Research map / Label-efficient and robust learning / Self-supervised and contrastive
Transcriptomics-guided Slide Representation Learning in Computational Pathology
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 TANGLE, a slide-expression multimodal contrastive pre-training framework that aligns ABMIL slide embeddings with bulk gene expression profiles, evaluated on few-shot, prototype, and retrieval tasks in rat liver and human breast/lung.
Problem
Learning task-agnostic slide-level embeddings from giga-pixel whole-slide images with self-supervised learning is hard, as patch-level augmentations are costly and intra-slide heterogeneity weakens training signals.
Approach
Use bulk gene expression profiles as a complementary modality: align ABMIL slide embeddings with MLP-encoded expression embeddings via a symmetric cross-modal contrastive loss, instead of relying on slide augmentations or masking.
Key results
- On rat liver few-shot lesion classification at k=10, TANGLE reaches 84.7% macro-AUC versus 78.8% for iBOT-Tox+ABMIL (+5.9%).
- On human breast and lung subtyping at k=10, TANGLE reaches 84.8% and 91.5% macro-AUC, versus 73.8% and 85.3% for CTransPath+ABMIL.
- In TG-GATEs slide retrieval, TANGLE retrieves on average 2.88/4 relevant slides in top-10 and 3.44/4 in top-20.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| align slide and gene-expression embeddings with symmetric cross-modal contrastive learning | TANGLE · TG-GATEs liver, k=10 | Macro-AUC: +5.9% | Helped |
| align slide and gene-expression embeddings with symmetric cross-modal contrastive learning | TANGLE · In-house breast, k=10 | Macro-AUC: +11.0% | Helped |
| align slide and gene-expression embeddings with symmetric cross-modal contrastive learning | TANGLE · In-house lung, k=10 | Macro-AUC: +6.2% | Helped |
| replace cross-modal contrastive alignment with expression reconstruction (ablation) | TANGLE · Few-shot classification, larger k | Macro-AUC | Hurt |
| replace cross-modal contrastive alignment with expression reconstruction (ablation) | TANGLE · Few-shot classification, k=1 | Macro-AUC | Mixed |
| add the TANGLE-REC objective (ablation) | TANGLE · TG-GATEs liver | AUC: +0.05% AUC | Helped |
| add INTRA on top of the TANGLE-REC objective (ablation) | TANGLE · TG-GATEs liver | AUC: -0.8% AUC | Hurt |
| add INTRA without TANGLE-REC (ablation) | TANGLE · TG-GATEs liver | AUC: -2.0% AUC | Hurt |
| replace cross-modal contrastive loss with an L1 objective (ablation) | TANGLE · TG-GATEs liver | AUC: -6.7% AUC | Hurt |
| replace cross-modal contrastive loss with an L2 objective (ablation) | TANGLE · TG-GATEs liver | AUC: -7.0% AUC | Hurt |
| use intra-modality slide SSL instead of transcriptomics-guided pre-training (ablation) | TANGLE · In-house breast, k=10 | Macro-AUC: 84.8 -> 71.9 | Hurt |
| use intra-modality slide SSL instead of transcriptomics-guided pre-training (ablation) | TANGLE · In-house lung, k=10 | Macro-AUC: 91.5 -> 75.4 | Hurt |
Surprising findings
- Replacing the cross-modal contrastive loss with L1 or L2 regression objectives drops AUC by about 6.7% and 7.0%.
- Reconstruction-based TANGLE-REC performs surprisingly well at k=1 but is outperformed at larger k.
Limitations
- Relies on paired slide and bulk expression data, which are costly to obtain and limit training cohort size.
- Intra-modality INTRA objective may overfit to staining differences between train and test cohorts.
Experimental setup
Datasets: TG-GATEs TCGA-BRCA TCGA-NSCLC in-house BRCA in-house NSCLC
Benchmarks: TG-GATEs liver lesion classification TCGA-BRCA subtyping TCGA-NSCLC subtyping in-house BRCA and NSCLC
Metrics: macro-AUC Recall@k
Baselines: ABMIL TransMIL HIPT ResNet50+Avg. CTransPath+Avg. iBOT-Tox+Avg. INTRA
Builds on: attention-based MIL iBOT CLIP-style symmetric contrastive loss HIPT CTransPath
Closest papers by meaning
- Unsupervised Foundation Model-Agnostic Slide-Level Representation Learning (CVPR 2025 · similarity 0.90)
- Modeling Dense Multimodal Interactions Between Biological Pathways and Histology for Survival Prediction (CVPR 2024 · similarity 0.87)
- Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching (ICML 2025 · similarity 0.87)
- A Cross Modal Knowledge Distillation & Data Augmentation Recipe for Improving Transcriptomics Representations through Morphological Features (ICML 2025 · similarity 0.87)
- MERGE: Multi-faceted Hierarchical Graph-based GNN for Gene Expression Prediction from Whole Slide Histopathology Images (CVPR 2025 · similarity 0.86)
- Multimodal Prototyping for cancer survival prediction (ICML 2024 · similarity 0.86)
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.