← Back to Home

Reading DNA Methylation Clues, AI Attempts to Classify 81 Types of Blood and Lymphoid Tumors

ALMA 3 showed strong tumor detection and classification performance in independent testing, but training scale does not equal clinical evidence; whether it can consistently classify each subtype still requires more comprehensive validation.

By SURL BioNews

For cancers of the blood and lymphatic systems, detecting a tumor is only one step in diagnosis; determining its subtype provides a clearer understanding of prognosis and treatment direction. When conventional methods struggle to distinguish different diseases, researchers are exploring DNA methylation patterns for another set of classification clues, allowing molecular information to complement pathological assessment.

At the 2026 Association for Molecular Pathology (AMP) Annual Meeting in Seattle, a research team presented the AI foundation model ALMA 3. According to an AMP announcement and reports on the same event by ecancer and LabMedica, the model was trained on more than 110,000 patient samples, and the study covered 81 subtypes of blood and lymphoid tumors, including categories that are difficult to distinguish using conventional diagnostic methods.

The study was led by Pedro Orsini of ALMA Genomics, with participating institutions including the University of Pennsylvania and the Massachusetts Institute of Technology. AMP listed it as abstract H071, focusing on the development of a hierarchical classification model for hematolymphoid tumors under the WHO and International Consensus Classification (ICC) frameworks. Methylation refers to chemical marks on DNA; ALMA 3 uses the distribution patterns of these marks to help identify tumors and their categories.

The independent testing reported in the study included two distinct metrics: sensitivity for detecting hematolymphoid tumors was 99.5%, and the positive predictive value for assigning cancer subtypes was 90.9%. The former describes the proportion of tested samples with tumors that the model detected; the latter reflects the proportion of the model’s subtype assignments that were correct. They answer different questions and cannot be combined into a claim of “99.5% diagnostic accuracy.”

Another distinction is that the figure of more than 110,000 refers to the number of training samples, not a count of individuals in independent testing disclosed in the announcement. Existing reports did not provide the total number of test samples, the distribution of cases across subtypes, confidence intervals, or performance for each category. It is therefore not yet possible to determine whether rare subtypes are equally reliable, nor can the overall figures establish consistent performance across all 81 subtypes. Specificity for tumor detection was also not disclosed in these reports; high sensitivity alone does not indicate how often samples were incorrectly classified as tumors.

These data make ALMA 3 a pathology support tool worthy of further validation, but they do not yet establish that it can improve decision-making in routine clinical care. Clinical use will require clarification of its performance across different specimens and laboratory conditions, as well as how to integrate information when the model’s classifications disagree with existing tests. Public reports also did not provide information on testing costs, turnaround time, or regulatory approval.

AMP stated that further clinical validation is underway. The next key question is whether the model can consistently provide useful classification clues in challenging cases encountered in practice and help physicians make more reliable judgments. This will determine whether the strong test results can translate into practical help in patients’ diagnostic pathways.

References

  1. ecancer
  2. Association for Molecular Pathology
  3. LabMedica International