Zero-shot prediction of drug responses using biologically informed neural networks trained on phosphoproteomic timeseries
Fig 1
Extension of the LEMBAS framework to time-resolved phosphoproteomics.
A. Illustration of a minimal signaling network including ligand (L), receptor (R), kinases (K1–K4), non-kinase signaling proteins (P1, P2) and kinase inhibitor (D). B. The model architecture mirrors this structure, ligand and drug inputs drive an RNN whose signal propagation is constrained by the prior-knowledge network (encoded as an adjacency matrix with trainable weights); the hidden state is feed into a layer that maps signal states to phosphosites, which in turn is mapped to specific timepoints. C. The phosphosite mapping layer connects each signaling node only to its own phosphosites that are identity-coded with a trainable a low-dimensional embedding (five dimensions), and a multilayer perceptron is used to (non-linearly) transform the signal back to a single phosphosite output. D. The time-point mapping layer: a set of learnable anchors is constrained to be positive, monotonic and with the first fixed at the first RNN step. Soft indexing of the two nearest integer RNN steps interpolates the value at each non-integer anchor. E. Generated phosphosite intensities across time points for EGF-stimulation data, only displaying variable sites (with standard deviation > 0.001). F. Ablation results across model configurations. Circles are the five cross-validation folds; X’s correspond to zero-shot evaluation. Average training times are also displayed.