Elucidating the Heterogeneity of Immunotherapy Response and Immune-Related Toxicities by Longitudinal ctDNA and Immune Cell Compartment Tracking in Lung Cancer Open Access

Despite the routine use of immunotherapy for patients with metastatic non–small cell lung cancer (NSCLC), intrinsic and acquired resistance to immunotherapy regimens are common. Although PD-L1 expression guides the selection of first-line therapy, it does not always predict clinical response (1). A multitude of tumor and immune biomarkers—from tumor mutational burden to immune-related gene-expression signatures—have been evaluated as predictors of immunotherapy response (2, 3); however, with the exception of microsatellite instability, most of these putative biomarkers fail to consistently predict outcomes. Optimal patient selection for immunotherapy represents a critical challenge that is intensified by the heterogeneity of clinical responses and inefficiency of imaging to fully and timely capture therapeutic response. As such, for the majority of patients treated with an immunotherapy-containing regimen—representing >70% of patients with NSCLC—clinical equipoise remains between the available options, even with preferences based on histology and PD-L1 expression (4).

Minimally invasive liquid biopsy approaches have been shown to enable early and accurate therapeutic response assessment and long-term response monitoring during immunotherapy (5, 6). An increasing number of studies have shown that monitoring of early on-therapy ctDNA kinetics, using the mean (7, 8) or maximum mutant allele fraction (MAF; refs. 9–12) of tumor-derived alterations can enable real-time tumor burden assessment and have supported the role of ctDNA molecular response as an early endpoint for response to immunotherapy (10–17). Collectively, these studies support the clinical utility of ctDNA response monitoring for immunotherapy-treated NSCLC and importantly highlight areas where ctDNA evaluation can be particularly informative and require further investigation. Given the challenges with imaging to capture the timing and magnitude of therapeutic response in the context of immunotherapy (18), ctDNA dynamics may be useful in identifying patients with radiographically stable disease that have differential outcomes. To date, such associations have only been evaluated in a small number of studies in lung cancer including few patients with radiologic stable disease (11, 15). Previous studies have also highlighted the value of longitudinal ctDNA response monitoring throughout the treatment course (10–12, 19–21), which would increase the clinical sensitivity of ctDNA response for predicting therapeutic response.

In addition to the identification of patients most likely to attain clinical benefit and those at risk for disease progression, which may open a therapeutic window of intervention for the latter and potentially spare unnecessary interventions for the former, minimally invasive blood-based analyses could be valuable for prediction of emergence of immune-related adverse events (irAE; refs. 22–24). Baseline T-cell characteristics, including the abundance of activated CD4⁺ memory T cells and the diversity of the TCR repertoire, have been linked to the development of severe irAEs (22). Similarly, multiple studies have evaluated immune cell subsets or cytokines/chemokines as predictors of irAEs (11, 25–27); however, no prospectively evaluated biomarkers are used in clinical practice today. Although recent studies have shown a correlation between early clonal expansions of T-cell populations and the timing and severity of irAEs, on-treatment biomarkers for therapeutic response assessment and prediction of irAEs to immunotherapy-based regimens remain limited (22).

To study the heterogeneity of immunotherapy response together with the value of peripheral blood immune repertoire analyses in capturing irAEs, we designed a longitudinal study of patients with NSCLC who received IO-based therapies. We assessed ctDNA, peripheral TCR dynamics, and plasma protein expression profiles from serial blood specimens to capture timing, depth, and quality of tumor and immune responses in the context of immunotherapy.

A total of 30 patients were selected for inclusion in cell-free DNA analyses performed in this study using the following criteria: (i) advanced or metastatic NSCLC who received at least one cycle of immunotherapy or chemoimmunotherapy treatment; (ii) at least two plasma samples evaluable for cfDNA sequencing; (iii) at least one buffy-coat sample for white blood cell (WBC) genomic DNA sequencing and/or with either clinical tumor mutational profiling (TMP) or whole-exome sequencing (WES) of tumor tissue; and (iv) had clinical follow-up including imaging through to time of death or at least 6 months from time of treatment initiation. Twenty-four patients treated with standard-of-care pembrolizumab (n = 13) or chemotherapy with pembrolizumab (n = 11) were identified from the Immunobiology Blood and Tissue Collection of Upper Aerodigestive Malignancies biospecimen collection protocol (NCT05995821) and registry (IRB00100653 approved by the Johns Hopkins University Institutional Review Board (IRB)). This included 1 patient from a prior study (Anagnostou and colleagues; ref. 11) from whom we collected and analyzed additional samples from 3 on-treatment timepoints (Supplementary Table S1). Blood samples were collected under the NCT05995821 protocol at prespecified timepoint intervals that coincided with each cycle of single-agent immunotherapy or chemoimmunotherapy combination regimen until disease progression or treatment discontinuation due to toxicity. Of 30 patients included in ctDNA analyses, the majority (n = 23) received immunotherapy until clinical disease progression (n = 15) or toxicity (n = 8), 6 were lost to follow-up, whereas only 1 patient continued on immunotherapy up to 2 years after which therapy was electively discontinued. Clinical data were abstracted from electronic health records and clinical tissue molecular profiling data, including variants, were retrieved from available clinical next-generation sequencing outputs. Six patients were enrolled at the Netherlands Cancer Institute (NKI), in accordance with institutional regulatory approvals and received second-line single-agent immunotherapy with nivolumab. An independent cohort of 49 patients with advanced (stage III/IV) NSCLC were enrolled at the Allegheny Health Network Cancer Institute under institutionally approved regulatory protocols and received treatment with either immune-checkpoint monotherapy (n = 6) or immunotherapy in combination with chemotherapy (n = 43). All patients were immunotherapy naïve prior to receiving immune-checkpoint monotherapy or chemoimmunotherapy combination. PD-L1 expression, assessed by IHC, was abstracted from medical records and evaluated as follows: high expression of PD-L1 on tumors was defined as ≥50% tumor cell staining and low or negative expression was defined as <50% tumor cell staining. Primary toxicities attributed as irAEs were assessed from medical records according to the CTCAE version 5.0 grading system (ref. 28; Supplementary Table S1). Radiographic response assessments were retrieved from the medical records. For patients with a best overall response of stable disease, changes in tumor burden were quantified using RECIST 1.1 criteria. One such patient (CGLU528), who only had lymph node disease, was excluded due to challenges in performing accurate RECIST evaluations. Progression-free survival (PFS) and overall survival (OS) were assessed as well as durable or nondurable clinical benefit (DCB or NDB). DCB was defined as lack of progression or death at 6 months; NDB was defined as progression or death within 6 months from immunotherapy initiation. Studies were conducted in accordance with the Declaration of Helsinki, approved by the IRB, and patients provided written informed consent for samples and clinical data for research purposes.

Plasma and whole blood samples were collected at times when standard-of-care clinical venipuncture was indicated and processed for targeted error-correction sequencing (TEC-Seq) as previously described (11). Cell-free DNA was isolated from plasma using the Qiagen Circulating Nucleic Acids Kit (Qiagen GmbH). For 29 patients, matched WBC DNA was extracted from buffy-coat obtained at baseline and/or after treatment initiation, using the Qiagen DNA Blood Mini kit (Qiagen GmbH) and sheared to a target fragment size of 200 bp. TEC-Seq DNA libraries were prepared from 6.5 to 125 ng cfDNA and 75 ng WBC DNA, followed by targeted capture using a custom set of hybridization probes [PlasmaSELECT-63 or -64 panels, Personal Genome Diagnostics (PGDx); Supplementary Table S2] and sequenced using 100 bp paired-end runs on Illumina HiSeq 2000/2500 instruments. TEC-Seq characteristics are described in Supplementary Table S3.

Matched tumor/normal WES was performed for 17 patients, and WES data were used to identify tumor-specific mutations in plasma. Briefly, tumor DNA was extracted following pathologic review and macrodissection of formalin-fixed paraffin-embedded (FFPE) tumor tissue using the Qiagen DNA FFPE kit (Qiagen GmbH). Tumor and matched WBC DNA were sheared to a target fragment size of 200bp and processed for WES as described previously (10). Hybrid capture of exonic regions was performed using Agilent SureSelect in-solution capture reagents and SureSelect XT Human All Exon V4 probes (Agilent). Captured libraries were sequenced using 100 bp paired-end runs on Illumina HiSeq 2500. WES technical metrics are described in Supplementary Table S4.

TEC-Seq data were analyzed as previously described using a genomics annotation pipeline including primary processing by Illumina CASAVA software (v1.8), alignment to the human reference genome (hg19/GRCh37) by Novoalign, and variant calling by VariantDx (29). Supermutant counts were defined as the total number of barcoded read families in which ≥95% of members contained the same variant. Variants in plasma cfDNA were defined as detectable at a supermutant count of at least 3, whereas variants in WBC specimens—that were also detected in plasma cfDNA—were defined as detectable at a supermutant count of at least 1. The variant allele fraction for a given plasma variant was determined from the percentage of distinct mutant versus total cfDNA variant reads at a given locus.

The origin of plasma cfDNA variants was classified using all available sequencing data, including matched tumor (clinical TMP or WES) and all WBC DNA sequencing, as tumor-derived, germline or clonal hematopoiesis (CH)-derived, based on a tiered approach. Patients with matched tumor NGS and detectable tumor-specific mutations in plasma (n = 21) were evaluated using a tumor-informed approach for the identification of ctDNA variants. Given the limitations of single-region tissue NGS for capturing the molecular heterogeneity of individual tumors, patients with matched tissue sequencing but no detectable tumor-specific mutations in plasma (n = 3) were further evaluated using a WBC-informed approach (10) to identify any candidate ctDNA variants missed by tissue NGS. Patients with no detectable tumor-derived mutations in plasma according to both approaches were subsequently classified as undetectable. Similarly, for 6 patients without matched tumor tissue NGS available (n = 6), we used the WBC-informed approach to characterize variant origin in plasma, and cases with no detectable tumor-derived mutations were classified as undetectable (n = 1). For patients with multiple WBC DNA samples sequenced (n = 11), the intersection of all alterations detected across sequenced WBC samples was used.

Plasma variants were cross-referenced against the Catalogue of Somatic Mutations in Cancer (COSMIC) v95 database for the annotation of hotspot alterations using OpenCRAVAT (30). A conservative COSMIC frequency threshold of 25 hits was used to define a lung cancer hotspot with high confidence (10, 31). Of the variants that were not classified as a hotspot, those with a variant allele frequency ≥25% in all plasma and available WBC samples from a patient were classified as germline. Plasma variants detected in genes commonly altered in clonal hematopoiesis-CH, namely, DNMT3A, TET2, ASXL1, PPM1D, TP53, JAK2, RUNX1, SF3B1, SRSF2, IDH1, IDH2, U2AF1, CBL, ATM, and CHEK2 (CH blacklist; Supplementary Table S5) and a supermutant count ≥1 in matched WBC TEC-seq were considered CH-derived. Plasma variants that were not detected in matched WBC sequence data or could not be evaluated as matched WBC sequencing was not performed were classified as follows: variants within a CH-blacklisted gene and an alteration frequency ≥5% in COSMIC hematologic malignancies were classified as CH-derived, providing that they were not lung cancer hotspots or reported in matched tumor tissue sequencing. Plasma variants that could not be verified by either matched tumor tissue sequencing or as lung cancer hotspots, and variants that were classified as a lung cancer hotpot within a CH-blacklisted gene, were assigned as variants of unknown origin. For WBC-informed classifications, all non-germline variants were further compared against matched WBC DNA sequence data to determine cellular origin. Plasma variants within a CH-blacklisted gene and a supermutant count ≥1 in matched WBC were considered CH-derived. Given the small possibility of buffy-coat contamination by circulating tumor cells, resulting in the misclassification of tumor-derived alterations as CH-derived, hotspot variants that were detected in matched WBC within a non-CH-blacklisted gene were classified as tumor-derived. Plasma variants within a CH-blacklisted gene and an alteration frequency ≥5% in COSMIC hematologic malignancies were classified as CH-derived providing that they were not lung cancer hotspots. A summary of the branched logic architecture used for variant origin classification is shown in Supplementary Fig. S1.

The MAF of the most abundant tumor-derived mutation at each serial plasma sample (referred to as the maximum MAF) was used as a measure of circulating cell-free tumor load (cfTL) and a reduction of cfTL to 0% indicated cfTL clearance (10). Patients who had complete cfTL clearance from baseline to the final timepoint analyzed were classified in the molecular response category. Patients with cfTL clearance followed by a subsequent increase were classified in the molecular response followed by the recrudescence category. Patients with cfTL persistence across all time points analyzed were assigned a classification of molecular progression. Molecular response assessments were performed in a unified manner for all study patients.

TCR Vβ NGS was performed for 79 patients. This cohort was comprised of 23 patients included in ctDNA analyses and 7 patients from Anagnostou and colleagues; Set 1; ref. 11). In addition, peripheral blood samples were analyzed from an independent cohort of 49 patients with advanced (stage III/IV) NSCLC enrolled at the Allegheny Health Network Cancer Institute (Set 2). Biospecimens were collected from both baseline and early on-treatment time points prior to the clinical diagnosis of an irAE for 15/17 patients included in the study cohort. For the remaining 2 cases, the first on-therapy blood sample was collected shortly after the clinical presentation of irAEs, as follows: for patient CGLU383, the first on-therapy blood sample was collected 1 week following the clinical presentation of irAE, at which time the patient was started on corticosteroid treatment. For patient CBUAD 1124, the first on-therapy blood sample was collected 3 weeks following the initial diagnosis of irAE, for which the patient underwent enhanced monitoring.

WBC genomic DNA was isolated from serial peripheral blood samples collected prior to treatment initiation (n = 58) and during the course of treatment (n = 199). Following genomic DNA isolation, TCR-β CDR3 regions were amplified using either the deep ImmunoSeq assay by multiplex PCR specific to TCR Vβ, Dβ, and Jβ regions (Adaptive Biotechnologies) or the AmpliSeq for Illumina TCR beta-SR Panel (Illumina). Productive TCR clonotypes were identified and analyzed at the amino acid sequence level. TCR clones were categorized as increasing or decreasing between baseline and each on-treatment timepoint and relative percentages of these clones were compared using Fisher exact test with P value adjustment (P_adj < 0.01) using the false discovery rate (FDR; Benjamini–Hochberg procedure). A minimum abundance stratification was used to select for clones with an abundance ≥0.01% in on-treatment blood samples. This was followed by aggregation of clone dynamic measures per subject and statistical comparisons of patient subgroups of interest using the nonparametric Mann–Whitney U test. Aggregate clone dynamics did not appear to be influenced by the significant expansion of one or few high-abundance clones, as we did not observe a directionality bias in the subset of clones that did not statistically significantly change in abundance between serial time points toward either a general increase or decrease (Supplementary Table S6).

Unique TCR-β CDR3 sequences were clustered according to anticipated specificity for epitope recognition and HLA-allele associations using the GLIPH2 (Grouping of Lymphocyte Interactions by Paratope Hotspots 2) method (32). GLIPH2 identified global TCR-β clusters based on the similarity of CDR3 sequences as follows: TCRs that shared the same length CDR3 differing by up to seven amino acids with matching global motifs ≥4 amino acids in length were grouped into the same cluster. Clusters identified by GLIPH2 were considered for statistical analyses based on the following parameters: detected in ≥3 distinct patients, number of unique CDR3 ≥ 2, vb_score (enrichment of V genes within cluster) <1, length_score (CDR3 length distribution within cluster) <0.05.

Expression levels of 92 proteins in the Olink Immuno-Oncology panel (Olink Proteomics AB) were measured in plasma samples (n = 88) from 28 patients using a proximity extension assay. Quality control and data normalization were performed based on internal and external controls. Internal controls consisting of an immuno-based nonhuman antigen control, an extension control comprised of an antibody conjugated to a unique oligonucleotide DNA pair for immediate proximity-dependent hybridization and extension, and a detection control comprised of a synthetic double-stranded DNA amplicon to monitor amplification were spiked into each sample well to assess intra-assay variability (Olink Proteomics AB). External sample controls, used to evaluate assay precision, were also included on the assay plate and consisted of negative controls (buffer-only) and interplate controls (pool of 92 antibodies conjugated to each oligonucleotide pair). Protein expression values were generated as log2 of normalized protein expression (NPX) as follows: counts of known barcode sequences were first normalized to the extension control, followed by log2 transformation and intensity normalization against the interplate control. Mean protein expression values were calculated for each sample, as well as the difference in values between timepoints and compared across different patient groups using the Mann–Whitney U test.

For survival analyses, median point estimates were estimated using the Kaplan–Meier method, and survival curves were compared by log-rank testing. Univariate and multivariate Cox proportional hazards regression analyses were used to evaluate the impact of molecular responses on overall and progression-free survival. For categorical comparisons, Fisher exact test was utilized. For nonparametric comparisons, the Mann–Whitney U test was utilized. Unless otherwise specified, differences across comparisons were considered significant at P less than 0.05. All analyses were conducted using R Statistical Software (version 4.1.2) as described above (33).

Next-generation sequence data can be retrieved from the European Genome-Phenome Archive (EGA; study accession number EGAS00001007291).

We performed deep targeted error-correction sequencing (TEC-seq) of 162 serial plasma cell-free DNA (cfDNA) and 47 WBC genomic DNA (gDNA) samples from a real-world cohort of 30 patients with advanced or metastatic NSCLC treated with immunotherapy or chemoimmunotherapy regimens (Fig. 1A and B). Plasma cfDNA sequencing was completed from baseline samples prior to treatment in 26 patients (median 0 weeks; range, −0.7–0 weeks; Supplementary Table S3). When baseline samples were insufficient or unevaluable (n = 4 patients), the next available samples were assessed, prior to the second treatment cycle (median 3 weeks, range, 3–3.14 weeks; Supplementary Table S3). Sequence alterations were analyzed in plasma at baseline and longitudinally after treatment initiation.

A total of 461 variants in 37 genes were detected in plasma cfDNA (Fig. 2A; Supplementary Table S5). The cellular origin of cfDNA variants was determined using all available plasma, WBC DNA, and tumor tissue sequencing data according to a tiered approach (Materials and Methods); using this classification approach, tumor-derived sequence mutations (43% of all variants with resolved cellular origin) were identified in 26 patients (Fig. 2B; Supplementary Table S5; Supplementary Fig. S2). This included 2 patients with no detectable tumor tissue-specific variants in plasma, in whom we identified KRAS G12C, G12V, and TP53 G245C ctDNA mutations using a WBC-informed approach (Materials and Methods). Notably, 25 mutations (6% of all variants with resolved cellular origin) were of germline origin and 205 variants (51% of all variants with resolved cellular origin) were CH-derived (Fig. 2C and D; Supplementary Fig. S2). CH variants were detected across genes both canonically and noncanonically associated with clonal hematopoiesis, including DNMT3A, TP53, BRCA2, and EGFR (Fig. 2C; Supplementary Table S5) and were removed from downstream analyses of molecular response.

At first sampling, 24 of 26 patients had ≥1 tumor-derived ctDNA mutation, at a median MAF of 5.40% (range, 0.11%–19.70%; Supplementary Table S5). Evaluation of the landscape of tumor-derived mutations revealed frequent nonsynonymous alterations in TP53 [altered in 14 of 26 (54%) patients] and KRAS [9 of 26 (35%) patients]. Tumor-derived mutations in ERBB2 (4/26 patients; 15%), ATM (2/26 patients; 8%), BRAF (2/26 patients; 8%), CDKN2A (2/26 patients; 8%), and CTNNB1 (2/26 patients; 8%) were also detected at >5% prevalence among patients (Fig. 2B). Comparison between the maximal MAFs of ctDNA mutations prior to therapy revealed a difference in baseline MAFs by outcome in patients treated with immunotherapy and chemoimmunotherapy. For patients on single-agent immunotherapy, a lower cfTL at baseline was significantly associated with durable clinical benefit (Supplementary Fig. S2; P = 0.029 Mann–Whitney U test). Further analyses of the distribution of baseline ctDNA maximal MAFs and outcome revealed that patients on single-agent immunotherapy with baseline MAFs <1% had significantly longer PFS compared with individuals with MAFs ≥1%. However, patients on chemoimmunotherapy had similar distributions of ctDNA MAFs independent of therapeutic response and outcome (Supplementary Fig. S2).

To evaluate changes in ctDNA dynamics under the selective pressure of therapy, we used the maximum MAF of tumor-derived sequence alterations detected at each timepoint as a measure of cfTL for each patient, and tracked cfTL during the course of single-agent immunotherapy or chemoimmunotherapy treatment. Using this approach, we were able to track cfTL dynamics for 26 patients with detectable ctDNA mutations in plasma and assign a molecular response classification. We defined molecular response (mR) as the complete clearance of cfTL following treatment initiation. Patients who displayed cfTL persistence across all sampled time points were assigned a classification of molecular progressive disease (mPD). A subset of cases displayed a transient molecular response followed by a subsequent increase in cfTL that was driven by either recrudescence of baseline ctDNA mutations or the emergence of mutations that were undetectable at baseline and were classified as molecular response followed by recrudescence or emergence.

Overall, 6 patients were classified in the molecular response category, with a time to molecular response as soon as 2 weeks from treatment initiation (median 4.5 weeks, range, 2.29–14.86 weeks), whereas 13 patients were classified as molecular progressors (Supplementary Table S5; Fig. 3A–D). Six patients had molecular response followed by recrudescence with a median time to molecular response of 7 weeks (range, 3–33 weeks; Supplementary Table S5; Fig. 3B). One patient had molecular response followed by emergence (Supplementary Table S5; Fig. 3C). Molecular response was not concordant with first radiographic response (P = 0.123, Fisher exact test), whereas a trend was noted toward a correlation between molecular response and best overall radiographic response (P = 0.056, Fisher exact test; Fig. 4A–C). To determine whether this limited concordance was attributed to the larger fraction of patients with radiographically stable disease in this cohort, we compared the first and best overall radiographic response assessments and found that 3 patients with a first response assessment of stable disease were reclassified as partial responders at the time of best overall response assessment (Supplementary Table S1). cfTL analyses revealed that 2 of these 3 patients had a molecular response followed by recrudescence, with a time to molecular response as soon as 3 weeks (median 4.5 weeks; range, 3–6 weeks; Supplementary Table S1). In these cases, ctDNA molecular response preceded best radiologic assessments by an average lead time of 21 weeks (range, 8–33 weeks). Conversely, 2 patients with a first radiographic response assessment of progressive disease were subsequently reclassified as stable disease at the time of best overall assessment. Analysis of ctDNA molecular response classifications revealed one patient to be a molecular progressor, consistent with short PFS and OS times of 1.4 and 6.3 months, respectively. In contrast, the second patient was assigned a classification of molecular response followed by recrudescence, consistent with an extended PFS and OS of 23.0 months (Supplementary Table S1). Notably, molecular responses were concordant with durable clinical benefit (P = 0.006, Fisher exact test, Fig. 4D), further supporting the accuracy of ctDNA molecular responses for capturing clinical outcomes. Collectively, these findings highlight the limitations of imaging-based response assessments for accurate evaluation of therapeutic benefits from immunotherapy and the association between ctDNA molecular response and durable therapeutic responses.

Finally, we assessed whether ctDNA molecular responses were associated with OS and PFS (Fig. 4E and F; Supplementary Figs. S3 and S4). Patients with molecular response or molecular response followed by either recrudescence or emergence had a significantly longer OS and PFS compared with individuals with molecular progression (median PFS 18.51 vs. 17.82 vs. 2.70 months, P = 0.001 and median OS 18.51 vs. 23.01 vs. 5.75 months, P = 0.04; Supplementary Fig. S3). Stratification of patients according to the presence of molecular response across any timepoint analyzed confirmed these findings (median PFS of 18.51 vs. 2.70 months, P = 0.0003 and median OS of 23.01 vs. 5.75 months, P = 0.01, respectively; Fig. 4E and F). These associations were consistent across patients who received single-agent immunotherapy and patients who received chemoimmunotherapy combination regimens (Supplementary Fig. S4), and were independent of PD-L1 expression (Supplementary Fig. S5). Furthermore, the relationship between molecular response and OS (HR = 0.15, 95% CI = 0.04–0.58, P = 0.006) and PFS (HR = 0.18, 95% CI = 0.06–0.56, P = 0.003) was statistically significant after accounting for clinical covariates in a multivariate Cox proportional hazards regression model (Supplementary Tables S7 and S8).

Importantly, we repeated these survival analyses in the subset of 11 patients with the best overall radiographic response classification of stable disease and detectable cfTL and found that molecular response was significantly associated with PFS (median PFS 23.01 vs. 5.75 months, P = 0.0075), with a trend toward longer OS (median OS 23.01 vs. 6.28 months, P = 0.09; Fig. 5A–D). We further evaluated the heterogeneity of radiographic stable disease by quantifying changes in tumor burden by RECIST 1.1 criteria. We assessed the association of ctDNA molecular response with changes in the sum of tumor longest diameter (SLD) in 10 patients with stable disease. These analyses revealed that the maximum percent change in SLD was significantly associated with molecular response (P = 0.02, Mann–Whitney U test; Fig. 5E). All 3 patients with molecular progression had an increase in SLD from baseline (mean +13.45%, range +8.97% to +18.87%). Two of 7 patients with a ctDNA molecular response had a minor increase in SLD from baseline, ranging from +1.3% to +2.3%. Of the remaining 5 molecular responders, 1 patient had 0% SLD change, whereas 4 patients had reductions in SLD, ranging from −7.1% to −28.2% (Fig. 5E). Although our findings indicate that the numeric SLD change better correlates with ctDNA dynamics, such that decreases in SLD are seen together with ctDNA molecular responses, 30% of the patients evaluated showed molecular response while attaining no change or an increase in SLD. Collectively, these findings indicate the utility of cfTL monitoring and molecular response classifications for capturing the magnitude of therapeutic responses in patients with radiographically stable disease.

To assess peripheral T-cell repertoire dynamics in relation to clinical response as well as the emergence of irAEs, we performed TCR sequencing analyses of 177 serial peripheral blood samples from 79 patients in two independent cohorts (Materials and Methods and Supplementary Tables S9 and S10). Paired analyses of peripheral TCR sequencing at baseline and on-treatment time points revealed no significant changes in TCR clonotype abundances according to either ctDNA molecular response or clinical benefit (Fig. 6A and B); however, significant on-treatment changes in circulating TCR clonotypes were observed in patients who experienced an irAE during therapy (Fig. 6C–F; Supplementary Figs. S6 and S7; Supplementary Tables S10 and S11). Patients who experienced an irAE harbored a higher fraction of both expanding and regressing TCR clones on-therapy, which was evident at the first timepoint sampled following treatment initiation (P = 0.02 and P = 0.02, respectively in set 1; P = 0.050 and P = 0.045, respectively in set 2; Fig. 6C–G; Supplementary Fig. S7; Supplementary Table S11), compared with individuals who did not develop an irAE. Among the 17 patients who developed an irAE, the timing of first on-therapy blood collections (median 3 weeks; range, 1–12 weeks), where significant TCR clonotypic expansions and regressions were observed, was prior to the clinical presentation of toxicity (median 18 weeks; range, 3–60 weeks) in 15 patients, with an average lead time of 20 weeks (Fig. 6H; Supplementary Fig. S8). In the remaining 2 patients, first on-therapy blood samples were collected shortly after clinical diagnosis of toxicity, precluding accurate lead time comparisons.

We next sought to characterize shared specificity among TCR clones and performed clustering based on global similarity in CDR3 sequences, using the GLIPH2 method and evaluated cluster dynamics across the combined (Sets 1 and 2) cohort (Materials and Methods). Expansions of multiple TCR CDR3 clusters were observed in on-treatment samples from patients who experienced an irAE (Supplementary Tables S12 and S13). Overall, 35 unique TCR clusters were significantly expanded across on-treatment samples compared with baseline in patients who developed an irAE, compared with individuals who did not develop immune-related toxicity. Analyses of global cluster dynamics in individual patients revealed additional clusters that were significantly expanded on-treatment on a per-patients basis (Supplementary Table S13). Next, we evaluated whether cluster dynamics could be used to identify similarities between TCR clonotypes that expanded for specific irAEs. As the most prevalent toxicity type across both cohorts was pneumonitis (n = 11 patients), we performed a targeted analysis of cluster dynamics across patients who developed pneumonitis compared with individuals who developed other irAEs (n = 16) and patients who did not develop any irAEs (n = 52) during immunotherapy treatment. These analyses revealed that 19 clusters (g.LRGTE, g.LSGANV, gG.GSQP, gR.GQGTE, gRD.NE, gRV.NTE, gSGG.NTE, gSLERN.P, gSLG.GNSP, gSLGGG, gSLIG.TE, gSLS.ANV. gSLSGAN., gSLT.TGE, gSPGQV.YG, gSPRG.TGE, gSRDS.TE, gSSRG.YG, and gSST.TGE) were significantly expanded on-therapy across patients who developed pneumonitis versus patients who did not develop toxicity, of which 4 (g.LRGTE, gSLERN.P, gSPGQV.YG, and gSSRG.YG) were preferentially expanded across patients with pneumonitis versus patients who developed other irAEs (Fig. 6I–J; Supplementary Fig. S8). Taken together, these findings suggest that TCR clonotypic dynamism in peripheral blood may be an indicator of the emergence of irAEs during immunotherapy and highlight the po—ntial for clustering of TCR CDR3 sequences to identify families of homologous TCR clonotypes that may serve as a biomarker to predict early onset of specific irAEs.

To further investigate phenotypic differences in peripheral immune cell subsets that may be associated with the onset of irAEs, we evaluated serial plasma protein expression profiles from 28 patients. Comparison between baseline and on-therapy samples revealed increased expression of multiple mediators of adaptive immune responses on-therapy, including interferon (IFN)-gamma (P = 0.0056, Mann–Whitney U test), TNF (P = 0.0003), and IL10 (P = 0.0001) cytokines, GZMA (P = 0.001), GZMH (P = 0.0003), the interferon-stimulated chemokine CXCL9 (P < 0.0001) and PDCD1 (P < 0.0001; Supplementary Table S14; Supplementary Fig. S9). Furthermore, in patients who experienced an irAE, we identified a significantly higher expression of proinflammatory mediators, including CCL19 (P = 0.007), GZMA (P = 0.017), and IL6 (P = 0.029), and T-cell markers, including CRTAM (P = 0.017) and the T-cell costimulatory molecule CD28 (P = 0.017), at baseline sampling (Supplementary Table S14). Similarly, at on-therapy time points, patients who experienced an irAE had significantly higher expression levels of inflammatory markers, including IL5 (P = 0.04), CCL23 (P = 0.01), MMP7 (P = 0.005), ANGPT2 (P = 0.02), and the inhibitory checkpoint PD-L2 (P = 0.02), compared with individuals who did not develop toxicity (Supplementary Table S14). Collectively, our findings indicate that differences in the phenotypes of immune cell subsets, reflected in both baseline and on-therapy plasma protein expression profiles, may serve as sentinel features to identify individuals most likely to develop immunotherapy-related toxicities.

The challenges of predicting clinical response to immunotherapy in lung cancer mandate the development and clinical integration of minimally invasive real-time biomarkers to capture clinical response and guide clinical decision-making (1–3, 5, 6, 34). Here, by analyzing cfTL dynamics during immunotherapy-containing regimens in patients with advanced NSCLC, we report that ctDNA molecular responses captured durable clinical benefit and differential clinical outcomes. These findings were independent of PD-L1 expression (35, 36) and treatment regimen. Focusing on patients with radiographically stable disease, we demonstrated that ctDNA molecular responses more accurately captured survival outcomes and the magnitude of therapeutic benefit in these cases, compared with imaging. Furthermore, our complementary assessment of the immune cell compartment revealed that peripheral blood T-cell characteristics and plasma proteomic profiles are associated with the development of immune-related toxicities, with potentially important implications for clinical cancer care.

ctDNA-driven early on-therapy response evaluation can enable real-time tumor burden assessment, with a growing number of studies supporting a potential role of ctDNA molecular response as an early endpoint for immunotherapy response (10, 11, 15, 37). Leveraging several smaller-sized independent cohorts, the Friends of Cancer Research ctDNA for Monitoring Treatment Response (cMoniTR) project showed a consistent-across-cohorts association between ctDNA response and clinical outcomes (7). Despite an increasing number of retrospective studies, there are several areas that require further investigation, starting with a unified definition of liquid biopsy-based molecular response. On-treatment ctDNA decreases based on ctDNA clearance or ctDNA level reduction below given cutoffs have all been associated with clinical response (37–39). Molecular progression—or the increase or lack of significant decrease in ctDNA variants on-treatment—has potentially greater sensitivity in predicting clinical progression (10, 11, 39). Our study supported the notion that cfTL clearance (or ctDNA decrease below the level of detection of the assay used in this study) is robustly associated with clinical outcomes. Furthermore, the importance of time to molecular response assessment in the setting of immunotherapy—ranging from as early as 3 weeks (39) in NSCLC to up to 12 weeks (37) in various solid cancers—remains unsettled. With respect to the NGS strategy, we employed a rigorous tumor and WBC DNA informed NGS strategy, to circumvent limitations related to biological noise due to clonal hematopoiesis (10, 12, 40). Although current commercial clinical liquid biopsy testing mostly follows a plasma-only approach for NSCLC genotyping purposes, incorporation of complementary sources of NGS sequencing has demonstrated benefit in NSCLC (41).

ctDNA approaches hold unique promise in evaluating outcomes for patients with radiographic stable disease, which represents a heterogeneous entity with respect to clinical outcomes. Our findings demonstrate the value of ctDNA-based molecular response assessments in distinguishing between differential clinical outcomes in patients with radiographic stable disease. Despite the observed association between numeric RECIST SLD changes and ctDNA dynamics, 30% of patients with stable disease in this study had a molecular response while attaining no change or an increase in RECIST SLD. These findings add to the challenges with radiographic response assessments, where misclassification rates of up to 30% have been reported for progressive disease at the current threshold of ≥20% SLD increase (42). RECIST measurement variability can have a greater impact on the overall accuracy of radiographic response assessment in patients with smaller tumor burdens, highlighting the value of using molecularly informed approaches to capture tumor burden.

Furthermore, and as seen in 2 patients in our cohort who remained on immunotherapy despite a first radiographic assessment of disease progression, ctDNA dynamics may be informative in providing prompt clarification of indeterminate imaging where radiographic assessments alone may fail to accurately capture tumor regression, such as in cases of pseudoprogression (43). Similarly, treatment de-escalation in patients who attain a molecular response indicated by ctDNA clearance may reduce the personal and financial burden of overtreatment. Elective immunotherapy discontinuation past 2 years is well supported (44–46) and proof-of-concept studies have shown that sustained ctDNA clearance may identify candidates for immunotherapy de-escalation (47); nevertheless, there are currently no molecularly informed approaches to guide treatment de-escalation. Translating these findings to the clinic, early on-therapy ctDNA persistence prior to and independent of the radiographic response can identify patients not attaining therapeutic benefit, who can be rapidly and accurately guided to alternate therapies. In the targeted therapy space and for patients with EGFR-driven NSCLC, the randomized phase II APPLE trial showed that early intervention based on longitudinal ctDNA dynamics over radiographic imaging does translate into a significant survival benefit (48). These early findings support the survival benefit following ctDNA intervention and the outcomes from ongoing prospective clinical trials will be essential to determine whether ctDNA molecular responses can provide similar clinical and regulatory utility as an early endpoint of response in the context of immunotherapy. To this end, several ongoing ctDNA-driven interventional clinical trials are investigating whether ctDNA molecular responses can be used to guide therapy for patients with advanced or metastatic NSCLC receiving immunotherapy (NCT04093167, NCT04166487, and NCT04966676). The BR.36 clinical trial, in particular, specifically investigated the concordance between ctDNA molecular responses and radiographic responses in stage 1 (12) and will soon start recruiting patients for the second ctDNA interventional randomized stage. Furthermore, ctDNA clearance early on during therapy could be implemented in drug development and inform go/no-go decisions in early-phase clinical trials.

In addition to the monitoring of clinical responses to immune-checkpoint blockade, accurate and timely identification of immunotherapy-related toxicity to guide early intervention remains an unmet clinical need. Biological drivers underpinning the development of irAEs remain poorly characterized and there are no standard clinical biomarkers to identify high-risk patients. We demonstrated that peripheral blood TCR repertoire reshaping and elevated plasma protein expression of proinflammatory cytokines and chemokine markers are associated with the emergence of irAEs, including pneumonitis. The degree of TCR clonotypic expansions and regressions early during treatment was driven by clusters of TCRs sharing similar CDR3 sequences and potential specificity for target antigens, consistent with the possibility of an underlying self-reactive immunologic mechanism. Further investigation of these results in larger pan-cancer cohorts is required to assess whether peripheral TCR features can distinguish between different irAE grades (22, 49), determine the relative contributions of different immune cell subtypes to these dynamics, and evaluate whether these findings generalize to the risk of immunotherapy-related toxicity in other solid tumor types.

Ultimately, monitoring TCR clonotype dynamics following treatment initiation may enrich for patients at risk for the development of severe immune-related toxicities and provide an adjunct for current clinical management strategies that are based on intervention following symptom recognition and the onset of irAEs (49). The early manifestations of irAEs can be subtle, often presenting challenges for accurate diagnosis. In such cases, the detection of increasing TCR clonotype frequencies early during immunotherapy treatment may serve as an indicator to enable appropriate monitoring and management, which will be important to curtail the morbidities associated with severe irAEs (24, 50). Timely intervention may further shorten the overall duration of immunosuppressive treatment required to reverse the effects of immune-related toxicity. A growing number of studies have reported a potential association between the onset of irAEs and the overall efficacy of immune-checkpoint inhibitors (51–53). Notably, in our study, although TCR repertoire shifts were detected in a few patients who attained clinical benefit and also developed an irAE, overall, peripheral T-cell dynamics did not correlate with clinical response.

Our study has several limitations including its retrospective design, sample size, and heterogeneity of treatment regimens. Furthermore, 4 of 30 patients (13%) were not evaluable for ctDNA molecular response due to undetectable ctDNA. While bespoke tumor-informed liquid biopsy approaches may provide greater sensitivity for ctDNA detection compared with tumor-agnostic hybrid capture panel NGS methods, such approaches may not be feasible in the metastatic setting where tissue specimens are limited, especially in the context of longitudinal ctDNA monitoring. Finally, rates of pneumonitis in this cohort appeared to be numerically higher compared with previously published data (54); this was driven by the longer follow-up time of our cohort (pneumonitis occurring between 1 and 13 months in those affected), likelihood of additional comorbidities modifying risk, and an institutional “IR Tox Team” with multidisciplinary support and rapid evaluation of irAEs (55), wherein pneumonitis was identified as the most common irAE.

Taken together, our study affirms the role of ctDNA-based molecular response as a robust predictor of clinical outcomes with immunotherapy in patients with NSCLC. Furthermore, we highlight the value of monitoring the peripheral T-cell repertoire and immunoproteomic profiling in assessing a patient's risk of immunotherapy-related toxicity. Building on these findings, we anticipate that combinatorial liquid biopsy-based tests can aid clinical decision-making toward adaptive treatment regimens that improve patient outcomes, by rapidly and accurately capturing clinical outcomes while at the same time limiting untoward side effects.

J.C. Murray reports grants from the Conquer Cancer Foundation (ASCO) during the conduct of the study and personal fees from Janssen and MJH Life Sciences outside the submitted work. J.R. White reports other support from Resphera Biosciences LLC outside the submitted work. B. Weksler reports grants from Atricure and other support from AstraZeneca and Intuitive Surgery outside the submitted work. N. Bahary reports personal fees from BMS, Astra Zeneca, Taiho, Pfizer, and Incyte outside the submitted work. J. Phallen reports other support from Delfi Diagnostics outside the submitted work; in addition, J. Phallen has a patent for detection of cancer pending, a patent for cell-free DNA for assessing and/or treating cancer pending and issued to Delfi Diagnostics, a patent for detection of lung cancer using cell-free DNA fragmentation pending, licensed, and with royalties paid from Delfi Diagnostics, a patent for matched white blood cell and cell-free DNA analyses for prediction of therapeutic response in patients with cancer pending, a patent for methods and compositions for analyses of cancer pending, a patent for noninvasive detection of response to a targeted therapy pending and licensed to UCSD, a patent for noninvasive detection of early-stage cancers pending, and a patent for genomic alterations in the tumor and circulation of pancreatic cancer patients pending and issued. A. Leal reports personal fees from Delfi Diagnostics, Inc. outside the submitted work. K.A. Marrone reports grants and nonfinancial support from Bristol Myers Squibb, grants and personal fees from Mirati, and personal fees from Regeneron, Janssen, Daiichi Sankyo/Lilly, Merck, Puma Biotechnology, AstraZeneca, and Amgen outside the submitted work. J. Naidoo reports grants and other support from Merck, AstraZeneca, BMS, and Roche/Genentech, other support from Elevation Oncology, Bayer, Regeneron, AbbVie, and Daiichi Sankyo, and grants and other support from Mirati, Amgen, and Kaleido Biosciences during the conduct of the study. B. Levy reports personal fees from Janssen, Daiichi Sankyo, AstraZeneca, Pfizer, Merck, Elli Lilly, Genentech, Takeda, Guardant 360, and BMS outside the submitted work. S. Rosner reports other support from MJH, EMpartners, Axion, ASCO Advantage, Life Sciences FGI, and Cardinal Health and grants from Conquer Cancer outside the submitted work. C.L. Hann reports grants and personal fees from AstraZeneca, BMS, and Amgen and personal fees from Puma, Daiichi Sankyo, and Janssen during the conduct of the study. S.C. Scott reports grants from Mirati Therapeutics and Janssen and personal fees from Genentech, AstraZeneca, Tempus, Foundation Medicine, and Regeneron outside the submitted work. J. Feliciano reports grants from Astra Zeneca, other support from Astra Zeneca and Eli Lilly, and grants from Bristol-Myers, Regeneron, Takeda, Merck, Coherus, and Pfizer outside the submitted work. V.K. Lam reports personal fees from Anheart Therapeutics, Takeda, Seattle Genetics, BMS, and AstraZeneca outside the submitted work; and has received research funding from GlaxoSmithKline, Bristol Myers Squibb, Astra Zeneca, Merck, and Seattle Genetics. P.B. Illei reports personal fees from AstraZeneca, Sanofi, AbbVie, Janssen, Roche, and Genentech, grants and personal fees from Bristol Myers-Squibb, and grants from Erbe Medical Systems outside the submitted work. K. Monkhorst reports personal fees from Lilly, Bayer, and Amgen outside the submitted work. R.B. Scharpf reports grants and personal fees from Delfi Diagnostics outside the submitted work. J.R. Brahmer reports personal fees from Merck and grants and personal fees from Bristol Myers Squibb, AstraZeneca, Regeneron and Genentech/Roche during the conduct of the study; personal fees from Amgen, Eli Lilly, Incyte, Sanofi, Janssen, Incyte, and GlaxoSmithKline outside the submitted work. V.E. Velculescu reports grants and personal fees from Delfi Diagnostics during the conduct of the study; in addition, V.E. Velculescu has a patent for various licenses and with royalties paid from Delfi Diagnostics; and is a founder of Delfi Diagnostics, serves on the Board of Directors and as an officer for this organization, and owns Delfi Diagnostics stock, which is subject to certain restrictions under university policy. Additionally, Johns Hopkins University owns equity in Delfi Diagnostics. V.E.V. divested his equity in Personal Genome Diagnostics (PGDx) to LabCorp in February 2022. V.E.V. is an inventor on patent applications submitted by Johns Hopkins University related to cancer genomic analyses and cell-free DNA for cancer detection that have been licensed to one or more entities, including Delfi Diagnostics, LabCorp, Qiagen, Sysmex, Agios, Genzyme, Esoterix, Ventana and ManaT Bio. Under the terms of these license agreements, the university and inventors are entitled to fees and royalty distributions. V.E.V. is an advisor to Viron Therapeutics and Epitope. These arrangements have been reviewed and approved by Johns Hopkins University in accordance with its conflict-of-interest policies. P.M. Forde reports grants and personal fees from AstraZeneca, Novartis, BMS, and Regeneron, grants from BioNTech, and personal fees from Amgen, AbbVie, Genentech, Merck, Inovox, Seagen, Ascendis, Curevac, G1, Genelux, Gritstone, Janssen, Sanofi, Fosun, Teva, Synthekine, Iteos, and Tavotek outside the submitted work; in addition, P.M. Forde has a patent for PCT/US2022/079403 pending. V. Anagnostou reports grants and personal fees from Astra Zeneca, grants from Bristol Myers Squibb, Delfi Diagnostics, and Personal Genome Diagnostics, and personal fees from Neogenomics and Foundation Medicine outside the submitted work; in addition, V. Anagnostou has a patent for 63/276,525 issued, a patent for 17/779,936 issued, a patent for 16/312,152 issued, a patent for 16/341,862 issued, a patent for 17/047,006 issued, and a patent for 17/598,690 issued. No disclosures were reported by the other authors.

J.C. Murray: Conceptualization, formal analysis, investigation, writing–review and editing. L. Sivapalan: Data curation, formal analysis, investigation, writing–original draft, writing–review and editing. K. Hummelink: Writing–review and editing. A. Balan: Data curation, formal analysis, writing–review and editing. J.R. White: Formal analysis, writing–review and editing. N. Niknafs: Formal analysis, writing–review and editing. L. Rhymee: Investigation. G. Pereira: Investigation. N. Rao: Investigation. B. Weksler: Data curation. N. Bahary: Data curation. J. Phallen: Writing–review and editing. A. Leal: Investigation. D.L. Bartlett: Writing–review and editing. K.A. Marrone: Writing–review and editing. J. Naidoo: Writing–review and editing. A. Goel: Data curation. B. Levy: Writing–review and editing. S. Rosner: Writing–review and editing. C.L. Hann: Writing–review and editing. S.C. Scott: Writing–review and editing. J. Feliciano: Writing–review and editing. V.K. Lam: Writing–review and editing. D.S. Ettinger: Methodology, writing–review and editing. Q.K. Li: Methodology, writing–review and editing. P.B. Illei: Resources, methodology, writing–review and editing. K. Monkhorst: Resources, methodology, writing–review and editing. R.B. Scharpf: Resources, writing–review and editing. J.R. Brahmer: Resources, supervision, writing–review and editing. V.E. Velculescu: Conceptualization, resources, supervision, funding acquisition, writing–review and editing. A.H. Zaidi: Resources, writing–review and editing. P.M. Forde: Conceptualization, resources, supervision, funding acquisition, writing–review and editing. V. Anagnostou: Conceptualization, resources, formal analysis, supervision, funding acquisition, visualization, writing–original draft, writing–review and editing.

We would like to thank Kellie Smith and Li Zhang in the FEST and TCR Immunogenomics Core as well as the Lung Cancer Precision Medicine Center of Excellence Investigators, Sharon Penttinen, Michael Conroy, Durrant Barasa, Mary Anderson, Mariam Ghobadi-Krueger, Kenneth Harkness, Kerry Smith, Sreenivasa Idamakanti, and Vivian Altiery De Jesus for their contributions. We would also like to acknowledge the contributions of the Hopkins Thoracic and Allegheny Health Network Cancer Institute Pan-Tumor Biorepository groups. This work was supported in part by the NIH grants CA121113, CA062924, and CA006973 (V. Anagnostou, R.B. Scharpf, and V. Velculescu), the Bloomberg-Kimmel Institute for Cancer Immunotherapy (V. Anagnostou, J.R. Brahmer, and P.M. Forde), the ECOG-ACRIN Thoracic Malignancies Integrated Translational Science Center grant UG1CA233259 (V. Anagnostou, V. Velculescu), the Emerson Collective Cancer Research Fund (V. Anagnostou), the V Foundation (V. Anagnostou and V. Velculescu), the LUNGevity Foundation (V. Anagnostou and V. Velculescu), the Hopkins-Allegheny Health Network (AHN) Cancer Research Fund (V. Anagnostou and A. Zaidi), the Conquer Cancer Foundation (J.C. Murray), a MacMillan Scholars Award (J.C. Murray), the Maryland Cigarette Restitution Fund (J.C. Murray), the International Lung Cancer Foundation (L. Sivapalan), and the Robyn Adler Fellowship Award (L. Sivapalan).

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.