一种基于原发肿瘤转录组学和组织病理形态学预测前列腺癌转移的多模态模型

Prostate cancer (PCa) is the most commonly diagnosed non-cutaneous malignancies in men in the United States, and metastatic dissemination remains the principal driver mortality. While advanced disease is typically managed with androgen deprivation therapy, most patients eventually progress to castration-resistant PCa, underscoring the need for early identification of tumors predisposed to metastasis. Clinicopathological metrics, including PSA, tumor stage, and Gleason grade, provide only coarse estimates of risk and fail to capture the molecular and spatial heterogeneity underlying disease progression. To address this limitation, we developed Met-Score , a transcriptomic signature derived from meta-analysis of primary tumor expression profiles from 1,239 PCa patients. Using a rigorous training (n=1000) and independent validation (n=239) design, we computed gene-level Hedges' effect sizes across cohorts, pooled them using a random-effects model, integrated evidence across datasets using Fisher's and log-sum methods, and applied false discover rate correction to identify genes most strongly associated with metastatic progression. In the validation cohort, Met-Score achieved an AUC of 0.72 for predicting metastasis, and maintained independent prognostic value for overall, metastasis-free, and progression-free survival across two independent datasets, even after adjustment for Gleason score, demonstrating its clinical relevance beyond standard pathology. To investigate whether Met-Score reflects morphological phenotypes encoded in routine histopathology, we quantified morphometric features from digitized H&E slides of prostate biopsies and prostatectomies. Met-Score showed significant associations with multiple architectural and cytological descriptors, suggesting that the signature captures transcriptional programs tightly coupled to microanatomical patterns observable on standard tissue sections. Building upon this link between molecular and morphological readouts, we trained a multimodal risk model that integrates Met-Score with an image-derived Morph-Score , and benchmarked its performance against unimodal predictors. Our preliminary analysis indicates that the multimodal framework substantially enhances risk discrimination and more accurately identifies patients likely to develop metastasis compared to transcriptomics or histopathology alone. This work highlights the potential of combining deep molecular features with interpretable morphometric signatures to improve prognostication in localized PCa, and establishes a broadly applicable strategy for multimodal biomarker development in solid tumors.