Title: scTranslation: A Holistic Benchmark for Single-Cell Multi-Omics Modality Translation

Abstract:

The ability to simultaneously measure multiple omics modalities within single cells has empowered researchers to achieve a deeper comprehension of cellular states and regulatory mechanisms. Nevertheless, the high costs associated with experiments, substantial noise levels, and partial modality coverage have driven the emergence of various computational techniques for modality translation in recent years. While translation models have advanced, there remains a critical absence of systematic benchmarking regarding datasets, evaluation metrics, and the factors that influence performance.

To bridge this gap, we introduce scTranslation, a robust benchmark designed for single-cell multi-omics modality translation tasks. This resource encompasses a wide array of translation datasets, incorporates state-of-the-art models, and offers a thorough suite of evaluation metrics. Furthermore, we evaluate model efficacy across distinct scenarios, including feature selection, feature quality, and few-shot conditions. Although these elements profoundly impact model performance, they have rarely been investigated in a systematic manner previously. By utilizing this benchmark, we execute a large-scale analysis of existing methods, uncovering numerous insightful findings that pave the way for future advancements. To support ongoing research, the benchmark is open-sourced, with anonymous code available at https://github.com/Bunnybeibei/scTranslation.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC