Interpretable Deep Learning for Elucidating Translation Elongation Dynamics

Gene expression regulation is a fundamental biological process governing protein abundance and cellular function. Translation, particularly elongation, is a key regulatory step whose dysregulation is implicated in metabolic disorders and cancer. Despite its importance, the sequence, structural, and contextual determinants of elongation dynamics remain incompletely understood. In this thesis, we develop three complementary interpretable machine learning frameworks to systematically elucidate translation elongation. First, we propose Riboclette, a conditional transformer-based model that achieves state-of-the-art performance in predicting ribosome density profiles across multiple amino acid deprivation conditions. Leveraging a novel double-head architecture, Riboclette isolates condition-specific effects. Through attribution and perturbation analyses, we uncover codon- and motif-level drivers of ribosome stalling, including a previously uncharacterized deprivation codon ramp effect. Second, we present RiboGL, a graph-based framework that integrates mRNA sequence and secondary structure by modeling transcripts as graphs. By combining graph neural networks with sequence models, RiboGL captures long-range structural interactions and explains stalling phenomena missed by sequence-only approaches. Finally, we develop BIO-Distiller, a model-agnostic knowledge distillation framework that transfers representations from biological foundation models into efficient and interpretable predictors. Applied to translation-related tasks, it enables motif-level insights while maintaining high predictive performance. Together, these frameworks provide a unified perspective on translation elongation, bridging local sequence features, global structural context, and scalable representation learning. By uncovering mechanistic drivers of ribosome dynamics across conditions and scales, this work advances our understanding of translational regulation and establishes a foundation for identifying regulatory elements, characterizing disease-associated dysregulation, and enabling the design of targeted therapeutic strategies.