AI驱动的三维空间转录组学用于前列腺腺癌的三维肿瘤微环境图谱绘制

Prostate adenocarcinoma (PRAD) is a multifocal and highly heterogeneous malignancy, marked by the coexistence of multiple Gleason grade glands within the same tumor microenvironment (TME). Spatial Transcriptomics (ST) has recently emerged as a powerful technology for characterizing the TME, yet its application is largely limited to two-dimensional (2D) histological sections, which typically capture less than 0.1% of the three-dimensional (3D) tumor context available upon resection. While promising, recently proposed in-situ approaches for 3D ST remain confined to reduced patient cohorts due to lengthy processing times and substantial costs. To enable scalable 3D morphomolecular analysis of PRAD TME, we devise an AI framework that obtains 3D spatial molecular maps in a cost-effective and efficient manner. Our framework, VORTEX (Volumetrically Resolved Transcriptomics EXpression), leverages 3D high-resolution tissue morphology from 3D pathology imaging modalities, and minimal 2D ST to predict 3D ST. By pretraining on diverse 3D morphology-transcriptomic pairs from heterogeneous tissue samples and then fine-tuning on minimal 2D ST data from a specific volume of interest, VORTEX captures both generic tissue-related and sample-specific morphological correlates of gene expression. Our framework leverages pathology and single-cell foundation models, and integrates a cross-modal registration pipeline for accurate learning of morphomolecular links. To evaluate our approach, we apply VORTEX to a cohort of 23 3D pathology volumetric images of PRAD specimens across 17 patients, acquired with micro computed tomography (microCT, 11 volumes) and open-top light-sheet microscopy (OTLS, 12 volumes). We additionally collect 88 sections of Visium ST across these samples and public cohorts, resulting in 243,682 spots with corresponding morphology. We demonstrate that VORTEX accurately predicts 3D ST and we identify two major trends: incorporating 3D morphology enhances the ability to learn morphomolecular links compared to 2D morphology alone, and including 2D ST data from the volume of interest further improves the performance by accounting for inter-patient heterogeneity. We observe that VORTEX captures intra-tumoral and inter-tumoral heterogeneity, with genes such as AZGP1 or GLO1 showing different expression profiles across patients and Gleason Grades. Furthermore, by analyzing 3D ST from multiple genes through Hallmark's pathways, we identify hidden structures in 2D views, including the 3D invasive tumor front. In summary, VORTEX, by leveraging AI, 3D tissue morphology and 2D ST, generates 3D spatial molecular maps in a reliable, efficient and scalable manner. The combined morphomolecular analysis of the 3D tumor context can provide a novel perspective of the TME for improved characterization of PRAD heterogeneity.