Normal Hematopoietic Stem Cells in Leukemic Bone Marrow Environment Undergo Morphological Changes Identifiable by Artificial Intelligence

TL;DR

This study shows that normal stem cells in a leukemic bone marrow environment can be identified by AI based on their unique morphological changes, which could help assess treatment responses and prognosis.

Contribution

The study demonstrates that AI can detect morphological changes in normal hematopoietic stem cells within a leukemic environment with high accuracy.

Findings

Non-JAK2V617F-expressing HSCs are distinguishable from LSCs in the same bone marrow environment by AI with high accuracy (>96%).

Non-JAK2V617F-expressing HSCs from PV mice are morphologically distinct from normal HSCs by AI with accuracy >98%.

AI-recognizable morphological changes in non-leukemic HSCs suggest potential for assessing therapy responses and prognosis.

Abstract

Leukemia stem cells (LSCs) in numerous hematologic malignancies are generally believed to be responsible for disease initiation, progression/relapse and resistance to chemotherapy. It has been shown that non-leukemic hematopoietic cells are affected molecularly and biologically by leukemia cells in the same bone marrow environment where both non-leukemic hematopoietic stem cells (HSCs) and LSCs reside. We believe the molecular and biological changes of these non-leukemic HSCs should be accompanied by the morphological changes of these cells. On the other hand, the quantity of these non-leukemic HSCs with morphological changes should reflect disease severity, prognosis and therapy responses. Thus, identification of non-leukemic HSCs in the leukemia bone marrow environment and monitoring of their quantity before, during and after treatments will potentially provide valuable information for correctly handling treatment plans and predicting outcomes. However, we have known that these morphological changes at the stem cell level cannot be extracted and identified by microscopic visualization with human eyes. In this study, we chose polycythemia vera (PV) as a disease model (a type of human myeloproliferative neoplasms derived from a hematopoietic stem cell harboring the JAK2V617F oncogene) to determine whether we can use artificial intelligence (AI) deep learning to identify and quantify non-leukemic HSCs obtained from bone marrow of JAK2V617F knock-in PV mice by analyzing single-cell images. We find that non-JAK2V617F-expressing HSCs are distinguishable from LSCs in the same bone marrow environment by AI with high accuracy (>96%). More importantly, we find that non-JAK2V617F-expressing HSCs from the leukemia bone marrow environment of PV mice are morphologically distinct from normal HSCs from a normal bone marrow environment of normal mice by AI with an accuracy of greater than 98%. These results help us prove the concept that non-leukemic HSCs undergo AI-recognizable morphological changes in the leukemia bone marrow environment and possess unique morphological features distinguishable from normal HSCs, providing a possibility to assess therapy responses and disease prognosis through identifying and quantitating these non-leukemic HSCs in patients.

Figures