Coordinated Immune Cell Networks in the Bone Marrow Microenvironment Define the Graft Versus Leukemia Response with Adoptive Cellular Therapy

Donor lymphocyte infusion (DLI) is an established therapy for relapsed acute myeloid leukemia (AML) after hematopoietic stem cell transplant (HSCT), but response rates are poor (~20%). Interactions between leukemia and immune cells within the leukemia microenvironment may determine responsiveness to adoptive cellular immunotherapies. We hypothesized that systematic characterization of the leukemic marrow microenvironment over treatment course with DLI would define leukemia-immune cell interactions critical to response and hence provide mechanistic understanding of the graft-versus-leukemia (GvL) effect. In particular, we anticipated that use of multi-modal single-cell sequencing across patients and time points before and after therapy accounting for temporal dependencies would facilitate disentangling of the complex microenvironment of leukemic marrow. To infer cell-cell interactions in a temporally resolved fashion from single cell RNA-sequencing (scRNA-seq) data, we developed DIISCO, a Bayesian model using a Gaussian process regression network. In total, we profiled 30 bone marrow aspirates (14 pre- and 16 post-DLI) from 9 patients (5 responder [R], 4 nonresponder [NR]) with post-HSCT relapsed AML treated with DLI by scRNA-seq, single cell TCR sequencing (scTCR-seq) and surface protein characterization (scCITE-seq). We obtained 51,371 high-quality transcriptomes from the marrow samples along with 43,918 cells from the DLI products of these patients (4 R, 3 NR) for a total of 95,289 cells. Cells were clustered with Phenograph resulting in 57 clusters, including 2 CD4+ T cell, 6 CD8+ T cell (C0, C5, C21, C25, C26, C40), 3 NK, 10 B cell, 3 AML leukemia, and several myelo-erythroid clusters. DIISCO revealed a cascading response post-DLI that centered around CD8 C0, CD8 C5, AML, and a B cell cluster in R, not observed in NR. Furthermore, C0 appeared to expand after DLI only in R. We observed a strong negative interaction from CD8 C0 to AML after DLI, suggesting an anti-leukemia immune response, supporting the notion that immune cell populations formed a coordinated interactome associated with effective GvL. Evaluation of inferred interacting pairs using DIISCO and with the tool CellPhoneDB pointed toward an activation circuit in R between CD226 on CD8 C0 and PVR/NECTIN2 on AML. Conversely, inhibitory interactions between the exhaustion marker TIGIT on CD8 C0 and AML were observed in NR. Because DIISCO pinpointed CD8 C0 as a central hub for response, we charac