Hist2Sig predicts exposure to 30 COSMIC signatures from a simple H&E blade. It does better than a model that knows only the type of tumor in 19 out of 29.
Mutational signatures have direct clinical value: tumors deficient in mismatch repair respond better to immunotherapy, tumors with homologous recombination deficiency are sensitive to PARP inhibitors, and POLE mutated tumors often carry a high mutational burden that weighs on the response to treatment. But establishing them requires complete genome or exome sequencing: expensive, complex, slow.
Hist2Sig takes the problem from the other end. This deep learning framework predicts exposure to 30 SBS signatures from the COSMIC catalog from whole slides stained with hematoxylin-eosin, the most mundane examination of pathology. The training is based on 7,063 patients from the TCGA database covering 29 types of cancer, for which complete sequencing and histology are matched. Important methodological point: the comparison is made against a reference using only the type of tumor, in order to distinguish what the morphology actually brings from what comes from the simple tissue of origin.
Unequal gains depending on the tumor
In internal cross-validation, the model found the relative composition of signatures for most tumor types and exceeded the reference on the average recovery of the first three signatures in 19 out of 29 cases, with the clearest differences on colorectal cancer, prostate, glioblastoma and endometrial cancer. It also recurrently retrieved signatures whose etiology remains poorly understood, including SBS8, SBS12, SBS39 and SBS40a. In an external CPTAC cohort of 193 patients and five tumor types, concordance remained moderate, with an advantage over the reference for glioblastoma and pancreatic adenocarcinoma. The authors do not derive from this a substitute for sequencing, but a way to sort before sequencing, or to support where the complete genome is not accessible.
Read the full article on the original source
Go to the source →