Researchers have developed ProteinTalks, an artificial intelligence-powered virtual cell model designed to accelerate in silico drug discovery. The model utilizes extensive, time-resolved proteomics data from perturbed breast cancer cell lines to predict drug efficacy and identify potential drug combinations.
Central to ProteinTalks is a new pretraining framework that learns transferable dynamical latent representations from temporal proteome trajectories. The team generated over 38 million temporal protein-abundance measurements. By modeling protein responses to various perturbations, the model can predict drug efficacy, discover new drug combinations, probe drug resistance mechanisms, stratify patient responses, and prioritize drug candidates for patient organoids.
The model demonstrates robust transferability, extending its predictive capabilities beyond cell lines to patient-derived organoids and clinical biopsies, consistently outperforming benchmark implementations. Raw mass spectrometry proteomics data are available via iProX (IPX0007409000), and the PTDS protein matrix and associated resources are accessible through db.prottalks.com for academic and non-commercial use. Additional data sources include the Kyoto Encyclopedia of Genes and Genomes, Metascape, STRING, and Ingenuity Pathway Analysis. The project’s data analysis codes are available on GitHub at https://github.com/guomics-lab/PTV-1.
Read the original coverage
💬 Comments
📜 Comment Policy