ContiTE: continuous manifold MoE for few-shot cross-tissue mRNA translation efficiency prediction

Predicting mRNA translation efficiency across diverse tissues is a critical yet challenging task due to the phenomenon of concept shift, where identical genomic sequences exhibit distinct functional profiles across varying cellular environments. Existing static models are often limited by fixed parameterization, struggling to capture these continuous regulatory variations without requiring computationally expensive fine-tuning. To address these limitations, we propose ContiTE, a continuous manifold mixture-of-experts (MoE) framework explicitly designed for efficient few-shot cross-tissue adaptation. Unlike traditional MoE architectures that rely on discrete output-space mixing, ContiTE operates on a continuous parameter manifold by dynamically synthesizing domain-specific weights through a context-aware hyper-router that linearly combines shared atomic basis kernels. This paradigm enables smooth interpolation of translational rules and provides a flexible mechanism to model complex, context-dependent biological regulation. Furthermore, we introduce a gradient-based test-time adaptation strategy that allows the model to rapidly align to new, unseen tissues by solely optimizing low-dimensional context embeddings while keeping the backbone parameters frozen. Experimental results on a comprehensive human and mouse atlas demonstrate that ContiTE significantly outperforms state-of-the-art methods in few-shot scenarios, improving average Pearson correlation coefficient by 12.09% and $R^{2}$ by 18.22%. By mitigating negative transfer and requiring only limited target-domain data for calibration, ContiTE provides a computational framework for tissue-specific mRNA translation-efficiency prediction and demonstrates a parameter-reconstruction strategy that may be extensible to other cross-domain sequence-learning tasks, subject to task-specific validation.