HDTree: Generative Modeling of Cellular Hierarchies for Robust Lineage Inference

Changxi Chi; Chang Yu; Jingbo Zhou; Stan Z. Li; Wenzhe Li; Yongjie Xu; Zelin Zang; Zhen Lei

arxiv: 2506.23287 · v3 · pith:7RYG2CLYnew · submitted 2025-06-29 · 💻 cs.LG · q-bio.QM

HDTree: Generative Modeling of Cellular Hierarchies for Robust Lineage Inference

Zelin Zang , WenZhe Li , Yongjie Xu , Chang Yu , Changxi Chi , Jingbo Zhou , Zhen Lei , Stan Z. Li This is my paper

Pith reviewed 2026-05-21 23:30 UTC · model grok-4.3

classification 💻 cs.LG q-bio.QM

keywords single-cell analysislineage inferencegenerative modelsdiffusion modelshierarchical structurescellular differentiationWaddington landscape

0 comments

The pith

HDTree models cellular hierarchies in a single latent space using a unified codebook and quantized diffusion for lineage inference.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces HDTree to address limitations in modeling single-cell differentiation trajectories. Traditional VAE methods require branch-specific modules that limit scalability and cause instability like posterior collapse. HDTree instead uses a unified hierarchical codebook in a latent space and a quantized diffusion process to capture tree structures and continuous transitions. This alignment with the Waddington landscape aims to produce more stable, scalable, and biologically plausible lineage inferences, as shown in tests on single-cell and general datasets where it outperforms prior approaches in accuracy and consistency.

Core claim

HDTree captures tree relationships within a hierarchical latent space using a unified hierarchical codebook and employs a quantized diffusion process to model continuous cell state transitions. By aligning the generative process with the Waddington landscape, this method improves stability and scalability while enhancing the biological plausibility of inferred lineages.

What carries the argument

Unified hierarchical codebook in a latent space paired with quantized diffusion process for modeling cell state transitions along branching trajectories.

Load-bearing premise

A single unified hierarchical codebook together with a quantized diffusion process can faithfully represent branching cellular trajectories and their alignment to the Waddington landscape increases biological plausibility.

What would settle it

A test on a dataset with well-characterized branching where HDTree produces less accurate lineages than methods using branch-specific modules would falsify the claim of superior performance and plausibility.

read the original abstract

In single-cell research, tracing and analyzing high-throughput single-cell differentiation trajectories is crucial for understanding biological processes. Key to this is the robust modeling of hierarchical structures that govern cellular development. Traditional methods face limitations in computational cost, performance, and stability. VAE-based approaches have made strides but still require branch-specific network modules, limiting their scalability and stability, while often suffering from posterior collapse. To overcome these challenges, we introduce HDTree, a generative modeling framework designed for robust lineage inference. HDTree captures tree relationships within a hierarchical latent space using a unified hierarchical codebook and employs a quantized diffusion process to model continuous cell state transitions. By aligning the generative process with the Waddington landscape, this method not only improves stability and scalability but also enhances the biological plausibility of inferred lineages. HDTree's effectiveness is demonstrated through comparisons on both general-purpose and single-cell datasets, where it outperforms existing methods in lineage inference accuracy, reconstruction quality, and hierarchical consistency. These contributions enable accurate and efficient modeling of cellular differentiation paths, offering reliable insights for biological discovery.\footnote{Code is available at https://github.com/zangzelin/code\_HDTree\_icml.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

HDTree swaps branch-specific modules for a shared hierarchical codebook plus quantized diffusion, but the abstract gives no numbers to show whether the gains in stability and consistency actually appear.

read the letter

HDTree combines a unified hierarchical codebook with quantized diffusion to model cellular lineages in a single generative framework. This setup is meant to sidestep the scalability and stability problems that come with VAE models that need separate networks for each branch and often run into posterior collapse. The alignment with the Waddington landscape is presented as the inductive bias that keeps the generations biologically plausible. Code release on GitHub is a straightforward plus for anyone who wants to inspect the implementation. The main new piece is the specific integration of the hierarchical codebook and quantized diffusion for lineage tracing in single-cell data rather than a brand-new paradigm. It does a clean job stating the motivation from prior limitations and positions the method as more scalable than the branch-specific alternatives. The soft spot is the missing evidence. The abstract asserts outperformance on lineage inference accuracy, reconstruction quality, and hierarchical consistency across general and single-cell datasets, yet supplies no metrics, error bars, or protocol details. Without those, the central performance claims stay unverified. The concern about branch separation is reasonable on the current text. A single shared codebook must actively preserve distinct trajectories during quantization and diffusion; otherwise cross-branch leakage could erase the reported consistency gains. The abstract relies on Waddington alignment to handle this, but does not show how the update rules enforce separation. This paper is for computational biologists already working on trajectory inference who might want to test the architecture. A reader in that niche could extract useful ideas even if the results section needs closer checking. I would send it for peer review to get the methods, ablations, and quantitative results examined properly.

Referee Report

2 major / 2 minor

Summary. The paper introduces HDTree, a generative modeling framework for robust lineage inference from single-cell differentiation trajectories. It models tree relationships in a hierarchical latent space via a unified hierarchical codebook and a quantized diffusion process for continuous cell state transitions. By aligning the generative process with the Waddington landscape, the method aims to improve stability, scalability, and biological plausibility over VAE-based approaches that rely on branch-specific network modules and suffer from posterior collapse. The abstract claims superior performance on general-purpose and single-cell datasets in lineage inference accuracy, reconstruction quality, and hierarchical consistency, with code released at a GitHub repository.

Significance. If the central claims hold under rigorous evaluation, HDTree could advance single-cell trajectory modeling by offering a scalable alternative that avoids per-branch modules while mitigating posterior collapse through quantization and hierarchical codebooks. The alignment with the Waddington landscape provides a potentially useful inductive bias for biological plausibility. The public code release is a positive step toward reproducibility and allows independent verification of the quantized diffusion and codebook mechanisms.

major comments (2)

[Abstract] Abstract: The central performance claims state that HDTree 'outperforms existing methods in lineage inference accuracy, reconstruction quality, and hierarchical consistency' on both general and single-cell datasets, yet the abstract supplies no quantitative metrics, error bars, baseline descriptions, or evaluation protocol details. Without these, the outperformance assertion cannot be assessed and is load-bearing for the paper's contribution.
[Framework description] Framework description (hierarchical codebook and quantized diffusion): The unified hierarchical codebook is presented as capturing tree relationships without branch-specific modules, but no derivation or update rule is shown demonstrating that quantization enforces separation of distinct developmental trajectories rather than permitting cross-branch averaging or leakage in the latent space. This assumption underpins all claims of improved hierarchical consistency and biological plausibility via Waddington alignment.

minor comments (2)

[Abstract] The footnote correctly notes code availability, but the main text should include a brief reproducibility statement summarizing key hyperparameters and dataset preprocessing steps used in the reported comparisons.
[Methods] Notation for the unified codebook and quantization operator could be introduced with an explicit equation early in the methods to improve clarity for readers unfamiliar with vector quantization in diffusion models.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and positive assessment of the work's potential. We address each major comment point by point below and will revise the manuscript accordingly to improve clarity and rigor.

read point-by-point responses

Referee: [Abstract] Abstract: The central performance claims state that HDTree 'outperforms existing methods in lineage inference accuracy, reconstruction quality, and hierarchical consistency' on both general and single-cell datasets, yet the abstract supplies no quantitative metrics, error bars, baseline descriptions, or evaluation protocol details. Without these, the outperformance assertion cannot be assessed and is load-bearing for the paper's contribution.

Authors: We agree that including key quantitative results would make the abstract more informative and allow readers to better evaluate the claims. In the revised manuscript, we will incorporate specific metrics (e.g., lineage inference accuracy improvements with error bars), brief baseline descriptions, and an outline of the evaluation protocol while maintaining conciseness. revision: yes
Referee: [Framework description] Framework description (hierarchical codebook and quantized diffusion): The unified hierarchical codebook is presented as capturing tree relationships without branch-specific modules, but no derivation or update rule is shown demonstrating that quantization enforces separation of distinct developmental trajectories rather than permitting cross-branch averaging or leakage in the latent space. This assumption underpins all claims of improved hierarchical consistency and biological plausibility via Waddington alignment.

Authors: We acknowledge the value of an explicit derivation to justify how the hierarchical codebook and quantization prevent cross-branch leakage. In the revision, we will add a mathematical derivation and update rules in the methods section (or appendix) demonstrating trajectory separation in the latent space, along with supporting analysis for the Waddington alignment and hierarchical consistency claims. revision: yes

Circularity Check

0 steps flagged

No circularity: framework claims rest on empirical comparisons and novel architectural elements rather than self-referential reductions.

full rationale

The abstract and framework description introduce HDTree as a new generative model employing a unified hierarchical codebook and quantized diffusion process aligned with the Waddington landscape. Performance is assessed via direct comparisons against existing methods on external general-purpose and single-cell datasets, with reported gains in accuracy, reconstruction, and hierarchical consistency. No equations, parameter-fitting steps, or derivations are shown that would make any prediction equivalent to its inputs by construction. The central claims do not reduce to self-citations, ansatzes smuggled via prior work, or renaming of known results; they rely on the proposed architecture's behavior on held-out data. This is the most common honest outcome for papers whose core contributions are architectural and empirically validated.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the domain assumption that cellular differentiation trajectories form tree structures that can be captured by a single hierarchical codebook and that diffusion can be quantized to match continuous state transitions without additional branch-specific components.

axioms (2)

domain assumption Cellular differentiation trajectories form tree-like hierarchical structures that can be represented in a unified latent codebook.
Invoked when the paper states that HDTree captures tree relationships within a hierarchical latent space using a unified hierarchical codebook.
domain assumption Quantized diffusion aligned to the Waddington landscape produces biologically plausible continuous cell state transitions.
Stated when the method is described as aligning the generative process with the Waddington landscape to enhance biological plausibility.

pith-pipeline@v0.9.0 · 5762 in / 1400 out tokens · 41955 ms · 2026-05-21T23:30:25.657867+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

HDTree captures tree relationships within a hierarchical latent space using a unified hierarchical codebook and employs a quantized diffusion process
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

aligning the generative process with the Waddington landscape

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.