Cancer Research · eu
Tracing the Primary Tumor to the Brain and Lungs: AI Spatial Analysis Identifies a Suspected Metastatic Cell Population
Samples collected from a melanoma patient over eight years allowed researchers to compare cell morphology with more than 6,000 proteins region by region and trace tumor populations that may have contributed to distant metastasis; however, this remains a single case and is still far from becoming a clinical prediction tool.
Not all cancer cells within the same tumor pose the same level of risk. If the small number of cells more likely to metastasize or resist treatment could be identified before the primary tumor spreads, treatment might be able to target the true threat earlier. A melanoma case study sought to use artificial intelligence to interpret tissue sections and then apply spatial proteomics to track where these cells subsequently went.
The subject was a woman diagnosed with scalp melanoma at age 24. Over eight years, she experienced recurrence in the cervical lymph nodes, lung metastases, and multiple brain metastases, and underwent surgery, treatment with BRAF and MEK inhibitors, immune checkpoint inhibitors, and radiotherapy. The research team retrospectively analyzed preserved paraffin-embedded tissue from the primary tumor, three lung metastases, and one brain metastasis to reconstruct the course of disease progression.
The team first used a deep-learning model trained on tissue sections from this patient to analyze conventional H&E-stained images. The model separated two spatially distinct cell populations, PT1 and PT2, within the primary tumor; in validation using held-out image regions from within the primary tumor, overall classification accuracy was approximately 94%. The model identified more than 170,000 cells in total, and PT1 cells had larger nuclei on average, indicating that the two populations differed not only in location but also in distinguishable morphology.
When the same model was applied to expert-annotated sections of the lung and brain metastases, tumor-cell morphology at both sites was predominantly more similar to PT1, while PT2-like cells were almost absent. The researchers then used the AI-mapped regions to isolate thousands of cells through laser microdissection for mass spectrometry analysis. A profile comprising more than 6,000 proteins likewise showed that the molecular features of PT1 were more similar to those of the lung and brain metastases than to PT2. The morphological and protein evidence therefore both indicated that PT1 may have been the cell population with greater metastatic capacity in this patient, although a direct lineage relationship between them cannot yet be established.
Protein signals also provided clues about tumor behavior. PT1 and the metastatic tissues showed more active glycolysis, oxidative phosphorylation, and proliferation-related signaling; elevated levels of SRC and YES1 in the metastases may have helped cancer cells bypass BRAF/MEK inhibition. These findings can be used to formulate hypotheses involving metabolic inhibition or multipathway combination therapy, but they are insufficient to prove that any drug would be effective for this patient or other patients with melanoma.
The more practical immediate value of this approach is to use pathology images as a navigation map to precisely select regions warranting further molecular analysis, rather than allowing AI to make treatment decisions directly. The study included only one patient, and the model's training and validation data also came from tissue from the same individual; it would need to be retrained for another case. The study also did not test in a large independent population whether PT1 characteristics could predict recurrence or metastasis. Before entering clinical practice, the approach must still demonstrate robustness across patients, laboratories, and scanning equipment, and prospective studies must confirm that its interpretations can indeed improve treatment decisions.