This is a single-arm, multicenter phase II trial designed to assess the combination of avelumab with cetuximab and mFOLFOX6 in patients with previously untreated RAS/BRAF wild-type (wt), MSI independent mCRC. Main inclusion criteria were: ≥18 years of age; Eastern Cooperative Oncology Group performance status 0 or 1; and no previous chemotherapy for metastatic disease (adjuvant chemotherapy allowed if terminated more than 6 months before trial start). The trial was conducted at 10 centers in Germany after approval by local ethics committees and competent authority and registered. All participants provided written informed consent. Patients or public were not involved in the design of the trial. The AVETUX regimen applied avelumab at a dose of 10 mg/kg intravenously over 60–90 min (biweekly from cycle 2 onward), cetuximab at a dose of 250 mg/m² intravenously over 60–90 min (weekly, first dose 400 mg/m²) and a modified FOLFOX6 with oxaliplatin at a dose of 85 mg/m² intravenously (day 1), 5-FU 400 mg/m² intravenously bolus (day 1) and 5-FU 2400 mg/m² intravenously continuous infusion (days 1–2), and LV at a dose of 400 mg/m² intravenously. The primary endpoint was progression-free survival (PFS) rate at 12 months (according to RECIST V.1.1); secondary endpoints were overall response rate (ORR), early tumor shrinkage, progression-free and overall survival (OS) as well as toxicity. A sample size of 41 patients achieved 80% power to detect a positive signal of improvement by 17% compared with the PFS assumption for standard treatment at a one-sided alpha error level of 0.1.

Twenty mL peripheral blood (STRECK cell-free DNA BCT tubes) were obtained every 4–16 weeks for translational research. Paraffin-embedded tissue obtained during surgical removal or biopsy before treatment initiation as well as—if applicable—on treatment was collected.

Immunohistochemistry (IHC) for PD-L1 was performed with the Bond Polymer refine detection Kit (Leica) to deduce tumor proportion score (TPS) and immune cell (IC) score. As described elsewhere,12 IHC was performed with antibodies directed against PanCK, CD3 and PD-L1 or CD56 and CD16, and multispectral imaging with antibodies against CD3, CD8, CD20, CD163, Foxp3 and PanCK (online supplemental table S1).

Amplification of the T cell receptor beta chain (TRB) repertoire from circulating or tumor-infiltrating lymphocytes (TiLs) was done as described elsewhere.13–19 Sequencing and demultiplexing was performed on the Illumina MiSeq platform with 2×301 cycles at a coverage of 80 000 reads per sample. Analysis of the TRB locus was computed using the MiXCR analysis tool V.3.0.5.20. Analyses were performed using R21 and the package tcR.22 Peripheral blood T cell diversification was calculated as follows: Shannon diversity index at week 4 minus Shannon diversity index at baseline. No delta or a negative delta were interpreted as not diversified, whereas a positive delta was interpreted as diversified.

For mutational profiling of tumor or circulating tumor DNA (ctDNA; liquid biopsy), up to 100 ng of tumor Formalin-Fixed Paraffin-Embedded (FFPE) tissue DNA or ctDNA were used. Target genes with known pathogenic or resistance variants (cBio Cancer Genomics Portal23) as well as PD-L1 were selected for the gene panel, and sequencing libraries were constructed using QIAseq Targeted DNA Custom Panels, Qiagen (online supplemental table S2). Sequencing was performed on the Illumina NextSeq or HiSeq platform with 2×151 cycles at a coverage of 26 500 reads per target region. Variant calling of unique molecular identifier based data was performed using smCounter2 as described elsewhere.24 Tumors were classified as mutation positive, if the mutation was found in either tumor tissue or liquid biopsy at a variant allele frequency (VAF) exceeding 10% or if the mutation was concordantly found in tumor tissue and liquid biopsy independently of VAF. Follow-up liquid biopsies were classified mutation positive if the same mutation was found repeatedly in one patient independently of VAF. All variants that were found are listed in online supplemental table S3.

ddPCR custom assays were designed by and ordered from IDT (Iowa, USA). To increase specificity locked nucleic acid bases were incorporated in the probes. Probes were HPLC purified and contained either a 5′-HEX (wildtype probe) or a 5′-FAM (mutant probe) reporter dye and a 3′ Iowa Black Fluorescent quencher. Probe and primer sequences are listed in online supplemental table S4.

Each reaction (22 µL) contained 11 µL 2xddPCR SuperMix for probes (no UTP) (Biorad), template DNA and primers and probes at a final concentration of 1.8 µM and 500 nM, respectively. After droplet generation (AutoDG, Biorad), the plate was sealed with a pierceable foil heat seal (PX1 PCR plate sealer, Biorad), and the PCR was performed on a C1000 Touch thermal cycler (Biorad). The cycling conditions were as follows: 95°C, 10 Min; [94°C, 30 s; 59°C, 1 min] ×45; 98°C, 10 min; 12°C, 30 min; 4°C hold. Prior to the read-out on QX200 droplet reader (Biorad), the plate was left at room temperature for 10 min.

ddPCR data were analyzed using the QX Manager 1.0 Standard Edition (Biorad).

Assay specificity was tested on gBlocks (IDT). Thresholds for positive/negative droplets were set manually for each assay according to the positive controls (gBlocks, IDT, online supplemental figure S1).

Novel PD-L1 variants were functionally validated in five cell lines, including Ba/F3 (PD-L1 negative murine pro B cell line), UT-SCC-14 (tongue squamous cell carcinoma), UT-SCC-29 (laryngeal squamous cell carcinoma), HT-29 (colorectal adenocarcinoma, KRAS wt) and DLD-1 (colorectal adenocarcinoma, KRAS G13D) (German Collection of Microorganisms and Cell Cultures GmbH) after exclusion of Mycoplasma contamination. Endogenous PD-L1 was depleted using the CRISPR plasmid pU6 containing gRNAs complementary to nts 508–527 in exon 4 (gRNA1) and nts 339–358 in exon 3 (gRNA2) of PD-L1 (NM_014143.4) and verified by immunoblotting and flow cytometry. For ectopic re-expression of PD-L1 variants, the wt coding sequence was cloned into the Lentiviral Gene Ontology (LeGO) vector LeGO-iC2-Puro+via AsiSI/EcoRI. The pLeGO-PD-L1 wt construct was used as template to generate the PD-L1 L88S (c.263T>C) by site-directed mutagenesis. The PD-L1 K162fs variant was synthesized and cloned into pLeGO-iC2-Puro+. Generation and transduction of lentiviral particles as well as selection of transduced cell was performed as described elsewhere.25 Cell lysates were generated using RIPA buffer supplemented with protease and phosphatase inhibitors (both Roche). Membrane fractions were generated using the ProteoExtract Native Membrane Protein Extraction Kit (Merck). Selective enrichment of membrane proteins was controlled via detection of prohibitin 1 and 2 (membrane) as well as NFκB subunit p65 (non-membrane) proteins (online supplemental table S1).

UT-SCC-14 cells were incubated with 2 µg of avelumab and complement for 3 hours. Cetuximab in combination with matuzumab (1 µg each) served as positive control. Propidium iodine flow cytometry staining was used to analyze CDC.

NK cell degranulation was evaluated in co-culture assays with PD-L1 variant cell lines by flow cytometry detection of CD107a.26 NK cells from healthy volunteers were isolated via negative selection (RosetteSep Human NK Cell Enrichment Cocktail, Stemcell Technologies) and cultivated overnight in medium containing 1 ng/mL IL-15. NK cells were added to the respective target cells at 1:1 ratio in the presence of 1.2 µL CD107a-PE-Cy7, 1 ng/mL IL-15, 5 µg/mL brefeldin A and 2 µg avelumab, 2 µg cetuximab or 2 µg human IgG (Intratect) for 5 hours, washed and stained with anti-CD3-FITC/anti-CD56-PE (online supplemental table S1). NK cell-mediated target cell killing was also quantified using the Cytotoxicity Detection Kit (LDH) from Roche following the manufacturer’s instructions.

T cell mediated tumor cell killing was assessed using the Incucyte Caspase-3/7 Reagent (Essen BioScience) according to the manufacturer’s instructions. Untouched CD8+ T cells were isolated from pooled PBMCs of five healthy donors using the CD8+ T Cell Isolation Kit (Miltenyi) and quantified using a Vi-CELL XR (Beckman Coulter). Five thousand HT-29 cells were seeded in triplicates in a 96-well flat bottom plate and cocultured with 50 000 CD8+ T cells. For T cell activation, the medium was supplemented with 10 ng/mL recombinant IL-2 (BioLegend), 100 ng/mL anti-CD3 (BioLegend) and 100 ng/mL anti-CD28 (BioLegend) (online supplemental table S1). Read out of dead (=green-fluorescent) target cells was performed on an Incucyte S3 after 24 hours.

Total RNA was isolated using the Quick-RNA kit (Zymo Research) and reverse transcribed with SuperScript III (Thermo Fisher). Target amplification was performed on the CFX96 System (Bio-Rad) using the SYBR Select Master Mix CFX and primers for PD-L1 (forward: TGCTGAACGCATTTACTGTCAC; reverse: TCTGTCTGTAGCTA CTATGCTG), mCherry (forward: AGGAGGATAACATGGCCATCAT; reverse: ACCC TTGGTCACCTTCAGCT) and HPRT1 (forward: TGACACTGGCAAAACAATGCA; reverse: GGTCCTTTTCACCAGCAAGCT). The relative expression levels were normalized to HPRT1 and calculated according to the comparative Ct (∆∆Ct) method.

To assess the relevance of NMD for transcript turnover of the detected PD-L1 variants, NMD was inhibited using 20 mM caffeine for 4 hours. Selective enrichment of NMD-targeted PD-L1 transcripts was detected (semi)quantitatively using (q)RT-PCR as described previously.

Cycloheximide (CHX) and tunicamycin (Tm) were obtained from Santa Cruz Biotechnology, Compound C/dorsomorphin and MG132 from Selleck Chem. To analyze degradation kinetics of PD-L1 variants, 2 million cells were seeded in 6-well format and either incubated with 100 µM CHX for 3, 6 and 9 hours, with 10 µg/mL Tm for 18 hours, with 10 µM Compound C for 2 and 4 hours or with 20 µM MG132 for 4 hours. Proteins were isolated as described previously and subjected to immunoblotting.

Between June 2017 and July 2018, the AVETUX trial enrolled 43 patients, 41 thereof with MSS or MSI-low tumors. Patients 22 and 26 were MSI-high. Four patients were excluded after central tissue review with (mostly lower frequency) RAS (n=3) or BRAF mutations (n=1) (figure 1A). For eligible patients (n=39), median age was 62 (range 29–82), 36 patients (92%) had left sided primary tumors and 30 patients (77%) had liver metastases. Prior adjuvant treatment rate was high with overall 31% (12/39) and oxaliplatin in 23% (9/39) of patients. Patient characteristics are summarized in table 1.

Median follow-up was 21.2 months. The median number of cycles administered (at least one study component) was 18 cycles (range 1–35), with a median of 8 cycles of oxaliplatin, 12 of cetuximab and 16 of avelumab. The combination of avelumab and cetuximab was applied for a median time of 5.4 months.

The Kaplan-Meier estimation of PFS from the time point of protocol treatment start is shown in figure 1B, which is based on a total of 31 observed events (79%) in the intention-to-treat population of 39 patients. The PFS rate at 12 months was 40% and the median PFS was 11.1 months (95% CI 9.6 to 13.5 months). Based on the high rate of prior oxaliplatin (23%) and early non-treatment related drop-outs, an exploratory analysis excluding these patients showed a PFS of 13.2 months. The overall response rate in the RECIST evaluable population (n=37) was 81% (95% CI 65% to 92%) (MSS/MSI-low 83%) with four complete responses (11%) and a disease control rate of 89% (figure 1C,D). Early tumor shrinkage of at least 20% during the first 8 weeks was observed in 30 patients (81%). The current OS Kaplan-Meier curve estimates are shown in figure 1E, which plateau at 77% at a median follow-up of 21.2 months. Until the data cut-off in September 2019, only eight events occurred. Secondary resection of metastases was achieved in 6 out of 39 patients (15%) of which five stopped treatment after surgery. Treatment-related adverse events are summarized in table 2, revealing no unexpected findings or overlapping toxicities.

Since ICB represented the experimental component of the treatment regimen, we studied potential immunological biomarkers of response.

We did not observe any association of PD-L1 TPS with PFS in our patient cohort (online supplemental figure S2). Yet, this analysis was hampered due to low tumor cell count (>30% of samples with less than 100 tumor cells), lack of the invasion front (>25% biopsy specimens)27 28 (online supplemental figure S3) and discordant PD-L1 staining with different antibody clones in 15% of samples (online supplemental table S5).

Next, we explored blood-circulating and tumor-infiltrating T cells (TiLs) by next-generation immunosequencing to deduce potentially predictive immune signatures based on diversification (related to an increase in Shannon diversity index) of the peripheral blood T cell space.13 29 30 In contrast to prior data from monotherapy trials, our combinatorial approach did not reveal any association with PFS (online supplemental figure S4). Tissue deep sequencing revealed a broad range of T cell infiltration levels across the cohort (online supplemental figure S5A). Since no defined threshold is available in literature, we defined 0.5% as cut-off for high (n=14) versus low (n=28) TiL levels. Interestingly, four of five patients with febrile reactions to ICB belonged to the subset of patients with high levels of T cell infiltration, and high TiL levels showed a trend toward improved PFS (p=0.07, online supplemental figure S5B). The TiL composition was rather heterogeneous both across the cohort and in different portions of the same tumor as analyzed by multispectral imaging with some patients showing CD8+ biases, some CD4+ biases and others showing T cell infiltration biased toward Tregs (online supplemental figure S6).

The complete study cohort was subjected to tumor tissue and liquid biopsy testing at baseline using a gene panel covering the most frequent driver and resistance mutations in mCRC (online supplemental table S2). In 40 out of 43 patients, mutations in the baseline tumor tissue and liquid biopsy screening were identified, which overlapped in 36 patients in at least one gene mutation (figure 2A,B). The vast majority of these patients turned negative or clearly reduced their circulating tumor DNA (ctDNA) levels to <1% after 4 weeks of treatment (figure 2C). Next, serial liquid biopsies were assessed as a predictor of disease progression. Of 26 cases with ctDNA positivity and progressive disease during the observation period, reappearance or increase of ctDNA preceded disease progression by a median of 8.5 weeks (mean 13 weeks) in those who initially reduced their ctDNA levels to <1% or turned ctDNA negative (figure 2D).

Beside baseline RAS mutations excluding the application of EGFR antibodies, recent studies on EGFR inhibition found high rates of acquired secondary resistance mediating downstream RAS mutations or EGFR ectodomain mutations disrupting the antibody’s epitope.25 31–36 While EGFR ectodomain mutations may appear in up to 20% of patients during EGFR directed treatment, acquired RAS mutations have been reported more frequently (32%–44% of cases).37–40 In our cohort, five patients showed (mainly low level) RAS mutations at baseline confirmatory central testing (patients 5, 12, 15, 34 and 42), leading to the post hoc exclusion due to violation of selection criteria in 2 patients (>10% mutant allels in tissue). Patients 5, 34 and 12 cleared their KRAS mutant ctDNA from blood in the course of treatment with patients 5 and 12 even achieving early CR (figure 2E). Patient 42’s KRAS Q131H mutation (VAF 0.14% in liquid biopsy at baseline) did not steadily expand on treatment but showed a rather fluctuating course not exceeding VAFs of 10% and the patient achieved early CR and long-term disease control (figure 2E). Patient 43 secondarily developed a classical NRAS mutation shortly after withdrawal of oxaliplatin. While the initial liquid biopsy was negative for NRAS Q61L, this mutation steadily increased on treatment, and within 7 weeks from first detection of this mutation, the patient progressed clinically (figure 2E). In line with the exceedingly low rate of emerging RAS mutations (only one in 43 patients with emerging NRAS mutated ctDNA levels and progression), there was no evidence for acquired EGFR ectodomain mutations on this protocol.

To search for potentially emerging avelumab-resistant tumor subclones, we used a liquid biopsy panel including genes known to be involved in resistance41 (β2-microglobulin (B2M), Janus kinase 1 (JAK1)). Moreover, proposing direct antitumor effects as mechanism of action of avelumab (ADCC), we covered the whole coding region of PD-L1 to be able to detect potential escape variants that may evade avelumab binding by epitope disruption or impaired membrane expression. In addition, we included regions within the Fcγ receptor (FcγR) genes with known single nucleotide polymorphisms (SNPs) associated with high-affinity toward IgG1 that have been described to allow more potent NK cell ADCC.11

Indeed, NGS of ctDNA or on-treatment tumor samples (n=5) identified patients with mutations in the previously mentioned genes (figure 3A). Patient 22 showed both a B2M L15fs and a JAK1 H183Y mutation in on-treatment tumor biopsies. The B2M L15fs mutation was truncating, and the JAK1 H183Y was compatible with loss-of-functionality as in previously reported resistance variants (figure 3B). Interestingly, both on-treatment tumor samples (primary and metastasis) showed the B2M L15fs mutation, while the JAK1 H183Y mutation was only detectable in the metastasis and with a different VAF. In patient 12, a B2M L15fs mutation was detectable by NGS already prior to treatment.

Moreover, we found two selected PD-L1 mutations in three patients of our cohort. Both mutations were located within the extracellular region of PD-L1, outside both the PD-L1/PD-1 interaction site and the avelumab epitope on PD-L1 (figure 3A,B). The PD-L1 L88S variant was a missense mutation, while the PD-L1 K162fs mutation led to a frameshift with premature stop codon before the transmembrane domain. The truncated variant PD-L1 K162fs was found in patient 21 by liquid biopsy analysis with allele frequencies at the detection limit in the course of treatment. VAF increased substantially at disease progression suggesting that this mutation was involved in resistance to treatment (online supplemental table S3). Interestingly, the PD-L1 L88S mutation occurred in patient 22 who also showed B2M L15fs and JAK1 H183Y mutations. It was undetectable in the tumor tissue by NGS analysis before treatment initiation as well as in leukocyte DNA of this patient. In patient 30, the same missense mutation PD-L1 L88S was found at end of treatment at the time of disease progression. Distribution of all mutations selected on treatment are summarized in figure 3C.

We used ddPCR to validate our findings and could confirm all four variants using specific custom assays (online supplemental table S6 and online supplemental figure S1). Since this method is more sensitive than NGS, we exemplarily reanalyzed the baseline tumor samples of patient 22 who showed three ICB resistance mutations in on-treatment tumor tissue. Interestingly, while the baseline tumor sample of patient 22 was negative for these mutations by NGS, we identified small frequencies of the PD-L1 L88S, B2M L15fs and JAK1 H183Y variants by ddPCR suggesting the presence of these small subclones already at baseline and a subsequent selection of those clones on the AVETUX regimen (figure 3D).

Of note, patients 21 and 22 with PD-L1 mutations were the only patients in the cohort homozygously expressing the V-allele of FcγR3a (rs396991) that is associated with high IgG1 affinity while patient 30 expressed the V-allele in a heterozygous fashion (figure 3C; online supplemental table S7). Association of the SNP with PD-L1 mutations as well as a trend toward improved PFS in patients homozygous or heterozygous for the V-allele (p=0.07, figure 3E) suggested that this may be a subset of patients experiencing strongest selective pressure by avelumab, in addition to the known higher susceptibility to cetuximab.42 43 In patients 21 and 30, tumor material or ctDNA was available after progression on the AVETUX protocol several months after avelumab withdrawal. Interestingly, in both patients, PD-L1 mutant subclones (but not the tumor defining mutations) decreased in size or fully disappeared suggesting low biological fitness of these clones when the selective pressure exerted by avelumab was withdrawn (figure 3F).

Together, their selection on treatment, their localization outside the avelumab epitope and their association with the FcγR3a V-allele suggested that the identified PD-L1 mutations may mediate immune escape toward avelumab through loss of the PD-L1 antigen on tumor cells. Since in patient 22 on-treatment tumor tissue was available, we performed PD-L1 immunohistochemistry (IHC) on this tissue and compared the results with pretreatment tumor tissue of the same patient (harboring only a minimal frequency of PD-L1 mutated clones). In line with our hypothesis, this patient showed a selective loss of PD-L1 surface staining on tumor cells while immune cells were clearly PD-L1 positive after avelumab exposure (figure 4). To definitively exclude epitope disruption by the mutation, we performed western blot analyses using clone QR1 used for tissue staining on cell lines strongly overexpressing PD-L1 L88S. This experiment confirmed that this antibody clone—recognizing an intracellular epitope of PD-L1—recognizes the PD-L1 L88S variant (online supplemental figure S7).

To further explore the functional role of the detected PD-L1 alterations, we generated five different cell models expressing these variants using CRSIPR/Cas9 and lentiviral transduction (figure 5A,B). Ectopic expression of PD-L1 L88S led to a transcription-independent reduction of full-length glycosylated PD-L1 protein in subcellular membrane fractions as compared with the wt, while cell lines expressing the truncating K162fs mutation did not yield a PD-L1 signal on immunoblots (figure 5B–D; online supplemental figure S8). Consequently, PD-L1 L88S cells showed reduced, K162fs cells no avelumab surface binding, which was not attributable to epitope disruption (figure 5E; online supplemental figure S9).

In accordance with previous reports,44 CDC effector functions were not detectable (online supplemental figure S10). In contrast, cells expressing the L88S variant were markedly less able to induce degranulation of primary NK cells in the presence of avelumab as compared with the wt (figure 5F). The 100% positive staining rate of patient tumor material with antibodies against CD56 and/or CD16 clearly showed that all of the tumors were infiltrated with cells able to mediate ADCC under in vivo conditions (data not shown). PD-L1 K162fs transduced cells failed to mount an avelumab-mediated NK cell response (figure 5F). NK cell degranulation mediated by cetuximab was independent of PD-L1 status or variant (online supplemental figure S11). Direct LDH release as a measure for target cell killing confirmed this pattern (online supplemental figure S12).

To test whether the K162fs or L88S variants affect T cell suppression, we cocultured HT-29, DLD-1 and UT-SCC-14 cells expressing these variants with primary cytotoxic CD8+ T cells and monitored immune cell killing over time. While cells expressing wt PD-L1 were more resistant to T cell mediated immune cell killing than L88S cells, cells expressing the K162fs variant were most effectively targeted by T cells (figure 5G).

Further experiments were performed to explore the mechanisms underlying loss of PD-L1 membrane expression in these mutants. As demonstrated with inhibition experiments, the K162fs mRNA product is constantly degraded by the NMD machinery (online supplemental figure S13). Since reduced PD-L1 protein levels in cells expressing PD-L1 L88S were independent of transcription, we reasoned that this mutation had a lower stability and/or higher turnover as the wt. In accordance with this notion, PD-L1 L88S exhibited a shorter lifetime after blocking protein biosynthesis with cycloheximide (figure 6A,B) or N-glycosylation (figure 6C), which is a key determinant of PD-L1 protein stability,45 with tunicamycin. Although the change from leucine to serine at position 88 might affect intrinsic protein stability, it also creates a potential phosphorylation site within the extracellular domain (ECD). Since homeostatic PD-L1 turnover includes phosphorylation-dependent degradation by the proteasome,45 46 we hypothesized that an additional phosphorylation site might enhance this process. To test this, we focused on the PD-L1 interacting AMP-activated protein kinase (AMPK), which phosphorylates serines within the ECD of PD-L1.46 AMPK inhibition using compound C (dorsomorphin) caused a doubling of PD-L1 L88S protein levels, which was not observed when serine 88 was substituted with alanine (figure 6D). Notably, the turnover of L88A variant was identical to the wt (figure 6A–C). Blocking of the proteasome with MG132 causes an enrichment of the glycosylated and non-glycosylated PD-L1 variants that was substantially stronger in the PD-L1 L88S transduced cell lines (figure 6E). Taken together, our data suggest that the truncating mutation PD-L1 K162fs is degraded by NMD, while the PD-L1 L88S variant creates a new phosphorylation site leading to loss of protein stability and proteasomal degradation.