scCoLLM: An LLM-Based Model Using Context-Space Adaptation for Single-Cell Analysis under Domain Shift

Deep learning has been adopted throughout single-cell analysis with remarkable success. However, once the target data to be analyzed depart from the distribution a single-cell model was trained on, current single-cell models exhibit marked performance degradation. Such domain shift is the rule rather than the exception in practice: newly profiled data routinely come from unfamiliar cell types, tissues, and donors. We attribute much of this degradation to a missing conditioning signal: existing methods predict each query cell from its own expression profile alone, ignoring the target-domain-specific biological context carried by the target dataset itself. Closing this gap by training or fine-tuning on the target domain is often infeasible, since every such remedy requires an additional training run for each new target, its supervised variants further require ground truth annotations that a newly profiled dataset does not carry, and the strongest models may not release their weights. We therefore recast target-domain adaptation from parameter space to context space. We introduce **s**ingle-**c**ell **Co**ntext **LLM** (**scCoLLM**), a training-free framework that performs **Context-Space Adaptation** on top of general-purpose LLMs such as GPT. Given a query cell, scCoLLM retrieves a query-conditioned biological context set from the unlabeled target data, selected to balance query-level local relevance against target-domain global representativeness, yielding a set of cell–profile pairs through zero-shot reasoning. The resulting cell–profile pairs form the **Bio-Context**. Conditioned on this Bio-Context, scCoLLM analyzes the query cell while aware of the biological context of the target domain, and thereby adapts to that domain without gradient updates, ground-truth labels, or access to model parameters. Under three domain shift tasks—cross-cell type, cross-tissue, and cross-donor—scCoLLM attains state-of-the-art performance, improving on the strongest of 12 baselines by up to 9%. Further experiments confirmed the soundness of the Bio-Context design and its decisive role in improvement, as well as the robustness and transferability of scCoLLM.