← Back to Home

Identifying Blood Cancers Through DNA Methylation: ALMA 3 Study Covers 81 Subtypes

The challenge of diagnosing blood cancers lies not only in detecting tumors, but also in distinguishing their types. ALMA 3 has reported independent testing results that show potential for aiding classification, but strong overall performance does not yet establish whether it is equally reliable for every rare subtype.

By SURL BioNews

For cancers of the blood and lymphatic systems, identifying a tumor’s subtype affects prognosis assessment and treatment choices. Yet some subtypes remain difficult to distinguish using traditional methods. At the 2026 Association for Molecular Pathology (AMP) Annual Meeting in Seattle, researchers presented an approach to supporting diagnosis: using artificial intelligence to find classification clues in DNA methylation profiles.

According to an AMP announcement and coverage of the same meeting by Clinical Lab Products and MedLabPortal, ALMA 3, jointly developed by ALMA Genomics, the University of Pennsylvania, the Massachusetts Institute of Technology, and other institutions, was trained on more than 110,000 patient samples. Led by Pedro Orsini of ALMA Genomics, with Francisco Marchi serving as corresponding author, the study covered 81 subtypes of cancers of the blood and lymphatic systems, including some categories that are difficult to classify using conventional methods.

The model reads DNA methylation patterns—the distribution of chemical marks on DNA. The study aims to use these molecular features to identify tumors and further determine their subtypes. The central problem it addresses is adding evidence for classification alongside existing pathological diagnosis, rather than simply providing a single number to indicate whether cancer is present.

The research team reported that, in independent testing, ALMA 3 had a sensitivity of 99.5% for detecting hematolymphoid tumors and a positive predictive value of 90.9% for assigning cancer subtypes. These two metrics answer different questions: the former measures how many samples with tumors were identified, while the latter measures how many of the model’s positive subtype assignments were correct. A sensitivity of 99.5% therefore cannot be directly interpreted as equivalent classification accuracy across all 81 subtypes.

Understanding these results also requires knowing how the testing was designed. The announcement and reports cited above currently do not specify the number of samples in the independent test, the distribution of cases across subtypes, confidence intervals, or performance for each individual subtype. The figure of more than 110,000 refers to the training sample size and cannot be treated as the number of cases in independent validation. Without these details, it remains difficult to determine whether the overall results also apply to categories with few cases that are most easily confused.

AMP positioned the findings as a potential tool to complement traditional pathological diagnosis, and both external reports also noted that further clinical validation is underway. What needs to be clarified next is whether the model can maintain its performance across different patient populations and real-world testing workflows, and whether the classification information it provides can improve diagnostic decisions. At this stage, the conference results provide a direction for validation but are not sufficient to demonstrate improved treatment outcomes for patients.

References

  1. Association for Molecular Pathology
  2. Clinical Lab Products
  3. MedLabPortal