A General Single-Cell Analysis Framework via Conditional Diffusion Generative Models

Wenzhuo Tang, Renming Liu, Hongzhi Wen, Xinnan Dai, Jiayuan Ding, Hang Li, Wenqi Fan, Yuying Xie, Jiliang Tang

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Single-cell analysis, Diffusion generative models, AI for science

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We developed a conditional diffusion generative model for single-cell analysis.

Abstract: The fast-growing single-cell analysis community extends the horizon of quantitative analysis to numerous computational tasks. While the tasks hold vastly different targets from each other, existing works typically design specific model frameworks according to the downstream objectives. In this work, we propose a general single-cell analysis framework by unifying common computational tasks as posterior estimation problems. In light of conditional diffusion generative models, we introduce scDiff through the proposed framework and study different conditioning strategies. With data-specific conditions, scDiff achieves competitive performance against state-of-the-art in various benchmarking tasks. In addition, we illustrate the flexibility of scDiff by incorporating prior information through large language models and graph neural networks. Additional few-shot and zero-shot experiments prove the effectiveness of the prior conditioner on scDiff.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4159