Myelodysplastic syndromes (MDS) and secondary acute myeloid leukemia (sAML) are related blood cancers that arise when bone marrow cells accumulate mutations and lose the ability to produce healthy blood cells normally. MDS covers a spectrum from relatively indolent conditions with near-normal life expectancy to aggressive diseases that rapidly progress to leukemia. For decades, these cancers have been classified primarily by the appearance of bone marrow cells under the microscope and by broad clinical features.
Recent advances in genomic sequencing have revealed that the mutations driving MDS and sAML are far more varied and informationally rich than microscopic examination can capture. Different combinations of mutated genes create distinct molecular disease states that respond differently to therapy and carry very different prognoses -- even when the cells look similar under the microscope. Standard classification systems have not fully incorporated this molecular complexity.
This study applied unsupervised machine learning to the genomic mutation profiles of 3,588 MDS and sAML patients to discover whether the molecular landscape of these diseases naturally organizes into distinct subgroups -- and if so, whether these subgroups correspond to clinically meaningful differences in survival and treatment response. The goal was to let the data reveal the true structure of the disease rather than imposing predefined categories.
Unlike supervised machine learning, where an algorithm is trained on labeled examples (for example, learning to distinguish cancer from non-cancer by exposure to labeled images), unsupervised learning finds patterns in data without being told what the correct categories are. The approach used here was an autoencoder neural network -- a type of deep learning model that learns a compressed representation of the input data by reconstructing it through a narrow bottleneck layer.
The autoencoder was trained on the genomic mutation profiles of all 3,588 patients, learning which combinations of gene mutations tend to co-occur. The compressed representation in the bottleneck layer captures the most important structure in the mutation data. Patients with similar mutation combinations will be placed near each other in this compressed space. The researchers then applied clustering algorithms to this compressed representation to identify discrete groups of patients -- the molecular clusters.
This approach is superior to simple mutation-by-mutation analysis because mutations do not occur independently: specific gene mutations tend to cluster together in particular patient types, and the biological meaning of a mutation in one gene may depend on the presence or absence of mutations in other genes. The autoencoder captures these complex interaction patterns in a way that standard statistical methods cannot.
The analysis identified 14 distinct molecular clusters (MC1 through MC14) within the combined MDS and sAML patient population. MC2 is the largest cluster, containing approximately 26% of all patients, and is characterized by a specific set of commonly mutated genes. MC3 is the smallest, comprising only about 2% of cases and likely representing a rare but biologically distinct disease subtype.
The distribution of secondary AML (sAML) cases across clusters was highly non-random: sAML was concentrated in MC2 (28% of cluster cases) and MC13 (18% of cluster cases), suggesting that these molecular configurations are particularly prone to progression from MDS to frank leukemia. This kind of cluster-level insight -- identifying which molecular disease states carry the highest transformation risk -- could eventually guide more intensive monitoring or treatment for high-risk patients.
Each cluster had a distinctive genomic signature: specific combinations of mutations in genes like SF3B1 (a splicing factor commonly mutated in MDS), TET2, DNMT3A, ASXL1, TP53, and others. Clusters with TP53 mutations, for example, clustered separately from those driven by splicing factor mutations, consistent with the known clinical differences between these mutation classes in terms of treatment response and prognosis.
To assess whether the 14 molecular clusters were clinically meaningful rather than just statistically distinct, the researchers compared overall survival across clusters. Survival differences between clusters were striking and statistically robust, with some clusters showing median survivals several times longer than others. Importantly, these survival differences persisted even after adjusting for the IPSS-M score -- the current gold-standard prognostic system for MDS -- meaning the molecular clusters provide survival-predictive information beyond what the existing scoring system captures.
Similarly, treatment response differed across clusters. The response to hypomethylating agents (HMA) -- drugs like azacitidine and decitabine that are standard first-line therapies for higher-risk MDS -- varied significantly by molecular cluster. Clusters characterized by certain mutation combinations showed substantially higher or lower HMA response rates, suggesting that molecular cluster assignment could help predict which patients are most likely to benefit from this standard therapy versus needing alternative approaches.
These clinical associations suggest that the 14-cluster molecular framework is not merely a research taxonomy but could eventually serve as a clinically actionable classification: informing prognosis conversations with patients, guiding treatment selection, and identifying patients who should be prioritized for clinical trials of novel agents. Clusters with the worst prognosis and lowest HMA response rates represent the highest-priority populations for therapeutic innovation.
A critical test of any new disease classification system is whether it reproduces reliably when applied to new, independent patients who were not used to develop the system. The researchers validated their 14-cluster framework in an external cohort of 412 additional MDS and sAML patients whose genomic data was not used during model development. Cluster assignments were made for these new patients using the trained autoencoder model.
Cluster reproducibility was quantified using the Adjusted Rand Index (ARI), a statistical measure where 1.0 indicates perfect agreement and 0 indicates only chance agreement. The minimum ARI across cross-validation folds was 0.85, indicating excellent reproducibility: the same 14 clusters re-emerged consistently when the model was trained on different subsets of patients. A web-based tool (drmz.shinyapps.io/mds_latent) was created to allow other researchers and ultimately clinicians to assign new patients to the molecular clusters.
The researchers also demonstrated that cluster assignments were robust to the specific mutations included in the analysis and to different clustering algorithm parameters -- an important test of stability, since fragile clusters that change dramatically with minor methodological decisions would not be reliable enough for clinical use. The consistency of the 14 clusters across these sensitivity analyses supports their biological reality rather than being artifacts of particular modeling choices.
Current MDS classification systems -- including the WHO classification and the IPSS-M prognostic score -- were developed primarily by expert consensus based on morphological features and selected genetic markers. While these systems are clinically useful, they were not derived by systematically asking which groupings of patients emerge naturally from comprehensive genomic data. This study takes the opposite approach: let the molecular data define the categories, then ask whether those categories are clinically meaningful.
The finding that 14 molecular clusters exist -- and that each has distinct survival outcomes, treatment responses, and genomic signatures -- suggests that MDS and sAML are more molecularly heterogeneous than current classification captures. Some of the 14 clusters likely correspond to recognized MDS subtypes, while others may represent previously unrecognized disease entities that have been lumped together under existing categories despite having meaningfully different biology.
This kind of data-driven reclassification has a precedent in other hematologic cancers: diffuse large B-cell lymphoma (DLBCL) was long treated as a single disease but was ultimately found to contain multiple molecular subtypes with different prognoses and treatment responses, leading to refined classification systems and eventually subtype-specific clinical trials. A similar evolution in MDS classification could ultimately yield more targeted treatments and better patient outcomes.
The 14-cluster molecular framework represents a foundation, not a finished clinical tool. Future work will need to expand the training dataset to ensure all 14 clusters are represented by sufficient patient numbers for robust clinical characterization. Some clusters, particularly the smallest ones, may represent disease subtypes so rare that individual centers cannot accrue enough patients to study them -- making international data sharing and consortium studies essential.
Prospective validation studies are needed to confirm that cluster assignment at diagnosis can reliably guide treatment decisions. The most immediate clinical application may be in the context of clinical trial design: stratifying patients by molecular cluster could improve the interpretability of trial results and help identify which patient subsets benefit most from experimental therapies. This kind of biomarker-driven trial design is already standard in solid tumor oncology and should become standard in MDS as well.
The web-based assignment tool (drmz.shinyapps.io/mds_latent) is a meaningful step toward democratizing access to this classification system -- allowing any center that performs MDS genomic sequencing to assign their patients to the 14 molecular clusters without needing specialized bioinformatics infrastructure. As sequencing becomes more routine in MDS diagnosis globally, this tool could enable systematic data collection that further refines the molecular classification and links clusters to treatment outcomes across diverse patient populations worldwide.