Zero-shot prediction of drug responses using biologically informed neural networks trained on phosphoproteomic timeseries

Cellular signaling is driven by complex, dynamic phosphorylation networks that control growth and survival, and their dysregulation underlies diseases such as cancer. Although modern mass spectrometry enables large-scale quantification of phosphoproteomic responses over time, these measurements remain descriptive and cannot by themselves predict how signaling will evolve under perturbations. Here, we extend a biologically informed recurrent neural network framework (LEMBAS), to learn time-resolved phosphoproteomic trajectories. We introduce two interpretable modules; a phosphosite mapping that links signaling nodes to measured phosphorylation sites and a monotonic time mapping that aligns continuous experimental times to discrete signaling steps. Using synthetic benchmarks and an EGF-stimulation dataset with inhibitor treatments, the model accurately interpolates unseen time points and predicts drug-induced phosphoproteomic responses in a zero-shot setting, outperforming naïve and fully connected baselines. Importantly, the model identifies both canonical and non-canonical signaling effects, including modulation of the transcription factor FOXO3:S7 (from the PI3K/AKT pathway) by drugs affecting PTPN11 (from the RAS/ERK pathway). By combining mechanistic priors with deep learning, our framework provides a scalable approach to interpret and predict dynamic drug responses from phosphoproteomic data.Author summary: Cells constantly adjust their behavior in response to signals from their environment, and many drugs work by altering these communication networks. Measuring these changes directly is expensive and time-consuming, so we set out to build a computer model that can make accurate predictions without needing new experiments for each drug. We trained a neural network using large datasets that track protein modifications over time after cells are stimulated. The model uses prior biological knowledge about how proteins interact, allowing it to connect molecular events to drug effects. Remarkably, it can make “zero-shot” predictions; that is, it can predict the effect of drugs it has never seen before. We show that the model can capture both expected and surprising drug responses, and it can even suggest new links between signaling proteins. Our approach demonstrates how combining biological knowledge with modern machine learning can improve the prediction of cellular responses and may ultimately accelerate drug discovery and personalized medicine.