Immune infiltrate diversity confers a good prognosis in follicular lymphoma

The study was conducted with approval from the North–West Multi-centre Ethics Committee (03/08/016), according to the Declaration of Helsinki. Examination of 350 FL patients’ electronic records in a random order from the archives of The Christie NHS Foundation Trust, Manchester, UK, identified 262 patients meeting the inclusion criteria: adult; diagnosed from incisional or excisional biopsy; non-primary cutaneous; and treated at first presentation with radiotherapy, watchful waiting, or rituximab-based systemic therapy. Pre-treatment diagnostic biopsies were requested for 262 patients, of which 131 had sufficient tissue. The 131 patients were diagnosed between 1998 and 2015, with a median follow-up of 114 months (range 3–199 months). Histological diagnosis of FL was re-confirmed by an expert haemato-pathologist (R.B). Regions of interest were identified by the haemato-pathologist, and cores were extracted in triplicate from formalin-fixed, paraffin-embedded (FFPE) blocks to construct five tissue microarrays (TMA). Core diameter was 1.2 mm. Follicular and extrafollicular regions were both selected for inclusion in the TMA. A section of each TMA stained with H&E is provided [28], to demonstrate the morphology of selected regions. Some cores were excluded because of poor quality, leaving 349 cores from 130 patients. The OS distribution was similar between the 130-patient cohort and the patients with insufficient tumour material (Supplementary Fig. 1, log rank test p = 0.37). Furthermore, three patients with FL grade 3b were excluded from survival analyses, as their disease progression and treatment pathways resemble more closely Diffuse Large B-cell Lymphoma and grade 3b FL is generally considered a separate disease entity [29]. Finally, 342 cores from 127 patients remained available for further analyses.

OS was recorded for all patients (N = 127). Progression-free survival (PFS) and disease progression within 24 months of starting treatment (POD24) were recorded only for patients treated with rituximab-containing immuno-chemotherapy at first presentation (N = 67). The events of disease progression and relapse were defined using the Lugano criteria [30]. PFS was calculated from diagnosis until the first observed progression event (or disease-specific death) or, if no events were observed, until the date of last follow-up.

A panel of antibodies was selected to identify non-neoplastic immune infiltrate subsets. CD68 was selected to observe monocytic cells and particularly macrophages [10, 11]. CD8 was used to observe cytotoxic T cells (CD8⁺) [8]. T helper cells were identified by inclusion of the CD4 marker, as the subset that expressed CD4, but not CD68 [8]. Additionally, FOXP3 was used to identified T regulatory cells (Tregs [CD4⁺FOXP3⁺]) [26]. The CD21 marker was used to identify follicular dendritic cell areas [31]. While CD21 might also be expressed in B-cells, follicular dendritic cells in follicular lymphoma can be identified by their characteristic meshwork staining pattern [32]. PD-1 detected CD4⁺PD-1⁺ T follicular helper cells and CD8⁺PD-1⁺ lymphocytes [20]. Finally, DAPI (4′,6-diamidino-2-phenylindole) was the nuclear counterstain. No B-cell tumour marker was used, as we aimed to study the diversity of the non-neoplastic microenvironment.

A single 4-μm section was cut from each of the five TMAs of the cohort. These sections were stained with a 6-plex immunofluorescent assay using tyramide signal amplification, the Opal 7 colour kit (Akoya Biosciences, CA, USA), and the Ultra Discovery auto-stainer (Roche, Switzerland), as described in the experimental protocol of Tsakiroglou et al. [33], and are summarised in Supplementary Table 1. The OPAL detection system allows for repeated cycles of staining and stripping of anti-mouse and anti-rabbit antibodies. In each cycle, the primary antibody and secondary conjugated to horseradish peroxidase (HRP) are incubated on the tissue, followed by fluorophore-labelled tyramide. The catalytic reaction that occurs binds the fluorophore strongly on the tissue epitopes. Thereafter, heat stripping is applied to remove the antibody before applying the next one, but the fluorophores remain on the tissue. We quantitatively validated good agreement between single-plex and multiplex immunofluorescent assays (Supplementary Material Sect. 2, Supplementary Fig. 2–5).

Stained sections were scanned with the Vectra 3.5 microscope (Akoya Biosciences). Initially, a low-resolution scan (10x) was performed to manually annotate the TMA core locations. Then, a multi-spectral image of each core was acquired at 20 × magnification (0.49 μm/pixel). Spectral unmixing [34] was performed using inForm 2.4 software (Akoya Biosciences). To separate the fluorophore signals, a spectral library was pre-built; the individual spectrum of each fluorophore, DAPI, and auto-fluorescence was acquired from single-plex controls. After unmixing, the images consisted of 6 channels, each containing the intensities of a different fluorophore, plus two channels for DAPI and auto-fluorescence (\(2420\times 2420\times 8\) pixels). All prepared images (example in Fig. 1) are publicly available at http://​dx.​doi.​org/​10.​17632/​274xbhc5rx.​3. Areas containing artefacts were manually excluded.

Nuclear detection in these images was challenging, because of densely packed and overlapping cells. Therefore, label-free cell detection, as offered by inForm software, was insufficient. Instead, we trained a deep learning model on the DAPI channel using the “StarDist” method [35] (code available: https://​github.​com/​mpicbg-csbd/​stardist). The model was trained on 67,991 nuclear outline annotations, was fine-tuned using an additional validation set of 906 nuclei, and achieved average precision of 83% when tested on 883 unseen nuclei. The training, validation, and test sets were created for the purpose of this study by a trained non-expert, under supervision from a pathologist (R.B.). More details on the performance validation can be found in the supplementary material (Supplementary Material Sect. 3, Supplementary Fig. 6, and Supplementary Table 2). The nuclear annotations are publicly available (http://​dx.​doi.​org/​10.​17632/​nb46s9trx3.​1). After nuclear detection, simulated membranes were grown around the nuclei by maximum 1.5 μm to represent whole cells (Supplementary Fig. 7) and measurements are taken of the median intensity for all stains and each cell compartment (nucleus, membrane).

A single threshold was selected to mark positive cells for each stain across the entire dataset by observing the median stain intensity on the relevant compartment (nucleus for FOXP3 and membrane for all others). To select positivity thresholds, two annotators examined 10 multiplex images (Supplementary Material Sect. 4 and Supplementary Table 3), and the average threshold values were chosen. Cell density was subsequently measured for each cell phenotype of interest by dividing the number of positive cells by the tissue area.

When testing the prognostic significance of individual cell subsets, the total number of cells expressing a stain were observed to facilitate comparisons with prior FL studies. For example, when observing the cell subset that expressed CD68, all CD68⁺ cells were counted, regardless of the expression of other markers. In this way, the findings can be compared with previous FL studies that used single-plex assays to observe the prognostic value of CD68 stain [9‐12].

Additionally, the total immune infiltrate was measured as the number of cells expressing any of the CD4, FOXP3, CD8, CD68, or PD-1 markers. The immune infiltrate ratio was subsequently calculated by dividing the immune infiltrate cells by the number of all non-immune cells that expressed only DAPI. This ratio can be used to represent the extent of total immune infiltration in the microenvironment of FL.

To quantify the extent of CD21⁺ follicular dendritic meshwork areas, manual annotations were drawn around them for all samples (Supplementary Fig. 8), under supervision of a pathologist. Subsequently, the area covered by these meshwork patterns was measured and expressed as a proportion of the total tissue area.

Immune infiltrate diversity was assessed by computing Shannon’s entropy for cell populations. Shannon’s entropy was calculated as:where \(N\) is the number of all possible cell phenotypes and \({p}_{i}\) the proportion of each phenotype \(i\) in a sample.

The diversity of spatial interactions in each sample was additionally quantified by applying the HID methodology [27]. HID performs a pair-wise examination of cell types identified during cell scoring and counts their spatial interactions, i.e. their frequency of co-occurrence within a pre-specified distance (see Fig. 1k–l). Further implementation details can be found in Rose et al. [27]. The distance parameter was selected as 30 μm similar to the study of Tsakiroglou et al. [36], which represents a neighbourhood of 3–4 cells. A co-occurrence between each pair of phenotypes \(i\) and \(j\) was considered a unique type of spatial interaction. The proportion of all interactions belonging to this type \({p}_{i,j}\) could then be calculated. If we consider each type of interaction as a separate “species”, Shannon’s entropy diversity index for the distribution of interactions in a sample can be derived:

In the current study all \(N\) phenotypes that were observed in the samples across the entire dataset were assessed, while cells expressing only DAPI were ignored. Intuitively, interaction entropy quantifies the diversity of co-occurrences between immune cell subsets.

A summary of the methodology steps is given in Fig. 1j.