Finally, to look at how these features could be combined together, as well as to compare modalities, we built multivariate models to predict outcome from each data type at each timepoint. We found large differences across both assay types and sample timepoints! (7/x)

When we looked at the specific features we defined, we found some that were temporally dependent, with good predictive power at one timepoint but poor predictive power at another timepoint (6/x)

We then tested which of the 800+ features from SpaceCat could predict response to immunotherapy, finding numerous strong predictors. Interestingly, features defined in specific regions of the tumor did an especially good job at predicting outcome (5/x)

To help us make sense of this spatially-resolved data, we built SpaceCat, an algorithm to quantify and summarize the key features from spatial datasets. SpaceCat can be applied to processed imaging data from any multiplexed imaging platform! (4/x)

We then generated highly multiplexed imaging data using an antibody panel of 37 antibodies. This allowed us to identify 22 cell types across the more than 650 TMA cores we imaged from 117 total patients (3/x)

Our awesome collaborators at NKI put together a unique cohort spanning primary disease, pre-treatment metastases, and on-treatment metastases from triple negative breast cancer patients enrolled in the TONIC clinical trial (2/x)