Adult T-cell leukemia/lymphoma (ATLL) is a T-cell malignancy with poor prognosis caused by human T-cell lymphotropic virus type 1 (HTLV-1). ATLL accounts for 9.6% of mature T-cell neoplasms. ATLL is endemic in several areas in the world and there are geographical differences in the frequency of ATLL. As conventional chemotherapy is not fully effective for ATLL patients, new therapeutic strategies are required. ATLL tumor cells are highly antigenic because of viral infection and tumor microenvironment is important for survival and progression of ATLL. We analyzed lymph nodes of ATLL patients by immunohistochemistry and found that overexpression of programmed cell death ligand-1 (PD-L1) and loss of major histocompatibility complex class I or II in tumor cells were adverse prognostic factors. In addition, we also examined expression of 15 immune checkpoint molecules in lymph nodes of ATLL patients by immunohistochemistry and found that increased number of stromal cells expressing PD-L1, OX40 ligand, and T-cell immunoglobulin mucin-3 (Tim-3) were favorable prognostic factors. Of note, ATLL patients with both MHC class II expression in tumor cells and increased number of PD-L1-positive stromal cells showed favorable prognosis compared to other patients. In conclusion, immune escape and interaction between tumor cell and specific stromal cells are important for tumor microenvironment. Our results may provide information on new therapeutic strategies for ATLL patients.
 

Background: Self-supervised learning is a versatile deep learning methodology which can encode histological features as quantitative vectors. This approach requires no up-front hypothesis or manual annotation. Our pipeline, histomorphological phenotype learning (HPL), breaks whole slide images (WSIs) into small tiles and learns features from each tile, encoding them numerically. These vector representations from each tile are clustered, to construct groups of morphological similar tiles. Each patient's tumour can then be expressed as a proportion of clusters. Crucially, representative images from each cluster can be reviewed by eye, providing a layer of interpretability.

Purpose: We demonstrate the ability to construct a dictionary of recurrent phenotypes spanning the spectrum seen in lung adenocarcinoma. Furthermore, we integrate spatially resolved morphological features with multiplex immunofluorescence.

Methods: We trained HPL using 4427 WSIs from 1007 patients. These were consecutive surgical resections of lung adenocarcinoma from a single centre. We used Cox proportional hazards to generate risk scores for overall and recurrence-free survival. We ran images of 23 TMAs through the trained model to assign cores to clusters and applied an immune-focused multiplex immunofluorescence panel on a serial section.

Results: There were 64 morphological clusters. Some of these are defined by classical growth patterns, while others by recurrent stromal appearances. Appearances with high fibroblast burden were associated with poor outcome and those with increased lymphocyte burden with favourable outcome. For each cluster we calculated cell phenotype fractions and linked these to prognosis.

Conclusions: Self-supervised learning is a versatile deep learning methodology which can enable discovery of clinically useful prognostic features and tumour biology. Our approach demonstrates a way to identify and quantify histological features and integrate them with spatial proteomic data.

This work is supported by a Pathological Society of Great Britain & Ireland and Jean Shanks Foundation clinical PhD fellowship.
 

Background: Specific HLA-DQ molecules, presenting deamidated gliadin peptides (with origin in dietary gluten) underpin the pathogenesis of coeliac disease and promote T-cell mediated duodenal injury. Current diagnostic methods require gluten consumption and use observation by pathologists of small intestinal biopsies, a process sometimes affected by subjective interpretation, leading to poor interobserver concordance (73% using only duodenal biopsies and 80% using duodenal biopsies with IgA tissue transglutaminase and haemoglobin data). The difficulty in ensuring patients with coeliac disease are diagnosed is exacerbated by the diverse and non-specific presentations of coeliac disease including anaemia, intestinal symptoms, osteoporosis, lymphoma, intestinal cancer and sometimes the lack of obvious signs or symptoms.
Purpose: With the increasing prevalence of coeliac disease and rate of diagnosis around 36%, there is an unmet need for a new gold-standard of testing.
Methods: We investigated the use of T-cell receptor (TCR) repertoire characteristics as a diagnostic tool. To build an interpretable machine learning model, we sequenced mucosal CD4+ T-cell TCR repertoires from 20 patients (12 with coeliac disease of which 5 were on gluten-free diets; 8 healthy controls) as a training dataset. We tested the diagnostic potential of our machine learning model using independently published TCR sequence data.
Results: Diagnosing coeliac disease showed a training accuracy of 100% (using 44 TCR alpha sequences alone and when combined with 28 TCR beta sequences), including patients on a gluten-free diet (using TCR alpha sequences alone). Testing accuracy was 80% (using TCR alpha sequences alone).
Conclusions: We identified the set of 20 CD4+ TCR sequences (10 TCR alpha and 10 TCR beta) with the greatest potential to discriminate between duodenal biopsies from patients with coeliac disease and healthy controls, thus developing a prototype for a more objective, gluten-independent, diagnostic test for coeliac disease.
A Path Soc grant supported some of this research.
 

Background: Parkinson’s Disease (PD) and multiple system atrophy (MSA) are two α -synucleinopathies caused by different strains of abnormally folded α-synuclein (α-syn) (1,2) and present with differential neuronal vulnerability. In PD, the substantia nigra, amygdala and locus coeruleus are consistently severely affected with characteristic neuronal Lewy bodies, which may also be detectable in other brain regions depending on the stage of the disease. While in MSA, neurones in the caudate nucleus and putamen or the inferior olivary nucleus (ION) and pontine base are selectively severely affected with other brain regions affected to a lesser degree (3). However, the exact mechanisms underlying differences regional vulnerability across different forms of α -synucleinopathies remain elusive.
Purpose: Our study aims to identify the molecular basis of neurons of the ION, which are severely affected in MSA, but spared in PD using cellular and spatial resolutions of a novel spatial transcriptomics platform to analyse the differentially affected cells of the ION.
Methods: We analysed fresh-frozen human post-mortem brain tissue from the Queen Square Brain Bank using the Nanostring GeoMx spatial transcriptomic platform (5) to profile the transcriptome of neurons, astrocytes and microglia in the ION from cases of MSA and PD. Briefly, regions of interest (ROIs), were manually marked and appropriately segmented for each cell type, and automated molecular profiling performed in the cells for each ROI by photocleaving index oligomers from respective probes.
Results: Our data reveal a successful enrichment of regional cell-specific transcriptomes and that there is significant differential expression of genes between ION cells across control, PD and MSA cases indicating that specific disease mechanisms are active in each disease.
Conclusions: These novel findings greatly improve our understanding of the pathomechanisms underlying disease pathogenesis and how these are affected in two different α -synucleinopathies.