Acute myeloid leukemia (AML) is a highly heterogeneous blood cancer in which 70% of adults die within five years of diagnosis. Despite eight new drugs being approved for AML in recent years, most patients still receive the same standard chemotherapy regimen because oncologists lack tools to reliably match individual patients to the treatments most likely to help them specifically.
Artificial intelligence offers a potential solution by analyzing drug screening data — experiments where living tumor cells from a patient are tested against dozens or hundreds of drugs simultaneously to measure sensitivity — combined with genetic sequencing to identify which mutations are present in each tumor. In principle, this combination should allow AI to recommend personalized treatments based on each patient's unique molecular profile.
The core challenge is that most high-performing AI models are "black boxes" — they produce a recommendation without explaining the logic behind it. A doctor cannot verify whether the recommendation is biologically sound, a regulatory body cannot easily approve it, and the result cannot be incorporated into simple clinical guidelines. This opacity has severely limited the translation of AI drug predictions into actual patient care.
Explainable Artificial Intelligence (XAI) addresses this by requiring that the model's reasoning be transparent and human-understandable. The authors of this 2022 paper designed a new XAI method specifically for AML precision medicine — one that produces clear, interpretable treatment rules based on a small number of genetic markers that any clinician can apply.
The authors developed the Multi-dimensional Module Optimization (MOM) algorithm, designed to systematically link drug sensitivity data with genetic biomarkers while producing results that are both interpretable and guaranteed to be mathematically optimal. MOM takes two inputs: the drug sensitivity scores from ex-vivo tumor experiments and the patient's genetic variant information from whole-exome sequencing.
The first step of MOM addresses a subtle but critical problem in drug selection: standard drug sensitivity measurements (IC50 — the dose needed to kill half the cells) favor uniformly toxic drugs over ones that are selectively effective. A drug that kills both tumor and healthy cells at low doses looks better on paper than one that works strongly only on a specific genetic subgroup. MOM introduces a corrected score called IC50* that rewards drugs with differential effects across patients — prioritizing specificity over brute toxicity.
The second step uses a statistical algorithm called Independent Hypothesis Weighting (IHW) to link individual genetic variants to specific drug sensitivities while carefully controlling for the risk of false positives that arise when testing thousands of possible gene-drug combinations. IHW also assigns a weight to each gene variant reflecting how likely it is to be a genuinely informative biomarker.
The third and final step uses Mixed Integer Linear Programming (MILP) — a mathematical optimization method — to combine the individual gene-drug associations from step two into a decision tree that maximizes overall patient benefit across the entire population. Crucially, MILP is deterministic: given the same inputs, it always produces the same optimal solution, which is a key requirement for regulatory approval of AI-driven medical decisions.
MOM was applied to the BeatAML cohort, which includes drug sensitivity data for 122 drugs and whole-exome sequencing from 319 AML patients — one of the largest ex-vivo AML datasets available. The algorithm identified three key genetic biomarkers — FLT3 mutation, inv(16)/CBFbeta-MYH11 fusion, and NRAS mutation — that together stratify all 319 patients into four clinically distinct subgroups, each with a recommended treatment.
The largest subgroup (103 patients, 32%) carries a FLT3 mutation and is assigned Quizartinib, a second-generation FLT3 inhibitor that has shown increased overall survival in clinical trials. The second subgroup (15 patients) has wild-type FLT3 but carries the inv(16) chromosomal abnormality and is assigned Trametinib, a MAPK pathway inhibitor that prevents cell replication. The third subgroup (42 patients) carries NRAS mutation without the preceding markers and is assigned Selumetinib, another MAPK inhibitor. The remaining 159 patients without any of these markers are assigned Crizotinib, an ALK and MAPK inhibitor approved for lung cancer that showed promising results in high-risk AML in earlier studies.
The IHW analysis also identified which genetic variants across the entire dataset were associated with general drug sensitivity or resistance. FLT3 and NPM1 mutations were associated with increased sensitivity to many drugs, while KRAS, NRAS, and TP53 mutations were associated with general drug resistance — findings consistent with known AML biology and helping validate that MOM was capturing real biological signals rather than statistical noise.
The treatment strategy recommended by MOM was validated not just within the BeatAML dataset using 10-fold cross-validation (where the algorithm is trained on 90% of the data and tested on the remaining 10%, repeated ten times), but also in three entirely independent large-scale experiments that used different technologies and different cell lines.
The three validation datasets were: CERES (14 AML cell lines, CRISPR-Cas9 gene knockout screening), DEMETER2 (18 AML cell lines, RNA interference gene knockdown), and GDSC (23 AML cell lines, drug sensitivity measurements). Each of these experiments was performed independently of BeatAML and uses different experimental platforms, making successful replication across all three a strong indicator of robustness.
The differential sensitivity predicted by MOM's biomarker subgroups was statistically significant in all three independent experiments — with p-values of 5.5x10-5, 6.8x10-6, and 5.5x10-4 for CERES, DEMETER2, and GDSC respectively. These very low p-values indicate that the biomarker-based treatment assignments are capturing real biological differences in drug response, not statistical coincidences.
Cross-validation within BeatAML showed that 90% of the ten data splits produced decision trees sharing at least 4 out of 5 biomarkers with the final result. The FLT3 and NRAS subgroups appeared in all 10 folds, and the inv(16) subgroup appeared in 3 out of 10, confirming that the treatment strategy is stable and reproducible rather than dependent on specific data points.
A central argument of this paper is that explainability in AI is not merely an academic nicety — it has direct regulatory and clinical implications. Regulatory agencies like the FDA and the European Medicines Agency (EMA) are more likely to approve AI-driven treatment recommendations when the companion biomarkers are clearly identified and biologically justified. A black-box model that simply says "give this patient drug X" without showing why will face much higher barriers to approval than one based on identifiable genetic markers.
The MOM decision tree is structured equivalently to how clinical treatment guidelines already work: a physician checks whether FLT3 is mutated, then whether inv(16) is present, then whether NRAS is mutated, and assigns the corresponding drug. This mirrors the logic of existing guideline flowcharts and can be implemented without any additional software or ML infrastructure at the point of care.
The fact that MOM's output is both deterministic (same inputs always give the same output) and optimal (mathematically proven to maximize patient benefit within the data) are properties directly relevant to regulatory review, clinical audit, and medicolegal accountability — qualities that most deep learning approaches cannot currently guarantee.
This study demonstrates that explainable AI can match the sophistication of black-box approaches while producing results that are transparent, interpretable, and directly actionable in clinical settings. By using only three biomarkers to stratify 319 AML patients into four treatment subgroups — each validated across independent datasets — MOM shows that simplicity and clinical utility are achievable alongside genuine scientific rigor.
The specific finding that Crizotinib — an FDA-approved lung cancer drug — may benefit the largest subgroup of AML patients without FLT3, inv(16), or NRAS mutations is particularly noteworthy as an example of drug repositioning: finding new clinical applications for existing approved drugs. If validated in prospective trials, this could rapidly benefit a large proportion of AML patients who currently have no targeted therapy options.
The authors conclude that XAI represents the most viable path for moving AI-based precision oncology from research databases into actual clinical care. By designing AI that explains its reasoning, scientists and clinicians can work together to validate, refine, and ultimately trust AI recommendations in a way that opaque models simply cannot support.