Reproduce “HDTree: Generative Modeling of Cellular Hierarchies for Robust Lineage Inference” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zelin Zang, WenZhe Li, Yongjie Xu, Chang Yu, Changxi Chi, Jingbo Zhou, Zhen Lei, Stan Z. Li
A generative modeling framework called HDTree was developed, integrating hierarchical vector quantization with a quantized diffusion process to robustly infer cellular differentiation lineages from single-cell data. This framework achieved superior performance in metrics like Dendrogram Purity and temporal consistency on various single-cell datasets, notably surpassing TreeVAE by over 8 percentage points in DP on CIFAR-10 and showing improved temporal consistency on C. elegans data.
Key insights
- Decoupling tree complexity from the neural network architecture through a unified hierarchical codebook significantly improves scalability, stability, and generalization for modeling deep hierarchies, even with sparse data.
- Utilizing a quantized diffusion process for data generation allows for modeling continuous cell state transitions more robustly, which aligns with biological differentiation processes and avoids issues like posterior collapse common in VAE decoders.
- A composite loss function incorporating soft contrastive learning, hierarchical quantization, and diffusion-based reconstruction effectively balances the preservation of local semantic information, multi-level hierarchical structure, and faithful data generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.