<p>The efficient and safe delivery of messenger RNA (mRNA) therapeutics remains a critical challenge for clinical translation, driving the need for advanced carrier design. Ionizable amphiphilic Janus dendrimers (IAJDs) represent a promising class of carriers; however, their structural complexity and limited available datasets hinder systematic exploration and optimization. In this study, we established a tailored machine-learning framework to investigate the structure-function relationships of IAJDs under a constrained data regime (<i>n</i>=231). Conventional molecular fingerprints were found to be suboptimal for representing these macromolecules, motivating the adoption of count-based descriptors and systematic ablation analyses to disentangle the contributions of the substructural features. These experiments identified key functional motifs underlying transfection performance and provided interpretable insights into the IAJD design principles. Complementing these handcrafted descriptors, we further applied deep learning-based molecular embeddings, which captured higher-order chemical semantics and significantly improved predictive accuracy. Collectively, these advances demonstrate that both refined fingerprinting and representation learning approaches can overcome data limitations, enabling the reliable prediction of IAJD activity while offering mechanistic interpretability. This study illustrates the potential of data-driven strategies as hypothesis-generation and prioritization tools for the design of next-generation mRNA delivery systems.</p>