Research map / Vision-language and multimodal / Cross-modal learning
PETRI: Learning Unified Cell Embeddings from Unpaired Modalities via Early-Fusion Joint Reconstruction
Ryan W Conrad, Ethan Weinberger, Saradha Venkatachalapathy, Yuwen Chen, Darshini Shah, Bay Johnson, Max R Salick, Vaishaali Natarajan, Emily Fox
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multimodal cell documents (early fusion) to unimodal transcriptomics model | PETRI · HepG2 | StringDB edge TPR@5%FPR: 0.144 -> 0.242 | Improved |
| add multimodal cell documents (early fusion) to unimodal image model | PETRI · HepG2 | StringDB edge TPR@5%FPR: 0.116 -> 0.242 | Improved |
| use PETRI instead of ViT MAE | PETRI · HepG2 | StringDB ROC AUC: 0.549 -> 0.628 | Improved |
| use early-fusion PETRI instead of late-fusion max cosine aggregation | PETRI · HepG2 | StringDB edge TPR@5%FPR: 0.219 -> 0.242 | Improved |
| use PETRI instead of CLIP early-fusion baseline | PETRI · Perturb-Multi | Guide Consistency: 0.000 -> 0.208 | Improved |
| group cells by perturbation instead of randomly (permuted data) (ablation) | PETRI · Perturb-Multi | Guide Consistency: 0.163 -> 0.208 | Improved |
| group cells by perturbation instead of randomly (permuted data) (ablation) | PETRI · HepG2 | StringDB edge TPR@5%FPR: 0.255 -> 0.242 | Hurt |
| group cells by perturbation instead of randomly (permuted data) (ablation) | PETRI · HepG2 | Guide Consistency: 0.274 -> 0.278 | No clear effect |
| use encoder outputs before the multimodal set transformer instead of after layer 4 (ablation) | PETRI · HepG2 | Guide Consistency: 0.221 -> 0.260 | Improved |
| use encoder outputs before the multimodal set transformer instead of after layer 4 (ablation) | PETRI · HepG2 | StringDB edge TPR@5%FPR: 0.186 -> 0.242 | Improved |
| apply robust center scaling and PCA whitening to embeddings (ablation) | PETRI · HepG2 | StringDB edge TPR@5%FPR: 0.031 -> 0.242 | Improved |
| vary latent tokens per cell and set size (ablation) | PETRI · HepG2 | StringDB and Guide Consistency | No clear effect |
Closest papers by meaning
- Comprehensive View Embedding Learning for Single-Cell Multimodal Integration (AAAI 2024 · similarity 0.86)
- A Cross Modal Knowledge Distillation & Data Augmentation Recipe for Improving Transcriptomics Representations through Morphological Features (ICML 2025 · similarity 0.86)
- ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy (ICML 2025 · similarity 0.84)
- CellCLIP - Learning Perturbation Effects in Cell Painting via Text-Guided Contrastive Learning (NeurIPS 2025 · similarity 0.84)
- Multimodal Lego: Model Merging and Fine-Tuning Across Topologies and Modalities in Biomedicine (ICLR 2025 · similarity 0.83)
- Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images (NeurIPS 2026 · similarity 0.83)
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.
