Super-Jump: State Intent Inference and Lineage Tracking through Semi-Supervised Jump Diffusion Modeling in Normal Hematopoiesis and Acute Myeloid Leukemia

Background: Hematopoiesis is arranged in a hierarchy of stem and progenitor cells capable of producing distinct cell fates. Blockade of these differentiation processes is a critical feature of myeloid leukemia development, with different mutation classes resulting in distinct maturation states. Understanding the cellular and transcriptional context for differentiation blockade and biases may offers insights into maximizing differentiation therapy and tracking disease evolution.

Single cell RNA-seq (scRNA) have provided unprecedented resolution into cellular and clonal heterogeneity. Computational approaches have made recovering differentiation dynamics tractably possible. However, many of these tools were developed through the lens of embryogenesis, a setting not well tuned for the hierarchical, multi-potency of hematopoiesis. Current approaches do not evaluate discontinuous differentiation processes present in malignant leukemia. To address these gaps, we developed SuperJump: a jump-diffusion based supervised cell-fate model. We deploy this approach in 1) normal murine, aged hematopoiesis, and 2) murine models of acute myeloid leukemia with distinct orders of mutation. We identify cells poised for differentiation, and their underlying transcriptional networks.

Approach: SuperJump is split into 3 phases. In phase 1: we use a semi-supervised method to select positive gene expression signatures from known terminal cell types. We aggregate these signatures as a differentiation entropy score. Here we query multiple cell fates in one model, whereas other strategies aim to rediscover the biological hierarchy for each fate separately.

In phase 2, we develop a cell-cell transition matrix using a stochastic model-based strategy. Our innovative jump-diffusion model allows for discontinuous interactions between cells to determine the transition probabilities, allowing for a broader reach of neighboring cell states.

In phase 3, we aim to identify cell transition likelihood using an absorbing Markov chain. We specify cell types as sources and sinks of differentiation. Sink cells are identified as those with the highest membership to their own cell-type. The likelihood for a cell to transition to any given sink is calculated and termed “fate propensity”. Differentiation-poised cells are identified within each population through an outlier detection statistic. Finally, poising is then associated with RNA expression and inferred transcription factor activity.

Results: First, we ran SuperJump on a well-established model of myeloid fate-bias in normal hematopoiesis from young and middle-aged mice (n=52967 cells). We queried B cells, neutrophils and monocytes as sink cells, and sought to identify poising bias with aging. In the middle-aged sample, we observed myeloid-poised cells as early as HSC (0.08%), a modest expansion in MPPs (0.25%) and a significant degree of poising in middle-aged GMPs (18.9%). This was in stark contrast to young mice which showed no myeloid poising at any stage. We found that myeloid-poised MPPs and GMPs were associated Irf8 expression (MPPs: p=7.52x10^-6,GMPs: p= 7.13x10^-39) and MPPs poised for B-cells were associated with Pax5 activity (MPPs: p=2.84x10^-2). These data highlight early stages of differentiation poising in HSPCs to specific cell fates.

We next investigated a multi-recombinase murine leukemia model mutant for Dnmt3a, Npm1 and Flt3 (n=5215 cells). In this model we varied the order of mutation between Npm1 and Flt3 and report how mutation order influences cell-fate trajectories. When Flt3 is the last mutation, we observe an expansion of common dendritic progenitors (CDP) which are poised for plasmacytoid dendritic cell differentiation. These cells were enriched for Tcf4 (p=2.200x10^-16) and Socs2 (p=5.023x10^-16), potentially linking Stat5 signaling and active proliferation towards pDC fate. How mutation order relates to cell fate will be further explored.

Conclusion: Here we deploy a new approach for identifying differentiation trajectories in normal and malignant hematopoiesis. Our approach is innovative in the use of supervised differentiation signatures, deployment of jump-diffusion to model discontinuous differentiation, and the identification of poised cell states. Our approach allows for insights into early progenitor fates and underlying mechanisms driving them.