Find the papers that actually matter

Developing effective cancer treatments requires identifying novel therapeutic agents with high biological activity. Fucoidan, a sulfated polysaccharide from brown algae, is a promising natural scaffold for anticancer drug development. In this study, oversulfated fucoidans (SFU and SFU-1) were derived from natural fucoidans (FU and FU-1), and the effects of this structural modification on anticancer efficacy were investigated comprehensively. Chemical characterizations of FU/FU-1 and SFU/SFU-1 were performed, and a systematically generated experimental dataset across multiple cancer cell lines was compiled. Using these data, several machine learning (ML) algorithms were applied to predict the anticancer efficacy of fucoidan-based molecules. Specifically, k-Nearest Neighbors Regression (kNNR), Least-Squares Boosting (LSBoost), Support Vector Regression (SVR), Decision Tree Regression (DT), Random Forest Regression (RFR), Linear Regression (LR), and Gaussian Process Regression (GPR) were evaluated with five-fold cross-validation. Model performance was assessed utilizing the Mean Absolute Error (MAE), the Root Mean Square Error (RMSE), and the Coefficient of Determination (R2). The results show that FT-IR analysis of the oversulfated derivatives, SFU and SFU-1, confirmed successful modification. For SFU, the appearance of a shoulder peak at 820 cm-1 (equatorial C-2), alongside the characteristic 840 cm-1 peak, verified site-specific sulfonation. In SFU-1, new peaks at 1243 cm-1 (SO stretching) and 583 cm-1 (OSO deformations) were identified. These spectral changes demonstrate the effective integration of sulfate groups into the molecular frameworks. Cytotoxicity assays against five human cancer cell lines revealed dose-dependent inhibition, with the SFU derivative exhibiting the most potent activity, particularly reducing HeLa and MDA-MB-231 cell viability to 23.93% and 25.13% at 2 mg/mL. GPR model achieves superior predictive performance compared to other methods, with the lowest MAE and RMSE (8.4627 and 11.5692, respectively) and the highest R2 (0.7039) values. The findings reveal that models that capture the nonlinear relationship between sulfation degree and anticancer efficacy, especially GPR, are powerful tools for the preliminary evaluation of natural product-based drug candidates. This study demonstrates that integrating chemical modification, experimental validation, and classical ML can accelerate the rational assessment of naturally derived therapeutics in oncology.

Tumor-infiltrating lymphocytes (TILs) are an important indicator of immune activity in breast cancer, yet scoring them consistently on H&E slides remains challenging in routine pathology. This work presents a modular deep learning pipeline that delivers fully automated and continuous stromal TILs (sTILs) scores in line with the International Immuno-Oncology Biomarker Working Group (IIOBWG) guidelines. The pipeline combines three components: a TIL segmentation model refined through pathologist-guided active learning, a robust stroma segmentation network based on an enhanced DeepLabV3+, and a lightweight regression module that learns how TILs distribute within stromal regions. A new adaptive aggregation strategy integrates patch-level predictions into a single, clinically meaningful score while accounting for heterogeneous infiltration. The system was evaluated on two independent datasets (60 and 112 WSIs) with expert-annotated ROIs, achieving strong agreement with pathologists (Pearson of 0.814; ICC of 0.808). Importantly, the pipeline is interpretable: each stage produces human-readable outputs (stroma masks, TIL-in-stroma maps), and SegGradCAM visualizations confirm that predictions rely on biologically relevant tissue regions. These findings demonstrate the pipeline's potential as a reliable and clinically adaptable tool for standardized, fully automated TILs quantification in breast cancer pathology. The source code and pretrained models are publicly available at https://github.com/Shrief-Abdelazeez/TILs-Scoring.

Computed tomography (CT) is essential for finding and diagnosing kidney tumors and cysts because good lesion segmentation enables accurate diagnosis, appropriate therapy planning, and disease monitoring. Renal mass (Tumor and cyst) shapes and sizes are diverse and complex, particularly around diffuse borders, and low-intensity contrast and heterogeneous morphology make effective segmentation difficult. To overcome these challenges, introduced internal dilated guided self-attention model based on dilated residual inception U-Net (IDGSA-DRIU-Net), a deep learning model for segmenting renal tumors and cysts. The architecture features a newly built Internal Dilated Guided Self-Attention (IDGSA) module, which combines Dilated Multiscale Position Attention and Channel Attention introducing an attention mechanism that leverages internal guidance to capture spatial dependencies across multiple dilation scales and emphasizes key feature channels, allowing for more effective utilization of local and global contextual information for better multi-scale feature aggregation. A Dilated Residual Inception (DRI) module improves multiscale contextual feature recovery while preserving structural features. In post-processing, a Conditional Random Field (CRF) is utilized to eliminate false positives and refine bounds. The proposed model was evaluated using the KiTS19, KiTS21, and KiTS23 datasets and obtained a Dice Similarity Coefficient (DSC) of 98.27% for the tumor on KiTS19, 96.31% and 94.82% for the tumor and cyst on KiTS21, and 94.98% and 93.64% for the tumor and cyst on KiTS23. The results suggest that IDGSA-DRIU-Net with CRF outperforms the popular state-of-the-art models, indicating that successful in kidney tumor and cyst segmentation.

Treatment outcome prediction plays an important role in realizing personalized cancer therapy. In triple-negative breast cancer (TNBC), neoadjuvant chemotherapy (NAC) is widely used to downstage tumors and improve surgical outcomes. In head and neck cancer (HNC), early prediction of lesion progression can assist treatment planning. However, inter-patient heterogeneity in treatment response and tumor behavior limits the effectiveness of generalized treatment strategies. To address this issue, we developed an evidential reasoning rule-enabled deep neural network (ER2-DNN) for reliable outcome prediction in cancer therapy. The ER2-DNN combines convolutional neural network (CNN) based image feature extraction with data augmentation, Monte Carlo dropout, test-time augmentation and evidential reasoning rule (ER2) fusion for generating uncertainty-aware prediction. Across both TNBC and HNC datasets, the model showed consistent predictive performance with well-calibrated confidence estimates. The ER2-DNN provides a framework for supporting individualized oncology decisions through reliable image-based modeling.

To develop and validate a contrast-enhanced CT (CECT) habitat radiomics model focusing on assessing independent predictive contribution of each intratumoral subregion to improve prediction of lymph node metastasis (LNM) in esophageal squamous cell carcinoma (ESCC). This retrospective study included 246 consecutive patients with confirmed ESCC undergoing preoperative CECT from two centers. Patients from Center 1 (n = 194) were randomly divided into training (n = 136) and internal-validation (n = 58) cohorts, and an external-validation cohort comprised 52 patients from Center 2. Conventional radiomics features were extracted from the whole-tumor. Habitat radiomics features were from three habitat subregions using K-means clustering. The selected core features were used to develop the corresponding support vector machine (SVM) classifiers followed with logistic regression (LR), k-nearest neighbors (KNN), and Light Gradient Boosting Machine (LightGBM) classifiers. Predictive performance of models was compared using area under the receiver operating characteristic curve (AUC). For SVM classifiers, habitat 3 (with high-entropy) based model showed better predictive performance than the conventional whole-tumor model, and habitats 1 and 2 based models (AUC: 0.925 vs. 0.911, 0.865 and 0.882; 0.898 vs. 0.871, 0.741 and 0.781; and 0.844 vs. 0.754, 0.744 and 0.812 in the training, internal-validation and external-validation cohorts, respectively, all p-values < 0.05). Using the same preprocessing, three types of other classifiers, including LR, KNN and LightGBM, reproduced superiority of habitat 3 with consistent predictive performance, supporting algorithms' robustness. Habitat with high-entropy based models provided the strongest performance for predicting LNM in ESCC, with SVM performing best and outperforming the whole-tumor models.

Nanozymes, engineered nanomaterials with enzyme-mimetic activities, have emerged as versatile platforms for ferroptosis-based cancer theranostics. Ferroptosis, an iron-dependent form of regulated cell death driven by lipid peroxidation, has emerged as a promising strategy to overcome resistance to conventional cancer therapies. By catalyzing redox reactions, nanozymes can generate reactive oxygen species (ROS) and promote ferroptotic lipid peroxidation, thereby triggering cell death in tumor cells that evade apoptosis-based treatments. In parallel, non-redox activities of nanozymes, including hydrolase- and phosphatase-like functions, enable them to remodel the tumor microenvironment (TME), modulate biomolecular signaling, and support targeted therapy. This review provides a systematic and design-oriented overview of nanozymes that interface with ferroptosis. We summarize how redox and non-redox nanozyme activities converge on key ferroptosis-related processes, such as ROS production, glutathione depletion, iron metabolism disruption, and TME regulation. We then highlight rational engineering strategies, including single-atom and multimetallic catalytic centers, biodegradable coordination frameworks, stimuli-responsive architectures, and protein corona engineering, that enhance catalytic specificity, tumor targeting, and biosafety. Theranostic implementations are discussed with emphasis on multimodal imaging-guided platforms and combination regimens that integrate chemotherapy, radiotherapy, phototherapy, and immunotherapy. Finally, we outline major translational challenges and future opportunities, including AI and computation-guided nanozyme design and adaptive, corona-informed systems tailored for personalized cancer therapy. This review aims to serve as a roadmap for developing clinically translatable nanozymes that unify diagnosis and treatment through ferroptosis-oriented precision oncology.

Lung adenocarcinoma (LUAD) is the predominant pathological subtype of non-small cell lung cancer. Its considerable tumor heterogeneity and drug resistance present major clinical obstacles, resulting in unfavorable patient outcomes. Protein palmitoylation is known to be a key factor in tumorigenesis; however, its cell-specific expression patterns and prognostic value in LUAD remain incompletely characterized. Using two independent datasets, TCGA-LUAD and GSE68465, and scRNA-seq data from 10 LUAD samples, we analyzed the expression of palmitoylation-related genes. Through single-cell clustering, CNV analysis, and palmitoylation activity scoring, malignant epithelial cells were identified. 10 machine learning algorithms were applied to construct prognostic models based on differentially expressed genes. RT-qPCR was used to detect mRNA expression of prognostic marker genes in clinical samples. In vitro experiments validated SEC61G's role in regulating drug sensitivity. A subset of malignant epithelial cells with high palmitoylation activity was identified. A 5-gene signature (UBE2S, SEC61G, CCT6A, GAPDH, HLA-DRA) was established by the integrated CoxBoost+SuperPC method, showing robust predictive efficacy in both GSE68465 and TCGA-LUAD. High-risk samples carried higher mutation burden, greater genomic heterogeneity, and a stronger tumor immunosuppressive microenvironment than the low-risk group. Clinical sample testing revealed upregulation of UBE2S, SEC61G, CCT6A, and GAPDH in LUAD patients and downregulation of HLA-DRA. SEC61G expression inversely correlated with AZD3759 sensitivity. In vitro, SEC61G knockdown or AZD3759 alone suppressed LUAD proliferation and induced apoptosis; no synergy was observed with combination therapy, indicating that SEC61G modulates AZD3759 sensitivity in LUAD cells. Our study comprehensively reveals the cellular heterogeneity of palmitoylation, establishes a robust palmitoylation-related prognostic model, and identifies SEC61G as a promising therapeutic target in LUAD, offering a novel perspective for LUAD precision stratification and treatment studies.

To investigate the potential risk of nicotine exposure on glioblastoma multiforme (GBM). A two-sample Mendelian randomization (MR) approach was employed to assess the causal association between serum cotinine levels, a metabolite of nicotine, and GBM. GBM datasets were obtained from the GEO database, and differentially expressed genes were analyzed. Nicotine-related targets were screened using databases such as BATMAN and CTD, followed by gene ontology (GO), Kyoto encyclopedia of genes and genomes (KEGG), and protein interaction analyses for core targets. Machine learning, GSEA analysis, GEPIA2, HPA database, and CNGB single-cell sequencing database were employed to screen and validate core targets. CB-Dock2 and iMODS were used for molecular docking and structural dynamics analysis to validate screening results. MR analysis revealed a causal relationship between serum cotinine and GBM. Network toxicology analysis identified 194 potential target genes, with KEGG analysis indicating pathways related to viral infection, immunity, cancer, and metabolism. Machine learning identified core targets including NFκB1, HIF1α, CD4, FN1, MMP2, and GSK3β, whose mechanisms were validated through GSEA, GEPIA2, HPA and CNGB databases, molecular docking, and structural dynamics analysis. This study employed multiple methodologies to investigate the association between genetically predicted serum cotinine levels and GBM risk, identifying core targets including NFκB1, HIF1α, CD4, FN1, MMP2, and GSK3β. These findings provide novel insights for future research on the association between nicotine-related exposure and GBM risk.

Artificial intelligence-based radiomics offers a potential adjunct to the current clinical management of paediatric brain tumours by enabling prediction of key diagnostic features. However, despite promising research, integration into clinical practice remains limited. This may in part reflect limitations in the underlying evidence base. Previous research has identified inconsistent methodological quality in primary studies and systematic reviews, as well as inadequate risk of bias assessment of primary studies within reviews. Given these concerns, rigorous appraisal of the available evidence base is essential to determine whether current findings are reliable and suitable for clinical translation. This study aimed to characterise the methodological appraisal approaches used in systematic reviews of artificial intelligence-based radiomics research in paediatric brain tumour imaging, with particular attention to how review authors select and apply tools for risk of bias, methodological quality, and reporting completeness. We performed a meta-research study of systematic reviews. PubMed, Embase, Web of Science, Scopus and Medline were searched in March 2024 without date limits. Study selection followed PRISMA 2020 guidelines. Reviews of artificial intelligence-based radiomics applications in predominately paediatric populations with primary brain tumours were included. We evaluated the methodological appraisal approaches used, including tools assessing risk of bias, methodological quality and reporting completeness. Seven systematic reviews met inclusion criteria. Among these, three reviews (3/7) did not employ any formal methodological appraisal tool, and only one provided a rationale for this omission. Four reviews (4/7) applied at least one formal appraisal tool using five distinct instruments in total. Two (2/5) of these instruments were designed specifically for artificial intelligence-based research. The Quality Assessment of Diagnostic Accuracy Studies 2 was the most used tool, applied across two reviews. Methodological appraisal is applied inconsistently in systematic reviews of artificial intelligence-based radiomics research in paediatric neuro-oncology. This may reduce confidence in the current evidence base and hinder clinical translation. While some reviews used formal appraisal tools, these were usually conventional tools that may not fully capture methodological considerations specific to artificial intelligence-based research. Greater use of consistent, transparent, and artificial intelligence-aware appraisal approaches, alongside improved reporting in primary studies, is needed to support more reliable evidence synthesis and translation of radiomics into clinical practice.

Immune checkpoint inhibitors (ICIs), especially PD-1/PD-L1 blockade, have transformed cancer therapy; yet objective response rates to anti-PD-(L)1 monotherapy remain ∼20-30% and resistance is common. Meanwhile, multi-omics, spatial profiling, imaging, and clinical datasets are expanding faster than our ability to extract mechanistic insight, creating a "data-rich but mechanism-poor" bottleneck in immuno-oncology. Conventional biomarkers such as tumor mutational burden and PD-L1 expression lack spatiotemporal resolution, while purely data-driven artificial intelligence models often suffer from limited causal interpretability and black-box behavior. To address these challenges, Biologically Informed Machine Learning (BIML) offers a new paradigm by embedding biophysical principles (e.g., pharmacokinetic ordinary differential equations) and biological priors (e.g., protein-protein interaction networks) into predictive models. Recent applications of BIML have enabled integrative decoding of tumor immune microenvironment heterogeneity, quantitative characterization of T-cell exhaustion dynamics, and identification of spatial barriers such as fibroblast-mediated immune exclusion. Crucially, to bridge the translational gap between computational inference and clinical reality, we emphasize the mandatory integration of orthogonal ex vivo validation (e.g., patient-derived organoids and microphysiological systems). Ultimately, by transforming static spatial snapshots into testable dynamic trajectories, this computation-experiment closed-loop aims to generate actionable insights and prioritize rational combination strategies safely under expert oversight.