AN EXPLAINABLE ARTIFICIAL INTELLIGENCE mDELRelapseNet MODEL TO PREDICT RELAPSE IN DIFFUSE LARGE B-CELL LYMPHOMA BASED ON THE SPATIAL ORGANIZATION OF MYC+BCL2+BCL6− Cells

Abstract

Introduction: The overexpression of MYC and BCL2 categorizes diffuse large B-cell lymphoma (DLBCL) termed double-expressor lymphoma (DEL) with worse survival after chemoimmunotherapy. The clinical utility of DEL is limited by controversy in cut-offs for positivity of these markers, and a possible protective effect of BCL6 expression. Using single-cell resolved imaging, we showed (Hoppe et al., Cancer Discovery 2023) that survival is robustly associated with the fraction of malignant cells that co-express the oncogenes MYC and BCL2 in the absence of BCL6 (M+2+6−), refining the DEL definition. Here we present a follow up study evaluating the clinical significance of the spatial distribution of M+2+6− cells, and its potential clinical applicability through an XAI model “mDELRelapseNet”.

Methods and Results: M+2+6− cells display non-random spatial organization within a tumour. To evaluate the significance of these patterns, we employed Geyers point process analyses- a method widely used in ecology and geography- to understand the spatial distribution of M+2+6− cells within DLBCL using x-y coordinate information from multiplexed fluorescent immunohistochemistry (mfIHC) images. Cases could be divided into two groups based on these: one with “clustered” and another with “dispersed” M+2+6− cell distribution. Interestingly, cases with “dispersed” pattern of M+2+6− cells consistently had shorter survival in all analyzed cohorts (p < 0.05 in 4 independent cohorts; N = 449 patients), the first description to our knowledge that the spatial organization of a subset of tumour cells influences clinical outcomes in cancer.

We then aimed to harness this spatial information from MYC, BCL2 and BCL6 staining to develop an XAI model to predict for relapse in DLBCL. A challenge however was the lack of versatile analysis tools to handle diverse image formats and marker combinations. We addressed this by developing a unified ground-up deep learning model “mDELRelapseNet”; to accept any standard image format with any combination of these markers, eliminating the need for specialized tools for each scenario. Our model achieved a validation accuracy of 70% and was trained on mfIHC and pseudo IHC images from two DLBCL cohorts (n = 253) and validated on a third (n = 18). Through backtracking, we saw that the early layers learnt from M+2+6− hotspots for prediction. We refined the model by providing the cell of origin classification, as we noted dispersed M+2+6− cells to be enriched in ABC DLBCL. This improved performance to 94.8% validation accuracy in ABC and 94.1% in GCB DLBCL.

Conclusions: We show that survival in DLBCL is linked not only to the numbers of M+2+6− cells but also their spatial organization. We created a web app (https://mdel-relapse-net.streamlit.app) that uses histopathological images of MYC/BCL2/BCL6 to identify DLBCL at high risk of R-CHOP failure, with potential applicability for patient selection in clinical trials of novel agents.

Research funding declaration: Anand D. Jeyasekharan was supported by the Singapore Ministry of Health’s National Medical Research Council Clinician Scientist Award (MOH-000715-00). Work in ADJ’s laboratory is funded by the core grant from the Cancer Science Institute of Singapore, National University of Singapore through the National Research Foundation Singapore and the Singap

Keywords: bioinformatics; computational and systems biology; tumor biology and heterogeneity; aggressive B-cell non-Hodgkin lymphoma

Potential sources of conflict of interest:

A. D. Jeyasekharan

Consultant or advisory role: Anand D. Jeyasekharan has received consultancy fees from DKSH/Beigene, Roche, Gilead, Turbine Ltd, AstraZeneca, Antengene, Janssen, MSD and IQVIA; r

Other remuneration: Anand D. Jeyasekharan has received research funding from Janssen and AstraZeneca.