Acute myeloid leukemia (AML) is an aggressive blood cancer in which immature myeloid cells in the bone marrow multiply uncontrollably. It can affect both children and adults, and treatment typically involves intensive chemotherapy. A critical goal after the first round of treatment is achieving complete remission (CR) - meaning leukemia cells can no longer be detected in the bone marrow. Failing to achieve CR significantly worsens prognosis.
The problem is that not all AML patients achieve complete remission after standard induction chemotherapy, and doctors currently cannot reliably predict before treatment begins which patients will respond and which will not. If a patient's pre-treatment gene expression profile could predict their likelihood of achieving remission, this information could guide more aggressive initial treatment choices, enrollment in clinical trials, or consideration of stem cell transplant earlier in therapy.
RNA sequencing (RNA-seq) measures the activity level of thousands of genes simultaneously by counting the messenger RNA molecules a cell produces. This creates a detailed molecular portrait of a cancer at the moment of diagnosis - before any treatment begins - capturing patterns of gene activity that may predict clinical outcomes. The challenge is identifying which patterns in this vast dataset are truly informative.
This study from National Cancer Institute researchers applied three different machine learning algorithms to pre-treatment gene expression data from 473 children and young adults with AML, testing whether patterns in gene activity could predict which patients would achieve complete remission after two courses of induction chemotherapy.
Bone marrow specimens were collected from 473 children and young adults (ranging in age from 8 days to 28 years) diagnosed with newly occurring AML, all enrolled in the Children's Oncology Group clinical trial AAML1031 - one of the largest pediatric AML trials ever conducted. An additional 20 bone marrow samples from healthy individuals served as normal controls for comparison. Clinical information about whether each patient achieved complete remission was available for 414 patients.
RNA sequencing was performed using the Illumina HiSeq2000 platform, generating comprehensive gene activity measurements across tens of thousands of genes per sample. The data was carefully normalized to ensure comparability across samples. Of the 414 patients with remission status information, 331 were used to train the machine learning models, and 83 were held back as an independent test set to evaluate how well the models performed on data they had never seen before.
Three distinct supervised machine learning algorithms were compared: the K-Nearest Neighbors (K-NN) algorithm (which classifies new samples by comparing them to the most similar training samples), Support Vector Machine (SVM) (which finds an optimal mathematical boundary separating the two classes), and Random Forest (RF) (an ensemble of decision trees that votes on classifications). Each algorithm has different strengths and mathematical foundations, so comparing them helps identify the most suitable approach for this problem.
A critical methodological challenge was feature selection - identifying which genes (out of tens of thousands) are most informative for predicting remission. Several strategies were used including statistical t-tests to identify the most differentially expressed genes, Hill Climbing (sequential addition of the most helpful genes), and Randomized Lasso (a statistical selection approach based on model weights). Performance was measured using the area under the ROC curve (AUC), a value between 0.5 (no better than chance) and 1.0 (perfect prediction).
Among the three algorithms tested, the K-Nearest Neighbors (K-NN) algorithm with K=27 (meaning each sample is compared to its 27 most similar training samples) outperformed both the Support Vector Machine and Random Forest approaches. On the training data, K-NN with Hill Climbing feature selection achieved an AUC of 0.84 - meaning the model correctly distinguished patients who would achieve remission from those who would not about 84% of the time.
When this trained model was applied to the completely independent test set of 83 patients it had never encountered during training, it achieved an AUC of 0.81 using the top 50 most predictive genes. This modest drop from training to testing performance is expected and actually suggests the model generalized reasonably well rather than simply memorizing the training data - a common failure mode in machine learning called overfitting.
The Random Forest algorithm achieved an AUC of 0.74 on the training data, meaningfully below K-NN's performance. The SVM also underperformed K-NN on this dataset, even though SVM is often competitive or superior in high-dimensional data settings like genomics. The superior performance of K-NN may reflect the specific structure of the AML gene expression data - a non-linear decision boundary that K-NN handles better than the linear separations SVM tends to find.
A principal component analysis (PCA) of all samples showed that AML patients did not naturally cluster into visually obvious remission vs. non-remission groups - meaning the predictive signal in the data is real but subtle, not simply a matter of samples being obviously different from each other. This is precisely the type of complex, non-obvious pattern where machine learning adds genuine value over simple visual inspection or single-gene analysis.
The 50 most predictive genes were subjected to pathway enrichment analysis to understand what biological processes they represent. The most statistically significant pathway identified was the GDP-fucose biosynthesis pathway (adjusted P value = 0.0092). Two genes in this pathway - TSTA3 (GDP-L-fucose synthase) and FPGT (fucose-1-phosphate guanylyltransferase) - were specifically highlighted as predictive of remission outcome.
Fucosylation is a type of protein modification called glycosylation, where sugar molecules (specifically fucose) are attached to proteins on the cell surface. These sugar modifications profoundly affect how cells interact with each other and with the immune system. Fucose-containing molecular structures play critical roles in how white blood cells attach to blood vessel walls (selectin-mediated adhesion) and in multiple immune signaling processes - functions directly relevant to how leukemia cells evade immune destruction.
Altered fucosylation has been documented in many cancers and is associated with poor prognosis. Experimental studies have shown that impairing fucosylated molecule expression in mice causes abnormal proliferation of myeloid progenitor cells (the same type of cells that become malignant in AML) and overproduction of granulocytes - mimicking some features of AML itself. Changes in specific blood group antigens (H and Lewis antigens), which are also affected by fucosylation, have been linked to poor prognosis in cancer patients.
Among the other top predictive genes, BATF (a transcription factor involved in immune cell differentiation) and RAC2 (involved in cell signaling and immune function) were associated with decreased IgM (a type of antibody) levels - and low IgM has previously been identified as an adverse prognostic factor in childhood leukemia. Meanwhile, TSTA3 and RAC2 together were linked to increased neutrophil counts, another clinically relevant finding in AML prognosis.
N-glycosylation - the process of attaching sugar chains to proteins at specific sites - is one of the most common and important protein modifications in biology. It profoundly affects protein folding, stability, trafficking within cells, and interactions with other molecules. The GDP-fucose pathway identified in this study feeds directly into the N-glycosylation process, suggesting that fundamental differences in how AML cells modify their surface proteins may determine treatment response.
Cancer cells are known to display abnormal patterns of glycosylation compared to normal cells. In AML specifically, leukemia cells with certain glycosylation patterns may be better able to evade the immune system, survive chemotherapy, or self-renew - all factors that would prevent achievement of complete remission. The finding that genes controlling this process are differentially expressed in patients who do vs. do not achieve remission provides a plausible biological mechanism.
The pathway identified - N-glycan precursor synthesis and transfer - represents a potential therapeutic target. If AML cells that fail to respond to chemotherapy depend on abnormal glycosylation for their survival, then combining standard chemotherapy with drugs that target glycosylation processes could potentially improve remission rates. Several glycosylation-targeting compounds are already in early-stage clinical development for various cancers.
Importantly, this study demonstrates that gene expression differences between patients who will vs. will not achieve remission exist at diagnosis - before any treatment has been given. This means the predictive information is intrinsic to the biology of each patient's leukemia, not a result of treatment effects. This strengthens the case that genomic profiling at diagnosis could genuinely guide treatment decisions, not just reflect outcomes after the fact.
An AUC of 0.81 on independent test data represents a clinically meaningful predictive performance. While not perfect, this level of prediction - substantially better than chance - could meaningfully inform treatment decisions for AML patients. Doctors currently use cytogenetic risk groups and mutation status to categorize AML patients, but these categories are imperfect predictors of individual patient outcomes. Adding a gene expression-based remission prediction score could refine risk stratification.
In clinical practice, such a tool could potentially identify patients at high risk of not achieving remission after standard induction therapy. These patients might benefit from more intensive initial treatment, early consideration of allogeneic stem cell transplant (transplanting bone marrow from a matched donor), or enrollment in clinical trials testing novel agents. Conversely, patients predicted to respond well might be spared the additional toxicity of unnecessarily intensive treatment.
The study used data from the Children's Oncology Group trial AAML1031 - a rigorously conducted, prospective multi-institutional trial with standardized treatment protocols. This gives the gene expression data high quality and the clinical outcome data high reliability. However, validation in additional, independent AML patient cohorts (particularly non-pediatric patients) will be essential before this approach could be clinically implemented.
Compared to other studies applying machine learning to cancer gene expression, an AUC of 0.81 is competitive. Similar studies in large B-cell lymphoma and hepatocellular carcinoma achieved comparable results. The growing library of such validated gene expression predictors across cancer types suggests that RNA sequencing at diagnosis, combined with machine learning analysis, may eventually become a standard component of cancer workup alongside traditional pathology and cytogenetics.