Formalin-fixed paraffin-embedded (FFPE) tissue blocks were obtained from 22 patients from the Dunedin Colorectal Cohort under ethical approval from the NZ Health and Disability Ethics Committee (HDEC 14/NTA/033). All samples were surgical samples excised prior to any adjuvant treatment. Previous H&E staining of these tissue blocks had revealed at least 3 highly-defined immune aggregates in each sample. Patient characteristics are shown in Table 1 .

5µm sections were derived from these blocks by the Histology Services Unit at the University of Otago. Sections were deparaffinised, antigen retrieved (AR6 solution, Akoya Bioscience) and blocked with 10% human serum (ThermoFisher). Slides were stained for markers using the 7-plex Opal Polaris kit (Akoya Biosciences) according to the manufacturer’s instructions, using the primary antibodies and Opal dyes listed in Table 2 ; Foxp3 was stained with a cocktail of two antibodies as shown. Slides were mounted and imaged with the PhenoImager HT instrument (Akoya Biosciences).

Due to the variable staining patterns of CD4, observed across a range of cell types including myeloid and epithelial cells that express varying levels of CD4, we did not quantify CD4 expression. In CRC, double positive or negative T-cells are present in low frequencies (27), which include NKT cells (~0.7%) (28) and γδ T-cells [~0.5% of all CD3+ cells (29)] in cancer-associated TLS. In the absence of confirmation of CD4 positivity, we defined CD3+CD8- cells as ‘most likely’ representing CD4+ T cells. Hence our analysis utilised 5 markers in addition to the nuclear stain DAPI: CD3, CD8, Foxp3, CD20, Ki-67.

For H&E images, slides were deparaffinised and stained with Gills haematoxylin (Merck/Sigma Aldrich) counterstained with Eosin Y, aqueous (Merck/Sigma Aldrich) before dehydration and mounting with Eukitt Quick-hardening mounting medium (Merck/Sigma Aldrich).

Image unmixing, cell segmentation and positive channel marker thresholding were undertaken using the InForm (version 2.6 AKOYA BioSciences) program. Phenotyping and analysis of InForm-derived data was undertaken using R (version 4.0.4) in conjunction with custom analysis pipelines and the phenoptr package (version 0.3.2).

To improve comprehension of our results, we classified CD3+ CD8- cells as CD4 T-cells, since almost all T-cells present in CRC are single positive for either CD8 or CD4 (27, 28). Lymphocytes were therefore classified into 4 groups: B cells (CD20+ CD3–); CD8 T-cells (CD3+ CD8+); Foxp3– CD4 cells (CD3+ CD8– Foxp3–); Foxp3+ CD4 cells (CD3+ CD8– Foxp3+). Each of these 4 cell subsets were then sub-classified as Ki-67+ or Ki-67– to generate counts of 8 cell classifications for each TLS. A ninth category of cells shown in some analyses is “undefined” – cells that did not fall into the aforementioned categories, the vast majority of which were negative for all markers with the exception of their DAPI+ status.

Our strategy to determine TLS from other non-TLS lymphoid aggregates for downstream analyses is outlined in Supplementary Figure S1 . Lymphoid aggregates were initially identified in images by two researchers; where B-cell aggregates were close to each other and were connected by strands of lymphocytes (e.g. Figure 2J ; Supplementary Figure S2E ) these were considered as a single aggregate. The two researchers then re-assessed each other’s selections to determine which aggregates met defined criteria for classification as TLS domains. Cellular composition of each aggregate was phenotyped as described above, and only aggregates of ≥250 cells comprising >50% B-cells and/or T-cells were included as TLS domains.

Given the lack of consistency in the literature in how the peri-tumoural region is defined (1), TLS domains were not further sub-classified according to their distance from tumour margins, although all TLS domains analysed were either intra-tumoural or within ≈10mm of cancer cells.

Total tissue area measurements were made on Qupath (30) (version, 0.4.2) using the pixel classifier thresholder function. Briefly, the average of all channels were prefiltered (Gaussian) and a threshold was chosen that coved the tissue region. Tumour area measurements were made manually, aided by H&E stained sections from the same tumour annotated by a pathologist (T. Jeon). TLS tissue measurements used the pixel classifier thresholder function using CD3 and CD20 channels. Manual adjustments were made to the region where required.

Statistical analysis on individual TLS data was performed in R using the R package lme4 (v1.1-29) to apply a generalised linear mixed model with REML/residual maximum likelihood to account for grouping effects. The p-values were calculated with lmerTest (v3.1-3) using Satterthawite’s method. For the whole TLS tissue analysis, a linear regression with Spearmann’s correlation test was implemented using the stats package on R (v4.0.4). Values scaled in log10.

TLS with fewer than 50 cells of the assessed phenotypes were removed from analysis when performing correlation analyses, to ensure each datapoint was based on robust sample sizes.

We examined FFPE tissue sections from the tumours of 22 patients with CRC; patient characteristics are described in Table 1 . We used the criteria shown in Methods to identify 524 TLS domains in these combined tissue sections ( Supplementary Figure S1 ).

TLS domains stained for the B-cell marker CD20, with variable staining for the T-cell markers CD3, CD8 ( Figure 2 ) and Foxp3 ( Supplementary Figures S2A–D ). Subsets of both B-cells and T-cells stained for Ki-67 ( Figure 2 ). Most TLS domains were outside the tumour zones, with a minority in the peri-tumoural zones within 1mm from the invasive margin, and even fewer within the tumours (data not shown); however, in this study we did not quantify location of TLS domains with respect to cancer cells.

In images of these TLS domains, some resembled the canonical descriptions of mature TLS (1), with a well-defined GC surrounded by a compact, mostly ovoid-shaped B-cell zone and an associated T-cell zone ( Figure 2A ). However, many TLS domains displayed other geometries. TLS domains were not always ovoid ( Figures 2B, D ) and both the B-cell zones ( Figure 2B ) and T-cell zones ( Figure 2C ) adjacent to GCs varied in shape, with non-lymphoid cells and ECM often intercalated ( Figure 2C , arrows). Local tissue architecture such as epithelial crypts also impacted TLS domain morphology ( Figures 2D–F ), while channels and holes sometimes led to perforated appearances, probably due to blood or lymphatic vasculature ( Figures 2G, H ).

Some TLS domains were large and contained multiple B-cell regions ( Figure 2I ), consistent with the section plane passing through large multi-lobular TLS (26). Other TLS domains appeared in close proximity, connected by strands of lymphocytes ( Figure 2J , arrows and Supplementary Figure S2E ), raising the possibility that they might be part of the same TLS in 3D. Occasional TLS domains surrounded Ki-67+ proliferating cells that were negative for CD3 and CD20 ( Figure 2K , highlighted by arrow).

Collectively, our results show that TLS domains do not usually appear as classical ovoid structures and often lack well defined T-cell zones within the sectioning plane, even when GCs are clearly present ( Figures 2B, C, E, F, H ).

We sought to quantify and classify individual TLS domains within each CRC tissue sample, by counting circumscribed TLS domains and phenotyping their cellular constituents. The counts of TLS domains within each tissue sample varied greatly ( Figure 3A ). The lowest and highest TLS domain counts were 4 and 100, respectively, with a mean of 24.

The cellular composition of each TLS domain was analysed by digitally quantifying four different lymphocyte subsets, and the proportions of each of these subsets that expressed Ki-67 ( Figure 3A ). The proportion of B-cells expressing Ki-67 within TLS domains ranged from 0% to 94% ( Figure 3A ). Within each patient’s sample, the Ki-67+ proportion of B-cells also varied considerably ( Figure 3A ). However, when comparing the results between patients across their TLS domains, differences between patients become apparent. Some patients had higher proportions of Ki-67+ B-cells in most B-cell zones ( Figure 3A , upper row), while others had consistently low proportions of Ki-67+ B-cells ( Figure 3A , lower row). Hence, despite the variation in Ki-67 expression by B-cells in different TLS domains within individual patient samples, patients clearly differed in their B-cell proliferation once all their TLS domains were taken into account. Averaging the proportion of proliferating B-cells within all TLS domains in a tissue section allowed patients to be ranked for B-cell proliferation within their tissue sample ( Figure 3B ; Supplementary Figure S4A ; this ranking was also used to order patients in Figure 3A ).

Previous studies have suggested CRC patients with higher TLS numbers as well as those with a higher proportion of mature TLS have a better prognosis compared with patients with lower proportions (10, 23, 31). The numbers of TLS domains varied substantially even within the patients with the highest levels of B-cell proliferation ( Figure 3A , top row). Additionally, patients among those with the lowest B-cell activation also had some of highest TLS domain counts (such as P007, P010 and P021). Taken together, these observations confirm that counts of TLS domains from 2D sections do not adequately summarise TLS activity.

Patients that ranked highest for the proportion of Ki-67+ B-cells also exhibited high Ki-67 expression in T-cell subsets ( Figure 3B ), suggesting a potential correlation between B-cell proliferation and that of certain T-cell subsets. We therefore compared the Ki-67 expression in B-cells with that in different T-cell subsets within individual TLS domains across all patients ( Figure 4A ). This analysis confirmed a correlation between the proportion of B-cells expressing Ki-67 within each TLS domain and the Ki-67 expression of both Foxp3– CD4 T-cells and CD8+ T-cells ( Figure 4A ).

Foxp3+ CD4 T cells also expressed Ki-67 frequently, and this expression correlated with that in B-cells, CD8+ T-cells and Foxp3– CD4 T-cells ( Figure 4B ). Hence individual TLS domains with high proportions of Ki-67+ B-cells also had higher proportions of proliferation in all T-cell subpopulations, including Foxp3+ CD4 cells.

To test whether non-proliferating (Ki-67–) Foxp3+ CD4 T-cells might suppress T-cell proliferation, we plotted the correlation between the presence of this population (represented by their proportion of all lymphocytes within the TLS domain) and the Ki-67 expression in the other T-cell subsets ( Figure 4C ). The proportion of non-proliferating Foxp3+ CD4 cells per TLS domain did not show an inverse correlation with the presence of Ki-67 on other T-cell populations (nor B-cells) suggesting they were not suppressing T-cell proliferation within TLS.

Accurate measures of the total quantity of TLS within tumour sections might generate clinically-informative parameters, given that parameters such as T-cell counts within CRC tissue correlate with increased survival (32). The striking variation in the area of individual TLS domains (e.g. Figure 2 ), and the multi-lobular nature of TLS in at least some CRC samples (e.g. Figure 2J ), suggested that TLS count would poorly represent the total area of the TLS domains within a complete section. We therefore measured the total area of TLS within each section and compared this parameter with the canonical approach of counting the number of TLS domains.

Our data show that the count of TLS domains ( Figure 5A ) does not always reflect total TLS area within the section ( Figure 5B ). When the two parameters are compared against each other in a scatterplot, a weak correlation is observed with notable outliers ( Figure 5C ). For example, although P006, P013 and P004 are very similar in terms of TLS count ( Figure 5A , highlighted samples) they have very dissimilar total TLS area ( Figure 5B ) due to different average TLS domain areas, as is clearly evident in images from these subjects ( Supplementary Figure S3 ). Hence measuring total TLS area across an entire section provides an alternative measure of the total quantity of TLS domains that accounts for the substantial variation in area of individual TLS domains, which TLS domain counts alone do not.

Assessing total TLS area within a section needs to take account of total section size, and the proportion of the section that is occupied by tumour. Within our 22 samples, the proportion of the total tissue area occupied by TLS varied substantially ( Figure 5D ) and there was no correlation between total section area and the area occupied by TLS ( Figure 5E ). Total TLS area did not correlate with total tumour area ( Figure 5F ) nor the proportion of the total section occupied by tumour ( Figure 5G ). Hence total TLS area varied independently of tissue section size and the degree of tumour infiltration.

To compare whether the area of TLS within a section might correlate with TLS activation, we plotted the proportion of total section area occupied by TLS against the average proportion of B-cells expressing Ki-67 in each TLS domain ( Figure 5H ). No correlation was observed suggesting that the level of TLS activation represented by B-cell proliferation is not dependent on TLS area.

Measurement of activation or proliferation of B- and T-cells across entire tissue sections may help correct for the failure of the section plane to transect structures such as GCs, while also simplifying analysis. We therefore re-analysed our data by combining data from all TLS domains identified within each section, and then ranking patients by the proportion of cell subsets across all TLS. When compared with the previous method based on identifying individual circumscribed TLS ( Figures 3A, B ), this whole-section approach led to broadly similar rankings of patient TLS activity ( Figure 6A ), with the same 8 subjects top-ranked, but with significant re-positioning of some patients with low to moderate levels of B-cell proliferation ( Supplementary Figure S4 ). Plotting the proportions of B-cells expressing Ki-67 against the Ki-67+ proportions of T-cell subsets showed improved correlations based on this whole-section aggregated data compared with the previous method based on individual TLS zones ( Figure 6B compared with Figures 4A, B ). This simpler whole-section analysis confirmed that proliferation of B-cells is highly correlated with proliferation of all T-cell subsets.