Acute myeloid leukemia (AML) is a cancer of blood-forming cells in the bone marrow. It is highly heterogeneous, meaning that even within a single patient, the leukemic cells differ from one another in their genetic makeup and behavior. This diversity makes it extremely difficult to treat the disease effectively with a single drug.
Drug resistance is a major obstacle in AML treatment. Even when a drug initially works, cancer cells often find ways to escape or become resistant. For example, nearly 60% of patients with a specific mutation called FLT3 relapse after treatment with the targeted drug gilteritinib, partly because other cancer-driving pathways become activated.
To overcome resistance, researchers have long sought drug combinations that attack multiple cancer pathways at once. However, the number of possible two-drug combinations from a library of hundreds of compounds exceeds 100,000 - far too many to test systematically, especially when patient cell samples are scarce.
An additional challenge is toxicity: combinations that kill leukemia cells must spare healthy, non-cancerous cells. Older AML patients are especially vulnerable to treatment side effects, so finding combinations that are both effective and tolerable is critical.
This study introduces a functional precision medicine platform that combines cutting-edge laboratory methods with machine learning to design personalized drug combination therapies for individual AML patients. The goal is to find pairs of drugs that work synergistically - together killing cancer cells more effectively than either drug alone - while minimizing harm to healthy cells.
The approach starts with single-cell RNA sequencing (scRNA-seq), a technology that reads the gene activity profile of thousands of individual cells in a patient's bone marrow sample. This reveals the diverse mixture of cell types present, including different leukemic subpopulations and normal immune cells.
This molecular information is then combined with ex vivo drug sensitivity testing - exposing patient cells directly to hundreds of individual drugs in the laboratory to measure how each drug affects cell survival. These two data streams are fed into a machine learning model that predicts which drug combinations will be synergistic, effective against AML cells, and safe for normal cells.
The platform was applied to four bone marrow samples from three AML patients, each at different disease stages and with different genetic mutations, demonstrating its flexibility and personalized nature.
The computational core of this approach uses an algorithm called XGBoost - a powerful machine learning method that learns patterns from complex data. The model is trained separately for each patient using two inputs: the gene expression profiles of individual cells from scRNA-seq, and the measured responses of those cells to 456 individual drug compounds.
A key step involves converting detailed gene expression data into compound-target enrichment scores using a method called GSVA (Gene Set Variation Analysis). This process compresses thousands of gene measurements into a single score that reflects how strongly a drug's target proteins are active in each cell. This makes the data manageable and biologically meaningful.
To handle the uncertainty inherent in predictions, the researchers added a conformal prediction (CP) layer. This statistical approach assigns confidence levels to each prediction, automatically excluding low-confidence drug combination forecasts. Only combinations where the model was sufficiently certain were taken forward for experimental testing.
The model ranks predicted combinations by three criteria simultaneously: predicted synergy (drugs work better together than expected), predicted efficacy (potent killing of cancer cells), and predicted low toxicity (minimal harm to healthy immune cells like T cells and NK cells).
Across all four patient samples, the researchers tested 68 predicted drug combinations in the laboratory. Overall, more than 50% of predicted synergistic combinations were confirmed experimentally - a remarkably high success rate given that true drug synergy is estimated to occur in only about 5% of random drug pairs.
Two combinations showed broad synergy across all patient samples. Venetoclax (a Bcl-2 inhibitor) combined with vistusertib (an mTOR inhibitor) produced synergistic cancer cell death, consistent with prior evidence that simultaneously blocking survival and growth pathways triggers powerful apoptosis (programmed cell death) in leukemic cells.
The platform also identified patient-specific combinations - synergies unique to individual patients. For example, the combination of molibresib (targeting BET family proteins) and verdinexor (targeting nuclear export protein XPO1) was uniquely effective in patient AML1, where leukemic cells expressed abnormally high levels of BET protein targets. This synergy would have been impossible to predict without the integrated single-cell data.
The validation accuracy was especially high for shared combinations, with a true-positive rate of 79%. The false-negative rate - cases where synergy existed but was missed - was below 2%, demonstrating the model's high sensitivity.
A critical feature of effective cancer therapy is selectivity - killing cancer cells without harming normal cells. The researchers used high-throughput flow cytometry, a technique that can simultaneously identify and count different cell populations, to test whether the predicted combinations preferentially killed leukemic cells compared to healthy T cells and NK cells (natural killer cells), which serve as a stand-in for normal immune cells.
The results showed that the average inhibition of leukemic cells was 60%, compared to only 40% for nonmalignant immune cells - a statistically significant difference. Notably, 67% of the predicted and synergistic combinations showed low toxicity to T and NK cells, confirming that the model's toxicity predictions were accurate.
Some combinations involving venetoclax showed variable toxicity patterns depending on the specific combination partner and the patient. This highlighted that even widely used drugs behave differently across patients, reinforcing the need for personalized prediction rather than one-size-fits-all approaches.
The combination of saracatinib (an ABL/SRC kinase inhibitor) and cytarabine (a DNA synthesis inhibitor) was uniquely synergistic and selective in patient AML1, suggesting it targets cancer stem cells while preserving healthy immune function - an ideal property for a therapeutic combination.
For patient AML2 - a 68-year-old man with refractory (treatment-resistant) leukemia - single-cell sequencing revealed that monocytic blast cells expressed high levels of cytidine deaminase, an enzyme that inactivates the standard drug cytarabine. This molecular finding directly explained why the patient failed to respond to standard induction therapy.
The model predicted a unique combination for AML2: GSK2656157 (a PERK inhibitor) combined with docetaxel (a microtubule inhibitor). The predicted synergy was linked to overexpression of genes in the PERK-mediated stress signaling and microtubule pathways specifically in this patient's leukemic cells - not in other patients.
For patient AML3, who progressed from diagnosis to a treatment-refractory stage, the researchers tracked how the cell population changed over time. The proportion of erythroid-like cells (immature red blood cell precursors) changed between disease stages, and a new mast cell population appeared in the refractory sample. These shifts influenced which drug combinations were predicted to work.
Crucially, the model correctly predicted that azacitidine (the drug given clinically) would NOT be effective for AML3 at diagnosis - a prediction that matched the patient's actual clinical outcome. This demonstrates the platform's potential to guide treatment decisions and avoid ineffective therapies.
Single-cell RNA sequencing (scRNA-seq) was performed using 10X Genomics Chromium technology, targeting 4,000-6,000 cells per sample. Each cell's transcriptome - the full set of active genes - was captured, providing a detailed map of cell type composition in each patient's bone marrow.
Drug testing used a library of 456 oncology compounds, including both approved drugs and experimental agents. Each compound was tested at five concentrations in patient cells, with cell survival measured by a luminescent assay (CellTiter-Glo). This produced detailed dose-response curves for each drug-patient combination.
Combination validation used 8x8 dose-response matrices - a grid where both drugs are varied across eight concentrations each, yielding 64 data points per combination. Synergy was scored using the ZIP (Zero Interaction Potency) model, which identifies cases where the combined effect exceeds what would be expected if the drugs were simply additive.
The entire platform - from patient sample collection through scRNA-seq processing, drug testing, computational prediction, and experimental validation - takes less than three weeks per sample. This timeline falls within a clinically actionable window, making the approach potentially practical for real-world application as sequencing technology continues to speed up.
This work demonstrates that personalized combination therapy for leukemia is computationally and experimentally feasible using patient-derived cells. The platform moves beyond the current paradigm of choosing drugs based on average responses in large patient populations, instead tailoring treatment to the unique molecular landscape of each individual's cancer.
The approach has important advantages over existing strategies. Unlike genomics-only approaches (which look for mutations to guide treatment), this platform captures functional drug responses - what the cancer cells actually do when exposed to drugs - which better reflects therapeutic vulnerability than mutation data alone.
The method requires only 30-40 million patient cells total - a feasible amount from a bone marrow aspirate - making it practical for use in clinical samples that are often limited in quantity. The platform is designed to be adaptable to other blood cancers beyond AML.
The study also opens the door to using longitudinal sampling - testing patients at multiple time points as their disease evolves - to dynamically update treatment recommendations. As leukemia progresses, different drug vulnerabilities emerge, and the platform showed it can track these changes and adjust its predictions accordingly.
The study successfully demonstrates that machine learning can accurately predict personalized drug combinations for AML patients that are simultaneously synergistic, effective against cancer cells, and selective against healthy cells. This triple optimization - synergy plus efficacy plus low toxicity - had not previously been achieved in a systematic, patient-tailored manner.
A key finding is that even combinations that show strong synergy across multiple patients still need to be evaluated in a patient-specific context. Drug combinations behaved differently in different patients depending on their unique genetic mutations, cell type composition, and disease stage - confirming that individual molecular profiling is essential.
The researchers envision this platform as a foundation for next-generation functional precision medicine in hematological cancers. Future directions include expanding the approach to larger patient cohorts, incorporating additional data types (such as protein expression or epigenetic marks), and running prospective clinical studies to directly test whether model-predicted combinations improve patient outcomes.