Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Learning Biological Hierarchies in Single-Cell Foundation Models
Xiangyu Guo, Ricardo Henao
Read the paperCheck your idea against 137,217 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.
Introduces scOntoFM, a single-cell foundation model that embeds the hierarchical cell ontology into representation learning, using lowest common ancestor distances from the ontology.
Problem
Most single-cell foundation models ignore the taxonomic lineage relationships among cell types encoded in the cell ontology.
Approach
Embeds the hierarchical cell ontology into representation learning by using Lowest Common Ancestor (LCA) distances from the ontology to structure the learned cell representations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ontology-aware losses and offline sampling vs SCimilarity on zero-shot clustering | scOntoFM · zero-shot clustering, nine datasets | scIB: 0.7960 -> 0.7958 | No clear effect |
| ontology-aware pretraining vs baselines on structure preservation | scOntoFM · zero-shot clustering | scGraph-LCA / scGraph-RWR: higher on most tissues (no numbers in text) | Improved |
| scOntoFM at coarse annotation levels | scOntoFM · HIHA AIFI_L1 | scIB score: 0.760 (cellPLM) vs 0.681 | Hurt |
| scOntoFM at fine annotation levels | scOntoFM · HIHA AIFI_L2 | scIB score: 0.722 (cellPLM) -> 0.782 | Improved |
| fine-tune scOntoFM vs other scFMs on HIHA | scOntoFM · HIHA | accuracy: 0.907 (cellPLM) -> 0.915 | Improved |
| novel cell-type identification with ontology-derived centroids | scOntoFM · 99 unseen cell types | average rank: 3.24 (scCello) -> 2.85 | Improved |
| gene function prediction with scOntoFM embeddings | scOntoFM · GO gene function | F1 micro: 0.913 -> 0.917 | Improved |
| scOntoFM vs Geneformer on AP | scOntoFM · GO gene function | AP micro: 0.922 -> 0.921 | No clear effect |
| spatial gene imputation with scOntoFM embeddings | scOntoFM · breast cancer | Pearson: 0.369 -> 0.388 | Improved |
| remove ontology alignment loss (ablation) | scOntoFM · zero-shot clustering | scIB vs scGraph-LCA: scIB up, scGraph-LCA substantially down (no numbers in text) | Mixed |
| remove dynamic margin ranking loss (ablation) | scOntoFM · zero-shot clustering | scIB vs scGraph-LCA: scGraph-LCA up, scIB slightly down (no numbers in text) | Mixed |
| offline ontology-guided sampling instead of random online sampling | scOntoFM · 10% of pretraining data | zero-positive anchors: 46.78% of anchors have zero positives online | Improved |
Closest papers by meaning
- Cell ontology guided transcriptome foundation model (NeurIPS 2024 · similarity 0.90)
- Reconstructing Cell Lineage Trees from Phenotypic Features with Metric Learning (ICML 2025 · similarity 0.86)
- CellVerse: Do Large Language Models Really Understand Cell Biology? (NeurIPS 2025 · similarity 0.85)
- ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy (ICML 2025 · similarity 0.84)
- scPilot: Large Language Model Reasoning Toward Automated Single-Cell Analysis and Discovery (NeurIPS 2025 · similarity 0.84)
- CellPLM: Pre-training of Cell Language Model Beyond Single Cells (ICLR 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.
