Exploratory biomarker findings from regorafenib plus toripalimab in patients with refractory metastatic colorectal cancer (REGOTORI study)

Abstract

Background

The REGOTORI study showed that some metastatic colorectal cancer (mCRC) patients benefited from programmed death 1 (PD-1) antibody toripalimab plus anti-angiogenic tyrosine-kinase inhibitor regorafenib. However, biomarkers for this combined therapy in mCRC remain unclear. To address this gap, we performed an integrated multi-omics biomarker analysis in REGOTORI.

Methods

Next-generation sequencing was performed on formalin-fixed, paraffin-embedded (FFPE) tissue and paired plasma samples. Multiplex immunohistochemistry was performed to analyze the tumor microenvironment (TME) on FFPE samples. Variant allele frequency, somatic alterations, tumor mutational burden (TMB), circulating tumor DNA (ctDNA), and TME markers were jointly assessed from both FFPE and plasma samples to explore their associations with clinical outcomes under regorafenib plus toripalimab.

Results

A total of 35 patients were included. Patients with lower ctDNA maximum somatic allele frequency (maxAF), high-allele-frequency blood-based TMB (HAF-bTMB), blood-based intratumor heterogeneity (bITH), or HAF-bITH had longer survival time than their respective counterparts. Somatic alterations in SMAD4 (progression-free survival: 2.4 vs 1.6 months; overall survival: not reached vs 5.1 months) or PIK3CA (overall survival: 15.5 vs 5.7 months) were associated with shorter survival time. ctDNA dynamics appeared related to treatment response. High positive rates of CD3⁺, CD3⁺CD8⁺, CD3⁺CD8⁻, and PD-1⁺CD3⁺ T cells in the stroma area correlated with prolonged progression-free survival and favorable disease control; however, neither the positive rate of PD-L1 cells nor the density of it in the tumor area was associated with survival and response.

Conclusion

In this combination-therapy-focused multi-omics analysis integrating tissue genomics, ctDNA features/dynamics and TME profiling, we identified ctDNA maxAF, HAF-bTMB, bITH, HAF-bITH, mutational status of SMAD4 or PIK3CA and TME markers as predictive or prognostic biomarkers for regorafenib plus toripalimab in refractory mCRC.

Introduction

Colorectal cancer (CRC) is one of the most prevalent gastrointestinal malignancies worldwide [1]. For patients with metastatic CRC (mCRC), regorafenib, an anti-angiogenic tyrosine kinase inhibitor (TKI), is recommended as third-line or later therapy by the National Comprehensive Cancer Network (NCCN) guidelines [2, 3]. However, its clinical efficacy remains limited, and only a subset of patients achieves meaningful or durable benefit. Meanwhile, immune checkpoint inhibitors (ICIs) have demonstrated remarkable responses in patients with high microsatellite instability (MSI-H) or deficient mismatch repair (dMMR), but show little benefit in proficient mismatch repair (pMMR) or microsatellite stable (MSS) tumors, which represent the majority of mCRC.

Preclinical studies have suggested that combining anti-angiogenic agents with ICIs may exert synergistic effects by modulating the tumor microenvironment (TME) [4]. This rationale has been supported clinically, with several trials, including REGONIVO and EPOC1706, having reported encouraging outcomes in gastrointestinal malignancies with TKI–ICI combinations [5, 6]. Consistent with these findings, our team previously conducted the phase Ib/II study REGOTORI, which demonstrated that the combination of regorafenib and toripalimab improved clinical outcomes in refractory mCRC [7].

Despite these promising results, reliable predictive or prognostic biomarkers combined for PD-1 antibody plus regorafenib therapy in mCRC remain lacking. In addition, analyses of biomarkers from similar trials have reported inconsistent findings across different tumor types and treatment regimens [8–14]. For instance, PD-L1 expression, tumor mutational burden (TMB), and gene alterations such as PTEN or PIK3CA have demonstrated variable predictive value depending on cancer type and therapeutic context, and their relevance to this specific regimen in refractory mCRC is still uncertain.

Therefore, in this study, we performed a comprehensive multi-omics analysis of tumor tissues and paired longitudinal plasma samples from patients enrolled in the REGOTORI trial [7]. By integrating genomic profiling, circulating tumor DNA (ctDNA) dynamics, and TME characteristics, we aimed to identify potential biomarkers associated with treatment response and survival in patients with refractory mCRC receiving regorafenib plus toripalimab.

Materials and methods

Ethics approval and consent to participate

This study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Sun Yat-sen University Cancer Center, Guangzhou, China (approval ID: B2019-003-05).

Sample collection

Formalin-fixed, paraffin-embedded (FFPE) tumor tissue samples were collected before therapy initiation (baseline). Peripheral blood (PB) samples were collected at baseline and then every two treatment cycles. At each time point, 10 mL of PB was drawn into ethylenediaminetetraacetic acid (EDTA) tubes. Plasma and PB lymphocytes were separated within 2 hours after collection, aliquoted, and stored at −80°C until further analysis.

DNA isolation

Tissue DNA was extracted from FFPE tumor tissues by using the QIAamp DNA FFPE Tissue Kit (Qiagen, Hilden, Germany). Circulating cell-free DNA (cfDNA) was extracted from 4 to 5 mL of plasma by using the QIAamp Circulating Nucleic Acid Kit (Qiagen, Hilden, Germany). Sheared tissue DNA and cfDNA were purified to enrich fragments of 200–400 bp by using the Agencourt AMPure XP Kit (Beckman Coulter, CA, USA), followed by hybridization with capture probe baits, magnetic-bead–based enrichment, and subsequent amplification.

Capture-based targeted sequencing and whole-exome sequencing

Capture-based targeted sequencing was performed by using a commercial panel (OncoScreen Plus, Burning Rock Biotech, Guangzhou, China) covering 520 genes and spanning 1.64 Mb of the human genome. Fragment size and quality were assessed with the High Sensitivity DNA Kit on the Bioanalyzer 2100 system (Agilent Technologies, CA, USA). Indexed libraries were sequenced on the NextSeq 500 platform (Illumina, Inc., CA, USA) with paired-end reads, achieving an average depth of 1,000× for tissue samples and 10,000× for liquid biopsy samples. Additionally, tumor samples from 32 patients were analyzed by whole-exome sequencing using the Agilent SureSelect All Human Exon V5 (50 M) kit (Agilent, Santa Clara, CA, USA) to provide a reference dataset.

Sequence data analysis

Sequence data were aligned to the reference human genome (hg19) using Burrows–Wheeler Aligner (BWA) version 0.7.10, Genome Analysis Tool Kit (GATK) version 3.2, and VarScan version 2.4.3. Tissue and plasma samples were compared with matched white blood cell controls to identify somatic variants. Variants were filtered using the VarScan fpfilter pipeline, and loci with a sequencing depth <100 were removed. Variants with a population frequency >0.1% in ExAC, 1000 Genomes, dbSNP, or ESP6500SI-V2 were considered single-nucleotide polymorphisms (SNPs) and excluded from further analysis. The remaining variants were annotated by using ANNOVAR (release 2016–02-01) and SnpEff version 3.6. Structural variations (SVs) were analyzed by using Factera version 1.4.3. Copy number variations (CNVs) were inferred based on depth of coverage data across capture intervals.

Tumor mutational burden calculation

Tumor mutational burden (TMB) was computed as the ratio between the total number of non-synonymous mutations detected and the total coding region size of the panel, using the equation below.

$$\mathrm{TMB}=\frac{\mathrm{Mutation}\mathrm{\hspace{0.17em}}\mathrm{count}\mathrm{\hspace{0.17em}}(\mathrm{excluding}\mathrm{\hspace{0.17em}}\mathrm{CNV},\mathrm{\hspace{0.17em}}\mathrm{SV},\mathrm{\hspace{0.17em}}\mathrm{SNPs}\mathrm{\hspace{0.17em}}\mathrm{and}\mathrm{\hspace{0.17em}}\mathrm{hot}\mathrm{\hspace{0.17em}}\mathrm{mutations})}{1.003}$$

The mutation count included non-synonymous single-nucleotide variants (SNVs) and indels detected within coding regions and ±2 bp of flanking sequences, while excluding hot-spot mutation events, CNVs, SVs, and germline SNPs. Only mutations with an allelic fraction (AF) ≥2% in tissue samples and ≥0.2% in plasma samples were included. For reliable TMB estimation, maximum AF (maxAF) was required to be ≥5% for tissue samples and ≥1% for plasma samples. The total coding region size for the 520-gene OncoScreen Plus panel was 1.003 Mb.

Intra-tumor heterogeneity calculation

Intra-tumor heterogeneity (ITH) was calculated using the Shannon index as follows:

$$\mathrm{ITH}=-{\sum }_{t=1}^N{p}_i{\mathit{ ln p}}_i$$

The AF values ranged from 0 to 100, the bin width was set to 10 for tissue samples and 1 for plasma samples. Here, $p_i$ represents the distribution probability of AF within bin i.

Multiplex immunofluorescence staining and multispectral imaging

To identify cell subsets expressing CD3, CD8, PD-1, and PD-L1, multiplex immunofluorescence staining was performed by using the PANO 7-plex IHC kit (Panovue, Beijing, China). Different primary antibodies were applied sequentially, followed by incubation with horseradish peroxidase–conjugated secondary antibodies and tyramide signal amplification (TSA). After each TSA step, the slides were heated in a microwave oven to remove bound antibodies before the next round of staining. Following completion of all antigen labeling, nuclei were counterstained with 4’-6’-diamidino-2-phenylindole (DAPI; Sigma-Aldrich).

To obtain multispectral images, the stained slides were scanned by using the Mantra System (PerkinElmer, Waltham, MA, USA). This system captures fluorescent spectra at 20-nm wavelength intervals from 420 to 720 nm by using identical exposure times, and the resulting scans were combined to generate a single stacked multispectral image.

Image analysis

The images of unstained sections and single-stained controls were used to extract the spectra of tissue autofluorescence and each fluorophore, respectively. These spectra were used to establish a spectral library for multispectral unmixing with inForm image analysis software (PerkinElmer, Waltham, MA, USA). Using this library, reconstructed images were generated with autofluorescence removed, enabling accurate identification and quantification of marker-positive cell subsets.

Statistical analysis

Statistical analyses were performed by using the R programming language (version 4.0.2) with the “survival” and “survminer” packages. Fisher’s exact test was used to compare proportions for nominal variables. All tests were two-sided, and statistical significance was defined as P < 0.05.

Results

Patient characteristics

A total of 36 patients were eligible for biomarker analysis. Due to sample unavailability or quality issues, 35 patients were included in the final biomarker analyses. Baseline tumor tissues and PB samples were collected for downstream profiling, including whole-exome sequencing, targeted sequencing, and multiplex immunohistochemistry (mIHC); however, not every assay was available for every patient. The detailed workflow for sample processing and the assay availability across the cohort are summarized in Supplementary Fig. S1 and Supplementary Table S1.

Somatic mutational landscape at baseline

Baseline somatic alterations were assessed in tumor tissues (n = 22) and PB (n = 33), including 21 matched tumor–plasma pairs. On the 520-gene panel, every sample showed at least one detectable mutation (Supplementary Fig. S2). In the matched cohort, KRAS alterations were identified in 47.6% of tissue specimens and 57.1% of the paired PB samples, whereas NRAS mutations were less frequent overall, appearing in 4.8% of tissues but 14.3% of PB. BRAF V600E was detected at a comparable rate in both compartments (4.8% in tissues and 4.8% in PB) (Fig. 1; Supplementary Table S2). Overall, ctDNA and tissue sequencing demonstrated moderate concordance, although several gene-level discrepancies were noted between the two platforms.

Figure 1.

Mutational landscape of plasma and tumor tissue samples.

Association between ctDNA features and survival outcomes

Blood-based TMB (bTMB) and the maximum somatic allele frequency (maxAF) of ctDNA were obtained from plasma variant allele frequencies (VAFs) (Fig. 2A). Our findings indicated that patients with lower bTMB tended to have improved overall survival (OS) compared with those with higher bTMB (not reached [NR] vs 10.3 months; hazard ratio [HR] = 0.24, 95% confidence interval [CI]: 0.05–1.08). This difference was supported by a log-rank test (P = 0.043), although the Cox model showed only borderline significance (P = 0.063; Fig. 2B). In contrast, the patients with higher ctDNA maxAF had poorer OS than those with lower ctDNA maxAF (NR vs 10.3 months; HR = 0.18, 95% CI: 0.04–0.84; P = 0.014; Fig. 2C).

Figure 2.

Associations of different plasma biomarkers with survival. (A) Flow chart of plasma-based biomarker analysis in this study (VAF, variant allele frequencies; bTMB, blood-based tumor burden mutation; HAF-bTMB, high allele frequency-bTMB; ctDNA maxAF, circulating tumor DNA maximum somatic allele frequency; bITH, blood-based intratumor heterogeneity). (B) Overall survival (OS) of patients with bTMB-high and bTMB-low. (C) OS of patients with ctDNA maxAF-high and ctDNA maxAF-low. (D) OS of patients with HAF-bTMB-high and HAF-bTMB-low. (E) OS of patients with bITH-high and bITH-low. (F) OS of patients with HAF-bITH-high and HAF-bITH-low.

When focusing on high allele frequency (HAF)-filtered metrics, our results showed that patients with lower HAF-bTMB had longer OS than those with higher HAF-bTMB (NR vs 10.3 months; HR = 0.29, 95% CI: 0.09–0.97; P = 0.032; Fig. 2D). Next, we compared intra-tumor heterogeneity measured in tissue (tITH) with blood-derived heterogeneity (bITH) and observed a borderline correlation between these two indices (Spearman r = 0.413, P = 0.063; Supplementary Fig. S3A). Although this did not reach statistical significance, the direction and magnitude of the association suggest that tissue- and blood-based heterogeneity might be biologically connected and should be validated in larger cohorts.

Survival analyses showed that patients with lower bITH had significantly longer OS time than those with higher bITH (NR vs 10.3 months; HR = 0.10, 95% CI: 0.01–0.81; P = 0.008) (Fig. 2E). Comparatively, tITH was not associated with OS, as patients with low tITH and high tITH had similar OS (15.5 vs 10.4 months; HR = 0.47, P = 0.378; Supplementary Fig. S3B). In contrast, blood-based heterogeneity using high-allele-frequency variants was found to be associated with OS, as patients with high HAF-bITH lived shorter than those with low HAF-bITH (10.3 months vs NR; HR = 0.18, P = 0.014; Fig. 2F).

Association between WES-TMB and survival outcomes

In this cohort, whole-exome sequencing-derived tumor mutation burden (WES-TMB) was in the same range as that reported for TCGA rectal adenocarcinoma (READ) and colon adenocarcinoma (COAD) datasets (Supplementary Fig. S4A). The median WES-TMB was 1.73 mutations per megabase (Muts/MB) (Supplementary Fig. S4B).

Survival analyses did not identify a significant association between WES-TMB levels and clinical outcomes. Progression-free survival (PFS) was comparable between the TMB-high and TMB-low groups, with median PFS of 2.3 versus 2.0 months, respectively (HR = 0.87, 95% CI: 0.41–1.83; P = 0.718; Supplementary Fig. S4C). Similarly, OS likewise did not differ between strata, with OS being NR in the TMB-high group and 15.5 months in the TMB-low group (HR = 0.74, 95% CI: 0.21–2.64; P = 0.647; Supplementary Fig. S4D). These results indicate that WES-TMB, although a conventional biomarker for anti-PD-1 monotherapy, may have limited utility for predicting benefit in the context of anti–PD-1–based combination therapy.

Survival-associated genotypes and dynamic mutational changes during treatment

To further explore molecular events associated with OS, baseline mutation profiles were analyzed to identify gene- and pathway-level associations (Fig. 3A, Supplementary Table S2). Among these, SMAD4 mutations were associated with worse clinical outcomes. Patients with wild-type SMAD4 had an objective response rate (ORR) of 17.8% (5/28) and disease-control rate (DCR) of 42.9% (12/28), while no patients with SMAD4 mutations responded to the combination therapy (Fig. 3B). PFS and OS were also shorter in SMAD4-mutant patients than in wild-type patients (PFS: 2.4 vs 1.6 months, P < 0.001; HR = 6.32, 95% CI: 2.26–17.69; Fig. 3C; OS: NR vs 5.1 months; HR = 4.25, 95% CI: 1.29–13.98; P = 0.011; Fig. 3D).

Figure 3.

Altered genes or pathways are associated with clinical outcomes. (A) Flow chart of identifying survival-associated genes and pathways both in tissue and plasma. (B) The association between SMAD4 status (WT, wild-type; Mut, mutation) and response (PR, partial response; SD, stable disease; PD, progressive disease). (C) Progression-free survival (PFS) of SMAD4-mutated vs -wild-type patients. (D) Overall survival (OS) of SMAD4-mutated vs -wild-type patients. (E) OS of PIK3CA-mutated vs -wild-type patients. (F) OS of patients with mutated SMAD4 or PIK3CA vs or those wild-type in both genes.

Similarly, patients with PIK3CA mutations had shorter OS than those with PIK3CA wild-type (15.5 vs 5.7 months; HR = 4.52, 95% CI: 1.28–15.93; P = 0.01; Fig. 3E). In addition, when evaluating SMAD4 and PIK3CA together, we observed improved risk stratification as patients harboring either mutation were found to have significantly shorter OS than those wild-type for both genes (NR vs 10.3 months; HR = 4.66, 95% CI: 1.44–15.09; P = 0.005; Fig. 3F).

To characterize molecular changes during treatment, ctDNA was analyzed at baseline, at best response, and at disease progression (PD) or last follow-up (Supplementary Fig. S1). The most frequently mutated genes at baseline were found to be KRAS (10/16, 62%), APC (9/16, 56%), TP53 (8/16, 50%), and LRP1B (4/16, 25%). At PD, the most frequent mutations were TP53 (11/16, 69%), APC (10/16, 62%), and KRAS (7/16, 44%). In addition, NRAS was detected in 1/16 patients (6%) at both baseline and PD (Fig. 4A and B).

Figure 4.

Dynamic changes in circulating tumor DNA (ctDNA) during treatment. (A) Plasma mutational landscape at baseline (PR, partial response; SD, stable disease; PD, progressive disease). (B) Plasma mutational landscape at PD or last follow-up. (C) Genomic alterations detected from plasma samples collected at baseline and at PD or last follow-up. (D) The dynamic change of ctDNA concentration during treatment.

Among the 16 patients with paired baseline and PD samples, newly emerged mutations were commonly observed, such as TP53 (3/16, 18.8%), SMAD4 (1/16, 6.3%), KRAS (1/16, 6.3%), and APC (1/16, 6.3%) (Fig. 4C), indicating potential mechanisms of acquired resistance. Overall, eight patients (50%) developed at least one new mutation during treatment with regorafenib plus toripalimab.

Longitudinal ctDNA monitoring revealed time-dependent changes in tumor burden during treatment. Among patients who achieved PR (n = 4), ctDNA levels decreased at best response and rose again at PD or last follow-up, paralleling radiographic dynamics. Notably, one of the PR responders maintained ctDNA levels lower than baseline at the last follow-up, suggesting sustained molecular control (Fig. 4D, Supplementary Table S3). In contrast, ctDNA levels in patients with SD or PD (n = 16) were heterogeneous, with no consistent pattern across individuals. Nevertheless, one SD patient showed a persistent ctDNA decline through the last follow-up, indicating ongoing molecular response despite only stable imaging findings (Fig. 4D, Supplementary Table S3).

Immune-related TME factors and their association with clinical outcomes

Multiplex immunohistochemistry (mIHC) was performed on FFPE tumor samples from 24 patients to characterize immune cell subsets and assess their relationships with treatment outcomes (Fig. 5A, Supplementary Figure S5). The results showed that patients with higher positive rate of stromal CD3⁺ T cells experienced longer PFS than those with lower (2.3 vs 2.0 months; HR = 3.89, 95% CI: 1.34–11.26; P = 0.007; Fig. 5B), and also showed a trend toward improved OS (NR vs 10.3 months; HR = 4.03, 95% CI: 0.78–20.91; P = 0.075; Fig. 5C). Similar to these survival patterns, DCR tended to be higher in the positive rate of CD3⁺ T cells-high group than in the positive rate of CD3⁺ T cells-low group (7/17 vs 0/7; P = 0.065; Fig. 5D).

Figure 5.

Tumor microenvironment biomarkers are associated with clinical outcomes. (A) Flow chart of multiple immunohistochemistry analysis (FFPE, formalin-fixed and paraffin-embedded; IHC, immunohistochemistry). (B) Progression-free survival (PFS) of patients with different positive rate of CD3⁺ T cells in stromal area. (C) Overall survival (OS) of patients with different positive rate of CD3⁺ T cells in stromal area. (D) Association between different positive rate of CD3⁺ T cells in stromal area and radiographic response (PR, partial response; SD, stable disease; PD, progressive disease). (E) PFS of patients with different positive rate of CD3⁺CD8⁺ T cells in stromal area. (F) OS of patients with different positive rate of stromal CD3⁺CD8⁺ T cells. (G) The association between different positive rate of stromal CD3⁺CD8⁺ T cells and response. (H) PFS of patients with different positive rate of stroma of CD3⁺CD8⁻ T cells. (I) OS of patients with different positive rate of stromal CD3⁺CD8⁻ T cells. (J) Association between different positive rate of stromal CD3⁺CD8⁻ T cells and response. Created with BioRender.com.

A similar association was observed for the positive rate of CD3⁺CD8⁺ in the stromal area. Patients with higher positive rate of CD3⁺CD8⁺ T cells in the stroma showed longer PFS (3.3 vs 2.0 months; HR = 3.18, 95% CI: 1.17–8.67; P = 0.016; Fig. 5E), longer OS (10.3 vs 7.3 months; HR = 7.02, 95% CI: 0.80–61.26; P = 0.042; Fig. 5F), and a higher DCR (6/12 vs 1/12; P = 0.069; Fig. 5G) than those with lower positive rate of CD3⁺CD8⁺ T cells.

Patients with a higher positive rate of CD3⁺CD8⁻ T cells had likewise better outcomes those with a lower positive rate of CD3⁺CD8⁻ T cells, including longer PFS (2.3 vs 2.0 months; HR = 3.89, 95% CI: 1.34–11.26; P = 0.007; Fig. 5H), together with favorable trends in OS (10.3 vs 7.1 months; HR = 4.03, 95% CI: 0.78–20.91; P = 0.075; Fig. 5I) and DCR (7/17 vs 0/7; P = 0.065; Fig. 5J). Notably, consistent results were observed when the positive rate of CD3⁺CD8⁻ T cells or their density was evaluated across both tumor and stromal compartments (Supplementary Fig. S6A–I).

In addition to stromal T-cell density, the expression of immune checkpoint markers in the stroma was analyzed and showed associations with treatment outcomes. Patients with a high positive rate of stromal PD-1⁺ cells had longer PFS than those with a low positive rate of stromal PD-1⁺ cells (median PFS: 3.5 vs 2.0 months; HR = 2.67, 95% CI: 1.03–6.90; P = 0.037; Supplementary Fig. S7A). Although not statistically significant, patients with a higher positive rate of PD-1⁺ cells also showed numerically longer OS (15.5 vs 7.3 months; HR = 6.17, 95% CI: 0.73–51.94; P = 0.057; Supplementary Fig. S7B) and a higher DCR (5/10 vs 2/14; P = 0.085; Supplementary Fig. S7C) than those with a lower positive rate of PD-1⁺ cells. Consistently, patients with a higher positive rate of stromal PD-1⁺CD3⁺ T cells experienced longer PFS than those with a lower positive rate of stromal PD-1⁺CD3⁺ T cells (3.5 vs 2.0 months; HR = 3.79, 95% CI: 1.36–10.56; P = 0.007; Supplementary Fig. S7D), but OS did not differ significantly between the positive rate of PD-1⁺CD3⁺ T cells-high and -low groups (NR vs 10.3 months; HR = 2.32, 95% CI: 0.42–12.81; P = 0.319; Supplementary Fig. S7E). Besides, patients with a higher positive rate of stromal PD-1⁺CD3⁺ T cells also tended to have a better DCR than those with a lower positive rate of stromal PD-1⁺CD3⁺ T cells (6/12 vs 1/12; P = 0.069; Supplementary Fig. S7F). In contrast, the positive rate of PD-L1 cells or the density of PD-L1 cells within the tumor area was not associated with clinical outcomes, whereas the positive rate of PD-L1 cells in the stromal area showed a borderline trend toward shorter PFS (P = 0.05; Supplementary Fig. S5).

Relationship between genotypes and immunophenotypes or ctDNA features

To further explore the relationships between genotypes, immunophenotypes, and ctDNA characteristics, we conducted the correlation analyses. The results showed no significant association between gene mutation status and immune cell infiltration. Although patients with SMAD4 mutations tended to exhibit lower positive rates of CD3⁺ T cells, CD3⁺CD8⁺ T cells, CD3⁺CD8⁻ T cells, PD-1⁺CD3⁺ T cells, and PD-1⁺ cells than patients with SMAD4 wild-type, the differences were not statistically significant. Similar trends with lower immune infiltration were observed in patients with PIK3CA mutation and patients harboring a mutation in either SMAD4 or PIK3CA. Meanwhile, neither SMAD4 nor PIK3CA mutation was associated with the ctDNA features, for example, bITH, bTMB, ctDNA maxAF, etc. (Supplementary Figs S8 and S9).

Multivariate analysis based on metastatic sites and tumor burden

Our preliminary clinical study observed that all patients (3/3) with lung-only metastasis responded, whereas no patients (0/4) with liver-only metastasis responded. Therefore, we conducted univariate analyses of the site of metastases and the number of metastatic organs and found neither site of metastases (liver metastasis, lung metastases, and other metastasis) nor the number of metastatic organs (<3 vs ≥3) was associated with outcome events (OS and PFS) in univariate analysis (Supplementary Fig. S10A and B), despite patients with lung metastasis tended to have a non-significantly favorable OS (P = 0.075).

Nevertheless, to account for potential confounding by the site of metastases and the burden of the tumor, we conducted further analysis to adjust for these two variables. After adjusting for the site of metastases and the number of metastatic organs, the results including SMAD4, PIK3CA, and immune profiles were not essentially changed (Supplementary Fig. S11A–D).

Discussion

In this study, we performed multi-omics analyses to identify potential predictors of response and survival in patients with refractory mCRC treated with regorafenib plus toripalimab. We observed that blood-based ctDNA features, including bTMB, HAF-bTMB, HAF-bITH, and ctDNA maxAF, were associated with OS. At the genomic level, SMAD4 and PIK3CA mutations were associated with inferior response and shorter survival time, and combined SMAD4/PIK3CA status further improved risk stratification. Longitudinal ctDNA monitoring revealed dynamic molecular changes and the emergence of new mutations during treatment. In parallel, TME analysis demonstrated that patients with higher positive rates of or densities of CD3⁺, CD3⁺CD8⁺, CD3⁺CD8⁻, or PD-1⁺CD3⁺ T cells in the stromal or tumor area, as well as the higher positive rate of stromal PD-1⁺ cells, derived greater clinical benefit from toripalimab plus regorafenib.

Although regorafenib has been incorporated into routine practice, reliable biomarkers for anti-angiogenic tyrosine kinase inhibitors (TKIs) are still lacking [15], in contrast to ICI monotherapy, where PD-L1 expression, TMB, and MSI status have been explored as potential predictors. For ICI plus anti-angiogenesis combinations, however, predictive markers remain poorly defined. To address this gap, we integrated baseline genomic alterations, longitudinal ctDNA dynamics, and immune features of the TME in our cohort. Targeted sequencing of paired tumor and plasma samples from 35 patients confirmed recurrent genomic events, with tissue-based RAS and BRAF V600E mutation frequencies aligning with previous reports [16]. Notably, RAS alterations were detected more frequently in plasma than in tumor tissue [17], a discrepancy that may reflect the limited sample size and/or treatment-related temporal heterogeneity [18], particularly considering that most FFPE specimens were collected before initiation of first-line therapy and may not fully represent later evolutionary changes captured by ctDNA.

TMB has been associated with the efficacy of ICIs across various tumor types, and previous studies have suggested that patients with high TMB (TMB-H) often derive greater benefit from ICI therapy than those with low TMB (TMB-L) [19, 20]. However, in our study, WES-derived TMB did not correlate with either PFS or OS. Evidence regarding TMB in ICI-anti-angiogenesis combination regimens remains inconsistent, as some phase II studies, including camrelizumab plus apatinib in advanced solid tumors and REGONIVO in mCRC, reported a positive association between TMB and treatment efficacy [5, 8, 14], whereas other trials did not observe a similar relationship [9, 10, 12]. Interestingly, we observed a trend toward longer OS in patients with lower bTMB. This observation is consistent with prior reports suggesting that higher bTMB may, in part, reflect elevated ctDNA levels, which is a negative prognostic indicator [21]. Earlier studies using similar combination strategies reported that patients with bTMB-H tended to have shorter PFS [13]. Wang et al. [22] further proposed the metric of high-allele frequency bTMB (HAF-bTMB), defined by an increase in bTMB driven by ctDNA abundance (cut-off = 5%), and demonstrated its association with poorer survival. Consistently, in our cohort, patients with higher maximal ctDNA allele frequency (maxAF) experienced shorter OS. Collectively, these findings support the notion that ctDNA features, particularly allele frequency–adjusted measures such as HAF-bTMB, may provide clinically relevant biomarkers for prognostic stratification and for guiding immunotherapy-based anti-angiogenic combination treatment.

ITH has been linked to resistance to chemotherapy, radiotherapy, and targeted agents and is frequently associated with inferior clinical outcomes. Similarly, it can modulate immune surveillance and influence responsiveness to immunotherapy [23]. However, the prognostic relevance of ITH in combination regimens remains insufficiently explored. In our cohort, higher bITH was associated with shorter OS, and this relationship became even more evident when considering HAF-bITH, as patients with elevated HAF-bITH experienced the poorest survival, indicating that bITH and HAF-bITH derived from ctDNA may provide a more sensitive reflection of tumor heterogeneity relevant to survival in this treatment setting. Notably, the correlation between bITH and tITH was weak, which may be explained by regional sampling bias in heterogeneous tumors and by biological differences between treatment-naïve and refractory disease contexts, both of which could limit the ability of single-site tissue profiling to reflect systemic heterogeneity detected in blood [24].

We identified gene alterations associated with treatment response and survival in patients receiving toripalimab plus regorafenib. SMAD4, a tumor suppressor in the TGF-β pathway, is altered in approximately 13% of CRC cases [25]. Accumulating evidence indicates that SMAD4 mutation or loss is linked to adverse outcomes, including shorter OS, PFS, and recurrence-free survival [26, 27], increased chemoresistance [28], and reduced immune infiltration [29]. In addition, mutated SMAD4 may facilitate tumor progression and has been proposed as a potential target for T cell-based immunotherapy [30], suggesting a biologically relevant connection between SMAD4 status and antitumor immunity. However, its role in shaping sensitivity to ICIs, particularly when combined with anti-angiogenic TKIs, remains uncertain. In our cohort, patients with wild-type SMAD4 achieved better clinical responses and experienced longer PFS and OS, whereas those harboring SMAD4 mutations showed no clinical response and had significantly shorter survival. Notably, the median PFS in SMAD4-mutated patients was less than two months, suggesting SMAD4 mutation as a candidate biomarker of primary resistance to toripalimab plus regorafenib.

In addition to SMAD4, PIK3CA, which encodes the catalytic subunit of PI3K, is frequently mutated across human cancers. Prior studies have reported inconsistent relationships between PIK3CA mutation and immunotherapy benefit. Some investigations associated PIK3CA mutations with more favorable responses to ICIs [31, 32], whereas others observed attenuated immune-related gene signatures [33] and reduced CD8⁺ T-cell infiltration in PIK3CA-mutant tumors [34]. Huang et al. [14] also reported improved OS in PIK3CA-mutant cervical cancer patients treated with camrelizumab plus apatinib, and additional evidence suggests that combining ICIs with PIK3CA inhibitors may enhance antitumor activity [33]. Comparatively, in our cohort, PIK3CA mutations were associated with shorter OS, indicating that their prognostic or predictive impact may differ by tumor type and treatment setting. Notably, when SMAD4 and PIK3CA status were evaluated together, survival stratification was further improved, with patients harboring either alteration showing worse outcomes, suggesting that these two mutations may have an additive adverse prognostic effect in the context of toripalimab plus regorafenib.

Accumulating evidence has demonstrated that the TME was closely associated with immunotherapy [35]. In our analysis, we found that positive rates or densities of T cells or immune checkpoint markers were associated with the clinical outcomes of combination therapy. These findings are consistent with prior CRC studies showing that increased infiltration of CD3⁺, CD4⁺, or CD8⁺ T cells and higher PD-1 expression are associated with a more favorable prognosis [36–39] and with improved outcomes during immunotherapy [40–42]. Similarly, a meta-analysis demonstrated that stromal CD8⁺ T-cell infiltration is significantly associated with better response in patients treated with ICI monotherapy or combination regimens [43]. In addition, studies exploring biomarkers for combination therapy in mCRC have reported that higher cytotoxic T-cell infiltration is linked to improved clinical outcomes [44, 45]. Collectively, these findings indicate that a T cell–rich TME may help to identify patients more likely to benefit from toripalimab plus regorafenib.

However, several limitations to this study should be acknowledged. The cohort size was small, the trial was single-armed, and a small number of patients were unable to provide samples. Those may limit statistical power and reduce generalizability. Yet, we performed comprehensive profiling for all patients for whom a specimen was available.

In summary, we identified SMAD4 and PIK3CA alterations as candidate predictive biomarkers for patients with mCRC treated with toripalimab plus regorafenib. We also observed that bTMB, bITH, maxAF, and TME features, including the positive rate or density of T cells and immune checkpoint markers in the stroma or tumor, were associated with clinical outcomes, supporting their potential utility for biomarker-guided immunotherapy-based combination strategies. In the future, we will further expand the sample size and conduct prospective cohort studies to validate the clinical utility and robustness of the identified biomarkers in patients with pMMR/MSS refractory mCRC receiving this combination therapy. Moreover, we plan to conduct mechanistic studies to investigate how specific genetic alterations or tumor microenvironment features may affect treatment outcomes. These efforts may help refine patient selection strategies and contribute to the development of more precise and effective immunotherapy-based regimens for patients with refractory mCRC.

Author contributions

F.W. conceived and designed the study. Y.C.Y., M.M.H., Y.J., and M.W. acquired the data. Y.C.Y., M.M.H., M.W., Q.Z., T.H., and Y.X.C. analyzed and interpreted the data. Y.C.Y., M.M.H., and M.W. drafted the manuscript. F.W. critically revised the manuscript for important intellectual content. Y.C.Y., M.M.H., Y.X.M, Q.Z., and T.H. performed the statistical analysis. Y.C.Y., M.M.H., M.W., Y.X.M, and Q.Z revised the manuscript. M.M.H. and Y.J. obtained the funding. F.W. provided administrative, technical, or material support and supervised the study. All authors read and approved the final manuscript.

Supplementary material

Supplementary material is available at Gastroenterology Report online.

Funding

This work was supported by the National Natural Science Foundation of China (82203006, 82403897, 82425048 and 82473204), Beijing Xisike Clinical Oncology Research Foundation (Y-QL202202-0089, to FW), the New Cornerstone Science Foundation through the XPLORER PRIZE, Young Talents Program of Sun Yat-sen University Cancer Center (YTP-SYSUCC-0018, to FW).

Conflicts of interest

Ting Hou is employed by Burning Rock Biotech. The remaining authors declare no competing interests.