Acute myeloid leukemia (AML) is an aggressive blood cancer with poor survival rates, and combination drug therapy -- using two or more drugs together -- often works better than any single agent. When two drugs produce an effect greater than the sum of their individual effects, that combination is called synergistic. Identifying which drug pairs are synergistic for which patients could dramatically improve treatment selection.
The challenge is scale: with hundreds of approved drugs, the number of possible pairs and patient-specific combinations is astronomically large. Testing each combination experimentally across every patient subtype is impractical. Machine learning offers a way to predict synergy computationally, but these models are difficult to trust clinically unless researchers can explain which biological features drive the predictions.
Previous machine learning approaches for drug synergy prediction either lacked interpretability -- making them black boxes that clinicians could not trust -- or suffered from unstable feature attribution when the underlying data features were correlated, as gene expression measurements typically are. This study introduces the EXPRESS framework to overcome both limitations simultaneously.
EXPRESS (EnsemPle-based fRamEwork for SynergiStic drug response) is built on two core components. First, it trains large ensembles of XGBoost models -- gradient-boosted decision trees that consistently outperform other methods on tabular data -- each on a randomly sampled subset of patient samples. Second, it uses Shapley values, a game-theoretic method for attributing predictions to individual input features, to explain what the models have learned.
The key innovation is how ensembling interacts with explainability. Individual models trained on small data subsets may individually find different sets of features as important, especially when features are highly correlated. But averaging Shapley values across many models in an ensemble dramatically stabilizes the attribution: random noise in individual models cancels out, leaving only the consistent signal. The researchers quantify this stability improvement using cosine similarity, which increased from 0.77 for individual models to 0.98 for ensembles.
The framework was applied to the Beat AML dataset, a landmark resource containing gene expression profiles, drug sensitivity measurements, and clinical data from 285 AML patients treated with 133 different drug combinations involving 46 total drugs. This dataset provides exactly the kind of real patient variability needed to test whether the approach can discover biologically meaningful predictors of drug synergy.
Before settling on XGBoost as the base learner for EXPRESS, the researchers benchmarked multiple machine learning algorithms on the AML drug synergy prediction task. Competitors included elastic net (a regularized linear model), random forests, and neural networks. XGBoost outperformed all alternatives in 53 out of 60 direct comparisons across different drug combinations and evaluation metrics.
This result was not entirely surprising -- XGBoost has repeatedly been shown to be the leading method on structured, tabular datasets in domains ranging from genomics to finance. What made the benchmark informative here was the specific demonstration that neural networks, despite their flexibility, did not surpass XGBoost even in this high-dimensional genomic setting. This has practical implications: XGBoost models are faster to train, require less hyperparameter tuning, and produce Shapley values more efficiently than deep learning alternatives.
The benchmark also revealed an important cautionary finding: when input features are highly correlated -- as gene expression measurements typically are due to co-regulated gene networks -- all methods including XGBoost produce unreliable feature attributions. This means that results from any single model should be interpreted cautiously when the input data has this property, reinforcing the need for the ensembling strategy that EXPRESS provides.
When EXPRESS was applied to predict synergy across all drug combinations in the Beat AML dataset, the Shapley value analysis revealed a consistent set of genes whose expression best predicted synergistic drug responses. These genes -- MEIS1, CITED2, DLL3, IL-4, and STAT6 -- collectively define what the authors call a haematopoietic differentiation signature.
MEIS1 and CITED2 are transcription factors critically involved in maintaining blood stem cells in an undifferentiated state. High expression of these genes in AML cells indicates that the leukemia is maintaining stem-cell-like properties -- a feature associated with drug resistance and poor prognosis. DLL3 is a Notch signaling pathway component involved in cell fate decisions during blood cell development. IL-4 and STAT6 are immune signaling molecules involved in the differentiation of immune cells.
The biological coherence of this signature -- all genes relate to the theme of blood cell differentiation and maturation -- provides confidence that EXPRESS is not simply finding statistical artifacts but is uncovering a genuine biological mechanism. AML cells that are arrested at an early differentiation state appear to create a molecular environment in which certain drug combinations work synergistically, perhaps because both drugs are simultaneously attacking the signals that keep these cells undifferentiated.
One of the most technically important contributions of this paper is the demonstration of how ensembling improves not just prediction accuracy but also the reliability of the biological insights extracted from those predictions. When gene expression features are correlated -- as is almost universally true in transcriptomic data -- the Shapley values attributed to any individual gene by a single model are highly variable across different training runs. This instability makes it impossible to trust which genes the model considers truly important.
The researchers demonstrate mathematically and empirically that averaging Shapley values across ensemble members mitigates this instability. As more models are added to the ensemble, the cosine similarity between feature attribution vectors (a measure of how consistently the same features are identified as important) rises toward 1.0. The improvement from 0.77 to 0.98 represents a qualitative difference: attributions become reliable enough to use as biological hypotheses.
This finding has broad implications beyond AML drug synergy. Any domain where machine learning is applied to correlated, high-dimensional data -- which encompasses most genomic and transcriptomic applications -- could benefit from the ensembling-stabilized attribution approach. The EXPRESS framework is therefore presented not just as an AML-specific tool but as a general methodology for more trustworthy explainable machine learning in biology.
To confirm that the haematopoietic differentiation signature was not a statistical artifact of the specific data split or modeling choices, the researchers performed extensive validation experiments. These included cross-validation across different patient subsets, testing sensitivity to hyperparameter choices, and comparing the signature discovered by EXPRESS to gene sets identified by independent biological analyses of AML differentiation states.
The signature was consistent across these analyses, with the same core genes (MEIS1, CITED2, DLL3, IL-4, STAT6) appearing as top predictors regardless of the specific model configuration tested. This robustness is exactly what would be expected if the signature reflects a real biological mechanism rather than chance associations in the training data.
The authors also tested whether the discovered signature had predictive value beyond what could be achieved using clinical variables alone -- demonstrating that the gene expression information provides independent information about drug synergy that is not captured by patient age, disease stage, or other standard clinical features. This supports the potential clinical utility of transcriptomic profiling as a tool for personalizing combination drug selection in AML.
The EXPRESS framework points toward a practical clinical application: using a patient's AML gene expression profile at diagnosis to predict which drug combinations will be most synergistic for their specific leukemia. Rather than relying on population-average treatment protocols, oncologists could theoretically select drug pairs most likely to work together for each individual patient based on their tumor's molecular characteristics.
The haematopoietic differentiation signature identified by the study also suggests a biological hypothesis that can be tested in the laboratory: if blocking the pathways that maintain AML cells in an undifferentiated state (the pathways driven by MEIS1, CITED2, and DLL3) is what enables drug synergy, then treatments that force differentiation might be especially effective in combination. This provides a direction for follow-up mechanistic and preclinical studies.
Future work will need to validate EXPRESS predictions prospectively -- selecting drug combinations based on model predictions and measuring actual patient outcomes -- to determine whether computational synergy prediction truly translates to clinical benefit. The Beat AML dataset represents a rich starting point, but broader validation across diverse AML subtypes and treatment settings will be needed before this approach influences clinical practice at scale.