Transcriptome-Informed Multi-Modal AI Predicts Breast Cancer Treatment Response
結合轉錄體推論的多模態 AI：精準預測乳腺癌術前輔助治療反應
To overcome the scarcity of labeled oncology data, researchers developed a two-stage AI model. The first stage learns to infer transcriptome-wide gene expression directly from standard histopathology images (trained on 8,742 patients across 32 cancer types). The second stage uses these inferred expressions and clinical variables to predict pathological complete response (pCR) to neoadjuvant therapy. Evaluated on 1,412 patients across nine independent cohorts, the model achieved a pooled AUROC of 0.79, outperforming conventional histopathological biomarkers.
Key points
Two-Stage Architecture
First infers transcriptome from pathology slides, then combines it with clinical variables to predict pCR, bypassing labeled data scarcity.
Cross-Cohort Validation
Validated on 1,412 patients across 9 independent cohorts, demonstrating robust generalization with a pooled AUROC of 0.79.
Robustness on Minimal Tissue
Maintains highly stable predictions even with minimal biopsy tissue or despite intratumoral sampling variations.
No Gene Selection Constraints
Transcriptome-wide inference avoids the pre-selected gene constraints of standard genomic assays, enhancing clinical utility.
How it works
Why it matters
Traditional genomic assays are expensive, time-consuming, and constrained by tissue availability. This study demonstrates that 'biologically informed compression' allows AI to extract deep molecular features directly from routine, low-cost H&E pathology slides. By bypassing the need for physical genomic assays, this approach lowers the barrier to precision oncology and offers a scalable paradigm for other data-scarce cancer types.
Who it affects
- AI Researcher
- Enterprise Leader
- Product Manager
How to use it
- 1Predicting breast cancer patients' response to neoadjuvant therapy (e.g., chemotherapy, targeted therapy) to optimize personalized treatment plans.
- 2Serving as a low-cost, virtual transcriptome alternative in community hospitals or low-resource settings where physical sequencing is unavailable.
Limitations & caveats
- The transcriptome data is inferred by the AI model rather than directly measured, which may introduce inference bias.
- This is a retrospective study that requires further prospective clinical trial validation across more diverse patient demographics.
Related
Less Decoder is More Encoder: Extracting Robust 3D Geometric Representations via Novel View Synthesis
減少解碼器反而增強編碼器：從新視角合成中提煉強大三維幾何表徵
This paper reveals how expressive decoders dilute geometric learning in Novel View Synthesis, and proposes SNAP—a self-supervised framework that restricts decoders to force encoders to learn robust 3D representations.
4DCodeBench: Benchmarking AI Agents on 4D Inverse Graphics and Dynamic Scene Code Generation
4DCodeBench：評估 AI Agent 動態場景 4D 反向圖形學與程式碼生成能力的全新基準
4DCodeBench is a new benchmark designed to evaluate AI agents' ability to reconstruct 4D dynamic scenes from videos by generating executable graphics and physics code.
What Should World Models Forget? Stratified Retention for Continual Adaptation
世界模型該遺忘什麼？以「分層保留」實現持續適應環境的能力
This paper argues that world models must not avoid all forgetting, proposing 'stratified retention' to distinguish permanent physical laws from dynamic, environment-specific facts that require timely revision.