Pan-cancer analysis in the real-world setting uncovers immunogenomic drivers of acquired resistance post-immunotherapy

Abstract

Background

Immune checkpoint blockade (ICB) has revolutionized cancer therapy, yet resistance—both primary and acquired—remains a significant obstacle, affecting the majority of patients.

Methods

Here, we leverage a large-scale, real-world clinicogenomic dataset to systematically explore the molecular underpinnings of ICB resistance in the post-progression setting. We analyze over 5,000 pan-cancer patients with clinical and pre-/post-treatment genomic and transcriptomic data and systematically compare the clinical and molecular features of acquired versus primary ICB resistance.

Results

Post-ICB progression, acquired resistance showed extended survival compared to primary resistance across all cancer types. This clinical phenotype was paralleled by a universally immune-inflamed, albeit dysfunctional, tumor microenvironment (TME) at the onset of acquired resistance, with sustained or ICB-induced inflammatory and interferon responses. We confirm previously described mechanisms of acquired resistance, including B2M loss-of-function (LoF) in non-small cell lung cancer (NSCLC), and identify novel potential mediators, including LoF of TGFBR2 in NSCLC, CYLD in head and neck cancer, and RUNX1 in triple-negative breast cancer. Further supporting their involvement in resistance, these acquired ICB alterations associated with immune-escaped TMEs, characterized by active immunomodulatory oncogenic signaling, hyperproliferation and invasiveness, or altered tumor metabolism.

Conclusions

These findings emphasize the heterogeneity of molecular drivers of acquired resistance to ICB within and across cancers, and highlight the potential for personalized therapeutic interventions post-progression to improve patient outcomes.

WHAT IS ALREADY KNOWN ON THIS TOPIC.

WHAT THIS STUDY ADDS

• We provide the largest and first pan-cancer comparison of acquired versus primary ICB resistance, expanding resistance characterization to understudied indications. We uncover both universal and cancer-specific features of acquired versus primary ICB resistance in the post-progression setting, including extended survival, consistent immune inflammation, and selection for distinct genetic alterations.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Our findings may inform positioning of next-generation immunotherapies and improve patient stratification post-ICB progression.

Introduction

Immune checkpoint blockade (ICB) has deeply transformed the treatment landscape of cancer, increasingly becoming established as the standard of care (SoC) across many solid malignancies.^{1 2} While a subset of patients can perform exceptionally well on ICB, 40–90% of cases derive no benefit from treatment (primary resistance) and up to 70% of initial responders eventually relapse and progress, yielding ∼5–25% of patients with acquired resistance.^{3 4} Significant translational efforts have been put towards understanding molecular correlates of response and resistance to ICB, focusing primarily on analyzing pre-treatment tumor biopsies from responders and non-responders. Such efforts contributed to the development of robust response biomarkers, some of which are approved by the US Food and Drug Administration (FDA) for patient selection and stratification. These include high tumor mutational burden (TMB), microsatellite instability (MSI) or mismatch repair deficiency, or high protein expression of the programmed cell death ligand 1 (PD-L1) immune checkpoint.^{4 5}

Despite these efforts, our understanding of which molecular features emerge as patients progress on ICB remains very limited, particularly in the context of acquired resistance. This is notably due to challenges in accessing post-treatment clinical samples, as well as the lack of a clear and standardized definition of acquired resistance.^{3 6 7} So far, efforts to address this gap, including ours,⁸ focused on analyzing (un)paired post-ICB tumor biopsies from mostly non-small cell lung cancer (NSCLC) or melanoma patients, uncovering several mechanisms underlying acquired ICB resistance. These mechanisms include genetic defects in antigen processing and presentation machinery (APM) and/or interferon (IFN) signaling,8,13 chronic IFN gamma (IFNγ) signaling,⁸ cancer driver alterations promoting immune exclusion,^{14 15} or upregulation of alternative immune checkpoints following treatment.16,18 The diversity of these mechanisms underscores the heterogeneous nature of acquired resistance to ICB and suggests that additional mediators may have been overlooked due to limited sample sizes. Moreover, the extent to which these mechanisms are shared in other cancer types, in post-ICB primary resistance, or in acquired resistance to non-ICB regimens remains underexplored. Addressing these questions is crucial to direct therapeutic interventions post-ICB resistance towards more personalized approaches.

The establishment of ICB as part of routine care comes with an ever-increasing availability of clinical samples across various treatment settings, including the post-progression setting. Therefore, clinicogenomic real-world data (RWD) provides an exciting opportunity to dissect the molecular drivers of resistance as it emerges across multiple cancer types. Here, we leverage a real-world dataset from Tempus AI (tempus.com) to systematically analyze clinical and pre-/post-treatment multi-omic data from >5,000 patients with cancer. We uncover molecular determinants of acquired resistance across various disease settings, providing the largest-to-date and first cross-indication and multimodal comparison of acquired versus primary ICB resistance.

Results

Building a pan-cancer multimodal ICB cohort from real-world data

We leveraged Tempus AI’s US-based and de-identified clinicogenomic database to build a pan-cancer and multimodal real-world ICB (rwICB) cohort with follow-up until August 2023. Available data includes treatment history and outcomes from patient electronic health records (EHRs), panel DNA sequencing (DNA-seq), full transcriptome RNA-seq, and associated sample and clinical metadata (online supplemental figure 1A). Following a systematic informatics approach, we started from 22,952 patients spanning ten cancer types and applied a set of inclusion and exclusion criteria to build rwICB (online supplemental figure 1B, Methods). We considered patients who had (1) received at least one line of ICB (monotherapy or in combination with other drugs) in their treatment journey, (2) complete treatment information, (3) treatment outcomes, and (4) omic data (DNA and/or RNA) for analysis. All disease stages were considered. This allowed us to identify 3,559 patients across NSCLC (n=2,689), head and neck squamous cell cancer (HNC, n=516), and triple-negative breast cancer (TNBC, n=354) who received ICB between 2016 and 2023. To contrast our findings with other treatment modalities, we built a homologous and non-overlapping control cohort of 1,894 non-ICB-treated patients, consisting of chemotherapy for all three cancer types, as well as tyrosine kinase inhibitors (TKIs) for NSCLC (online supplemental figure 1B). The baseline patient characteristics of the ICB and non-ICB cohorts are summarized in table 1.

Table 1. Baseline patient characteristics of the ICB and non-ICB multimodal cohorts.

	ICB			Chemotherapy			TKI
	NSCLC (n=2,689)	HNC (n=516)	TNBC (n=354)	NSCLC (n=574)	HNC (n=462)	TNBC (n=503)	NSCLC (n=355)
Median age at treatment start	66 (21–83)	62 (26–87)	57 (29–84)	66 (31–83, 3 N/A)	60 (21–86)	57 (25–84, 1 N/A)	65 (30–87)
Median treatment start year	2019 (2016–2023)	2020 (2016–2022)	2020 (2016–2022)	2019 (2016–2022)	2020 (2016–2022)	2018 (1991–2022)	2019 (2016–2022)
Median follow-up (months)	18 (19 N/A)	13 (4 N/A)	16	19 (4 N/A)	15	25 (1 N/A)	26 (2 N/A)
Gender
Female	1,285 (47.8%)	121 (23.4%)	353 (99.7%)	300 (52.3%)	113 (24.5%)	500 (99.4%)	243 (68.5%)
Male	1,404 (52.2%)	395 (76.6%)	1 (0.3%)	274 (47.7%)	349 (75.5%)	3 (0.6%)	112 (31.5%)
Ethnicity
White	1,605 (59.7%)	293 (56.8%)	174 (49.2%)	325 (56.6%)	207 (44.8%)	232 (46.1%)	176 (49.6%)
Non-white	472 (17.6%)	69 (13.4%)	84 (23.7%)	87 (15.2%)	70 (15.2%)	122 (24.3%)	93 (26.2%)
N/A	612 (22.8%)	154 (29.8%)	96 (27.1%)	162 (28.2%)	185 (40%)	149 (29.6%)	86 (24.2%)
Smoking history
Never smoked	144 (5.4%)	62 (12%)	98 (27.7%)	39 (6.8%)	22 (4.8%)	46 (9.1%)	72 (20.3%)
Ever smoked	1,053 (39.2%)	139 (26.9%)	49 (13.8%)	182 (31.7%)	95 (20.6%)	53 (10.5%)	45 (12.7%)
N/A	1,492 (55.5%)	315 (61%)	207 (58.5%)	353 (61.5%)	345 (74.7%)	404 (80.3%)	238 (67%)
NSCLC histology
Non-squamous	1,963 (73%)	N/A	N/A	440 (76.7%)	N/A	N/A	350 (98.6%)
Squamous	726 (27%)	N/A	N/A	134 (23.3%)	N/A	N/A	5 (1.4%)
ECOG
0–1	420 (15.6%)	0	68 (19.2%)	80 (13.9%)	0	32 (6.4%)	50 (14.1%)
≥2	60 (2.2%)	0	3 (0.8%)	4 (0.7%)	0	1 (0.2%)	4 (1.1%)
N/A	2,209 (82.1%)	516 (100%)	283 (79.9%)	490 (85.4%)	462 (100%)	470 (93.4%)	301 (84.8%)
Stage
≤Stage III	658 (24.5%)	25 (4.8%)	39 (11%)	357 (62.2%)	66 (14.3%)	244 (48.5%)	16 (4.5%)
Stage IV	1,935 (72%)	406 (78.7%)	299 (84.5%)	168 (29.3%)	305 (66%)	227 (45.1%)	327 (92.1%)
N/A	96 (3.6%)	85 (16.5%)	16 (4.5%)	49 (8.5%)	91 (19.7%)	32 (6.4%)	12 (3.4%)
LoT
1st	1,448 (53.8%)	186 (36%)	89 (25.1%)	574 (100%)	462 (100%)	503 (100%)	355 (100%)
≥2nd	1,241 (46.2%)	330 (64%)	265 (74.9%)	0	0	0	0
ICB regimen
anti-PDx mono	720 (26.8%)	263 (51%)	28 (7.9%)	N/A	N/A	N/A	N/A
anti-PDx+chemo	1,331 (49.5%)	181 (35.1%)	265 (74.9%)	N/A	N/A	N/A	N/A
Other ICB	638 (23.7%)	72 (14%)	61 (17.2%)	N/A	N/A	N/A	N/A
TKI regimen
EGFR inhibitors	N/A	N/A	N/A	N/A	N/A	N/A	283 (79.7%)
ALK inhibitors	N/A	N/A	N/A	N/A	N/A	N/A	61 (17.2%)
Other TKI	N/A	N/A	N/A	N/A	N/A	N/A	11 (3.1%)

For age and treatment start year, both median and range (min-max) are reported. For TNBCs in the chemotherapy cohort, 51/503 (10%) patients received chemotherapy between 1991-2015. All other ICB and non-ICB patients received treatment between 2016-2023. For age, the number of patients missing age information (N/A) is reported in brackets. For follow-up, the number of patients missing or with inconsistent last known follow-up date and/or censoring information is reported in brackets (N/A). ‘anti-PDx mono’ designates anti-PD(L)1 administered as monotherapy, ‘anti-PDx + chemo’ designates anti-PD(L)1 administered concomitantly with chemotherapy, ‘Other ICB’ includes anti-PD1 combined with anti-CTLA-4, anti-PD(L)1 combined with other non-ICB drugs, or any anti-PD(L)1 part of a sequential drug combination.

anti-PDx, anti-programmed cell death (ligand) 1; ECOG, Eastern Cooperative Oncology Group performance status; HNC, head and neck squamous cell cancer; ICB, immune checkpoint blockade; LoT, line of treatment; mono, monotherapy; N/A, not assessed; NSCLC, non-small cell lung cancer; TKI, tyrosine kinase inhibitor; TNBC, triple-negative breast cancer.

We then used clinical follow-up and real-world outcomes from the Tempus database to annotate long-term response, acquired or primary resistance to treatment (figure 1A). Long-term response was defined as documented complete response (CR), partial response (PR) or durable stable disease (SD), with no report of progression throughout a follow-up period lasting at least 6 months. Primary resistance was defined as evidence of progression via progressive disease (PD), report of disease recurrence or metastasis, or presence of a subsequent line of treatment (LoT), with no evidence of initial response. Lastly, acquired resistance was defined as evidence of initial response (CR/PR/SD) followed by evidence of progression^{19 20} (PD, recurrence/metastasis, or subsequent LoT, see Methods for detailed definitions). We implemented these definitions for both ICB- and non-ICB-treated patients to compare and contrast our findings across cancer types and treatment modalities without definition-related biases. In both the ICB and non-ICB cohorts, long-term response was mostly evidenced by report of CR/PR, while primary resistance consisted mainly of PD or report of tumor recurrence or metastasis in the patient EHR (online supplemental figure 1C,D). Similarly, only a minority of acquired resistant patients had durable SD as their initial response or presence of a subsequent LoT as their evidence of progression (online supplemental figure 1E).

Figure 1. Mapping acquired and primary ICB resistance in real-world data. (A) Schematic patient timeline showing different timepoints of the treatment journey, including tumor biopsy collection and outcome events. Based on clinical follow-up and outcomes, patients were categorized as long-term responders, acquired or primary resistant to ICB. (B) Proportion of rwICB long-term responders, acquired, and primary resistant patients across ICB regimens and cancer types. Total number of patients (in brackets) and number of patients in each group are reported. “anti-PDx – chemo” designates anti-PD(L)1 administered as monotherapy, “anti-PDx+chemo” designates anti-PD(L)1 administered concomitantly with chemotherapy, “Other ICB” includes anti-PD1 combined with anti-CTLA-4, anti-PD(L)1 combined with other non-ICB drugs, or any anti-PD(L)1 part of a sequential drug combination. (C) Collapsed timelines of rwICB acquired (top) and primary (bottom) resistant patients showing distribution of pre/post-ICB sample collection, response events (acquired), and progression events (acquired and primary) relative to ICB start in each cancer type. Each point represents a patient (response/progression events) or a sample (pre/post-ICB sample collection). Dashed lines indicate median months from ICB start for that time point. anti-PDx, anti-programmed cell death ligand 1; CTLA-4, cytotoxic T-lymphocyte-associated protein 4; HNC, head and neck squamous cell cancer; ICB, immune checkpoint blockade; NSCLC, non-small cell lung cancer; PD1, programmed cell death protein-1; rwICB, real-world ICB; TNBC, triple-negative breast cancer.

Overall, rwICB reflected expectations of outcome to ICB,^{1 2 4 7} with <15% of patients showing long-term benefit from treatment, >50% experiencing primary resistance, and ∼22–34% experiencing acquired resistance, varying across cancer types and treatment settings (figure 1B, online supplemental figure 1F, online supplemental table 1). Addition of chemotherapy to the ICB regimen improved the proportion of responsive patients overall (figure 1B), consistent with the establishment of chemoimmunotherapy as the primary SoC across many solid tumors.^{2 4} This was also reflected in rwICB by the gradual adoption of combination ICB in real-world practice, becoming the predominant treatment option through the years (online supplemental figure 1G). However, the proportion of patients experiencing acquired resistance was consistently higher for those treated with chemoimmunotherapy (27–40%) as compared to those treated with anti-PD(L)1 monotherapy (7–27%, figure 1B, online supplemental figure 1H, online supplemental table 1).

In rwICB, initial response to ICB occurred mostly within 6 months of treatment start for both long-term responders (>76%) and acquired resistant patients (>87%), while progression was reported within a year of treatment start for most acquired resistant patients (66% in NSCLC, >80% in HNC and TNBC) and within 6 months of treatment start for most primary resistant patients (67% in NSCLC, 79% in HNC, and 72% in TNBC, figure 1C, online supplemental figure 1I, online supplemental table 1). In addition to annotating patient outcomes, we used biopsy collection dates from the database to annotate pre- and post-treatment tumor DNA/RNA sample availability across all cohorts (online supplemental figure 1A, online supplemental figure 1J). Nearly all post-treatment samples came from acquired or primary resistant patients, reflecting absence of tumor tissue in long-term responders due to effective tumor clearance following treatment. Most post-ICB resistant tumors were collected at or after progression for both acquired and primary resistant patients, while the majority of pre-ICB samples were collected up to 6 months before treatment start (figure 1C, online supplemental figure 1I, online supplemental figure 1K). Importantly, pre- and post-treatment samples in these cohorts were typically unmatched (ie, they come from different patients, online supplemental figure 1K).

Clinical features of acquired and primary resistance to ICB

Using the comprehensive clinicogenomic data available in rwICB, we sought to compare the baseline patient profiles of our three defined response groups across treatment modalities. First, we computed and compared the real-world overall survival (rwOS) across response groups. To mitigate survival overestimation inherent to real-world clinicogenomic databases,²¹ we performed risk set adjustment in all survival analyses using tumor sequencing date to account for delayed entries (Methods). We confirmed that rwOS following treatment was longest for long-term responders, followed by acquired then primary resistant patients, across all cancer types and regardless of stage or treatment modality (online supplemental figure 2A,B, online supplemental table 2). We then compared a set of established ICB biomarkers⁴ along with other clinical variables across patient groups (figure 2A, online supplemental table 3). In NSCLC, higher TMB, PD-L1 protein expression, and inferred immune infiltration associated with long-term benefit to ICB, while only higher PD-L1 expression and immune infiltration associated also with acquired versus primary ICB resistance. Interestingly, chemotherapy NSCLC responders also showed higher TMB and PD-L1 expression at baseline (online supplemental table 3), suggesting these biomarkers are not ICB-specific. In HNC and TNBC, there were no clear differences in TMB, PD-L1 expression, or immune infiltration between acquired versus primary ICB resistance. This suggests that currently used response biomarkers cannot generally predict acquired resistance to ICB.

Figure 2. ICB prolongs survival post-acquired resistance. (A) Proportions of ICB long-term responders, acquired and primary resistant patients across a range of baseline (pre-treatment) clinical features in each cancer type. Error bars represent the 95% CI of each proportion as computed using the one-sample proportions test with “prop.test” R function. PD-L1 protein expression was assessed via immunohistochemistry and the readout was calculated using the TPS approach. Immune-high designates an Immune Score²⁴ above the median within each cancer type as calculated using single sample gene set enrichment analysis⁸⁰ on RNA-seq data (Methods), only tumors biopsied from the corresponding primary organ site were used. In HNC only tumors with purity ≥50% were used due to baseline differences (online supplemental figure 2C). HLA class I haplotypes were derived from DNA-seq data and “HLA heterozygous” refers to heterozygosity at all three loci (HLA-A/B/C). Presence of liver/brain/bone mets at baseline was determined using pre-ICB tumor samples from rwICB (online supplemental figure IJ). Asterisks indicate statistical significance (p value ≤0.05) comparing the three patient groups with the χ² test. Additional features, breakdown across (non-)ICB treatment modalities, and statistical comparisons of acquired versus primary resistance are reported in online supplemental table 3. Risk set-adjusted Kaplan-Meier curves of the real-world 5-year post-progression survival probability for acquired and primary ICB-resistant (B) or chemotherapy-resistant (C) patients in each cancer type. The total number of patients used in the analysis is reported in brackets for each group. HRs, corresponding 95% CIs, and likelihood ratio test p values comparing acquired versus primary resistant patients are reported. Number of patients at risk at ten-month intervals is reported for each group. Only patients with complete last known follow-up and censoring information are included in the analysis. Note that the total number of patients is different from the number of patients at risk at T=0 because of delayed entry due to risk set adjustment (Methods). Analyses stratified by (non-)ICB treatment modalities and disease stage are reported in online supplemental table 2. chemo, chemotherapy; ethn., ethnicity; HLA, human leukocyte antigen; HNC, head and neck squamous cell cancer; ICB, immune checkpoint blockade; LoT, line of treatment; mets, metastases; MSI, microsatellite instability; muts/Mb, mutations per mega bases; NSCLC, non-small cell lung cancer; PD-L1, programmed cell death ligand 1; rwICB, real-world ICB; TMB, tumor mutational burden; TNBC, triple-negative breast cancer; TPS, Tumor Proportion Score.

Patient-level characteristics were generally similar across response groups, including germline zygosity of the human leukocyte antigen (HLA, consistent with²²), patient gender, or ethnicity (figure 2A, online supplemental table 3). Older age associated with durable ICB benefit, particularly in HNC, while smoking history and presence of distant metastases pre-ICB associated specifically with acquired resistance in TNBC. Despite ICB long-term responders tending to present with earlier stage disease, there was no difference in disease stage between acquired and primary resistant patients. However, among stage IV patients, those who experienced acquired resistance were more likely to have received ICB as their first ever treatment compared to primary resistant patients, especially in NSCLC as previously reported.⁸

Focusing only on acquired or primary resistant patients, we compared them prognostically after they progressed on ICB by calculating their risk set-adjusted real-world post-progression survival (rwPPS, ie, overall survival calculated from the onset of progression). Despite equally advanced stage, acquired ICB resistant patients showed significantly longer rwPPS than primary resistant patients across the three cancer types (figure 2B). In NSCLC, this observation was regardless of line of ICB administration (online supplemental table 2) and was even more striking in patients treated with anti-PD-(L)1 monotherapy (online supplemental figure 2D), which aligns with a previous report in the same patient setting.⁸ Interestingly, this prognostic difference was not observed in chemotherapy-treated or TKI-treated patients (figure 2C, online supplemental figure 2E), even when stratifying by stage (online supplemental figure 2F), as acquired generally survived as long as their primary counterpart post-progression. These results indicate a universal and ICB-specific prognostic advantage procured by an initial response that, although less effective, may still sustain patient survival after acquiring resistance. This therefore suggests ICB-specific differences at the molecular level between acquired and primary resistant patients.

The acquired resistant immune TME is consistently inflamed post-ICB

Given the consistently prolonged survival in acquired versus primary ICB resistant patients following progression, we hypothesized that differences at the molecular level underpinned these distinct clinical phenotypes. We leveraged the multimodal samples available in rwICB and sought to first compare the transcriptional profiles of post-ICB acquired versus post-ICB primary resistant tumors (figure 3A). We used RNA-seq data of tumors biopsied from the corresponding primary organ site of each cancer type (lung tissue for NSCLC, head/neck tissue for HNC, and breast tissue for TNBC) to prevent introducing variability due to distinct metastatic sites (online supplemental figure 3A). Since RNA-seq was performed with two different assays (online supplemental figure 3B), we accounted for the assay version as a co-variate in the differential expression analysis (DEA) for all cancer types, while also accounting for histology in NSCLC. To infer the immune composition of the tumor microenvironment (TME) and the level of oncogenic signaling across tumor groups, we ran gene set enrichment analysis (GSEA) on the DEA-ranked list of genes using 50 Hallmark gene sets from MSigDB²³ and 19 cancer-specific immune signatures from ConsensusTME²⁴ (Methods, all GSEA results are reported in online supplemental table 4).

Figure 3. ICB maintains or induces an immune-inflamed TME post-acquired resistance. (A) Schematic of the comparative analysis of the transcriptional profiles of acquired versus primary resistant tumors. The transcriptomes of post-ICB acquired and post-ICB primary resistant tumors were compared to infer differential enrichment in immune composition and/or oncogenic signaling. (B) GSEA NES heatmap of differentially enriched gene sets in post-ICB acquired versus primary resistance across the three cancer types. Rows and columns are clustered hierarchically. (C) TMB distribution in post-ICB acquired versus primary resistant tumor samples used in RNA-seq analysis. Distributions are compared within each cancer type using the Wilcoxon rank-sum test. The number of samples in each group is reported in brackets. In NSCLC, two acquired resistant samples and five primary resistant samples had no corresponding TMB as assessed using xTv2, xTv3, or xTv4 DNA-seq assays (Methods). (D) Proportion of post-ICB acquired versus primary resistant tumors with PD-L1 protein expression ≥1%. PD-L1 protein expression was assessed via immunohistochemistry and the readout was calculated using the TPS approach. Proportions are compared within each cancer type using Fisher’s exact test. Number of samples is reported for each group. All post-ICB samples with available PD-L1 data were used. (E) Distribution of tumor purity in post-ICB acquired versus primary resistant samples used in RNA-seq analysis. Distributions are compared within each cancer type using Wilcoxon rank-sum test. Number of samples in each group is reported in brackets. (F) Scatter plots of differentially enriched gene sets in pre-ICB or post-ICB acquired versus primary resistance across the three cancer types. For each gene set, the NES-signed GSEA log₁₀ FDR for the pre-ICB and post-ICB comparison is reported on the x and y axes respectively, with positive values indicating enrichment in pre/post acquired, and negative values indicating enrichment in pre/post primary. Gene sets are colored by functional category. ±Log₁₀ FDR density curves for immune and non-immune gene sets are shown at the top for pre-ICB comparisons and on the right for post-ICB comparisons. Diagonal dashed lines indicate the identity line. Horizontal and vertical dashed lines indicate significance at FDR=10%. (G) Paired comparison of the significance level (log₁₀ FDR) for immune-related enrichment in acquired versus primary resistance post-ICB versus post-chemo. Each point represents an immune gene set from MsigDB Hallmark²³ or ConsensusTME²⁴ (26 total gene sets). Points are colored by their NES. Positive log₁₀ FDR values indicate enrichment in post acquired versus post primary resistance. Dashed lines indicate significance level at FDR=10%. (H) Log-TPM distribution of representative immune-related genes in post-ICB acquired versus primary resistant tumors across the three cancer types. Number of samples in each group is reported in brackets. Distributions are compared for each gene within each cancer type using Wilcoxon rank-sum test. (I) GSEA NES barplots of differentially enriched literature-compiled gene sets in post-ICB acquired versus primary resistance across the three cancer types. Positive NES values indicate enrichment in post-ICB acquired, negative NES values indicate enrichment in post-ICB primary. Gene sets are grouped by functional category. (J) Post-ICB acquired versus primary DESeq2 log2-FC heatmap of genes from the immunotherapeutic gene set⁸ across the three cancer types. Columns are clustered hierarchically. FDR was computed using Benjamini-Hochberg approach in all analyses. For B, F, and I, only gene sets differentially enriched in at least one comparison in any cancer type are shown. See online supplemental table 4 for gene set abbreviations and all GSEA results. DEA, differential expression analysis; FC, fold change; FDR, false discovery rate; GSEA, gene set enrichment analysis; HNC, head and neck squamous cell cancer; ICB, immune checkpoint blockade; ICI, immune checkpoint inhibitor; muts/Mb, mutations per mega bases; NES, normalized enrichment score; NSCLC, non-small cell lung cancer; PD-L1, programmed cell death ligand 1; TMB, tumor mutational burden; TME, tumor microenvironment; TNBC, triple-negative breast cancer; TPM, transcripts per million; TPS, Tumor Proportion Score.

Post-ICB, acquired resistant tumors showed a remarkably consistent immune-inflamed TME compared to primary resistant tumors across all cancer types (figure 3B). This immune-inflamed profile was characterized by inferred infiltration of different lymphocytes and myeloid cells, as well as active IFNγ signaling and inflammatory responses (figure 3B). We checked that these results were not confounded by differences in TMB (figure 3C), PD-L1 protein expression (figure 3D), tumor purity (figure 3E), or overall stromal content as inferred from RNA-seq data (online supplemental figure 3C). We also checked that timing of tumor biopsy with respect to both ICB start and ICB progression did not correlate with overall immune infiltration in either acquired or primary resistant tumors (online supplemental figures 3D,E).

In NSCLC, post-ICB immune differences between acquired and primary resistance were consistent across treatment settings, although more prevalent in the anti-PD(L)1 monotherapy or in the ≥2^nd line advanced-stage settings (online supplemental table 4). Since squamous and non-squamous NSCLCs showed markedly distinct transcriptional profiles (online supplemental figure 3F), we stratified DEA and GSEA by histology to understand if immune profiles post-ICB resistance were histology-dependent or widespread. Again, acquired resistant NSCLCs showed an immune-inflamed profile versus primary resistant tumors post-ICB, regardless of histology (online supplemental figure 3G). Similarly, the acquired resistance immunophenotype was consistent in low-purity versus high-purity NSCLC samples (online supplemental figure 3H), further emphasizing its biological relevance.

To evaluate the post-ICB specificity of the immune-inflamed acquired resistant TME, and since most post-treatment samples do not have a matched pre-treatment sample (online supplemental figure 1L), we contrasted post-ICB GSEA results with pre-ICB comparisons. While pre-ICB acquired resistant NSCLCs and HNCs were also more immune-inflamed than pre-ICB primary resistant tumors, immune-related differences were exacerbated in, or sometimes exclusive to, the post-ICB setting across all cancer types (figure 3F). TNBC stands as a typical case, where no transcriptional differences were found between acquired and primary resistant tumors pre-ICB, yet a substantially different TME was observed post-ICB (figure 3F). Despite coming from different patients, we also compared pre- versus post-ICB tumors from acquired and primary resistant patients respectively. In line with our previous analysis of paired pre/post-ICB acquired resistant NSCLCs,⁸ we found stable or increased immune inflammation in acquired resistant tumors from pre- to post-ICB, including increased inflammatory response and tumor necrosis factor alpha signaling in NSCLC (online supplemental figure 3I). Contrastingly, primary resistant tumors consistently showed vigorous immune depletion from pre- to post-ICB, further supporting a fundamentally different post-ICB TME in the two resistance phenotypes. Moreover, no immune differences were found in post-chemotherapy acquired versus primary resistant HNCs or TNBCs, while post-chemotherapy (but not post-TKI) acquired resistant NSCLCs were moderately more inflamed than primary resistant tumors (figure 3G, online supplemental table 4). Taken together, these results indicate that post-ICB immunophenotypic differences between acquired and primary resistance are not merely a reflection of pre-existing baseline differences, but are rather specifically induced by ICB.

Different genes driving these immune differences were upregulated in post-ICB acquired versus primary resistant tumors across the three cancer types (figure 3H, online supplemental table 4). These include genes encoding components of immune inflammatory processes (OSM, CD70, IL16, CCL17), IFN signaling (STAT2/4, IL18BP, IRF4, IFNG, JAK3, ISG20, MX1, RIPK1), or immune cell markers (CD244, ITGAX, IL10RA, GPR15). To further characterize the post-ICB acquired resistant immunophenotype, we ran GSEA on a complementary list of immuno-oncology-focused gene sets we compiled from the literature.825,28 Gene sets describing immunotherapy targets, alternative immune checkpoints, pro-tumor cytokines, and antigen processing and presentation components were all enriched in post-ICB acquired versus primary resistant NSCLCs, with similar trends in HNC and TNBC (figure 3I–J). Cancer-related processes, including cell proliferation and Myc signaling, were on the other hand mainly enriched in primary resistant tumors, particularly in HNC and TNBC (figure 3I).

Altogether, our transcriptional profiling of post-ICB resistant tumors revealed a universally immune-inflamed TME in the acquired resistance setting across the three cancer types studied, suggestive of an ongoing, albeit impaired, ICB-induced immune response.

ICB-specific selection of mutations in the post-acquired resistance setting

The distinctive immunophenotype of acquired resistant tumors post-ICB prompted us to probe their genetic landscape for the presence of ICB-specific cancer alterations that may confer resistance to treatment. To this end, we leveraged panel DNA-seq data covering 592 cancer-related genes and used dndscv²⁹ to quantify the level of selective pressure acting on gene mutations from pre- to post-treatment. Briefly, dndscv identifies genes undergoing positive selection by quantifying the ratio of non-synonymous to synonymous mutations (dN/dS) using local (same gene, cross-samples) and global (cross-genes) information to estimate the background mutation rate. This approach presents many advantages, including sensitivity to low-frequency events and robustness against potential confounding factors in cancer genomics data, such as mutational signature biases and variability in tumor purity and/or TMB.²⁹

We used cross-organ site samples to run dndscv in each of pre-/post-treatment acquired/primary resistant tumor groups separately within each cancer type and for each of ICB, chemotherapy, and TKI treatment modalities (figure 4A, online supplemental figure 4A, full output of all dndscv runs is reported in online supplemental table 5). To exclude biases, we checked sample distribution by sequencing panel, NSCLC histology, organ site, and availability of matched tumor/normal sample (germline control) for DNA-seq (online supplemental figures 4B–E). We also checked that the global dN/dS estimates in the mutation datasets were ≥1 (online supplemental figure 4F), as expected.²⁹

Figure 4. ICB selects for mutations specific to the acquired resistance setting. (A) Schematic of dndscv analysis. For each cancer type and treatment modality (ICB, chemo, TKI), dndscv was run separately on short variant data of pre-/post-treatment acquired/primary resistant tumors to identify mutations under significant positive selection (dN/dS >1 and qval ≤10%). Observed dN/dS estimates of positively selected mutations in post-ICB acquired resistance were further compared against an expected distribution of dN/dS estimates generated by randomly shuffling pre/post acquired/primary tumor group labels within each cancer type of the ICB cohort. (B) Scatter plot of pre-ICB (x-axis) and post-ICB (y-axis) dN/dS estimates of all positively selected post-ICB acquired (upper panels) and primary (lower panels) resistance mutations across the three cancer types. Darker color shade indicates mutations identified post-ICB only, while lighter shade indicates mutations identified also pre-ICB (ie, both pre-ICB and post-ICB). Dot shape indicates mutation type (missense, non-sense, indels). Dashed lines represent the identity line (ie, pre-ICB dN/dS=post-ICB dN/dS). Stacked barplots showing the recurrence of positively selected post-ICB acquired resistance mutations in acquired and primary resistant tumors pre-/post-ICB or chemo (C) or pre-/post-TKI (D) across the three cancer types. Recurrence of both non-synonymous (darker color shade) and synonymous (lighter color shade) mutations is reported. Asterisks indicate mutations positively selected in that tumor group. (E) Distribution of the difference between the observed and expected dN/dS estimates (ΔdN/dS) for post-ICB acquired resistance mutations positively selected both pre-ICB and post-ICB (light orange) or post-ICB only (dark orange) in each cancer type of the ICB cohort. Each dot represents a gene mutation and each gene mutation appears 10,000 times, each time corresponding to one iteration of the random shuffling process performed to get an expected dN/dS estimate (Methods). ΔdN/dS values >0 indicate higher-than-expected dN/dS estimates. For C and D, the total number of samples in each tumor group (ie, denominator) is reported in online supplemental figure 4A. “Non-synonymous” refers to any non-synonymous substitution or indel. For D, PIK3CA is shown as a TKI positive control alongside post-ICB acquired mutations. alt, alternative allele; chemo, chemotherapy; Chr, chromosome; dN/dS, ratio of non-synonymous to synonymous mutations; HNC, head and neck squamous cell cancer; ICB, immune checkpoint blockade; NSCLC, non-small cell lung cancer; pos, position; ref, reference allele; smpl, sample; TKI, tyrosine kinase inhibitor; TNBC, triple-negative breast cancer.

To understand selection dynamics before and after treatment, we contrasted dN/dS estimates of mutations positively selected post-ICB with corresponding pre-ICB estimates for both acquired and primary resistant tumors (figure 4B). As expected, mutations selected for post-ICB included cancer-specific drivers that were also selected pre-ICB and in both acquired and primary resistance (figure 4B). This included STK11 in NSCLC, for which we confirmed stronger and specific pre-ICB selection in primary resistant tumors as compared to ICB-sensitive tumors³⁰ (online supplemental figure 4G). On the other hand, a subset of post-ICB mutations was identified in the post-ICB analysis but was not detected pre-ICB (ie, post-only mutations). This was the case for all cancer types and even in post-primary resistance (except primary resistant TNBCs), indicating that ICB alters the genetic make-up of tumors regardless of resistance phenotype (figure 4B).

In acquired resistant NSCLCs, we confirmed selection for B2M loss-of-function (LoF) mutations post-, but not pre-, ICB (figure 4B), as previously described.^{8 12 13} Interestingly, primary resistant NSCLCs and HNCs also showed selection for B2M mutations post-ICB, suggesting systematic immune selection affecting APM after treatment. We also identified mutations which have not been previously reported in the context of acquired resistance to ICB in NSCLC, including selection for TGFBR2 (TGF-β signaling), ERRFI1 (EGFR signaling), ARID1B (chromatin remodeling) and NFKBIA (NF-κB signaling) mutations post-ICB (figure 4B). Likewise, we found evidence for NOTCH1, FAT1, PTEN, and CYLD LoF mutations in acquired resistant HNCs and RUNX1 LoF and MAF missense mutations in acquired resistant TNBCs.

These post-only acquired ICB mutations were more recurrent post-ICB acquired resistance than both pre-ICB acquired resistance and, except FAT1 in HNC, post-ICB primary resistance (figure 4C). Further consolidating their involvement in ICB resistance, more than half of acquired ICB mutations were positively selected pre-ICB primary resistance but not in pre-ICB long-term responders, with the exception of B2M in NSCLC,³¹ (online supplemental figure 4H). Moreover, with the exception of PTEN and NOTCH1 in HNC, these acquired ICB mutations were not selected for post-chemotherapy acquired resistance (figure 4C), indicating treatment-specific selection. This was particularly striking for CYLD in HNC and RUNX1 in TNBC, despite these two cancer types being more statistically powered in the chemotherapy cohort (online supplemental figure 4A). Similarly, while we confirmed selection for PIK3CA mutations following acquired TKI resistance in NSCLC,^{32 33} none of the acquired ICB mutations were selected for post-TKI acquired resistance (figure 4D, online supplemental figure 4I).

To further evaluate the extent to which these mutations are specific to acquired ICB resistance, we compared their observed dN/dS estimates to an expected distribution of dN/dS estimates (figure 4A, Methods). Across all cancer types, while the observed post-ICB dN/dS estimates were as expected for mutations identified both pre- and post-acquired ICB resistance (eg, TP53 mutations), post-only acquired mutations had higher-than-expected dN/dS estimates overall (figure 4E) and at the single-gene level (online supplemental figure 4J). Taken together, these results emphasize the ICB-specificity of the identified acquired resistance mutations and suggest a role for these mutations in driving ICB resistance after initial clinical benefit.

Acquired ICB alterations associate with immune-escaped TMEs

The interplay between cancer alterations and the immune TME plays a crucial role in determining response to ICB.34,37 We next sought to investigate whether the observed selection of mutations post-ICB in the acquired resistance setting could be linked to ICB-induced immune escape. To address this, we characterized the phenotype of these mutations by comparing the transcriptional profiles of altered (ALT) versus wild-type (WT) post-ICB acquired resistant tumors. We first expanded the repertoire of genomic alterations to include copy number variants (CNVs) and focused on mutations with a potential detrimental impact on the encoded protein (Methods). After expanding to CNVs, the recurrence of acquired ICB alterations remained highest in post-ICB acquired resistant tumors, relative to both pre-ICB acquired resistance and non-ICB regimens (figure 5A, online supplemental figure 5A). Collectively, the identified acquired ICB alterations affected 25–55% of acquired resistant tumors post-ICB, with <20% of altered samples carrying more than one acquired ICB alteration in the corresponding cancer type (figure 5B). Consistent with overall cohort composition (table 1), most of the affected tumors came from stage IV patients treated in either the first (∼46%) or 2^nd+ line (∼54%) setting with chemoimmunotherapy in the TNBC cohort or mostly ICB monotherapy or other ICB regimens in the NSCLC and HNC cohorts (figure 5B, online supplemental figure 5B).

Figure 5. Acquired ICB alterations predict contrasted but escaped immune TMEs. (A) Stacked barplots showing the recurrence of acquired ICB alterations in acquired resistant tumors pre-ICB and post-ICB. Bars are annotated by alteration type for each gene. For NFKBIA, one sample carries multiple missense mutations and one truncating mutation and both mutation types are counted once in the barplot. The total number of samples in each tumor group (ie, denominator) is reported in online supplemental figure 4A. (B) Chord diagram of post-ICB acquired resistant tumors carrying acquired ICB alterations across the three cancer types. The diagram is split by cancer type and the proportion of post-ICB acquired resistant tumors carrying acquired ICB alterations in that cancer type is reported in brackets. Each fragment represents a sample carrying the labeled acquired ICB alteration and each arc links alterations co-occurring in the same sample. The alteration type, organ site, histology, disease stage, type of ICB received, line of treatment, PD-L1 and TMB status are annotated for each sample. For NFKBIA, the sample labeled with missense mutation carries multiple missense mutations as well as a truncating mutation; missense is labeled as it is the mutation type driving the dN/dS signal. TMB-high refers to tumors with TMB ≥10 mutations/Mbp. (C) GSEA NES heatmap of differentially enriched gene sets in post-ICB acquired resistant tumors carrying acquired ICB alterations (ALT) versus WT. GSEA was performed separately for each acquired ICB alteration, focusing only on post-ICB acquired resistant tumors in the corresponding cancer type. Bubbles are proportional to -log₁₀ FDR and shown only for significantly enriched gene sets (FDR ≤10%). FDR was computed using the Benjamini-Hochberg approach. Only gene sets differentially enriched in at least one comparison in any cancer type are shown. See online supplemental table 6 for all GSEA results, gene set abbreviations, and number of samples used in each comparison. (D) Distributions of ssGSEA ES of differentially enriched gene sets in ALT versus WT post-ICB acquired resistant tumors across the three cancer types. Distributions were compared between ALT versus WT samples using the Wilcoxon rank-sum test. The ES scale is divided by a factor of 1,000. Dashed lines indicate median ES for ALT and WT groups. The specific gene alteration is labeled for each altered sample. For NFKBIA, only non-squamous NSCLC samples were used as they represent the histology underlying the GSEA enrichment. Number of ALT/WT samples is reported in brackets. ALT, altered; anti-PDx, anti-programmed cell death ligand 1; chemo, chemotherapy; CN, copy number; ES, enrichment score; FDR, false discovery rate; GSEA, gene set enrichment analysis; HNC, head and neck squamous cell cancer; ICB, immune checkpoint blockade; LoT, line of treatment; mono, monotherapy; NA, not applicable/available; NES, normalized enrichment score; NSCLC, non-small cell lung cancer; nsq, non-squamous; PD-L1, programmed cell death ligand 1; ssGSEA, single sample gene set enrichment analysis; TMB, tumor mutational burden; TME, tumor microenvironment; TNBC, triple-negative breast cancer; WT, wild-type.

Comparing ALT versus WT tumors for each acquired ICB alteration, we found two main patterns of tumor-immune associations (figure 5C, GSEA results and number of samples used are reported in online supplemental table 6): one group of acquired ICB alterations that associated with increased immune inflammation/infiltration (ARID1B, B2M, TGFBR2, RUNX1, MAF) and another group that associated with immune depletion and/or active oncogenic signaling (ERRFI1, NFKBIA, NOTCH1, FAT1, CYLD, PTEN). Importantly, these contrasted phenotypes coexisted within the same cancer type, highlighting the heterogeneity of molecular pathways linked to acquired resistance to ICB within and across cancers. For instance, in NSCLC, B2M LoF associated with both decreased B2M expression and inferred infiltration of cytotoxic cells (figure 5C, online supplemental figure 5C), consistent with loss of APM due to immune selection.38,41 On the other hand, NFKBIA alterations, driven mainly by amplifications, associated with widespread immune depletion (figure 5C). In HNC, tumors carrying FAT1, NOTCH1, or PTEN alterations showed increased proliferation (E2F/G2M/Myc), active oncogenic signaling (mTORC1/WNT), and elevated metabolism (glycolysis and oxidative phosphorylation (OXPHOS)), while CYLD LoF associated with predicted infiltration of immunoregulatory dendritic cells (mregDCs)²⁷ and active epithelial-mesenchymal transition (EMT). Similarly, RUNX1-LoF TNBCs showed increased infiltration of dendritic cells (DCs) and M2 macrophages as well as active EMT (figure 5C).

Despite low alteration frequency and tumor sample heterogeneity, these phenotypes were generally consistent at the single-sample level (figure 5D, online supplemental figures 5D,E). This includes overall immune depletion in NFKBIA-altered non-squamous NSCLCs, contrasting with increased IFN signaling in TGFBR2-LoF tumors (figure 5D, online supplemental figure 5D). Likewise, FAT1, NOTCH1- or PTEN-LoF HNCs showed increased Myc, OXPHOS, glycolysis, and mTORC1 signaling, while CYLD-LoF tumors showed increased mregDCs and active EMT, and RUNX1-LoF TNBCs showed enrichment for M2 macrophages (figure 5D, online supplemental figure 5E).

Taken together, our findings suggest that ICB-induced immune responses select for acquired alterations that confer resistance to immune killing (eg, B2M/TGFBR2), promote immune exclusion (eg, NFKBIA/RUNX1), or enhance cancer cell-intrinsic phenotypes (eg, ERRFI1/FAT1) to support cancer progression after initial clinical benefit.

Discussion

More than a decade after the first approvals of ICB to treat advanced-stage cancers, tumor and immune characteristics that govern the emergence of acquired resistance remain incompletely understood. In this study, we aimed to address this gap by systematically analyzing patient and tumor profiles at baseline and at ICB progression, leveraging multimodal (clinical, DNA, and RNA) and large-scale RWD. Using deep clinical and molecular metadata, we built rwICB, a real-world and multimodal cohort of >3,500 ICB-treated patients with >2,600 pre-/post-ICB tumor samples across NSCLC, HNC and TNBC. To the best of our knowledge, rwICB constitutes the largest dataset of post-ICB resistant NSCLCs, and the first multimodal dataset of post-ICB resistant HNCs and TNBCs. Analyzing rwICB through a systematic and unbiased approach, we identified both common and cancer-specific patterns of acquired resistance, including extended survival following progression, consistent immune inflammation post-ICB, and selection for distinct acquired alterations within and across cancers.

Acquired resistance to ICB is a frequent phenomenon, estimated to occur in up to ∼70% of initial responders.³ Yet, to this date, no standard definition of this clinical phenotype is agreed on.^{3 7} While initiatives to reach a consensus definition in specific indications are ongoing,^{7 19 20 42} our use of RWD and consideration of distinct cancer types across distinct treatment settings represent additional layers of complexity. Aware of these challenges, we implemented definitions that reconcile clinical precision, cohort size needs, and RWD-inherent issues, including data missingness, incompleteness, and inconsistency.

A key asset of rwICB is the availability of both rw-outcomes and long-term clinical follow-up that retrospectively allows the distinction between long-term response and acquired resistance, while also capturing primary resistance. Therefore, we used rw-outcomes conjointly with clinical follow-up to define these three response groups in rwICB. While reflecting overall expectations of outcome to ICB, including higher response rates in chemoimmunotherapy-treated patients, the proportion of acquired resistant patients in rwICB was above previous estimations (22–34% vs ∼5–25%, overall).³ Besides longer patient follow-up, several factors can explain these differences. First, unlike more stringent definitions that consider only objective CR/PR as evidence of initial response,^{3 19} we considered documented rwSD of at least 6 months as a form of response, yielding ∼20% of SD patients in our acquired ICB resistant cohorts. Second, the proportion of acquired resistant patients varied substantially across treatment and disease settings. Notably, 27–40% of chemoimmunotherapy-treated patients experienced acquired resistance, as compared to 7–27% for those treated with anti-PD(L)1 monotherapy. The latter range aligns more with estimations provided in,³ as these were focused only on ICB monotherapy. Accordingly, the acquired resistance rate (ie, the proportion of acquired resistant patients over all responders) at 12 months ranged between 54-58% for chemoimmunotherapy and 33-48% for ICB monotherapy in the advanced-stage setting. Therefore, our findings indicate that the incidence of acquired resistance to ICB may have been so far underestimated, emphasizing the importance of treatment and disease setting when estimating the likelihood of developing acquired resistance later on.

Despite variability in incidence, acquired resistant patients consistently showed extended survival compared to primary resistant patients following ICB progression. This was the case for all cancer types and regardless of treatment or disease settings, while not observed in non-ICB modalities. We previously reported the same prognostic difference in ICB-treated patients with NSCLC.⁸ Here, we show that this pattern is universal and ICB-specific, indicating that an initial response to ICB is always beneficial, as its effect persists even after the onset of resistance.

Leveraging post-ICB bulk RNA-seq data and immune deconvolution, we found that acquired resistant tumors were consistently more immune-inflamed than primary resistant tumors. This immune-inflamed profile was either comparable to pre-ICB levels or exacerbated post-ICB, suggesting a sustained or enhanced immune response at the time of acquired resistance. Of note, alternative immune checkpoints, including LAG3,⁴³ were also upregulated post-ICB acquired resistance. This is in line with previous reports16,18 and suggests that alternative checkpoint inhibitors,⁴⁴ including bispecific antibodies,^{45 46} could reinstate an effective anti-tumor response in a subset of patients who acquired resistance to PD-(L)1/CTLA-4 inhibitors.

While the inflamed immunophenotype of acquired resistant tumors post-ICB aligns with our previous observations in a smaller cohort of matched pre/post-ICB NSCLCs,⁸ we note that Ricciuti and colleagues¹³ reported a significant decrease in T cell populations in acquired resistant NSCLCs post-ICB. These findings are not necessarily in contradiction with ours for several reasons. First, our use of bulk RNA-seq and the unavailability of matched pre/post-ICB samples limit our ability to detect specific immune cell changes at the single patient resolution. Second, immune TMEs can still be inflamed but T cell-excluded.^{37 47} This is partly what we are observing with enriched inflammatory responses and myeloid cell signatures in post-ICB versus pre-ICB acquired resistance, but no significant changes in predicted T cell composition. With our larger cohort size, this lack of evidence in T cell changes is likely confounded by a mix of acquired resistant patients with either increase or decrease in T cell infiltration post-ICB, as recently reported in NSCLC.¹⁸ In any case, these independent observations highlight the widespread interpatient immune heterogeneity in the context of acquired resistance to ICB.

Unlike acquired resistant tumors, primary resistant tumors showed significant immune depletion from pre- to post-ICB, supporting a radically distinct immunophenotype at the onset of acquired versus primary ICB resistance. This is also in radical contrast with the systematic immunogenic effect of chemotherapy,^{48 49} which we also observe in our data regardless of resistance phenotype. Therefore, our findings provide evidence for a highly conserved and ICB-specific immunophenotype in the context of acquired resistance.

Using dN/dS to assess causal selection from pre-treatment to post-treatment, we identified both known and novel drivers of acquired resistance. We confirmed selection for B2M LoF as a hallmark of acquired resistance to ICB in NSCLC,^{8 12 13} while also confirming selection for PIK3CA mutations post-TKI acquired resistance.^{32 33} Thus, these positive controls corroborate our cohort annotations and provide a solid foundation for the interpretation of our novel findings.

Post-ICB acquired resistance, we identified mutations that were selected for ubiquitously (ie, both pre-ICB and post-ICB), as well as mutations that were exclusive to the post-ICB setting (ie, post-only). These post-only mutations may have resulted from two non-mutually exclusive evolutionary scenarios, mirroring the framework we previously proposed in the context of acquired ICB resistance in NSCLC.⁸ First, they may represent pre-existing subclonal mutations that were present prior to treatment at variant allele frequencies below the detection limit of bulk panel DNA-seq. These rare clones may have later gained a selective advantage following ICB, leading to clonal expansion and subsequent detectability at progression. In the second scenario, a subset of post-only mutations may have arisen de novo during treatment as a combined consequence of ongoing mutagenesis and treatment-induced selective pressures that favor newly acquired resistance-conferring alterations.

All post-only acquired alterations were cancer-specific, and >80% (9/11) were not selected for post-chemo/TKI, indicating high context specificity. Further corroborating their involvement in resistance, these alterations affected distinct hallmarks of cancer, from dampened antitumor immunity to uncontrolled proliferation and dysregulated metabolism.⁵⁰ In NSCLC, TGFBR2 LoF associated with increased IFN signaling and immune infiltration, suggesting immune selection for the immunomodulatory role of TGF-β signaling.^{51 52} This is consistent with a CRISPR screen performed on immunocompetent lung cancer mouse models where loss of Tgfbr2 was significantly selected for in vivo versus in vitro.⁴¹ Interestingly, while Tgfbr2-knockout tumor cells lost sensitivity to TGF-β, the surrounding immune and stromal cells remained responsive to TGF-β and its immunosuppressive effects, likely exacerbated by its increased bioavailability in the TME.⁴¹ Taken together, these findings suggest targeting of stromal TGF-β as a potential therapeutic avenue post-ICB acquired resistance in TGFBR2-altered NSCLC.⁵² On the other hand, amplification of NFKBIA associated with widespread immune depletion, suggesting immune exclusion via NF-κB pathway alterations.⁵³ Intriguingly, NFKBIA encodes IκBα, a member of the IkappaB (IκB) protein family that is both a transcriptional target and an inhibitor of NF-κB signaling.⁵⁴ This suggests that increased dosage of NFKBIA in tumor cells disrupts this negative feedback loop, leading to aberrant NF-κB activity.⁵⁴ Thus, existing or future strategies targeting IκBα directly or through its regulators (eg, IκB kinase inhibitors)^{53 54} may reverse the immunosuppressed phenotype of NFKBIA-altered NSCLCs.

While not as immune-excluded as NFKBIA-altered tumors, ARID1B-LoF NSCLCs showed increased inflammation together with decreased overall immune infiltration and cell proliferation. This aligns with a CRISPR screen where loss of chromatin remodeling components, including ARID1B, showed no phenotype in vitro but was selected for in the presence of an adaptive immune system in vivo.³⁹ Conversely, ERRFI1-LoF NSCLCs showed a hyperproliferative phenotype. ERBB receptor feedback inhibitor 1, or ERRFI1, encodes MIG6, a negative regulator of EGFR signaling.^{55 56} In genetically engineered mouse models (GEMMs), loss of MIG6 accelerated both the formation and the progression of lung tumors,⁵⁷ thus supporting the highly proliferative phenotype we are observing in the context of acquired resistance to ICB. Of note, evidence for both chromatin remodeling and EGFR pathway alterations was previously found in the context of acquired ICB resistance in NSCLC, though via SMARCA4 and ERBB2 alterations, respectively.¹³ Therefore, our findings support the positioning of distinct tumor-targeting strategies, including chromatin remodeling therapeutics,⁵⁸ in a subset of patients with NSCLC who acquired resistance to ICB.

In HNC, besides active immunomodulatory oncogenic signaling, there was also strong evidence for a hypermetabolic phenotype in NOTCH1-LoF, FAT1-LoF, or PTEN-LoF tumors, with increased OXPHOS and glycolysis. Metabolic reprogramming in cancer cells can impact the tumor-immune cross-talk⁵⁹ and has been implicated in acquired ICB resistance in melanoma⁶⁰ and NSCLC.⁸ Therefore, therapeutic strategies targeting altered metabolism, including glycolytic and OXPHOS inhibitors,⁶¹ may reverse this phenotype and restore effective antitumor immunity in NOTCH1/FAT1/PTEN-driven acquired resistant HNCs. While lacking a hypermetabolic phenotype, CYLD-LoF HNCs showed enrichment for mregDCs²⁷ and active EMT. Cylindromatosis, or CYLD, is a HNC tumor suppressor that negatively regulates NF-κB signaling^{62 63} and inhibits EMT-mediated invasiveness.^{64 65} Moreover, loss of CYLD was found positively selected in the context of an adaptive immune response in an in vivo CRISPR screen, but showed no phenotype in vitro,³⁹ suggesting that the tumorigenic effect of CYLD LoF is likely due to its immunomodulatory role. Similarly, RUNX1-LoF TNBCs showed increased infiltration of M2 macrophages and active EMT. The association between RUNX1 alterations and macrophage infiltration in breast cancer was previously reported in the untreated setting.⁶⁶ Therefore, our findings point towards an involvement of the myeloid compartment in driving acquired resistance to ICB via cancer driver gene alterations. With an increasing number of clinical trials testing the use of myeloid-targeting immunotherapies,67,69 these alterations can inform their positioning in post-ICB acquired resistance and enrich for patients most susceptible to benefit from them.

Overall, our phenotypic characterization of acquired ICB alterations nuances the generally immune-inflamed profile of acquired versus primary resistant tumors and supports the co-existence of distinct immunophenotypes driven by specific cancer alterations. This indicates that genetic alterations are a major driver of heterogeneity in acquired resistance to ICB. Thus, each identified alteration can serve as a biomarker to monitor the emergence of resistance and stratify patients for personalized interventions.

We note that, with the exception of TNBC, primary resistant tumors also showed selection for post-only alterations, raising the question as to why ICB induces further immune depletion and genetic remodeling post-primary resistance. A hypothesis could be that, although some determinants of resistance were already present pre-treatment (eg, low TMB, low immune infiltration, and/or resistance mutations), additional escape mechanisms, including enhanced immunomodulatory oncogenic signaling,^{34 36} may be required to sustain tumor progression with increasing ICB-induced immune pressure. Therefore, therapeutic strategies aimed at rechallenging the immune system may not benefit cases with primary ICB resistance, who may instead benefit from therapies targeting diverse oncogenic signaling pathways, including Myc signaling.⁷⁰

Despite its systematic and comprehensive nature, our study presents several limitations, some of which were previously discussed. First, the unavailability of matched pre- and post-treatment tumor biopsies complicates the interpretation of our findings at the single patient level. Nevertheless, we addressed this limitation analytically by contrasting all our findings with (1) pre-ICB acquired versus primary resistance, (2) post-ICB primary resistance, and (3) non-ICB regimens. In any case, both translational and reverse translational studies leveraging longitudinal biopsies, GEMMs, patient-derived organoids, and/or humanized mouse models⁷¹ are needed to validate our findings and provide mechanistic insights into the immunomodulatory role of each identified driver. In particular, since most of our identified acquired resistance drivers undergo LoF alterations, coupling previously developed acquired ICB resistance mouse models^{8 72} with in vivo CRISPR technologies^{31 39 41} holds great potential in future follow-up studies. Second, the use of targeted DNA-seq limits the scope of identifiable drivers of resistance. While the depth of coverage (>500×) in our samples likely contributed to the identification of rare events within the covered genes, future studies using whole genome/exome sequencing will complement the repertoire of alterations underlying ICB resistance. Third, the TME features of acquired ICB resistance that we describe are inferred from deconvolution of bulk RNA-seq data, with all the caveats that this approach may entail.⁷³ The consistency of the acquired resistance immunophenotype across indications and its ICB specificity emphasize its biological relevance. Still, future studies leveraging single-cell/spatial transcriptomics and imaging modalities are warranted to provide a more refined landscape of the acquired resistant TME, particularly in the two less studied indications—HNC and TNBC. Finally, clinically refined resistance definitions accounting for oligoprogressive patterns, localized versus spread progression, pseudoprogression, or treatment discontinuation due to toxicity^{3 7} were precluded by the limited granularity of clinical annotations in rwICB. While our definitions effectively capture the bulk of each response group, future investigations could address whether distinct molecular features define these refined clinical subgroups.

Altogether, we provide a comprehensive overview of universal and cancer-specific ICB resistance patterns, generating clinically meaningful and actionable hypotheses to potentially guide treatment strategies post-progression and optimize outcomes. With an increasing awareness of the potential behind RWD⁷⁴ and the establishment of ICB as part of routine care, we anticipate that our understanding of primary and acquired ICB resistance will rapidly expand, ultimately benefiting a larger portion of unresponsive or relapsing patients.

Methods

Patient cohort building and response annotation

A sample of de-identified records from patients with cancer aged 18 years old or above with early-stage or advanced-stage disease was selected from the multimodal (clinical and DNA-seq/RNA-seq data) US-based Tempus AI real-world clinicogenomic database. Patients were enrolled in diverse oncology practices across the USA, including both academic medical centers and community practices. Clinical data (patient and disease characteristics, treatment journey, and response/progression outcomes) were abstracted from structured and unstructured longitudinal physician progress notes, pathology reports, and radiology reports, using Tempus AI’s proprietary abstraction application following data ingestion and cleaning.

Starting from a total of 22,952 selected patients across ten cancer types (NSCLC, breast cancer, head and neck cancer, bladder cancer, gastric cancer, small cell lung cancer, colorectal cancer, biliary tract/hepatocellular carcinoma, prostate cancer or ovarian cancer) and with follow-up until August 2023, we focused on those who:

• Received at least one LoT with complete treatment information:

  • Care plan database identifier.

  • Name and type of medication(s) received.

  • Treatment start date;

• Received at least one line of ICB-containing (anti-PD-1, anti-PD-L1, or anti-CTLA-4) regimen;

• Had at least one treatment-specific real-world response event (CR/PR/SD) and/or progression event (PD or report of recurrence/metastasis) for the first line of ICB received, and these events must have been:

  • Reported >14 days after treatment start to ensure minimum drug exposure, and

  • Reported before any subsequent LoT.

• In case of presence of response event(s) only, then the patient must have been followed up for at least 6 months (180 days) after treatment start.

No filtering on disease stage or on overall patient health profile was applied. These filtering steps yielded 4,394 ICB-treated patients across the ten cancer types, out of which the top three largest cancer types (NSCLC, HNC and breast cancer) represented >85% of the whole cohort (the next largest cancer type was bladder cancer (n=175), representing <4% of the cohort). We therefore considered these three cancer types for subsequent clinical and molecular analyses. For breast cancer, we further focused on TNBC, the subtype where ICB is indicated, to mitigate subtype-related biases in subsequent analyses. This yielded 2,689 patients with NSCLC, 516 patients with HNC, and 354 patients with TNBC, for a total of 3,559 patients in the ICB cohort.

In parallel to the ICB cohort, we built a non-overlapping cohort of non-ICB-treated patients across the same three cancer types. The non-ICB cohort consisted of the rest of the patients who had no record of ICB-containing regimen in their treatment journey and who had received either chemotherapy (for all three cancer types) or TKI (for NSCLC) as their first-line treatment. Applying the above filtering criteria yielded 929 NSCLC, 462 HNC, and 503 TNBC non-ICB-treated patients, for a total of 5,453 multi-modal patients in the built ICB and non-ICB cohorts. With the exception of 51/503 (10%) TNBCs from the chemotherapy cohort, all patients in the ICB and non-ICB cohorts (5,402/5,453=99%) received their treatment between 2016 and 2023, with a median treatment start year of 2020 in the ICB cohort, and 2019 in the non-ICB cohort.

To define response groups, we used treatment-specific response and progression events available for each patient and applied the following definitions for both the ICB and non-ICB cohorts:

• Long-term responders were defined as patients who had:

  • Documented CR/PR/SD, and

  • No report of progression, and

  • Follow-up ≥180 days from treatment start.

• Primary resistant patients were defined as those who had:

  • Documented progression (PD/recurrence/metastasis) as their first reported outcome, or

  • Report of SD followed by documented progression <180 days of treatment start, or

  • Report of (pseudo-)response events (CR/PR) rapidly followed by documented progression <6 weeks (42 days) of treatment start.

• Acquired resistant patients were defined as those who had:

  • Documented CR/PR followed by documented progression ≥42 days of treatment start, or

  • Documented SD followed by documented progression ≥180 days of treatment start.

In cases where the first outcome was SD, we always checked for and considered the subsequent non-SD event, if any, to implement these definitions. For primary resistant patients who had a progression event as their first reported outcome, no time restriction was applied on when progression was reported with respect to treatment start. This is to prevent introducing biases in survival analyses.

Annotation of clinical and sample metadata

For accurate disease staging, stage information was parsed from both primary_tumor_characterization (reporting tumor, node, metastases staging) and metastasis (reporting metastasis events) files and “metastatic” patients were grouped with Stage IV patients. In case of multiple stages reported at different time points for the same patient, we considered the stage reported closest to the treatment start date. The number of patients missing data for clinical features compared (gender, age, ethnicity, albumin level, smoking status, and stage) is reported in online supplemental table 3.

Lines of treatment were defined in terms of drug exposure, with “1^st LoT” referring to patients who had not previously received a drug-containing treatment, and ≥2nd LoT referring to patients who received non-ICB drug(s) before the ICB regimen. The non-ICB drug can be part of a completely different treatment regimen, or part of a sequential ICB combination therapy.

For all patients in the built ICB and non-ICB cohorts, we excluded blood-based/non-solid biospecimens and focused only on solid tumor tissue biopsied from primary or distant organ sites. Pre-treatment samples consisted of any sample collected before the treatment start date. All post-treatment samples were collected on or after the treatment start date and before any subsequent LoT. When more than one pre-treatment sample was available per patient, we chose the one collected closest to the treatment start date, or chose one arbitrarily in case of collection date ties. When more than one post-treatment sample was available per patient, we chose the one collected closest to the date of progression, or chose one arbitrarily in case of collection date ties or absence of progression. DNA-seq samples were further restricted to those sequenced with xT.v2, xT.v3 or xT.v4 target sequencing assays (see DNA-seq Methods section). These sample-level filtering steps yielded a total of 4,100 pre-/post-treatment tumor samples available from 3,910 patients in the built cohorts. Of these 4,100 samples, 3,247 had RNA-seq data, 3,743 had DNA-seq data, and 2,890 had both RNA-seq and DNA-seq data.

Hormone receptor status, tumor content, and PD-L1 quantification

For patients with breast cancer, hormone receptor status was derived from a mix of curated (eg, extracted from clinical notes) and native (EHR) data sources. Tumor content (purity) was determined by board-certified pathologists prior to DNA extraction from formalin-fixed paraffin-embedded (FFPE) tumor tissue and consisted of the proportion of tumor nuclei over all nucleated cells. PD-L1 protein expression status was either obtained externally or determined by Tempus clinical testing with the 22C3 anti-PD-L1 antibody (Agilent), as previously described.⁷⁵ Briefly, slides were scored by a pathologist using the Tumor Proportion Score, calculated as the percentage of tumor cells with complete or partial membrane staining. Where multiple PD-L1 readouts were reported for a patient, we used the one closest to treatment start.

Survival analyses

Survival analyses were performed for two endpoints: (1) 5-year rwOS calculated from treatment start, and (2) 5-year rwPPS calculated from progression. Dates of death were captured either from retrospectively abstracted patient medical records or from third-party data sources that come from obituary documentation that is augmented to the death master file from the Social Security Administration. The resulting death information was integrated with Tempus AI data via an encrypted token system. We excluded patients with missing LKD or death information (n=1/5,453) and those with follow-up or censoring discrepancies (no death event but reported death date, death event but no reported death date or death date different from LKD, n=29/5,453). To address survival overestimation in real-world clinicogenomic data, we performed risk set adjustment (RSA) using the patient’s first tumor sequencing date to account for delayed entries.²¹ In other words, a patient was considered at risk (ie, entry time) starting from the index date (treatment start for rwOS or progression date for rwPPS) and only after their tumor underwent next-generation sequencing. We further excluded patients whose sequencing was performed at or after LKD or more than 5 years after index date (n=979 for rwOS, n=855 for rwPPS). For rwPPS, we also excluded acquired/primary resistant patients whose progression date coincided with LKD (n=41). This yielded a total of 4,444 long-term responder, acquired and primary resistant patients for rwOS and 3,773 acquired and primary resistant patients for rwPPS.

We ran survfit function from “survival” R package V.3.3.1 to fit the following RSA model and generate Kaplan-Meier curves with survival medians and associated 95% CIs (based on the cumulative hazard or log(survival)):

Surv(time=tentry, time2=Ti, event=death_flag) ~ patient_group

where “tentry” is the delayed entry time (ie, index date if sequencing was performed before or on index date, or date of sequencing if sequencing was performed after index date), “Ti” is LKD, “death_flag” is the death event variable (1 for death, 0 for censored), and “patient_group” is the patient response label (long-term responder, acquired or primary resistant). If LKD was 5 years or more after index date, LKD was capped to 5 years and the patient was censored (ie, death_flag was updated to 0).

We then ran the coxph function from the same package on the above RSA model to fit a Cox proportional hazards regression model, compute HRs and associated 95% CIs, and perform likelihood ratio tests comparing all three patient groups (rwOS) or acquired versus primary resistant patients (rwPPS).

All survival analyses were run separately for each treatment modality and cancer type and without adjusting for any covariate. Instead, we stratified both rwOS and rwPPS analyses by disease stage (Stage IV vs Stage I-III), LoT (1st vs ≥2nd), and (non-)ICB treatment modalities within each cancer type.

RNA sequencing and data processing

RNA-seq was performed on FFPE tumor samples using the Tempus xR.v1 or xR.v2 exome-capture-based assays (Tempus AI, Chicago, Illinois, USA). All RNA extraction, library preparation, sequencing and data processing are described in.⁷⁶ Briefly, Tempus xR assays are whole-transcriptome next-generation sequencing assays spanning a 39 Mb target region of the human genome covering a total of 19,375 genes. Sequencing was performed with Illumina HiSeq 4000 (RS.v1) or NovaSeq 6000 (RS.v2) and reads were aligned to GRCh37. Gene-level expression was quantified as both raw counts as well as gene length-normalized, G+C content-normalized, and library size-normalized transcripts per million (TPM). TPM values were log₂-transformed and batch-corrected for assay version using an in-house approach.⁷⁶ All subsequent analyses used either raw count data (DEA) or batch-corrected log₂-transformed TPM data (single gene comparisons, Uniform Manifold Approximation and Projection (UMAP), and single sample GSEA (ssGSEA)).

Differential expression and gene set enrichment analyses

We used the DESeq2⁷⁷ R package V.1.32.0 to run a contrasted DEA model comparing the transcriptional profiles of pre-/post-treatment acquired/primary resistant tumors within each cancer type and treatment modality separately (chemo for all three cancer types; anti-PDx mono, anti-PDx+chemo, other ICB and TKI for NSCLC; any ICB for all three cancer types in addition to 1^st LoT Stage IV NSCLC, ≥2^nd LoT Stage IV NSCLC, squamous NSCLC, and non-squamous NSCLC). We used raw count gene expression data of tumors biopsied from the primary organ site for each cancer type (lung tissue for NSCLC, head/neck tissue for HNC, breast tissue for TNBC) and ran the following model accounting for RNA-seq assay version (RS.v1/2) and NSCLC histology (squamous/non-squamous), leaving all parameters to default:

design = ~ isolate_molecular_assay (+ histology) + sample_group

where “sample_group” is a concatenation of pre-/post-treatment and acquired/primary resistant sample type (ie, one of pre_acquired, post_acquired, pre_primary, post_primary). DEA results were then extracted specifying the contrast of each comparison of interest and setting the significance threshold (ie, “alpha”) to 0.05:

results(dds, alpha=0.05, contrast)

where “dds” is the fitted DEA object and “contrast” is one of:

• Pre-acquired versus pre-primary.

• Post-acquired versus post-primary.

• Pre-acquired versus post-acquired.

• Pre-primary versus post-primary.

The reference group of each comparison and the number of samples used are reported in online supplemental table 4. Log₂ fold changes (log₂FC) and associated p values were computed using Wald tests and p values were adjusted for multiple tests using Benjamini-Hochberg correction.

We then used DESeq2 results to run pre-ranked GSEA using the fgseaMultilevel function from the fgsea⁷⁸ R package V.1.20.0 with default parameters. Specifically, all 19,375 genes were ranked based on their DESeq2 log₁₀(FDR) signed by the log₂FC sign, with genes ranked on the top being the most significantly upregulated in the reference group, and the ones in the bottom being the most significantly downregulated. Genes whose DESeq2 FDR=NA were excluded. For each gene set, fgseaMultilevel walks through the ranked gene list and computes a running sum statistic-based enrichment score with an associated p value using the adaptive multilevel splitting Monte Carlo approach. Normalized enrichment scores were computed to account for differences in gene set sizes and were used in all comparisons. For each comparison, p values were adjusted for multiple testing across gene sets using the Benjamini-Hochberg correction. The following gene sets were considered for all comparisons and no filter on the gene set size was applied:

• 50 Hallmark gene sets from MSigDB.²³

• 19 cancer-specific immune cell signatures from ConsensusTME²⁴ (The Cancer Genome Atlas (TCGA)-LUAD (lung adenocarcinoma) for NSCLC, TCGA-HNSCC for HNC, and TCGA-BRCA (breast invasive carcinoma) for TNBC).

• 28 literature-compiled immuno/oncology-focused gene sets: 4 DC signatures (DC1, DC2, mregDC, cDCmaturation) from,²⁷ 2 macrophage signatures (M1 and M2 macrophages) from,²⁶ 9 TME signatures (major histocompatibility complex I (MHCI), MHCII, coactivation molecules, antitumor cytokines, protumor cytokines, checkpoint inhibition, myeloid-derived suppressor cells (MDSCs), cancer-associated fibroblasts (CAFs), Matrix) from,²⁸ and 10 tumor meta-program signatures (MP1 Cell Cycle—G2/M, MP2 Cell Cycle—G1/S, MP3 Cell Cycle HMG-rich, MP4 Chromatin, MP5 Stress, MP8 Proteasomal degradation, MP10 Protein maturation, MP11 Translation initiation, MP19 Epithelial Senescence, MP20 MYC) from.²⁵

For each comparison, GSEA was run together on Hallmark and ConsensusTME gene sets, and run separately on the literature-compiled gene sets. Gene sets with GSEA FDR ≤0.1 were considered significantly differentially enriched.

Single sample gene set enrichment analysis and ESTIMATE scoring

We used the gsva function from the GSVA⁷⁹ R package V.1.42.0 to run ssGSEA⁸⁰ for all gene sets on the batch-corrected and log₂-transformed TPM data of all RNA-seq tumor samples. All parameters were set to default except for ssgsea.norm which was set to FALSE (ie, run without normalization) to allow for absolute computation of enrichment scores (ES).

To estimate the overall stromal (non-cancer) content in bulk RNA-seq tumor samples, we used the ESTIMATE⁸¹ algorithm as implemented in the tidyestimate R package V.1.1.1 on the batch-corrected and log₂-transformed TPM data, setting the find_alias parameter to TRUE and the is_affymetrix parameter to FALSE.

Distributions of ssGSEA ES of specific gene sets and ESTIMATE scores were compared between tumor groups of interest using the Wilcoxon rank-sum test.

DNA sequencing and data processing

DNA-seq was performed on FFPE tumor samples and was available for a total of 3,981 FFPE samples from the built ICB and non-ICB cohorts. Sequencing was performed using the Tempus xT.v1-4 targeted panel (3,930/3,981=98.7% samples, 596–648 genes), xO targeted panel (18/3,981=0.5% samples, 1,714 genes) or xE.v1-2 whole exome sequencing (33/3,981=0.8% samples) assays (Tempus AI). We focused on samples sequenced with xT.v2, xT.v3, or xT.v4 assays (3,743/3,981=94% DNA-sequenced FFPE tumor samples) to maximize both the number of samples and commonly targeted genes across assays (n=595 commonly covered cancer-related genes across xTv2-4). A matched normal sample (ie, germline control) was available for 2,352/3,743 samples (63%) and consisted of 200 µl blood or 650 µl saliva. Sequencing with xT assays was performed to achieve an average on-target depth of 500×. All DNA extraction, library preparation, sequencing and data processing, including calling of somatic mutations and CNVs (for both normal-matched and tumor-only samples), classification of mutation functional impact, calculation of TMB, detection of MSI, and HLA typing, is described in.^{82 83} A total of 235,311 short variants (single nucleotide variants and indels) aligned to GRCh37 were detected across the 3,743 samples.

Given sensitivity of dN/dS analysis to germline contamination (see dndscv Methods section), and since a germline control was not always available for DNA-seq, we further cross-checked all mutation calls against germline variant databases (dbSNP V.154 and gnomAD V.2.1.1) and filtered out any germline variant with an allele frequency ≥0.1%. Out of the 235,311 mutations, 55,173 (23%) were filtered out, yielding a total of 180,138 mutations in the final mutation dataset, with a median mean coverage at mutated sites of 563× across samples (90% of samples with mean coverage ≥200×, 99% with mean coverage ≥100×).

Two versions of the mutation dataset were used in subsequent analyses: (1) an unfiltered version with all 180,138 mutations containing synonymous mutations and without filtering applied on variant allele frequency or mutation functional impact (version used for dndscv analysis), and (2) a filtered version with 67,759 non-synonymous and potentially more biologically relevant mutations after applying the filtering steps described in^{82 83} (version used for the functional characterization of acquired ICB alterations).

Analysis of mutation selection with dndscv

Using the unfiltered mutation dataset containing both synonymous and non-synonymous mutations (see DNA-seq Methods section), we ran dndscv²⁹ (https://github.com/im3sanger/dndscv) V.0.0.1.0 in each of pre-/post-treatment acquired/primary resistant and pre-treatment long-term responder tumor groups separately within each cancer type and for each of ICB, chemotherapy, and TKI treatment modalities, for a total of 35 dndscv runs. Because this analysis relies on cancer mutations, we used all tumor samples available in our annotated dataset, including those biopsied from non-primary (metastatic) organ sites, to increase statistical power.

We used the hg19 reference coding DNA sequence (refCDS) object downloaded from https://github.com/im3sanger/dndscv/blob/master/data/refcds_hg19.rda. Of the 595 commonly covered genes across xT.v2-4 assays, 592 genes were present in hg19 refCDS (ie, all but H19, HLA-DPB2, and HLA-DRB6). For CDKN2A, dndscv considers two isoforms separately: CDKN2A.p16INK4a and CDKN2A.p14arf. We used the following parameter settings for all dndscv runs: max_muts_per_gene_per_sample=10 (ie, the maximum number of mutations per gene per sample), outmats=T, constrain_wnon_wspl=T (ie, constraining non-sense and essential splice site dN/dS estimates for increased power), gene_list=common_panel_genes (ie, all 592 genes commonly covered by the three DNA-seq assays).

An important requirement for dndscv is the exclusion of all germline variants from the mutation dataset, as such variants typically bias dN/dS estimates towards lower values (<<1).²⁹ After cross-checking all mutation calls against germline variant databases as previously described, we checked that the global dN/dS estimates across all 592 genes (“wall”) were close to or slightly above neutral selection (≥1) for all dndscv runs (online supplemental figure 4F), as expected.

For each gene and for each mutation type (missense, non-sense, indels), dndscv computes a maximum-likelihood estimate (MLE) of the dN/dS ratio and a corresponding p-value from the likelihood-ratio test, in addition to a global p value integrating all mutation types. All p values are adjusted for multiple testing using Benjamini-Hochberg correction (FDR). A gene was considered under significant positive selection if at least one mutation type affecting the gene (missense/non-sense/indel) was under significant positive selection, and if the global selection FDR was below the significance threshold. More specifically:

• (wmis_cv>1 and qmis_cv≤0.1) or (wnon_cv>1 and qnon_cv≤0.1) or (wind_cv>1 and qind_cv≤0.1), and

• qglobal_cv≤0.1.

where wmis_cv, wnon_cv, and wind_cv are the dN/dS MLE estimates for missense, non-sense (truncating), and indel mutations respectively, qmis_cv/qnon_cv/qind_cv are the corresponding mutation-level selection FDRs, and qglobal_cv is the global selection FDR integrating all mutation types. Positively selected genes are represented by their mutation-level log-transformed dN/dS MLE from the corresponding dndscv run in figure 4B.

Selection for NFKBIA mutations in post-ICB acquired NSCLCs was driven by missense mutations in two samples: one sample that had one missense mutation, and another sample that had six missense and one non-sense mutation. We checked that the latter sample was not hypermutated (TMB <10 mutations/Mbp and MSI-stable), and all seven mutations affecting NFKBIA in that sample had a read depth ranging from 1,361 to 1,630× and a variant allele frequency ranging from 8% to 15% for a tumor purity of 20%. Restricting the dndscv analysis to a maximum of three mutations per gene per sample (ie, the default) in post-ICB acquired NSCLCs yielded a missense dN/dS MLE of 5.9 for NFKBIA with an uncorrected missense p value of 0.01 and an uncorrected global p value of 0.048.

Comparison of observed versus expected dN/dS estimates

To evaluate the specificity of mutations selected post- (but not pre-)ICB acquired resistance, we compared their observed dN/dS estimates to an expected distribution of dN/dS estimates, contrasting with the same comparison for mutations selected both pre-ICB and post-ICB. The expected distribution of dN/dS estimates was generated by randomly shuffling pre/post-ICB acquired/primary tumor group labels and rerunning dndscv on the pseudo-post-ICB acquired resistant group, repeating the process 10,000 times. More specifically, for each cancer type in the ICB cohort, we:

• Focused on pre_primary, post_primary, pre_acquired and post_acquired tumor samples (ie, four labels);

• Randomly shuffled the four labels then took the pseudo-post_acquired labeled samples (this way, we ensure that we randomly sample the same number of actual post_acquired samples at each iteration);

• Ran dndscv on the pseudo-post_acquired samples using the same parameter settings as previously described;

• Repeated this process 10,000 times to generate a null distribution of dN/dS estimates.

Then, for each positively selected post-ICB acquired mutation, we calculated the difference between its observed dN/dS MLE and all 10,000 expected dN/dS MLEs, obtaining ΔdN/dS distributions for mutations selected post-only or pre-ICB and post-ICB acquired resistance, as depicted in figure 4E. At the gene level, we assessed the significance of the difference between observed versus expected dN/dS estimates by computing an empirical p value assessing how often the observed dN/dS estimate was higher than expected across all iterations, that is, p value=n(ΔdN/dS>0)/10,000.

Functional characterization of acquired ICB alterations

Focusing on post-ICB acquired resistant tumor samples with both DNA-seq and RNA-seq data, we ran DEA and GSEA comparing ALT versus WT tumors for each acquired ICB alteration in each corresponding cancer type separately. To annotate ALT tumors, we integrated CNV data with the filtered mutation dataset (see DNA-seq methods) and considered only pathogenic (P), likely pathogenic (LP), or mutations with unknown significance (US), excluding all benign or likely benign mutations. A gene was considered altered (ALT) depending on its cancer driver role as annotated in⁸⁴ or on the mutation type driving dN/dS selection (ie. missense or truncating/indel mutations):

• If the gene was a tumor suppressor gene (TSG), then only LoF alterations were considered;

• If the gene was an oncogene (OG), then only gain-of-function (GoF) alterations were considered;

• If the gene had no clearly defined cancer driver role (TSG/OG), then LoF alterations were considered if the dN/dS selection was driven by truncating/indel mutations, otherwise any alteration (P/LP/US mutation or CN loss/gain) was considered for the gene.

Any CN loss, stop gain, frameshift, or start loss mutation was considered as LoF alterations, while CN gains were considered as GoF alterations. Missense mutations affecting TSGs were considered LoF alterations, and those affecting OGs were considered GoF alterations. We considered LoF alterations for ARID1B, B2M, TGFBR2, NOTCH1, FAT1, CYLD, PTEN, and RUNX1 (TSGs), as well as for ERRFI1 (no clearly defined cancer driver role but dN/dS selection driven by indels). We considered GoF alterations for MAF (OG). Since no clear cancer driver role was reported for NFKBIA and since dN/dS selection was driven by missense mutations, we considered all alteration types for this gene.

We used raw count gene expression data of tumors biopsied from primary or metastatic organ sites (cross-tissue sites) and ran the following DESeq2 DEA model accounting for RNA-seq assay version (RS.v1/2), NSCLC histology (squamous/non-squamous), and organ site (primary site/lymph node/liver/other), leaving all parameters to default:

design = ~ isolate_molecular_assay (+ histology) + tissue_site+sample_group

where tissue_site is the organ site and sample_group is one of ALT or WT. DEA results were extracted setting the “alpha” parameter to 0.05. We then ran pre-ranked GSEA using DEA results to find differentially enriched gene sets between ALT versus WT tumors, using Hallmark, ConsensusTME, and literature-compiled gene sets as previously described.

Statistical analyses

All analyses were performed and plots produced using R programming language V.4.1.0. Statistical tests used in each analysis/comparison are specified in the corresponding figure legends or Methods section. All p values were adjusted using Benjamini-Hochberg correction where applicable.

Data availability statement

No data are available. All data relevant to the study are included in the article or uploaded as supplementary information.