Daniel West, Susan Stepney, Y. Hancock. Unsupervised self-organising map classification of Raman spectra from prostate cell lines uncovers substratified prostate cancer disease states. Scientific Reports, 15:773, 2025. doi:10.1038/s41598-024-83708-6
This started out as a feasibility study, to see if Kohonen Self-Organising Maps (SOMs) could be used to cluster minimally preprocessed Raman spectroscopy data taken from individual cells. SOM an unsupervised learning approach, and can cluster high dimensional data (here, over 1000D) down into a 2D visual representation. We had Raman spectra of prostate cells, some cancerous, some not. Could a SOM distinguish these two classes?
We blinded the data, so that the system did not know which spectrum was in which class, to ensure this was truly an unsupervised exercise. After some fiddling about to understand what values several parameters should be, we fed the data in, and looked at the resulting map. We could see three clusters.
Had it worked? We unblinded the data, and yes, one of the clusters was the non-cancerous cells, and the other two clusters were cancerous cells. Why two clusters? Well, it turns out the mapping process had managed to discover two distinct classes of cancerous cells. Further research is underway to investigate these differences.
So yes, it works, and better than we had hoped!
No comments:
Post a Comment