Embedding-Guided Design Rules for Programmable Mitochondrial Protein Import and Modular Organelle Organization

Abstract

Mitochondria coordinate cellular energy conversion, metabolism, signaling, and survival through the coordinated action of thousands of proteins distributed across the outer membrane, intermembrane space, inner membrane, and matrix. Yet, a systems-level framework that links mitochondrial protein sequence space to sub-mitochondrial localization and import logic remains incomplete. Here, we use protein language model embeddings from ESM2 to analyze the human mitochondrial proteome curated in MitoCarta3.0. Low-dimensional projections of embedding space recover biologically meaningful structure, with proteins forming compartment-enriched groupings consistent with submitochondrial organization. Unsupervised clustering and bipartite network analysis identify oxidative phosphorylation, protein import, and mitochondrial translation as highly connected hubs that anchor mitochondrial functional architecture. To probe engineering principles, we perform in silico minimaledit mutagenesis of N-terminal targeting presequences and quantify shifts in embedding position and predicted import potential. These perturbation experiments show that modest changes in charge distribution, hydropathy, and helix propensity can tune import behavior, with a substantial subset of proteins exhibiting large, directionally consistent shifts in predicted targeting. Together, the results support a modular view of mitochondrial organization and indicate that targeting presequences behave as tunable control elements rather than fixed codes. This work provides an embedding-driven computational blueprint for programmable mitochondrial targeting, with applications in synthetic biology, therapeutic payload routing, and organelle engineering.