Research map / Probabilistic and scientific modeling / Molecules and proteins
Reconstructing Cell Lineage Trees from Phenotypic Features with Metric Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CellTreeQM, a transformer-based metric-learning method that embeds single-cell transcriptomes so that quartet four-point additivity holds, then reconstructs lineage trees with neighbor-joining, benchmarked on C. elegans lineage data and Brownian-motion simulations.
Problem
Reconstructing cell lineage trees from phenotypic (transcriptomic) features, where classical tree-fitting on raw expression data is poor because gene expression is noisy and only sparsely supervised lineage labels exist.
Approach
Learn an embedding with a transformer encoder and feature gating whose pairwise distances satisfy the tree four-point condition on quartets of leaves, supervised by known (or partially known) quartet topologies, then build the tree with Neighbor-Joining.
Key results
- On C. elegans Small in the supervised setting, CellTreeQM-G reaches test RF 0.226 vs. 0.923 for raw-data reconstruction (delta RF 0.757), outperforming Triplet (0.741) and Quadruplet (0.492 without gating).
- In the high-level partitioning setting at partition level 2, CellTreeQM reaches QD 0.129 and delta %QD 0.764, while Triplet and Quadruplet reach 0.326 and 0.286 QD.
- Under partially leaf-labeled setting with 80% known leaves, CellTreeQM achieves test RF 0.506 with delta %RF 0.448; at 30% known labels, all methods show only modest gains and negative delta %RF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CellTreeQM instead of Triplet loss | C. elegans Small supervised reconstruction · C. elegans Small | Test RF: 0.741 (0.01) -> 0.286 (0.05) | Helped |
| use CellTreeQM instead of Quadruplet loss | C. elegans Small supervised reconstruction · C. elegans Small | Test RF: 0.492 (0.03) -> 0.286 (0.05) | Helped |
| add feature gating (ablation) | CellTreeQM · C. elegans Small, supervised | Test RF: 0.286 (0.05) -> 0.226 (0.02) | Helped |
| permute the training tree labels (ablation) | CellTreeQM · C. elegans Small, supervised | Test RF: 0.286 (0.05) -> 0.566 (0.04) | Hurt |
| use CellTreeQM instead of Triplet loss | High-level partitioning · C. elegans Small, partition level 3 | RF: 0.702(0.01) -> 0.538(0.01) | Helped |
| use CellTreeQM instead of Quadruplet loss | High-level partitioning · C. elegans Small, partition level 3 | RF: 0.854(0.01) -> 0.538(0.01) | Helped |
| use partition level 2 rather than partition level 3 | CellTreeQM high-level partitioning · C. elegans Small | QD: 0.156(0.00) -> 0.129(0.01) | Helped |
| reduce the known leaf fraction from 0.8 to 0.3 | CellTreeQM partially leaf-labeled training · C. elegans Small | Delta%RF: 0.448(0.06) -> -0.023(0.02) | Hurt |
| use the Transformer encoder instead of fully connected layers (ablation) | CellTreeQM supervised model · C. elegans real datasets | reconstruction performance | Helped |
| provide no quartet constraints or partial subtree information | CellTreeQM · Real empirical data, unsupervised | lineage reconstruction performance | Hurt |
Surprising findings
- In the unsupervised setting, performance on real empirical data was not significant, while simulated data showed only partial success.
- At only 30% known leaf labels, all methods including CellTreeQM show negative delta %RF, indicating essentially no improvement over raw data.
Limitations
- Relies on Neighbor-Joining for final tree construction, which can fail under complex noise.
- Distortion regularization may impede learning when lineage signals are significantly distorted in real data; gating impact on real data is mild.
- Unsupervised performance on real data is not significant; benchmarks are limited to lineage-resolved, well-defined ground-truth datasets.
Experimental setup
Datasets: C. elegans Small C. elegans Mid C. elegans Large Synthetic Brownian-motion simulations
Benchmarks: Cell Lineage Reconstruction Benchmark (C. elegans Small/Mid/Large, synthetic)
Metrics: Robinson-Foulds distance quartet distance delta % RF improvement over raw data
Baselines: Triplet loss Quadruplet loss raw-data neighbor-joining
Builds on: triplet loss quadruplet loss Gumbel-Softmax neighbor-joining Brownian motion
Closest papers by meaning
- CellVerse: Do Large Language Models Really Understand Cell Biology? (NeurIPS 2025 · similarity 0.85)
- CellPLM: Pre-training of Cell Language Model Beyond Single Cells (ICLR 2024 · similarity 0.85)
- A Simple and Comprehensive Benchmark for Single-Cell Transcriptomics (AAAI 2025 · similarity 0.85)
- ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy (ICML 2025 · similarity 0.84)
- sciLaMA: A Single-Cell Representation Learning Framework to Leverage Prior Knowledge from Large Language Models (ICML 2025 · similarity 0.84)
- Cell2Sentence: Teaching Large Language Models the Language of Biology (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.