Introduction

Proteins are the functional workhorses of cellular life, carrying out essential tasks that maintain cellular health, growth, and survival. From gene expression to energy production, proteins are central to nearly every biological process. However, their ability to perform these roles is deeply dependent on one critical factor: where they are located within the cell. Some proteins must be in the nucleus to regulate transcription, while others localize to mitochondria to produce ATP, or to the plasma membrane to mediate signaling. These precise placements are directed by intrinsic signals within the protein sequences—like molecular “zip codes”—that guide them to the correct subcellular compartments.

    Understanding protein localization is not just a matter of cellular logistics—it’s essential to understanding disease. When proteins are mislocalized, they can lose their function or interfere with the function of others, often resulting in cellular dysfunction. This is especially relevant in cancer biology, where disrupted protein localization can contribute to uncontrolled growth, evasion of apoptosis, and treatment resistance. In the context of radiation therapy, which is commonly used to treat cancer, changes in protein localization after exposure may be part of how tumor cells adapt, survive, and eventually recur. Uncovering these patterns could offer insight into why some cancers resist radiation while others do not.

    This project uses machine learning (ML) to predict protein localization based on sequence-derived features, biochemical properties, and public datasets like UniProt. UniProt is a rich, curated resource containing detailed information about protein sequences, structures, and known localizations. By leveraging this data, machine learning models can be trained to recognize patterns—both obvious and subtle—that determine where a protein is likely to reside in the cell. These patterns include amino acid composition, sequence motifs, hydrophobicity, charge, and predicted secondary structures. Some models can also integrate data from protein-protein interaction networks and gene expression levels, offering a more context-aware prediction.

    Traditional laboratory methods such as fluorescence microscopy or subcellular fractionation, though accurate, are time-consuming, costly, and difficult to scale. Machine learning presents a faster, more scalable alternative, capable of analyzing thousands of proteins at once and making predictions based purely on sequence data. Modern deep learning approaches, in particular, can extract complex, nonlinear features from protein sequences—often capturing biological signals that traditional models may miss.

    Accurate protein localization prediction has far-reaching implications for biomedical research. In cancer, mislocalized proteins have been linked to numerous pathological processes. For example, a mutation that causes a tumor suppressor to be trapped outside the nucleus could render it ineffective, allowing cancer progression. Similarly, resistance to radiation therapy may involve stress-response proteins that translocate to new compartments after exposure, enabling the cell to repair damage or avoid cell death. By building better models for protein localization, we can better understand these mechanisms and potentially identify therapeutic targets to disrupt them.

    Just as AlphaFold has transformed the field of structural biology, the future of cellular biology will require models that go beyond static structures and consider the functional context of proteins—when, where, and under what conditions they operate. This project aims to contribute to that future by building predictive models that help uncover how protein localization contributes to disease progression and therapy response.

    Although predicting localization is a challenging task due to the dynamic nature of the cell and the complexity of molecular interactions, it is a solvable problem—especially with the right computational tools. By combining deep learning, biological insight, and publicly available datasets, this work seeks to make meaningful contributions to our understanding of cell function and cancer biology. Ultimately, this research could help uncover hidden patterns that drive tumor behavior and open up new avenues for precision medicine.