Fig. 5

Single-cell afatinib response prediction in TNBC cells. A Schematics of scASTRAL. Contrastive learning is used in the first step to build an embedding space where positive examples (two cells of the same class) are separated from negative examples (two cells of different classes). In the second step an SVM classifier is trained on the learned embedded space to predict the class (i.e. afatinib-tolerant or afatinib-sensitive) of novel cells. B Inter- and intra- cosine distance between sensitive and tolerant MDAMB468 cells before and after the training. Intra-distance is the cosine distance among cells of the same afatinib response class, while inter-distance is the distance among cells belonging to two different afatinib response classes C UMAP representation of 22,724 triple negative breast cancer cells from 16 cell lines. D Spearman Correlation Coefficient (SCC) between scASTRAL predicted resistance to afatinib and experimentally estimated IC50. E Histogram of SCC values computed between scASTRAL predicted sensitivity to afatinib and experimentally estimated IC50 using 374 random genes. 1,000 simulations were performed. Red arrow indicates SCC value obtained using the 374 afatinib response marker genes we identified with our retrospective lineage tracing approach. F UMAP representation of 41,189 triple-negative breast cancer (TNBC) cells extracted from treatment-naïve primary tumours of 16 patients. G Estimation of the proportion of cells sensitive to afatinib using DREEP and BeyondCell tools. H Spearman Correlation Coefficient (SCC) between scASTRAL-predicted afatinib sensitivity and estimated sensitivity obtained from DREEP and BeyondCell tools