We’ve got new results relating to the study of ovarian cancer! Previously, we’ve shown that we can distinguish tumor tissue from adjacent healthy control tissue using TCRs sequenced from tissue biopsies (the case and controls are from the same patient). This result was obtained using our approach for diagnosing disease from immune repertoires, which we have framed as a problem of multiple instance learning (MIL) [1,2]. We’ve continued working on computational methods for diagnosing disease from immune repertoires, and now we’ve got new results to share!
We’ve shown that we can distinguish ovarian tumor from ovarian tissue of non-cancer patients using TCRs sequenced from ovarian biopsies. In other words, the case and controls did not come from the same patients. It is remarkable our approach worked because we did not control for HLA type. Next, we obtained additional ovarian biopsies where we remained blindfolded to the cancer status. Using our model and the TCRs sequenced from these biopsies, we made predictions and then unblindfolded ourselves. We correctly distinguished tumor from non-cancer with 80% accuracy. It would appear our results truly generalizes to new patients. Here is a link to our PLoS publication (https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0229569).
In the future, we hope to turn this into a diagnostic test for ovarian cancer. These results help lay down the foundation.
- PMCID: PMC5588725
- PMCID: PMC6445742