Cellular deconvolution algorithms virtually reconstruct tissue composition by analyzing the gene expression of complex tissues. We present the decision tree machine learning algorithm, Kassandra, trained on a broad collection of >9,400 tissue and blood sorted cell RNA profiles incorporated into millions of artificial transcriptomes to accurately reconstruct the tumor microenvironment (TME). Bioinformatics correction for technical and biological variability, aberrant cancer cell expression inclusion, and accurate quantification and normalization of transcript expression increased Kassandra stability and robustness. Performance was validated on 4,000 H&E slides and 1,000 tissues by comparison with cytometric, immunohistochemical, or single-cell RNA-seq measurements. Kassandra accurately deconvolved TME elements, showing the role of these populations in tumor pathogenesis and other biological processes. Digital TME reconstruction revealed that the presence of PD-1-positive CD8⁺ T cells strongly correlated with immunotherapy response and increased the predictive potential of established biomarkers, indicating that Kassandra could potentially be utilized in future clinical applications.

细胞去卷积算法通过分析复杂组织的基因表达来虚拟地重建组织组成。我们展示了决策树机器学习算法 Kassandra，该算法对广泛收集的 >9,400 个组织和血液分选的细胞 RNA 谱进行了训练，这些 RNA 谱被整合到数百万个人工转录组中，以准确重建肿瘤微环境 (TME)。对技术和生物变异性、异常癌细胞表达包含以及转录表达的准确量化和标准化的生物信息学校正增加了 Kassandra 的稳定性和稳健性。通过与细胞计数、免疫组织化学或单细胞 RNA-seq 测量进行比较，对 4,000 个 H&E 载玻片和 1,000 个组织的性能进行了验证。Kassandra 准确地解卷积了 TME 元素，显示这些群体在肿瘤发病机制和其他生物学过程中的作用。数字 TME 重建显示存在 PD-1 阳性 CD8⁺ T 细胞与免疫治疗反应密切相关，并增加了已建立生物标志物的预测潜力，表明 Kassandra 可能在未来的临床应用中得到利用。