Physiological mapping and prediction reveal stress-responsive translation initiation landscapes
Translation initiation defines the protein output of the transcriptome, but its physiological dynamics in cells and tissues remain incompletely characterized. Here we develop TISCA2, an inhibitor-free method that captures eIF5B-associated initiating ribosomes to map translation initiation sites genome-wide. By directly identifying start sites rather than inferring translation from elongating-ribosome periodicity, TISCA2 improves the detection of short and overlapping translons. Translation of a subset of TISCA2-identified translons is supported by N-terminal proteomics and HLA-I immunopeptidomics. The arsenite-induced integrated stress response broadly redistributes start-site usage, reducing upstream initiation in favor of downstream initiation. TISCA2 further reveals tissue-specific initiation landscapes in vivo, with increased detection of downstream candidate start sites in the testis. Finally, we develop IncepTIS, a sequence-based deep learning model that predicts position-resolved and condition-dependent initiation landscapes. IncepTIS-guided optimization generates de novo mRNA 5' leaders with stress-responsive initiation patterns. Together, TISCA2 and IncepTIS provide an experimental and computational framework for mapping, predicting and engineering translation initiation in mammalian cells and tissues.