Introduction: Chronic myeloid leukemia (CML) is a type of cancer affecting the development of myeloid cells in the bone marrow. It results from translocation of the ABL proto-oncogene on chromosome 9 with the breakpoint cluster region on chromosome 22. The current first line treatments for CML are tyrosine kinase inhibitors (TKIs), which effectively prolong the lifespan of most CML patients when taken continuously (i.e. for life). Questions regarding long term impacts of TKIs on comorbidities or with respect to predisposing patients to other conditions remain unanswered. Since the FDA approved the first TKI in 2001, there has been a high volume of literature published in this domain We leveraged SemNet, a literature based discovery tool, to identify and agnostically rank prevalent comorbidities and medical conditions which may be predisposed by TKIs.

Methods: SemNet is an LBD tool which navigates over a knowledge graph created from United Medical Language System (UMLS) concepts and predications taken from SemMedDB, which is a repository of nearly 100 million relationships scraped from PubMed abstracts [1]. A knowledge graph consists of nodes and edges; the knowledge graph we utilize consists of nearly 300,000 nodes representing biomedical concepts, and approximately 20,000,000 edges, which are predications depicting relationships between the nodes. SemNet utilizes metrics such as Hetesim and novelty score to identify and rank relationships between user specified target nodes and source nodes retrieved from a query to the graph. Hetesim is a similarity metric which is effective for identifying pre-established relationships. Novelty scores down-weight highly connected, noisy nodes and emphasize understudied relationships. A simulation was run to identify diseases or syndromes related to CML. Source nodes were 396 total nodes representing diseases or syndromes and treatments related to CML; these were ranked based on Hetesim and novelty scores with respect to the target node “chronic myeloid leukemia”. Two subsets of 20 nodes were examined: those with the highest novelty rank and those with the highest Hetesim rank. Relationships between the target node and those in the subsets were confirmed based on literature review and subgraph visualization tools depicting highly occurring predications between sources and targets.

Results and Discussion: Notable nodes which fell in both the high-rank novelty and Hetesim (H) subsets included “aplastic anemia” (H:1, N:7), “pancytopenia” (H:2, N:9), “hypocellular bone marrow” (H:10, N:1), and “red-cell aplasia” (H:19, N:10); each of these nodes pertains to blood cell deficiency. Hypocellular marrow can also lead to other hematological malignancies like myelofibrosis. TKIs inhibit leukemic blood cell production, and by using subgraph relationship verification methods we identified that they can inhibit other types of blood cell production. As such, we formed the hypothesis that in some cases, TKIs may “overcorrect”. Follow up simulations with “tyrosine kinase inhibitors” and “aplastic anemia” and “pancytopenia” were performed to provide more details on the mechanism behind this phenomenon. Highly ranked nodes from these simulations suggested that TKIs inhibit biological substances such as interleukin-3 (H:2, N:7) and erythropoietin (H:17, N:10), which promote blood cell production and treat conditions such as pancytopenia and aplastic anemia.

Conclusions: Key findings from our use of SemNet literature mining study included relationships between nonCML related blood cell depletion (i.e. patients who had obtained a complete BCR-ABL1 cytogenetic response) and TKI usage. The enhanced risk of abnormal cell depletion with TKI use could enhance other TKI-related adverse events. For future studies, SemNet will be utilized to better identify risk factors which predispose patients to this potential over-correction phenomenon, such as brand of TKI, dosage, or other co-morbidities. Better understanding will enhance predictive medicine to determine optimal personalized TKI treatment.