Multimodal Atlas of Paired Diagnosis and Relapse AML Reveals Surface Antigens for Multi-Specific Immunotherapy

Note: A.H. and M.U. share co-first authorship.

AML is an aggressive hematologic malignancy with few treatment options beyond chemotherapy and hematopoietic cell transplantation (HCT). Previously, we introduced a multimodal AML atlas of paired patient primary bone marrow mononuclear cell (BMMC) samples collected at diagnosis and relapse from a large patient cohort, revealing extensive inter- and intra-patient blast antigen (Ag) heterogeneity. We have since expanded this atlas to include additional samples from healthy donors, allowing us to categorize blasts into hematopoietic cell lineages. We leverage this large dataset of blast heterogeneity across 68 samples to identify novel Ags that can be targeted with multi-specific immunotherapies. In addition, we determined the number of target Ag molecules presented on individual blasts (Ag count), since immunotherapies including CAR-T cells and antibody drug conjugates (ADC) show varying sensitivities to Ag counts for effective blast recognition and killing. Here, we identify Ag combinations at therapeutic thresholds amenable to multi-specific immunotherapy and validate their targetability with in vitro cytotoxicity assays using CAR-T cells and ADCs.

Nearest-neighbor projection of AML BMMCs to the annotated healthy donor reference revealed considerable heterogeneity in blast state composition across both patients and timepoints suggesting that multi-specific targeting can enable comprehensive targeting of blast populations. As described previously, all samples in our atlas include both feature barcoding with antigen derived tags (ADT) on 81 Ags and QuantiBRITE flow cytometric data, measured as antibodies bound per cell (ABC), on four well-characterized AML Ags: CD33, CLL-1, CD123, and ADGRE2. Using patient-matched data from these two assays, we trained random forests to model the relationship between ABC and ADT readouts for these Ags and applied it to estimate Ag count for the remaining 77 Ags on individual blasts for all samples. This represents a novel method for integrating sequencing and fluorescence data with machine learning to infer Ag counts.

Effective target Ag combinations are collectively expressed across blasts and co-expressed on individual blasts to enable recognition by multi-specific therapies and guard against Ag escape. To demonstrate this on established AML Ags, we developed novel multi-specific CAR-T cells with an “OR” logic gate directed against CD33 and CLL-1. The lead candidates significantly reduced tumor growth, induced T cell expansion, extended animal survival, presented greater persistence in hematological compartments, and showed osteotropic activity in HL-60-based in vivo models of AML.

As observed in our atlas, CD33 and CLL-1 cover a considerable fraction of blasts but 25% of relapse samples had <80% of blasts expressing either CD33 or CLL-1 at a threshold of 1000 Ags/cell, a sufficient therapeutic threshold for some immunotherapies. Thus, to increase the likelihood of targeting all blasts we ranked Ags in our atlas which, when combined together, or with CD33 and CLL-1 label the maximum number of blasts in the most patients. We identified four additional Ag candidates: LAIR1, ITGA4, LY75, and CD244 that expressed >1000 Ags/cell on >80% of cells in 22, 24, 14, and 7 of the 28 diagnosis samples, and 21, 25, 14, and 10 of the 28 relapse samples, respectively. As validation, we used QuantiBRITE to measure ABCs of these four Ags on four samples selected from our atlas repository. We observed high consistency between the measured ABC readout and the predicted ABC for all four Ags and samples (Pearson r = 0.87, p = 9.9e-6). Next, we evaluated therapeutic targeting of these Ags individually using in vitro ADC cytotoxicity assays with MOLM-13 cells, which provides a sensitivity measure of the ADC as a function of Ag expression. We observed >50% ADC-mediated cell killing at <0.1nM of primary antibody for all four Ags, supporting their therapeutic targetability.

In summary, we integrated rich multimodal data from our atlas with machine learning and performed experimental validation to identify promising single and combinations of AML Ag targets including CD33, CLL-1, LAIR1, ITGA4, LY75, and CD244. These novel combinations can be leveraged to develop new multi-specific agents (e.g. CAR-T cells or ADCs) that better address AML heterogeneity.