Pretreatment biopsy and post-treatment resection specimen archival blocks from patients who received SCRT or LCRT atNorth East – York Research Ethics Committee were retrieved (Research Ethics Committee No 08/H0903/62). Clinicopathological characteristics of the patients were collected. Patients with a non-standard interval between the end of radiation and surgery (longer than 20 and 80 days for SCRT and LCRT, respectively) were excluded.

All cases were reviewed and TRG was evaluated according to AJCC TNM Staging Manual (8th Edition)15 in resection specimens by a specialist gastrointestinal pathologist. In order to overcome the subjectivity inherent in the TRG system, a previously described quantitative method (TCD) was also evaluated.8 9 In brief, TCD evaluation involves quantification of 300±15 individual points distributed across the whole tumor area on digital H&E-stained tumor sections (figure 1A, B). The selection of 300 points is derived from earlier modeling studies which identified that at least 250 points were needed to accurately quantify the percentage of tumor cells. Baseline and post-treatment TCD were assessed by an independent observer, blinded to TRG evaluation.

Following demarcation of tumors on H&E slides by a specialist gastrointestinal pathologist, tumor areas were macrodissected from unstained slides to capture the whole tumor microenvironment. Nucleic acids were extracted using The RecoverAll Total Nucleic Acid Isolation Kit and quantified using Qubit Fluorometry (both from Thermo Fisher Scientific, UK), according to manufacturer’s instructions. The expression of 760 genes included in the NanoString panCancer Immune Panel (NanoString Technologies, Seattle, Washington) were measured according to manufacturer’s instructions and normalized using positive and negative controls and the housekeeping genes included in the panel. Only genes with non-zero expression in 75% or more of the samples were retained.16 The presence of a batch effect in log₂ transformed and normalized data was assessed using exploBatch17 and corrected using ComBat from SVA Bioconductor-based R package.18

Patients’ characteristics were summarized using descriptive statistics. The difference between resection specimen and baseline TCD (ΔTCD) was calculated; its association with TRG was assessed using the Kruskal-Wallis test. Next, ΔTCD was used to classify each patient’s tumor response into three categories (good, intermediate or poor) using k-means clustering. To determine GEP that were most relevant to differential radiotherapy responses, we excluded the intermediate response group and compared good versus poor responders in the majority of subsequent analyses (an H&E demonstration of good vs poor responders is shown in figure 2).

The Estimation of STromal and Immune cells in MAlignant Tumours using Expression data (ESTIMATE) algorithm19 was used to infer tumor purity and the fraction of stromal and immune cells in baseline biopsies of tumors with good versus poor radiotherapy response. Following this, significance analysis of microarrays (SAM) was used to identify individual genes with a significant difference in expression at baseline between good and poor responders.20 A comparison of intermediate responders’ GEP with those of the good and poor responders was also performed. Following this, functional annotation of the significant genes according to pathways (hypeR R package21) and cell types was performed (using an Enrichr-based gene enrichment analysis22), and two publicly available gene sets (ARCHS4 and the Human Gene Atlas).23 24 ARCHS4 is a published database of cell types derived by mining RNAseq data from public gene expression databases.23 For Human Gene Atlas, the authors created custom microarrays to measure gene expression of known and predicted gene proteins associated with tissues and cell types.24 We used Enrichr combined score analysis to perform enrichment of cell types in our radiotherapy-treated rectal gene expression cohort.22 The pathways to genes plots were done using multiple packages—clusterProfiler,25 msigdbr26 and enrichplot.27

Longitudinal (treatment-induced) significant changes in gene expression were evaluated in paired samples of good and poor responders in two separate SAM analyses.20 For pathway analysis, the Molecular Signature Database (MSigDB)’s hallmarks gene sets26 and hypeR R package21 was used to describe the biological pathways represented by differentially expressed genes. Similarly, longitudinal changes in the 18-gene T-cell-inflamed GEP,28 that is associated with prediction of immunotherapy response, were evaluated in good and poor responders. Of note, our panel included 16 of the 18 genes in the GEP. Finally, longitudinal immune cell type changes in good and poor responders were inferred using single-sample gene set enrichment analysis29 and immune enrichment signatures (IES) from Rooney et al.30

The use of different GEP platforms, non-overlapping genes and inconsistent measures of radiotherapy response across published reports poses challenges for independent validation of findings. To address this, meta-analysis at the level of gene sets/pathways was used to compare upregulated pathways according to response in our datasets with those in two independent cohorts.31 32 In addition, as our study cohort and the above two cohorts did not have survival outcomes available, the impact of differential gene expression on disease-specific survival (n=182) was assessed in a third independent and publicly available cohort—GSE87211 (n=188), where patients were treated with CRT.33

The first validation cohort consisted of GEP from 15 patients receiving preoperative radiotherapy. Radiotherapy response was defined using the Mandard tumor regression grade system where TRG1 and TRG2 were considered to be responders and TRG3, TRG4 and TRG5 were classified as non-responders. PrimeView Affymetrix arrays were used for GEP using RNA extracted from microdissected epithelial and stromal components of fresh frozen preradiotherapy tumor tissue biopsies.31 We used the pathway analyses data, from a supplementary table of the Goncalves-Ribeiro study, derived from dysregulated genes in the stromal component of radiotherapy non-responders (compared with responders).31

The second validation cohort consisted of 23 patients with rectal cancer treated with preoperative radiotherapy to a dose of 45 Gy in 25 fractions, 5 days per week for 5 weeks, ± 5.4 Gy boost. The AJCC TRG system was used to define radiotherapy response in surgical specimens; tumors were classified as “total responders” when assigned to TRG0, “partial responders” when TRG1 and TRG2, and “non-responders” when TRG3. Proteomics analysis was performed using high performance liquid chromatography separation, coupled to mass spectrometry by the authors. We used the list of 139 proteins (from the supplementary table of the original publication) that discriminated non-responder from total responders to preoperative radiotherapy from Chauvin et al.32 We applied enrichment analysis using “Investigate Gene Sets” from the MSigDB26 to discern the changes in pathways associated with the “hallmarks gene sets”.

The third validation cohort consisted of gene expression data (n=188) and disease-specific survival outcomes of 182 patients with locally advanced rectal cancer treated with preoperative radiotherapy to a dose of 50.4 Gy.33 GEP was carried out by the authors using the Human 4×44 K v2 array platform from Agilent Technologies (G4845A). Disease-specific survival measured from the time of surgery to death or last clinical follow-up was used.

To classify the samples from the third cohort into two classes that resemble poor and good responders from our cohort, average (mean) expression of 40 genes, significantly upregulated in poor responding samples from our cohort, was calculated for each sample. The optimal cutpoint for the mean expression was further calculated using the maximally selected rank statistics from “survminer” R package using “surv_cutpoint” command. The optimal cutpoint provided two classes with high (representing poor-like responders) and low (representing good-like responders) expressors. These expressor groups arising from poor responding genes were further used to perform disease-specific survival and further gene expression analysis.

In order to provide further protein-level confirmation of the above gene expression findings, dual staining of CD68 and CD163 using immunohistochemistry was carried out in a limited number of samples due to the constraints of tissue availability. The immunohistochemistry methods for this are described in the online supplemental methods.

Kruskal-Wallis, Fisher exact and t tests were applied where necessary. Survival analysis was carried out using Kaplan-Meier methods and the log rank test.

A total of 140 archival blocks from 70 patients were identified (figure 3A). Eleven patients with a prolonged interval between the end of radiotherapy and surgery were excluded. Fifty-three cases with successful RNA extraction from diagnostic biopsies were available for analysis, of whom 52 cases had TRG scores. Patients’ characteristics are tabulated in figure 3B. In line with the published literature,3 4 the median age at diagnosis was 67 years and 72% of the patients were male. At least 45/53 (85%) of tumors were staged greater than pathological stage I after neoadjuvant treatment.

TRG 2 (moderate response) was ascribed to 90.4% (47/52; leaving out one sample with TRG status unavailable) of the tumors; only 2 and 3 patients were scored as TRG 1 and 3, respectively, and no patients had a pathCR (TRG 0). Hence, the TRG, consistent with its known limitations,3 was not considered a suitable endpoint to identify differential response to neoadjuvant treatments. Conversely, the ΔTCD scores ranged from 12.3% to −47.7% and from 12.4% to −42.1% in patients treated with SCRT and LCRT, respectively.

Despite very few patients being categorized as TRG 1 or 3, a marginal association between TRG and ΔTCD was seen (figure 1C). The three patients with TRG 3 had minimal ΔTCD (mean ΔTCD 0.96%) whereas the two patients with TRG 1 had much more substantial ΔTCD (mean ΔTCD −22.8%). As ΔTCD scoring better represented heterogeneous responses to neoadjuvant therapy, ΔTCD was subsequently used to define response categories. Following k-means clustering into three response groups, 23 good, 16 intermediate and 14 poor responders were identified (figure 1D, E). An example of H&E staining preradiotherapy and postradiotherapy in good versus poorly responding tumors is shown in figure 2. Here, the poorly responding postradiotherapy resection specimen shows a high density of residual tumor cells. Conversely, in the resection specimen showing a good response, there are very few tumor cells and the tissue is predominantly replaced by reactive stroma.

A comparison of clinicopathological characteristics in good versus poor responders indicated that poor responders were significantly younger with higher pathological stage. Importantly, there was no significant difference between the two response categories and type of neoadjuvant treatment (SCRT or LCRT) or the interval between radiation and surgery (figure 3B). This reinforces the biological meaning of ΔTCD and suggests that significant associations with variation in radiotherapy technique are unlikely, at least in this cohort.

Use of the ESTIMATE algorithm, which infers cancer versus stromal/immune content at the gene expression level, showed that, at baseline, poorly responding tumors contained a significantly higher fraction of stromal and immune cells (ie, lower tumor purity) than tumors with a good response (figure 4A–C). When the immune GEP at baseline of good and poor responders (n=33) were compared using SAM analysis, 40 genes showed significantly higher expression in poor compared with good responders (figure 4D). These included genes involved in immune checkpoint inhibition (CD274 or PD-L1), immune regulation (IL11, IL15RA, IL6R) and integrin signaling (ITGA1, ITGA4) as well as markers of macrophages (CD163, ITGAM) and stromal fibroblasts (PDGFR-β). Eight genes showed significantly (FDR≤0.05) differential expression when the three responder groups—good, intermediate and poor were compared by SAM analysis. Of these 8 genes, 7 genes overlapped with the 40 poor responder genes suggesting that the intermediate responders were of an immune phenotype in between the good versus poor responders (online supplemental table 1).

Pathway analysis using the 40 upregulated genes in poorly responding tumors revealed 13 significantly (FDR<0.2) upregulated hallmark pathways, where more than 50% of the pathways were associated with direct immune regulation (figure 4E). The genes associated with four pathways showing the most significant upregulation in our cohort (IL6/JAK/STAT3 signaling, inflammatory response, interferon gamma response and allograft rejection) are shown as a pathways network with associated genes in figure 4F. More than 50% of these genes were shared by the four pathways representing similar mechanism(s) of antiviral-like immune regulation operating in the pretreated poor responding samples compared with good responding samples.

We used two independent cohorts of rectal cancer samples for validation using pathway analysis derived using gene and protein expression, respectively. A third independent cohort with disease-specific survival was used for validation. Upregulation of the epithelial–mesenchymal transition, and apical junction pathways were validated in both independent cohorts and pathways of complement and coagulation were validated in one independent cohort (figure 4G–I, online supplemental table 2). These two cohorts served as independent validation using both gene and protein expression profiles. The pathways from the first cohort were generated from microdissected stroma. Hence, the genes enriched in the poor responding samples are potentially from stroma. In the third validation cohort, there was a trend towards inferior disease-specific survival outcomes in patients (high expressors or poor like) whose tumors had increased expression of the 40 genes upregulated in poor responders (from our training cohort) compared with those patients with tumors not showing increased expression (low expressors or good like; figure 4J). The mean expression of the 40 upregulated genes in poorly responding tumors in high versus low expressors and their expression patterns as a heatmap in the third validation cohort are shown in online supplemental figure 1. These results show pathway-based and survival-based validation of the responder groups.

Functional annotation of the 40 genes using publicly available resources and enrichment analysis demonstrated their significant association with macrophages, monocytes and dendritic cells, as well as stromal cells including fibroblasts and smooth muscle cells (figure 5A, B, online supplemental tables 3–6). CD163 is marker of alternatively-activated (M2-like) macrophages which are thought to have tumor-promoting effects.34 CD163 gene was significantly upregulated in poorly responding tumors (figure 5C). Dual staining of CD68 and CD163 using immunohistochemistry was possible for a very limited number of biopsies due to tissue availability. Nevertheless, figure 5D,E demonstrate marked upregulation of both CD68 and CD163 staining in the baseline biopsy of a poorly responding tumor when compared with the baseline biopsy of the tumor showing a good radiotherapy response.

Significant longitudinal changes in immune gene expression during radiotherapy were evaluated with two separate SAM analyses in 15 paired samples (pretreatment biopsies and resection specimens) from patients with a good response and in 12 paired samples from poor responders. In good responders, 198 immune genes showed significant upregulation (figure 6A and online supplemental figure 2) in the postradiotherapy resection specimen when compared with the preradiotherapy biopsy. However, in poor responders, the expression of only seven genes was significantly modified by radiotherapy treatment (C7, CHIT1, CXCL12, and SPP1 were upregulated and CCL28, DMBT1 and CEACAM1 were downregulated; figure 6B).

Significant increases in coinhibition of T-cells, type I interferon response and macrophages were demonstrated in good responders (figure 6C–E, the genes used in these signatures are show in figure 6F). Conversely, no significant changes were observed in poor responders which, as described earlier, appeared to have a higher macrophage score at baseline (online supplemental figure 3A-C). Similarly, a significant increase in the T-cell-inflamed GEP signature was demonstrated in good responders, but not in poorly responding tumors, which appeared to have a more inflamed GEP signature at baseline than good responders (figures 6G and 4D). Evaluation of IES showed a consistent, significant increase in the expression of 25 (out of 56) genes associated with multiple immune cell types (including CD8+ T cells and perforin; PRF1, and macrophages; Figure 6F) and processes in good responders. Conversely, such immune activation was not present in poor responders (except for two significant genes—PDCD11LG2 and CD70; figure 6F), which showed increased macrophage score at baseline (figure 5C).

To understand the biological pathways represented by the 198 upregulated genes in tumors with a good response, pathway analysis was performed. A highly significant upregulation of interferon gamma response, allograft rejection and inflammatory response pathways was seen alongside several other significantly (FDR<0.2) upregulated pathways (figure 7A). Taken together, these suggest an immunostimulatory response to radiotherapy, together with radiation-induced replacement of tumor cells with reactive stroma. Figure 7B shows the six most significantly upregulated pathways and their associated genes in a network with potential activation of antiviral pathways. The downstream effect of this activation is an interferon gamma response. We observed a highly significant increase in interferon-gamma response pathways scores in good responders (figure 7C).