Machine Learning-Based Predictive Modeling Maximizes the Efficacy of mTOR/TP53 Co-Targeting Therapy Against Acute Myeloid Leukemia

The mammalian target of rapamycin (mTOR) is a serine/threonine kinase that plays a critical role in regulating cell growth, proliferation and survival. The mTOR pathway is often hyperactivated in acute myeloid leukemia (AML) and is associated with poor prognosis, making it a promising therapeutic target. However, mTOR inhibitors have shown heterogeneous effects in clinical trials, reflecting the biological heterogeneity of AML. Therefore, it is important to identify predictive biomarkers that can recognize AML subtypes affected by mTOR signaling. In addition, the development of combination therapies with mTOR inhibitors is also an important challenge to improve the therapeutic efficacy. We first investigated the gene alterations and genetic dependencies that affect the sensitivity of AML cells to the mTOR inhibitor rapamycin using the Cancer Cell Line Encyclopedia (CCLE), Cancer Dependency Map Portal (DepMap) and Beat AML databases. We found that AML cells with TP53 mutations were unresponsive to rapamycin, whereas those dependent on MDM2 (a negative regulator of TP53) were sensitive to rapamycin. These findings suggest that the anti-leukemic effect of mTOR inhibition is mediated in part through the TP53 pathway. We then assessed the combined effect of rapamycin and a TP53-activating drug, DS-5272, in human AML cell lines with wild-type TP53: MOLM13, MV4-11, and OCI-AML3. The combination of rapamycin and DS-5272 exhibited significantly higher cytotoxicity compared to either drug alone, with combination index values of less than 1. Mechanistically, mTOR inhibition and TP53 activation cooperatively induced downregulation of MYC and MCL1 at the protein level, thereby inducing cell cycle arrest and apoptosis. Thus, co-targeting of mTOR and TP53 synergistically inhibits the growth of AML cells with wild-type TP53. To further identify specific AML subtypes that are susceptible to mTOR inhibition, we next applied the Joint Non-Negative Matrix Factorization (JNMF) to CCLE datasets containing 31 human AML cell lines, gene expression profiles, and drug sensitivities. JNMF is an unsupervised clustering algorithm that retrieves underlying features from sparse and high-dimensional multimodal data. We retrieved 60 feature sets comprising genomic, transcriptomic, and morphological features relevant to specific AML subtypes as common pattern modules (CPMs) of JNMF. Interestingly, four CPMs (CPM-4, 14, 16 and 40) were associated with high sensitivity to rapamycin and were enriched in monocytic AML cell lines defined as FAB-M5. This finding suggests that monocytic AML, which is known to be preferentially dependent on MCL1 for energy metabolism and survival, is highly sensitive to mTOR inhibition. We then applied a statistical regression algorithm to relate gene expression to rapamycin sensitivity in human monocytic AML cell lines and patient samples using the most expressed top 20 genes in the CPM-14 that showed the highest sensitivity to rapamycin. Using the random forest regression, we detected the optimal 11-gene signature (Rapa-11 score), which is the weighted sum of the expression of the 11 genes in each cell line. Remarkably, our machine learning-based modeling predicted that low Rapa-11 scores are strongly associated with high rapamycin sensitivity in monocytic AML, which would be the ideal target AML subtype for the rapamycin-based therapies. Finally, we investigated the combined effect of rapamycin and DS-5272 using three mouse AML models driven by MLL-AF9, SETBP1/ASXL1 mutations, or RUNX1-ETO. MLL-AF9-expressing cells are monocytic AML cells with a low Rapa-11 score. AML cells expressing SETBP1 and ASXL1 mutations (which we termed cSAM cells) are also monocytic AML but have a relatively high Rapa-11 score. RUNX1-ETO-expressing cells are more primitive AML cells. Consistent with our in silico prediction, rapamycin treatment showed marginal or no effect on the in vivo growth of cSAM and RUNX1-ETO cells, respectively. In contrast, co-treatment with rapamycin and DS-5272 showed a dramatic in vivo effect on MLL-AF9 cells, curing 85% of the leukemic mice. In summary, we showed that mTOR/TP53 co-targeting therapy has a curative effect on monocytic AMLs with wild-type TP53 and low Rapa-11 score. The machine learning-based predictive approach has the potential to be applied to other drugs and tumors to develop optimally tailored treatments for individual patients.