We also ran Cox regressions for the features included in the random forest models to contextualize our predictive analysis, eg, finding that contact between B cells and M1 macrophages correlates with better patient outcomes. /8

For example, we find that features related to CD163+ cell interactions are among the most important features extracted by the model. We dig deeper into these and related results in the paper. /7

We then used feature importance scores to identify the subset of features that best predict outcomes. This strategy can be used to generate mechanistic hypotheses and inspire future lines of research into novel treatments for HGSC. /6

All models performed better than chance at predicting high/low progression-free survival. The best model (model 8, mean AUC = 0.71) used a combination of TIME composition and spatial features. /5

We used multiplexed ion beam imaging (MIBI) to image the tumor immune microenvironment (TIME) of samples, identifying more than 160,000 cells across 23 cell types, and used spatial networks to generate features related to their patterns of interaction. /3

HGSC is the most common form of ovarian cancer. Unfortunately, survival and recurrence rates and treatment options have not improved much in 30 years. Here, we use machine learning to generate new hypotheses about what microenvironment features predict patient outcomes. /2

I’m excited to share our new paper on the spatial structure of the tumor immune microenvironment in high-grade serous carcinoma, now out in Cancer Immunology Research, with Benjamin Bitler, @aaronclauset.bsky.social, and others. doi.org/10.1158/2326... 🧵/1