Extracellular Matrix-MYCAF Signatures Correlate with Resistance to Neoadjuvant aPD-L1 Immune Checkpoint Inhibition with Durvalumab + Metformin in HPV+ HNSCC

Abstract

Purpose: Immune checkpoint inhibitors (ICI) have demonstrated clinical benefit in head and neck squamous cell carcinoma (HNSCC); however, single-agent efficacy is limited, leaving significant unmet needs. Metformin may synergize with ICIs, offering promise to improve response rates. We leveraged multiomic data from a randomized, presurgical neoadjuvant trial (NCT03618654) evaluating a single infusion of the anti-PD-L1 ICI durvalumab with or without daily, standard dose metformin in previously untreated, nondiabetic patients with HNSCC to understand predictors of response and the effect of combination therapy.

Patients and methods: Clinical, pathologic, and correlative data were analyzed to investigate response and resistance mechanisms. We present an in-depth multiomic analysis of primary tumor specimens to study treatment response/resistance in human papillomavirus-positive HNSCC.

Results: Baseline samples revealed that myofibroblastic cancer-associated fibroblast and extracellular matrix signatures were enriched in durvalumab plus metformin nonresponders, which were localized to the leading tumor edge on spatial transcriptomics. In contrast, baseline responder samples were enriched for the Langerhans-like dendritic cell (DC) state and IFN signatures. Treatment increased intratumoral CD8+ T-cell and IFN signatures and peripheral blood CCL2 levels. Responders demonstrated macrophage and DC enrichment and antigen processing and presentation upregulation. Enrichment of cell cycle-related gene sets, specifically the MYC targets V1 hallmark gene set, correlated with nonresponse.

Conclusions: Early response and resistance dynamics for durvalumab plus metformin in human papillomavirus-positive HNSCC reveal baseline extracellular matrix-myofibroblastic cancer-associated fibroblast as predictive of nonresponse. In contrast, responders were distinguished by baseline enrichment in the Langerhans-like DC state and posttreatment antigen-presenting gene sets.

Figure 1.

Study design, imaging, histology, and response assessment. A, Study schema outlining patient enrollment, treatment groups, and intervention timeline. Patients were randomized 3:1 to receive DM or durvalumab alone (D). B, Representative pre- and posttreatment radiographic images illustrating tumor size reduction in a responder patient following treatment. C, Histologic analysis of tumor specimens. Hematoxylin and eosin staining of posttreatment samples from a responder and nonresponder (top). The arrows point to fibrotic stromal tissue in the responder and dense tumor tissue with less fibrosis in the nonresponder, indicating differences in the TME. Immunofluorescence image of DSP in a responder, highlighting PanK-Cy3 (tumor), CD45-Texas Red (immune), and CD163-Cy5 (stromal/macrophage) compartments (bottom). D, RECIST 1.1 waterfall plot depicting changes in tumor size following treatment, stratified by response category, pathologic response, and PD-L1 CPS. NA, not available. NR, nonresponder; PanK-Cy3, pan–keratin-Cy3; R, responder. (Created with BioRender.com. Llerena, P., and Samarah, H. [2025] https://BioRender.com/nf4luru.)

Figure 2.

Posttreatment vs. baseline unsupervised GSEA. Selected results from an unbiased, longitudinal GSEA via query with 28,531 gene sets, contrasting post- vs. pretreatment with DM in HPV⁺ HNSCC. A, Parallel analyses were performed on nonresponder and responder groups. ΔNES per gene set, between nonresponder and responder groups are listed (right column). Maximum differential gene expression was seen in the Anastassiou multicancer invasiveness signature, with a ΔNES of 5.5 (top). Supervised GSEAs by query with several different CAF gene sets revealed that some mirrored the enrichment characteristics seen with the multicancer invasiveness signature (three bottom images). LEs of enrichment for nonresponder and responder groups were pooled separately and each filtered to remove redundancies (bottom). Here, the negative enrichment seen in nonresponders equates to upregulated, pretreatment; the positive enrichment seen in responders equates to upregulated, posttreatment. B, Venn diagram demonstrating gene expression correlates of nonresponders and responders in CAF gene set pools. Enriched gene expression lists for nonredundant nonresponders and responders, n = 92 and 119, respectively, were parsed in the Venn diagram. C, Venn-derived gene lists unique to either nonresponders (n = 38) or responders (n = 65) were then queried in parallel for functional predictions via STRING analyses. Nonresponse correlates consisted of five ECM-based, protumor functions (left). In contrast, elastic fiber formation emerged as a response correlate, albeit as seen through the prism of CAF gene sets (right). D, Elastic fiber formation as a response correlate, as detected in unbiased GSEA. Shown are results of the Reactome gene set for elastic fiber formation (n = 44 genes) that had been included in our 28,531–gene set unbiased query. Enrichment achieved significance in responders, posttreatment (right), whereas it was without relevance to nonresponders (left). NR, nonresponder; post-Tx, posttreatment; pre-Tx, pretreatment; R, responder. (Created with BioRender.com. Llerena, P., and Samarah, H. [2025] https://BioRender.com/nf4luru.)

Figure 3.

Biomarkers of nonresponse to DM, sourced by using GSEA rank-ordered gene lists. Based on the magnitude of fold changes in gene expression, these lists are generated for each contrasting set of samples. Comparison of row assignments for individual genes on rank-ordered lists for nonresponder and responder groups enables the identification of genes with the maximal level of differential expression. In this manner, we defined predictive markers for nonresponders to DM in HPV+ HNSCC. A, Heatmap for a portion of rank-ordered gene list for nonresponders, illustrating the 50/19,258 most upregulated genes, pretreatment. Based on nonresponder and responder rank-ordered gene list assignments for these 50 genes, an 11-gene signature for nonresponders was developed, as indicated here (*) and as detailed in Supplementary Fig. S2A. B and C, Supervised GSEAs of nonresponder and responder groups by query with the 50 genes the expression of which was elevated in nonresponders. Juxtaposition of these longitudinal contrasts illustrates that enriched expression of the 50 genes correlating with nonresponder is not relevant to the responder group. D, Normalized read counts for the expression of specific genes in nonresponder and responder groups. Baseline nonresponders (red) were compared with pretreatment responders (blue), by using ordinary one-way ANOVA. *, P < 0.05; **, P < 0.01. Elevated expression of POSTN and ADAMTS5 correlated with resistance to therapy (E–H). Correlation between POSTN and ADAMTS5 expression and tumor purity (E and G, respectively) and CD8⁺ T-cell infiltration (F and H, respectively) in TCGA head and neck cancer data. I and J, GSVA scores and ROC analysis for the 11-gene resistance signature. K and L, Periostin IHC and quantification. Box plot showing percent staining of four baseline samples for each group which correspond to the samples analyzed by bulk RNA-seq (K). Images of a responder (left image) and nonresponder (right image) baseline sample stained with anti-periostin (L). Scale bars, 100 microns. NR, nonresponder; post-Tx, posttreatment; pre-Tx, pretreatment; R, responder. (Created with BioRender.com. Llerena, P., and Samarah, H. [2025] https://BioRender.com/nf4luru.)

Figure 4.

Baseline TME characterization and cell state composition in responders and nonresponders. A, Heatmap of baseline tumor samples displaying transcriptomic profiles of key TME features, including angiogenesis, immune infiltration, and EMT signatures, used to classify TME subtypes. B, NESs for hallmark cytokine pathways (IFNα, IFNγ, TNFα, and TGFβ) in bulk and spatial compartments (PanCKe, CD163e, and CD45e), comparing baseline samples between responders and nonresponders. Statistically significant differences (P < 0.05) are indicated by gray bars. C and D, EcoTyper analysis of baseline bulk RNA samples, showing cell state composition across immune and stromal populations, with box plots highlighting S01 epithelial cells (D). E and F, EcoTyper analysis of the tumor (PanCKe) compartment, showing cell state abundance, with box plots highlighting differences in S02 mast cells, S03 fibroblasts, S02 CD8⁺ T cells, and S03 NK cells (F). G and H, EcoTyper analysis of the stromal (CD163e) compartment, with box plots highlighting differences in S04 macrophages/monocytes, S02 NK cells, and S02 polymorphonuclear cells (H). Asterisks (* or **) indicate statistically significant differences, with P values provided. NR, nonresponder; PMN, polymorphonuclear; post-Tx, posttreatment; pre-Tx, pretreatment; R, responder; Treg, regulatory T cells. (Created with BioRender.com. Llerena, P., and Samarah, H. [2025] https://BioRender.com/nf4luru.)

Figure 5.

Posttreatment TME changes and cytokine enrichment in responders and nonresponders. A, Heatmap displaying transcriptomic profiles from posttreatment RNA-seq, illustrating changes in key TME features, including angiogenesis, immune infiltration, and EMT signatures. B, Sankey diagram showing shifts in TME classification from baseline (Pre) to posttreatment (Post) in responders and nonresponders, highlighting transitions between IE, IE-F, F, and depleted (D) phenotypes. C and D, NES for hallmark cytokine pathways (IFNα, IFNγ, TNFα, and TGFβ) in bulk RNA and spatial compartments PanCKe, CD163e, and CD45e, comparing baseline vs. posttreatment samples in responders (C) and nonresponders (D). Statistically significant differences (P < 0.05) are indicated by gray bars. E, Box plot showing the percent expression of CD1a at baseline between responders and nonresponders. Quantification of CD1a staining is based on percent CD1a-positive cells in three regions per sample (n = 15 responders; n = 12 nonresponders), with responders exhibiting a 2.25-fold higher average CD1a expression compared with nonresponders (*, P = 0.0294). F, Box plot showing the mean distances between CD8⁺ and FOXP3+ cells at baseline in responders and nonresponders. No statistically significant difference was observed (*, P = 0.74). G, Representative CD1a IHC staining images from a baseline responder (left) and nonresponder (right) biopsy sample. Scale bar, 50 microns. NR, nonresponder; Pre R, baseline responders; R, nonresponder. (Created with BioRender.com. Llerena, P., and Samarah, H. [2025] https://BioRender.com/nf4luru.)

Figure 6.

Antigen processing and presentation pathway enrichment in responders and nonresponders in PanCKe and CD163e compartments. A and D, Heatmaps displaying NESs for GOBP, KEGG, and Reactome antigen processing and presentation pathways, including MHC I and MHC II, in PanCKe (A) and CD163e (D) compartments, comparing responders and nonresponders. B, C, E, and F, Volcano plots highlighting genes in the LE of the GOBP antigen processing and presentation gene set, which was selected because of the greatest difference in NESs between responders and nonresponders: for responders (B and E) and nonresponders (C and F) in the PanCKe (B and C) and CD163e (E and F) compartments. AP, antigen presentation; APP, antigen processing and presentation; GOBP, GO biological processes; NR, nonresponder; R, nonresponder. (Created with BioRender.com. Llerena, P., and Samarah, H. [2025] https://BioRender.com/nf4luru.)

Figure 7.

Cytokine differential analysis, CCL2 expression, and its impact on survival in responders and nonresponders. A and B, Canonical plots displaying cytokine differential analysis, with F1 and F2 axes representing variance in cytokine expression. A, Colors indicate different treatment conditions: baseline (red), durvalumab alone (green), and DM (blue). B, Colors indicate different treatment response: nonresponders (red), responders (blue), and baseline (green). C, Box plots comparing pre- vs. posttreatment peripheral blood CCL2 titers in responders and nonresponders, showing a significant posttreatment increase. Asterisks (*) indicate significance threshold P < 0.01. D, Box plot showing significantly higher baseline transcriptomic CCL2 expression (log₂-normalized read counts, CD163e) in nonresponders compared with responders. Asterisks (**) indicate significance threshold P < 0.01. E, Box plots showing transcriptomic CCL2 expression (log₂-normalized read counts) in the PanCKe region, comparing pre- and posttreatment samples in responders and nonresponders. Asterisks (*) indicate significance threshold P < 0.01. F, Patients with high CCL2 expression and high M2 macrophage abundance (red) had significantly improved survival compared with those with high CCL2 but low M2 macrophage abundance (orange; HR = 0.519; P = 0.000551), with no significant survival difference between low CCL2 groups (HR = 0.903; P = 0.635). G, Scatter plots illustrating the correlation between CCL2 expression and tumor purity (left) and CAF infiltration (right) in HPV+ HNSCC samples. H–K, Box plots comparing baseline responders (Pre R) vs. posttreatment responders (Post R) cell state estimates in bulk, PanCKe, CD163e, and CD45e segments. NR, nonresponder; R, responder. Asterisks (*) denote location of significant P values. (Created with BioRender.com. Llerena, P., and Samarah, H. [2025] https://BioRender.com/nf4luru.)