Fig. 1: An integrated counterfactual optimization framework for discovering therapeutic strategies predicted to drive CD8⁺ T-cell infiltration in human tumours.

a, Overview of the Morpheus framework, which consists of first training a T cell predictor and then generating perturbations. b, Training a neural network classifier to predict the presence of CD8⁺ T cells from multiplexed tissue images where cells in the IMC images are pixelated and CD8⁺ T cells are masked (Methods). c, The trained classifier is then used to compute an optimal perturbation vector δ⁽ⁱ⁾ per patch by jointly minimizing three loss terms (L_pred, L_dist and L_proto). The perturbation δ⁽ⁱ⁾ represents a strategy for altering the level of a small number of signalling molecules in patch \({x}_{0}^{(i)}\) in a way that increases the probability of T-cell presence as predicted by the classifier. The optimization also favours perturbations that shift the image patch to be more similar to its nearest T-cell patches in the training data, shown as Proto. Each perturbation corresponds to adjusting the relative intensity of each imaging channel. Taking the median across all perturbations produces a whole-tumour perturbation strategy, which we assess by perturbing in silico tumour images from a test patient cohort and examining the predicted T-cell distribution after perturbation.