過去，乳腺癌通常被分為HR陽性HER2陰性、HR陽性HER2陰性、HR陰性HER2陽性、三陰性等若干大類進行分類治療，但是同一大類患者治療效果有好有壞、相差懸殊。隨后，同一大類乳腺癌又被進一步細分為若干亞型進行分型治療，但是同一亞型患者治療效果仍然有好有壞、相差懸殊。雖然這些分類分型越來越內卷，但是都將乳腺癌視為均勻物，其實大多數乳腺癌為混合物，可能由一部分HR陽性或陰性細胞、一部分HER2陽性或陰性細胞、一部分其他靶點陽性或陰性細胞組成，不同比例的多種惡性亞型細胞可能共存于同一腫瘤內，并影響療效，也就是說可以將乳腺癌細胞分類分型，但是難以將乳腺癌及其患者分為某一類某一型采用一刀切的治療方案。

2026年5月26日，歐洲腫瘤內科學會官方期刊《腫瘤學年鑒》在線發表美國國家癌癥研究所、西達賽奈醫學中心、以色列特拉維夫大學、拉賓醫學中心、韓國成均館大學、英國倫敦癌癥研究院、皇家馬斯登醫院、劍橋大學、中國臺灣成功大學的研究報告，利用人工智能技術，根據腫瘤基因轉錄組學和組織病理學，對乳腺癌進行反卷積分析，以量化其不同亞型組成比例，并預測術前治療效果。

該研究首先開發了計算機軟件BRIDGE對治療前混合腫瘤組織基因轉錄組數據進行解卷積，以估計不同分子亞型組成比例，并預測術前治療后的病理完全緩解。該研究采用10個轉錄組數據集進行訓練，并采用涵蓋不同亞型的24個獨立多中心數據集進行測試。此外，還分析了6個包含治療前蘇木精伊紅染色切片和治療效果數據的額外數據集，以評價根據腫瘤組織病理學的預測結果。

結果，通過對實測轉錄組學數據進行分析，BRIDGE對ER陽性HER2陰性乳腺癌術前治療病理完全緩解的預測能力優于現有商業化21基因、70基因、50基因檢測，并對HER2陽性和三陰性乳腺癌術前治療病理完全緩解的預測能力超越其他轉錄組學預測方法。

BRIDGE預測病理完全緩解真假陽性率曲線下面積：

• ER陽性HER2陰性乳腺癌：0.84（比值比：8）

• HER2陽性乳腺癌：0.77（比值比：8.3）

• 三陰性乳腺癌：0.73（比值比：3.1）

該研究隨后通過人工智能深度學習推斷轉錄組學，將BRIDGE用于治療前腫瘤組織病理切片，開發了BRIDGE切片版，性能優于直接的切片預測治療效果模型，凸顯其作為首個此類快速低成本生物標志物的潛力。

對采用不同術前治療方案的數據集進行探索性留一交叉驗證分析表明，該方法可推廣至接受免疫檢查點抑制劑治療的ER陽性HER2陰性腫瘤，但是仍然需要在更大規模的隊列中進行驗證。

最后，空間轉錄組學分析表明，BRIDGE亞型分類形成與典型分子特征相一致的空間聚集區域，進一步增強其生物學解釋能力。

因此，該研究結果表明，基于生物學原理的術前乳腺癌治療效果預測軟件BRIDGE已在大量不同患者隊列中得到驗證，其組織病理學切片版有望在術前治療中實現快速、低成本的治療效果預測，但是仍然需要進一步的前瞻測試和驗證，尤其對于中國大陸患者。

Ann Oncol. 2026 May 26. IF: 65.4

Predicting neoadjuvant breast cancer therapy response using BRIDGE from tumor transcriptomics and histopathology.

Cantore T, Hoang DT, Pal LR, Stemmer A, Chang TG, Dhruba SR, Shulman E, Campagnolo E, Lee JS, Levy J, Yao K, Liao I, Stemmer SM, Sammut SJ, Lipkowitz S, Rajagopal PS, Filipits M, Caldas C, Yuan Y, Nair NU, Ruppin E.

National Cancer Institute (NCI), National Institutes of Health (NIH), Bethesda, MD, USA; Cedars-Sinai Medical Center, Los Angeles, CA, USA; Tel Aviv University, Tel Aviv, Israel; Rabin Medical Center, Petah Tikva, Israel; Sungkyunkwan University, Suwon, Korea; The Institute of Cancer Research, London, UK; The Royal Marsden Hospital NHS Foundation Trust, London, UK; University of Cambridge, Cambridge, UK; Medical University of Vienna, Vienna, Austria; National Cheng Kung University, Tainan, Taiwan, China.

HIGHLIGHTS

• BRIDGE deconvolves transcriptomes of breast tumors to quantify their subtype composition and predict neo-adjuvant response.

• Validated across 24 independent multi-center cohorts, BRIDGE outperforms established molecular biomarkers.

• BRIDGE-Slide extends BRIDGE to pre-treatment histopathology by using deep learning to infer the tumor transcriptome.

BACKGROUND: While expression-based signatures inform adjuvant therapy in breast cancer (BC), no approved molecular biomarkers exist for the neoadjuvant setting, where early response prediction could inform treatment decisions. This challenge is compounded by intratumoral heterogeneity, as multiple malignant subtypes may coexist within a tumor and influence therapy sensitivity.

METHODS: We developed BRIDGE, a computational framework that deconvolves the pre-treatment bulk tumor transcriptome to estimate molecular subtype composition and predict pathological complete response (pCR) to neoadjuvant therapy. BRIDGE was trained on 10 transcriptomics datasets and tested on 24 independent ones spanning different sub-types. Six additional datasets with pre-treatment H&E slides and response data were analyzed to evaluate histology-based predictions.

RESULTS: Analyzing measured transcriptomics, BRIDGE outperformed surrogate implementations of established commercial signatures (Oncotype DX, MammaPrint, ROR-S) in ER+/HER2- tumors, and exceeded other transcriptomic predictors in HER2+ and TNBC disease. In ER+/HER2- patients, it yields an ROC-AUC of 0.84 with a high Odds Ratio (OR = 8); in HER2+ disease, an AUC of 0.77 (OR = 8.3); and in TNBC, an AUC of 0.73 (OR = 3.1). We further developed BRIDGE-Slide, which applies BRIDGE to pre-treatment histopathology slides via deep learning-inferred transcriptomics. BRIDGE-Slide outperforms direct slide-to-response models, underscoring its potential as a first-of-its-kind, fast, low-cost biomarker. Exploratory leave-one-dataset-out analyses across datasets treated with alternative neoadjuvant regimens suggest generalizability to ICB-treated ER+/HER2- tumors, pending validation in larger cohorts. Finally, spatial transcriptomics shows that BRIDGE-derived subtype assignments form spatially cohesive regions aligned with canonical molecular features, reinforcing its biological interpretability.

CONCLUSIONS: BRIDGE is a biologically grounded framework for neoadjuvant BC response prediction, validated on a rich set of different patients' cohorts. Its histopathology-based version opens the door for fast and low-cost prediction in the neoadjuvant setting, upon further prospective testing and validation.

KEYWORDS: Breast Tumors, gene expression, PRECISION MEDICINE, AI in Oncology

DOI: 10.1016/j.annonc.2026.05.700

來源：SIBCS

免責聲明：本公眾號旨在為醫療衛生專業人士提供腫瘤領域資訊參考，促進專業交流。內容僅供信息交流，不可替代專業醫療指導，亦非診療建議。 據此操作，風險自擔。轉載內容均已標注來源及作者。如有版權問題，請留言聯系，我們將及時處理。