Accurate prognosis prediction is pivotal for personalized cancer treatment. While pre-treatment gene expression data offers significant potential, existing predictive models are predominantly data-driven and often neglect critical biological prior knowledge, such as gene-gene interactions and adjacent tissue expression patterns, thereby limiting interpretability and generalizability. To address this, we propose KEGnet, a knowledge-enhanced graph attention framework that systematically incorporates biological priors into both feature selection and model prediction. KEGnet comprises two core components: (1) a knowledge-guided feature screening module (KGATV2) that leverages protein-protein interaction (PPI) network and tumor-normal expression contrasts to identify task-specific gene signatures; and (2) a prediction pipeline that integrates a stacked graph attention network and XGBoost, unified via logistic regression. Applied to key prognostic tasks in two major cancers, the breast cancer (BC) signature PNAC50 identified by KEGnet demonstrated superior performance to traditional signatures (OncotypeDX, PAM50, and HER2DX) in predicting pathological complete response (pCR) across six public datasets$(\mathrm{n}=1,316)$. Furthermore, leveraging prior lung adenocarcinoma (LUAD) signatures, KEGnet delivered more robust predictions of recurrence risk in a private LUAD dataset ($\mathbf{n}=\mathbf{1 1 9}$) compared to conventional approaches. Notably, KEGnet also demonstrates superior clinical and biological interpretability. Altogether, by fusing expression data with biological knowledge, KEGnet not only enhances prediction performance but also facilitates the discovery of high-impact gene signature biomarkers with potential.

MOTIVATION: Endometrial cancer is a prevalent gynecological malignancy that requires accurate identification of its molecular subtypes for effective diagnosis and treatment. Four molecular subtypes with different clinical outcomes have been identified: POLE mutation, mismatch repair deficient, p53 abnormal, and no specific molecular profile. However, determining these subtypes typically relies on expensive gene sequencing. To overcome this limitation, we propose a novel method that utilizes hematoxylin and eosin-stained whole slide images to predict endometrial cancer molecular subtypes. RESULTS: Our approach leverages a hierarchical foundation model as a backbone, fine-tuned from the UNI computational pathology foundation model, to extract tissue embedding from different scales. We have achieved promising results through extensive experimentation on the Fudan University Shanghai Cancer Center cohort (N = 364). Our model demonstrates a macro-average AUROC of 0.879 (95% CI, 0.853-0.904) in a five-fold cross-validation. Compared to the current state-of-the-art molecular subtypes prediction for endometrial cancer, our method outperforms in terms of predictive accuracy and computational efficiency. Moreover, our method is highly reproducible, allowing for ease of implementation and widespread adoption. This study aims to address the cost and time constraints associated with traditional gene sequencing techniques. By providing a reliable and accessible alternative to gene sequencing, our method has the potential to revolutionize the field of endometrial cancer diagnosis and improve patient outcomes. AVAILABILITY AND IMPLEMENTATION: The codes and data used for generating results in this study are available at https://github.com/HaoyuCui/hi-UNI for GitHub and https://doi.org/10.5281/zenodo.14627478 for Zenodo.

MOTIVATION: Differentiating 12 stages of the mouse seminiferous epithelial cycle is vital towards understanding the dynamic spermatogenesis process. However, it is challenging since two adjacent spermatogenic stages are morphologically similar. Distinguishing Stages I-III from Stages IV-V is important for histologists to understand sperm development in wildtype mice and spermatogenic defects in infertile mice. To achieve this, we propose a novel pipeline for computerized spermatogenesis staging (CSS). RESULTS: The CSS pipeline comprises four parts: (i) A seminiferous tubule segmentation model is developed to extract every single tubule; (ii) A multi-scale learning (MSL) model is developed to integrate local and global information of a seminiferous tubule to distinguish Stages I-V from Stages VI-XII; (iii) a multi-task learning (MTL) model is developed to segment the multiple testicular cells for Stages I-V without an exhaustive requirement for manual annotation; (iv) A set of 204D image-derived features is developed to discriminate Stages I-III from Stages IV-V by capturing cell-level and image-level representation. Experimental results suggest that the proposed MSL and MTL models outperform classic single-scale and single-task models when manual annotation is limited. In addition, the proposed image-derived features are discriminative between Stages I-III and Stages IV-V. In conclusion, the CSS pipeline can not only provide histologists with a solution to facilitate quantitative analysis for spermatogenesis stage identification but also help them to uncover novel computerized image-derived biomarkers. AVAILABILITY AND IMPLEMENTATION: https://github.com/jydada/CSS. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.