augment anti-PD1 efficacy, supporting an immunomodulatory role beyond direct cytotoxicity. A high-throughput screen of 1554 compounds identified gemcitabine as the most effective agent at inverting intratumoral Treg activity, attributed to its selective cytotoxicity toward intratumoral over peripheral Tregs and inhibition of TRPS1, a key regulator of Treg function²¹. Furthermore, gemcitabine treatment inhibited tumor growth in immunocompetent but not immunocompromised allografts²¹. In preclinical syngeneic subcutaneous MC38 colon cancer models, gemcitabine enhanced anti-PD1 efficacy by selectively depleting tumor-infiltrating Tregs²¹. In our study, the murine triple-combination regimen increased the intratumoral CD8 + T cell:Treg ratio, supporting an immunomodulatory role for gemcitabine in this context. Compared with prior reports, we did not observe increased CD8 + T-cell infiltration with the CXCR4i and PD-1 doublet, which may reflect later sampling in our study (6 days versus 24 h). However, our data do not establish Treg depletion as the sole mechanism underlying efficacy. In the clinical qmIF analyses, FOXP3+ cells were relatively sparse and we observed only a trend toward increased CD8 + T cell:FOXP3+ cell ratios, limiting definitive conclusions regarding Treg dependence. Instead, our findings support a broader model in which gemcitabine may facilitate CXCR4i/PD-1 blockade through multiple, non-mutually exclusive effects on the tumor microenvironment, including modulation of lymphoid and myeloid compartments.

Importantly, our correlative analyses also implicate myeloid biology as a potential mediator of resistance. While a pre-existing inflamed tumor microenvironment was associated with partial response, on-treatment biopsies from non-responders were enriched for myeloid cells, and progression was accompanied by co-expression of inhibitory checkpoint programs across both myeloid and T-cell compartments. These findings suggest that persistent or treatment-emergent myeloid states may counteract effective antitumor immunity during MCGN therapy. Notably, prior studies of CXCR4-directed therapy have similarly reported increased myeloid signatures and macrophage-associated suppressive programs, supporting myeloid remodeling as a broader feature of response and resistance in this setting^13,40. Thus, although gemcitabine was selected in part based on prior data supporting Treg-modulatory effects, our current data indicate that alternative mechanisms, including myeloid recruitment, remodeling or persistence, are also plausible contributors to response heterogeneity and resistance. These questions will be addressed prospectively in biospecimens from the ongoing randomized phase 2 trial, including analyses of MDSC populations and their modulation by gemcitabine-based chemotherapy.

Building on these prior insights and the encouraging pre-clinical data presented here, we conducted an investigator-initiated single-arm phase 2 clinical trial combining motixafortide, cemiplimab, gemcitabine, and nab-paclitaxel (MCGN). MCGN resulted in encouraging objective response and disease control rates of 64 and 91%, respectively³⁴. While cross-trial comparisons should be interpreted cautiously, the response rate and disease control rates with gemcitabine/nab-paclitaxel alone were 23% and 48%, respectively³⁴. Furthermore, responses in some patients were prolonged, with a median PFS of 9.7 months, compared with 5.5 months with gemcitabine/nab-paclitaxel³⁴. Despite prolonged follow-up, the narrow interval between median PFS and OS suggests that post-progression survival was short for many patients in this cohort. Responses varied in durability; five patients progressed within five months, while four remained progression-free at 12 months or longer, including one achieving pCR within the tissue obtained via pancreatoduodenectomy and metastatectomy and remains disease free for over 18 months since curative resection without systemic therapy. For context, a meta-analysis of neoadjuvant trials in pancreatic cancer reported a complete response rate of 4%, and 2.5% when rigorous tissue sampling was performed, as in our study⁴¹. These results suggest that MCGN can induce potent and prolonged disease control in a subset of patients.

Prior clinical studies of CXCR4 inhibition in PDA, particularly COMBAT/KEYNOTE-202, showed that motixafortide plus pembrolizumab combined with 5-FU/nanoliposomal irinotecan had only modest clinical activity in patients previously treated with gemcitabine-based therapy, with a median PFS of 3.8 months versus 3.1 months in NAPOLI-1 with 5-FU plus nanoliposomal irinotecan, although such cross-trial comparisons should be interpreted cautiously^19,42. These findings suggest that the contribution of chemotherapy in CXCR4i/PD-1-based regimens may not be strictly gemcitabine-specific, but may instead reflect broader immunologic and cytoreductive functions that facilitate CXCR4 and PD-1 blockade. However, important differences in treatment setting, prior therapy exposure, chemotherapy backbone and likely tumor microenvironment preclude direct comparison with our first-line study of MCGN. Accordingly, our findings support gemcitabine/nab-paclitaxel as an effective and mechanistically rational backbone for this strategy in treatment-naive metastatic PDA, rather than indicating that gemcitabine is uniquely required for benefit.

To glean ecosystem-wide insights into tumors from enrolled participants, we performed qmIF and snRNA-seq on sequential biopsies to interrogate the stromal, malignant, and immune compartments. Enrichment of CXCL12-expressing pCAFs strongly correlated with response to MCGN^43–46. These CXCL12⁺ pCAFs exhibited a distinct proaxonogenic transcriptional program, characterized by expression of RELN, NEURL1B, and SOX6, consistent with neurotropic fibroblast phenotypes previously described⁴⁷. Their enrichment in responders suggests that these tumors may be more dependent on CXCL12-CXCR4 signaling and therefore more vulnerable to disruption by MCGN. Consistent with this, the expression signature of CXCL12⁺ pCAFs was not associated with prognosis in the TCGA dataset and our post-hoc exploratory analysis suggests that their prevalence may predict therapeutic benefit in the context of MCGN therapy, warranting further investigation to validate their utility as a predictive marker for CXCR4i-based therapies. Moreover, our spatial and ligand-receptor analyses supported the notion that the CXCL12-CXCR4 axis, supported by CXCL14 expression from malignant cells, was disrupted during treatment at the time of response but reestablished at progression, potentially by CXCL14, reinforcing the axis’s central role in modulating immune exclusion and therapy response.

MCGN treatment induced increased intratumoral CD8⁺ T cell infiltration and enhanced their spatial proximity to malignant cells in patients achieving PR. This effect diminished or reversed at the time of progression. snRNA immune profiling highlighted that responders’ pre-treatment biopsies were enriched for memory B cells, monocytes, memory CD4 + T cells, and Tex cells, indicative of a partially primed immune microenvironment. However, Tex cells in non-responders exhibited stronger exhaustion gene signatures, implying deeper dysfunction downstream of the reversion to a stem-like T cell phenotype. Progression was associated with increased exhausted T cells expressing immune checkpoint regulators such as TIGIT, LAG-3, and VISTA, highlighting potential therapeutic immune targets in MCGN-resistant tumors^48–50.

Non-responder and post-progression samples were enriched for metaprograms within the malignant compartment that associated with EMT, chromatin regulation, chromosomal instability (CIN), and strongly correlated with the CIN70 gene expression signature, an established surrogate for chromosomal instability known to drive immune evasion and resistance. Notably, these resistant metaprograms were not explained solely by classical versus basal-like lineage assignment, suggesting that treatment resistance reflects adaptive malignant cell states beyond canonical PDAC subtypes. Indeed, prior reports showed that a higher rate of CIN is associated with poor prognosis in PDA, and because CIN is a key driver of tumor heterogeneity, this may have contributed to a higher degree of heterogeneity in patients without objective responses, as observed here⁵¹. MCGN therapy preferentially depleted tumor cells expressing MP2/MP6 programs, leaving behind tumor cells with greater transcriptional diversity but potentially less aggressive phenotypes. Indeed, non-responders saw increased expression of all metaprograms except for the EMT-related MP2 after treatment and progression. Consistent with this, high MP2/MP6 expression independently predicted poor overall survival in TCGA PDA datasets. This suggests that treatment-resistant tumors may arise from alternative malignant cell states not characterized by EMT/CIN programs.

Our findings suggest that increased resistance to treatment may be driven by: (1) absence or resistance (potentially by CXCL14) of a targetable CXCL12-CXCR4 axis with or without, (2) enrichment of malignant cells expressing transcriptionally aggressive and CIN-high programs (MP2/6), and (3) progressive immune dysfunction marked by checkpoint upregulation and T cell exhaustion.

We have initiated a multicenter, investigator-initiated phase 2 trial comparing MCGN to standard of care gemcitabine and nab-paclitaxel in treatment-naïve metastatic pancreatic cancer (2:1 randomization; n = 108). This trial incorporates pre-treatment and optional on-treatment biopsies to validate CXCL12⁺ pCAF abundance as a predictive biomarker and assess immune checkpoint expression dynamics at progression, guiding future therapeutic strategies.

The main limitations of this study are its single-arm design, small sample size, and enrollment to only two clinical sites. While the abundance of CXCL12+ pCAFs is associated with treatment response to MCGN, its value as a biomarker requires further evaluation.

In summary, the combination of MCGN demonstrates promising clinical activity in treatment-naïve metastatic PDAC, with response rates and PFS exceeding historical benchmarks for chemotherapy alone. By simultaneously blocking the CXCL12/CXCR4 axis, enhancing intratumoral T-cell infiltration, and reshaping immune and stromal states, MCGN appears to reprogram the tumor microenvironment and selectively eliminates aggressive tumor cell states. The enrichment of CXCL12⁺ CAFs in responders suggests a potential predictive biomarker and underscores the importance of stromal context. While preliminary, these findings warrant further exploration in randomized studies and highlight the value of integrating targeted stromal disruption into immuno-chemotherapy for PDAC.

Methods

All research was conducted in compliance with all relevant ethical regulations. Murine studies were approved by the Columbia University Institutional Animal Care and Use Committee and conducted in accordance with the NIH Guide for the Care and Use of Laboratory Animals. The clinical trial was approved by the Institutional Review Boards at Columbia University Irving Medical Center and Brown University Health, and was conducted in accordance with the Declaration of Helsinki and Good Clinical Practice guidelines

Mouse PDA model

All murine studies were approved by the Columbia University institutional animal care and use committee (IACUC), protocol number AC-AABS3610, and conducted in accordance with the NIH “Guide for the Care and Use of Laboratory Animals.” Kras^LSL.G12D/+; P53^LSL.R172H/+; Pdx1-Cre (KPC) mice were generated as has been previously described⁴³. KPC mice (ranging from 12 to 50 weeks of age) were monitored for tumor formation by weekly palpation. Upon positive tumor palpation, tumors were monitored with twice-weekly ultrasound until an average cross-sectional diameter of 4-7 mm was reached. Upon reaching this size, mice with tumors in the head of the pancreas were enrolled randomly into one of the seven treatment arms (n = 10 [5 females and 5 males] per arm) in the survival study which included combinations of gemcitabine, AMD3100 (CXCR4 inhibitor [CXCR4i]) and immune checkpoint blockade (ICB). AMD3100 was administered by continuous subcutaneous infusion using a 14-day Azlet osmotic pump. Pumps were serially replaced every 14 days throughout the survival study. Those mice with tumors in the tail of the pancreas were randomized into one of the seven treatment arms in the biopsy study. Tumor growth was monitored twice weekly after enrollment and volume quantification was carried out as previously described⁵².

A clinical performance scoring system incorporating physical and behavioral symptoms was used to monitor animal health and determine humane endpoints. Mice with a cumulative score ≥5 were euthanized. Scoring criteria included abdominal distension due to hemorrhagic ascites (5), mild difficulty breathing (5), cool to the touch (5), abdominal distension due to malignant ascites (3), >20% body weight loss since enrollment (3), failure of grasp test (3), jaundice or pallor (3), weak hold in grasp test (2), failure to interact with cage mates (1), hunched posture (1), and piloerection, failure to groom, or porphyrin staining (1). Overall survival (OS) was determined by the Kaplan–Meier Method. When detectable tumors have become large in diameter (exceeding 10% total body weight or 2 cm in diameter), interfere with postural adjustments, or if they become necrotic or ulcerated, animals were euthanized. This maximal tumor size was never exceeded. Mouse colonies were bred and maintained with standard mouse chow and water, ad libitum, under a 12-h light cycle, at 70-72 *F, with 35% humidity.

First stage of phase 2 clinical trial

This was the first stage of an open-label, multicenter, investigator-initiated, single-arm phase 2 study evaluating the combination of motixafortide, cemiplimab, gemcitabine and nab-paclitaxel (MCGN) in metastatic PDA (mPDA) using a Simon optimal 2-stage design (NCT04543071), with enrollment beginning on 9/30/2020 and ending 3/28/2023. The first stage, reported here, had a planned enrollment of 10 patients. The present report describes the first stage of a single-arm cohort and its prospectively collected translational correlatives, including mandatory pre-treatment and on-treatment tumor biopsies for qmIF and snRNA-seq analyses. The primary endpoint was response rate (RR). If 3 or more patients met the primary endpoint in the first stage, the study was planned to expand to a total of 40 patients. Secondary endpoints included safety, disease control rate (DCR), progression-free survival (PFS) and OS as determined by the Kaplan–Meier method. Median follow-up time was estimated by the reverse Kaplan–Meier method. In light of encouraging preliminary efficacy observed in the first stage of the single-arm phase 2 portion, the original Simon two-stage design was discontinued and the protocol was amended to a randomized phase 2 trial with a planned enrollment of 108 patients comparing gemcitabine/nab-paclitaxel with MCGN. Exploratory analyses, including single-nucleus profiling, were performed on participant biopsies as described below.

Sex was documented as a biological variable using clinical records available at the time of enrollment. Baseline demographic characteristics, including age, are summarized in Table 1. All participants provided written informed consent before enrollment. Participants did not receive financial compensation for study participation. Although sex was included as a baseline demographic variable, prespecified sex-based subgroup analyses were not performed because the study was not designed or statistically powered to evaluate outcomes by sex.

All participants included in the study were assessed for response to treatment and assigned to a response category of complete response (CR), partial response (PR), stable disease (SD) or progressive disease (PD) based on Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1. A confirmed partial response was defined as a partial response maintained on two consecutive imaging studies at least 4 weeks apart. RR was defined as the percentage of patients achieving CR or PR. DCR was defined as the percentage of patients achieving CR, PR or SD by 16 weeks.

Criteria for inclusion included pathologic confirmation of mPDA, ECOG performance status of 0 or 1, age ≥18 years, adequate hematological and end-organ function and negative screening for lower extremity deep venous thrombi. Exclusion criteria included prior systemic, surgical or radiation therapy for PDA, history of leptomeningeal or central nervous system metastases and uncontrolled ascites, pleural effusion, pericardial effusion or pain. Patients with a history of autoimmune disease other than controlled hypothyroidism, type 1 diabetes and autoimmune dermatologic conditions were excluded.

Enrolled patients began treatment with a 5-day priming period of daily motixafortide (1.25 mg/kg subcutaneously [SC]). After priming, patients received treatment with MCGN, consisting of gemcitabine (1000 mg/m² intravenous [IV]) and nab-paclitaxel (125 mg/m² IV) on days 1, 8, and 15 every 28 days, motixafortide twice weekly and cemiplimab (350 mg IV) every 21 days. Clinical response was evaluated by RECIST criteria with cross-sectional imaging every 8 weeks.

The counts and percentages of treatment-related toxicities are reported by type and severity according to the National Cancer Institute Common Terminology Criteria for Adverse Events, version 5 (NCI, 2017).

Correlative studies

All 11 patients underwent pre-treatment biopsies and on-treatment biopsies, which were scheduled on C1D18 of treatment, and 3 patients underwent on-progression biopsies performed 5 days prior to confirm progression (001-08); 7 days after clinical progression (001-04); and 23 days after radiological progression (001-03). Biopsy specimens included both fresh frozen and formalin-fixed paraffin-embedded (FFPE) tissue which were stored for translational analyses.

Quantitative multiplex immunofluorescence: procedures for staining and scanning

Quantitative multiplex immunofluorescence (qmIF) enables staining of multiple markers within the tumor environment, facilitating characterization of spatial relationships between tumor, immune and stromal cells⁵³. Prior to staining, H&E slides were viewed by a gastrointestinal pathologist to determine the presence of tumor. For murine studies, we optimized a qmIF antibody panel consisting of DAPI, CD3 (clone SP7; Spring Bioscience; 1:100 dilution), CD8 (clone 4SM15; eBioscience; 1:2000), CK19 (clone EPNCIR127B; Abcam; 1:450), FoxP3 (clone FJK-16s; eBioscience; 1:100 dilution), and PD-L1 (clone 22C3; Dako; 1:150). For human studies, the following optimized qmIF antibody panel was used: DAPI (nuclear marker), CD3 (T cell marker; clone LN10; Leica; 1:100), CD8 (cytotoxic T cell marker; clone 4B11; Leica; 1:100), granzyme B (T cell activation/cytotoxicity marker; clone 11F1; Leica; 1:40), FoxP3 (Treg marker; clone 236AE7; Leica; 1:200), CD20 (B cell marker; clone L26; Leica PA0200; 1:2), CD56 (natural killer cell marker; clone CD564; Leica; 1:500), and CK19 (epithelial/tumor cell marker; clone A53-B/A2.26; Cell Marque; 1:200).

qmIF: computational methods

A blinded gastrointestinal pathologist selected representative areas for multispectral image capture at 20 magnification using VECTRA® (PerkinElmer, now Akoya Biosciences). As capturing lymph node or lymphoid aggregates would potentially skew results, the gastrointestinal pathologist was tasked with identifying up to 10 representative areas per slide predominantly composed of tumor. We performed cell composition analyses using a qmIF platform, inForm®. To quantitatively assess patient response to the combinational treatment, we measured three key parameters within the tumor region: (1) immune cell proportion, (2) immune cell ratio within a tumor cell radius, and (3) median distance from tumor cells to the nearest immune cell.

The immune cell proportion is calculated as the raw percentage of each immune cell type within the selected samples. To determine the immune cell ratio within a tumor cell radius, we established a neighborhood with a radius of 40 µm centered each tumor cell. Within this specified area, we calculated the proportion of a particular immune cell type relative to the total number of cells present. All tumor cells were considered separately, and the proportion within each tumor cell’s radius was averaged together. This ratio quantifies the immune cell density in the immediate vicinity of each tumor cell, providing insights into local immune cell infiltration.

To calculate the median distance between tumor and immune cells, we first determined the distance from each tumor cell to its nearest target immune cell. A list of these distances was compiled, and the median value was calculated. This metric reflects the typical proximity of immune cells to tumor cells, which is useful for understanding the spatial distribution of immune cells in relation to the tumor.

The significance of differences between groups was assessed using the Mann–Whitney U test. This non-parametric test was chosen due to its robustness in comparing distributions without assuming normality. Python (v 3.10.2) was utilized to calculate spatial metrics and perform statistical analysis. Specifically, the NumPy (v 1.26.0) and pandas (v 2.1.1) packages were employed to compute spatial metrics. For plotting figures, the seaborn (sns) (v 0.13.0) and matplotlib (v 3.8.0) packages were used.

Immunohistochemistry

Tissue for immunohistochemistry was fixed overnight in 10% phosphate-buffered formalin, transferred to 70% ethanol, processed, and embedded in paraffin. Sections were cut at 5 µm, mounted on positively charged slides, deparaffinized, rehydrated, and subjected to heat-induced epitope retrieval in a 10 mM citrate buffer. Endogenous peroxidase activity was quenched with 3% hydrogen peroxide, and slides were blocked before overnight incubation at 4 °C with cleaved caspase-3 antibody (Cell Signaling, #9664S, 1:1000). Slides were then incubated with ImmPRESS HRP anti-rabbit IgG secondary antibody (Vector Laboratories, MP-7401), developed with 3,3′-diaminobenzidine, counterstained with hematoxylin, dehydrated, and mounted. Images were acquired using a Leica LMD6500 microscope. Cleaved caspase-3 quantification was performed in a blinded fashion by counting positive cells in 10 fields per sample at 40× magnification, and results were expressed as the mean value for each slide.

Single cell RNA sequencing: sample processing and sequencing

We obtained fresh frozen tissue specimens from the New York Presbyterian/Columbia University Medical Center under IRB approval (#AAAT6521, AAAS9513) and prepared them as previously described for sequencing^54,55. We embedded samples in Tissue-Tek Optimal Cutting Temperature (OCT) Compound (Sakura Finetek; #4583) and obtained 4–5 sections of 20 µm thickness per sample using a Leica CM1950 cryostat (Leica Biosystems, Wetzlar, Germany). We placed the samples in 5 mL tubes (Eppendorf, Hamburg, Germany), washed with ice cold PBS (Thermo Fisher, #10010023), centrifuged for 2 min at 400 g, then discarded the supernatant and resuspended in 1 mL ST buffer [146 mM NaCl, 10 mM Tris-HCl pH7.5, 1 mM CaCl₂, and 21 mM MgCl₂ in ultrapure water] with 0.03% Tween-20 (Sigma Aldrich, #P7949) and 0.1% BSA (New England Biolabs, #B9000S) plus or minus 40 U/mL RNAse OUT inhibitor (Thermo Fisher, #10777019). We pipetted each suspension 15 times with a 1 mL pipette to mechanically dissociate tissue, then incubated for 5 min on ice before repeating the pipetting and finally quenching with 4 mL ST buffer plus or minus 40 U/mL RNAse OUT (Thermo Fisher, #10777019). We pre-wetted sequential 70-µm nylon mesh filters (Fisher Scientific, #NC0446099), then filtered samples into a 50 mL conical tube and washed the filter with 5 mL ST buffer. We centrifuged for 5 min at 500 g and discarded supernatant to isolate the dissociated nuclei, which were resuspended in 100-400 µl ST buffer, filtered with a 40 µm mesh filter (Fisher Scientific), stained with 50 µg/ml Hoechst 33342 (Thermo Fisher, #H3570), and then counted with a Neubauer counting chamber (Bulldog Bio, Inc., Portsmouth, NH).

Single nucleus RNA library preparation and sequencing

We loaded 0.9-1.5e3 nuclei in ST buffer without RNAse inhibitor using a Chromium controller and 5’ chemistry v2 capture (10X Genomics, #1000263). We prepared cDNA libraries according to manufacturer instructions for reverse transcription and cleanup, and performed an extra cycle of cDNA amplification to compensate for the reduced total RNA content when analyzing single nuclei compared to single cells. We prepared final sequencing libraries with the 10X Library Construction Kit (#1000190) and Dual Index Kit TT Set A (#1000215) and quantified them using D1000 and D5000 reagents (Agilent, #5067-5583 and #5067-5589) on an Agilent 2200 TapeStation system. We then mixed samples and sequenced to a target depth of at least 20,000 reads per droplet using NovaSeq S4 (Illumina, San Diego, CA) with at least 2x100bp read length.

Single cell RNA sequencing: computational methods

Generation and filtering of single nucleus expression matrices

The demultiplexed fastq files from raw single nuclei RNA sequencing reads were aligned to GRCh38-2020-A transcriptome (human), followed by quantification of gene counts using Cell Ranger v7.1.0⁵⁶ ‘count’ with introns included (Single Cell 5’ PE chemistry). Single-nuclei experiments suffer from several background noise in the raw observations due to contamination by ambient transcripts, impurity of barcode beads, to name a few. To counter these issues, we ran ‘remove-background‘ function from CellBender v0.2.2²⁴ on a per sample basis, (total droplet number = 15,000, epochs = 300), and an expected number of total cells equivalent to 80% of the number estimated by the CellRanger output. As part of the downstream analysis, the gene expression count matrices were processed on a per-sample basis with R v4.2.0 using Seurat v4.3.0.1²⁶. For quality control purposes, we further removed doublets via Scrublet²⁵, using an estimated doublet rate for each sample based on multiplex rate estimations from the 10X Genomics website⁵⁷. Counts matrices were loaded into Seurat and filtered such that each barcode had a gene count between 200 and 12,000 genes and a mitochondrial gene load of <15%. Further filtering ensured that all barcodes were either predicted as a singlet by Scrublet²⁵, or had a cell cycle phase prediction in G2M/S phase as predicted by the ‘CellCycleScoring‘ function. Downstream processing was performed using the Seurat ‘NormalizeData‘ pipeline, followed by dimensional reduction using Principal component analysis (PCA) and uniform manifold approximation and projection (UMAP) with top 30 principal components.

Identification of malignant cells based on copy-number alterations

After running this ‘NormalizeData‘ pipeline, we used the SingleR library⁵⁸ with the built-in Blueprint ENCODE and HPCA datasets as a reference to generate first-pass cell annotations at the barcode level for each sample. To generate a non-malignant reference dataset for CNV analysis, we combined barcodes and their corresponding count matrices that were annotated by SingleR as T cells across all samples. This was used as the reference dataset for reference-based methods of malignant cell prediction. We predicted malignant cells on a sample-wise basis prior to integration. To estimate which cells were malignant, we used a consensus of several CNV-based methods including both reference-based and reference-free tools; when a reference was required, we provided the global T cell reference as described above. We further verified the validity of these predictions against a known epithelial gene signature⁵⁹.

We ran the reference-based tools Numbat v1.4.0 and InferCNV v1.3.3 as well as the reference-free tool CopyKAT v1.1.0^27,60,61, with Numbat providing the majority of high-quality malignant predictions used for final annotations. Numbat and InferCNV assign a malignant probability to each cell—to minimize false negatives, we applied a separate probability threshold for each sample that maximized the proportion of malignant cells within clusters in the sample; conversely, to minimize false positives, we applied a separate filter for minimum proportion of predicted malignant cells within these clusters. For samples that displayed low consensus across the different prediction methods, we labeled cells as malignant if they belonged to clusters that expressed the malignant epithelial signature and were predicted by at least one of the above methods to be malignant. Once tumor cells had been predicted for each sample, we performed integration and verified that all the predicted malignant cells indeed shared biological features in the combined dataset via PCA visualization (Fig. 4b).

Sample integration

After high-confidence malignant predictions were annotated for each sample, we combined all samples using the standard Seurat CCA integration pipeline, with integration features and anchors selected across all 22 samples. This was performed by first selecting a set of ranked features using the ‘SelectIntegrationFeatures’ function, followed by selecting n = 2000 anchors using the ‘FindIntegrationAnchors’ function between different samples based on canonical correlation approach’s top 30 dimensions, to specify the search space. The final integration was done using the ‘IntegrateData’ function based on the already computed anchors, followed by performing scaling using the ‘ScaleData’ function of Seurat. For all downstream analyses, metadata (Table 1) collected at the sample level was used to associate a given barcode with metadata categories such as patient demographics, treatment stage, response status, and tumor percentage change at peak radiographic response.

Nonnegative Matrix Factorization (NMF)

Following integration, all cells predicted as malignant were analyzed separately from the non-malignant cells for NMF and tumor diversity analysis. NMF is a method for performing dimensional reduction and feature extraction on ‘non-negative’ data. Conventionally, NMF decomposes a matrix A into two matrices with non-negative entries with smaller ranks, A ≈ WH, where $A\in R\left(n\times m\right)$, $W\in R\left(n\times k\right)$, $H\in R\left(k\times m\right)$. Herein, rows and columns of A represent features (genes) and samples, respectively. Based on specific context, W is interpreted as feature mapping, its rows representing meta-genes whereas columns of H represent samples (cells).

For each sample within the malignant compartment, the raw counts matrix was filtered to exclude mitochondrial and ribosomal genes, and then further filtered to include only the 7000 genes with the highest expression, as described by Gavish et al.²⁹ This resulting count matrix was re-normalized as described above, and rank estimation was performed for ranks 2 through 8 with a total of 10 runs per rank using the Brunet method of the NMF package^62,63. For each sample, the optimal rank was manually selected based on plotted values of the cophenetic correlation and silhouette scores in this estimation. We then determined the top 125 most variable genes and computed the Spearman correlation between each sample’s factors with respect to these top genes. This correlation matrix, combined with sample metadata, was plotted with the ComplexHeatmap package⁶⁴, and the optimal value of dendrogram clusters was estimated as k = 9 based on silhouette scores for each cluster. These clusters were then manually readjusted based on strength of correlation to arrive at 6 consensus metaprograms (Supplementary Fig. 6a–c).

For each metaprogram, we calculated the normalized contribution y of each gene i using the formula $y_i=\frac{{\sum }_{f=1}^r{x}_{f,i}}{\sqrt{r}}$ where r is the total number of factors contributing to a given metaprogram and x_f,i represents the value of the combined W matrix at factor f and gene i⁶⁵. We then took the top 20 genes for each metaprogram, excluding the highly non-specific genes ‘NEAT1’, ‘MALAT1’, ‘FP671120.4’, and ‘FP236383.3’, and visualized these normalized gene contributions towards each metaprogram. Each rank/factor in the dataset was associated with a subset of barcodes that most strongly contributed to those factors, allowing us to assay levels of the Chromosomal Instability (CIN)-related CIN70 gene signature⁶⁶ within each metaprogram based on the strength of the signature in the associated cells (Fig. 4c).

Survival analysis

Using the above methods to identify the top genes for each program, excluding highly non-specific genes, we defined a signature for each metaprogram based on the top 50 genes contributing to that metaprogram. Using the GEPIA2 web portal^67–69, we performed survival analysis on PAAD patients within the TCGA database by supplying the 50-gene signature for each metaprogram. We analyzed Overall Survival with a median (i.e., 50-percentile) group cutoff threshold, computing hazard ratio and log-rank tests between groups with high vs low expression each metaprogram signature.

Tumor heterogeneity

Within the malignant compartment, we estimated tumor heterogeneity within each sample using a method adapted from Ma et al.²⁸. We performed Principal Component Analysis on the entire malignant compartment and then computed the per-sample center of mass within the first 50 PCs. We then calculated the Euclidean distance from each cell to its respective sample centroid in PC space as a measure of transcriptomic distance. Aggregating these distance values by sample, we established the level of tumor transcriptomic heterogeneity for each sample.

Differential cell abundance analysis

We used the MiloR framework⁷⁰ to perform differential abundance analysis within each major cell population. This framework models differences in the cell state abundance among treatment stages (Pre, On-Treatment, On-Progression) and response categories (PR, SD, PD). This calculation is based on graph neighborhoods, which are calculated based on expressional similarity between cells in the analyzed compartment. For each of these predicted neighborhoods, hypothesis testing identified whether the neighborhood was differentially abundant in each of the aforementioned testing categories, using a weighted FDR procedure to account for the overlap of cells between distinct neighborhoods. A weighted mean $\underset{\_}{\pi }$ was calculated as a summary statistic for each group as a summary statistic, in which log fold change values were weighted by neighborhood size.

Differential expression

As described previously, the Seurat ‘FindAllMarkers‘ function was used to identify marker gene lists specific to each cluster. When it was desirable to perform differential expression of a specific subset of cells based on metadata categories such as treatment stage, we used the ‘FindMarkers‘ function with patient ID as a latent variable in the MAST regression model⁷¹. We validated the results of these single-cell differential expression tests using the DotPlot function to assay levels of given genes within different cell types and within different patient samples.

Ligand-receptor analysis

Contact tracing was used to identify ligand–receptor interactions across patients prior to and following MCGN treatment. For each potential ligand–receptor pair, a likelihood ratio test was conducted comparing receptor-expressing to receptor-null cells within a given cell cluster. Variability in these results was measured between pre- and post-treatment conditions across patients with PR and PD. Inputs for ligand–receptor interactions were pulled from curated protein–protein interactions in CellPhoneDB⁷² v2.1.4. A threshold fraction of minimum expression for receptors and ligands was set to 0.05 for a given cell type to filter ligand–receptor pairs.

Statistical analysis

Multiple-hypothesis correction was applied where appropriate for families of related comparisons, except where noted for comparisons that represent independent subgroups rather than a single family. Bonferroni-adjusted P values were used for analyses involving multiple prespecified comparisons, whereas standalone prespecified tests are reported with unadjusted P values together with effect-size measures. For MiloR analyses, false-discovery control was applied using the framework’s spatial FDR procedure at the level of cellular neighborhoods

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Conceptualization: G.A.M., B.I., K.P.O., and A.G.R. Data curation: G.A.M., B.I., K.P.O., A.G.R., E.K.D., M.S.M., S.Se., N.S., I.R., S.P., and S.W. Formal analysis: G.A.M., B.I., K.P.O., A.G.R. E.K.D., M.S.M., S.P., and S.W. Funding acquisition: G.A.M., B.I., and K.P.O. Investigation (preclinical, clinical, and correlative): G.A.M., B.I., K.P.O., A.G.R., E.K.D., M.S.M., K.L., S.T., P.S., I.P. C.P., S.Sa., K.L., P.S., S.T., Z.H.W., A.T., S.B., S.D., A.I., M.D.K., B.A.S., T.F., D.B., J-M.S., Y.P., R.H.M., S.B., L.W., H.H., L.L., A.C.T., and J.A.C. Methodology: G.A.M., B.I., K.P.O., A.G.R., E.K.D., M.S.M., S.P., S.W., C.P., S.Sa., K.L., P.S., S.T., Z.H.W., A.T., S.B., S.D., A.I., H.H., L.L., and A.C.T. Project administration: G.A.M., B.I., K.P.O., and A.G.R. S.Se., N.S., and I.R. Resources: G.A.M., B.I., K.P.O., A.C.T., and G.K.S. Supervision: G.A.M., B.I., K.P.O., A.G.R., E.K.D., and M.S.M. Validation: G.A.M., B.I., K.P.O., A.G.R., E.K.D., M.S.M., S.P., S.W., L.L., and A.C.T. Visualization: G.A.M., B.I., K.P.O., A.G.R., E.K.D., M.S.M., S.P., S.W., S.P., S.W., and A.C.T. Writing—original draft: G.A.M., B.I., K.P.O., A.G.R., E.K.D., and M.S.M. Writing—review and editing: G.A.M., B.I., K.P.O., E.K.D., M.S.M., I.P. K.L., P.S., S.T., Z.H.W., A.T.T., R.M., S.B., G.K.S., and J.C. All approved the final manuscript for publication.

Peer review

Peer review information

Nature Communications thanks Grainne Okane, who co-reviewed with Paula Lynch, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This work was supported by the National Institutes of Health (NIH), National Cancer Institute (NCI) grants, U01CA274312, R01CA266558 (to K.P.O. and G.A.M.), Herbert Irving Comprehensive Cancer Center Velocity grant and Conquer Cancer Foundation Young Investigator Award (to G.A.M.), NIH, NCI grants, R37CA258829, R01CA266446, R01CA280414 and U54CA274506 (to B.I.), F30CA298572 (to Z.H.W.), F30CA301910 (to E.K.D.), K38CA304069 (to M.S.M.), and additional support by the Burroughs Wellcome Fund Career Award for Medical Scientists, a Velocity Fellows Award, the Louis V. Gerstner, Jr. Scholars Program, a Tara Miller Young Investigator Award by the Melanoma Research Alliance (MRA), a Tara Miller Team Science Award for Brain Metastasis Research by the MRA, and the Pershing Square Sohn Cancer Research Alliance Award (to. B.I.), and a Melanoma Research Foundation (MRF) Medical Student Award (to Z.H.W.). E.K.D., K.L., and Z.H.W. are supported by NIH Medical Scientist Training Program grant T32GM145440. B.I. is a CRI Lloyd J. Old STAR (CRI5579). This work was additionally supported by the Herbert Irving Comprehensive Cancer Center (HICCC) Human Tissue Immunology and Immunotherapy Initiative (to B.I.). This work was supported by NIH/NCI Cancer Center Support Grant P30CA013696.

Data availability

The minimum dataset needed to interpret, verify, and extend the findings of this study is provided within the article and its Supplementary Information. The trial protocol and statistical analysis plan are available in the Supplementary Information. Individual participant-level clinical data are not publicly available because of patient privacy considerations and ethical restrictions. Deidentified individual participant data may be made available by the corresponding author, G.A.M. (gam2140@cumc.columbia.edu), upon reasonable request and with approval from the institutional review board. Any data sharing will require execution of an appropriate data-sharing agreement and will be conducted in accordance with applicable institutional and regulatory policies. Processed and de-identified data have been made available via GEO database under accession GSE316195 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE316195). Code for downstream analysis of this data is available on the Izar Lab GitHub (see below). Source data are provided with this paper.

Code availability

All code for single-cell analyses is publicly available at https://github.com/IzarLab/MCGN-for-PDA (https://doi.org/10.5281/zenodo.20516281).

Competing interests

G.A.M. received research support from BioLineRx Pharmaceuticals. In the last five years, G.A.M. has served as a consultant for or received honoraria from Exelixis, Roche/Genentech, Ipsen, Cend Therapeutics, EnGeneIC, Pfizer, Revolution Medicines, and has received research funding to Columbia University from Merck (Inst), Roche/Genentech (Inst), Plexxikon (Inst), Regeneron (Inst), BiolineRx (Inst), Arcus Biosciences, and has received travel accommodations from Roche/Genentech, Exelixis, Celgene, Revolution Medicines, DAVA Oncology. G.A.M. has stock ownership interests in EnGeneIC. B.I. has a sponsored research agreement with Pharming Group. In the last five years, B.I. has served as a consultant for or received honoraria from Volastra Therapeutics, Johnson & Johnson/Janssen, Novartis, GSK, Eisai, AstraZeneca and Merck, and has received research funding to Columbia University from Agenus, Alkermes, Arcus Biosciences, Checkmate Pharmaceuticals, Compugen, Immunocore, Regeneron, and Synthekine. B.I. is the scientific founder of Basima Therapeutics, Inc. K.P.O. has received research funding to Columbia University from Revolution Medicines. The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76559-4.