AI and Cancer Research ── Cancer biology ── 2026-11-09
Cancer molecular biology and AI
- Abstract
OBJECTIVES: Chemotherapy-associated steatohepatitis (CASH) increases the risk of postoperative liver failure and mortality in patients undergoing liver surgery. However, non-invasive diagnosis remains an unmet need. Radiomics has potential for characterizing liver parenchyma. This study evaluated the performance of a virtual liver biopsy for CASH prediction, comparing it with whole-liver radiomic analysis. MATERIALS AND METHODS: Two retrospective series of patients undergoing hepatectomy for colorectal liver metastases were analyzed: the public "Colorectal-Liver-Metastases" dataset (training and internal validation) and a monocenter series from Humanitas Research Hospital (external validation). Inclusion criteria were preoperative chemotherapy (≥ 2 months) and post-chemotherapy CT (thickness ≤ 5 mm). IBSI-compliant radiomic features were extracted from whole-liver parenchyma and from a virtual biopsy (manually segmented 2-mL cylinder of non-tumoral liver parenchyma). Estimated liver volume (eLV) was computed using three orthogonal diameters. Four XGBoost models were developed (70/30 split, 5-fold cross-validation): clinical; clinical + VB-radiomics (ViB); clinical + VB-radiomics + eLV (ViBE+); clinical + whole-liver radiomics (WholeLiver). RESULTS: One hundred eighty patients were included: 105 internal (63 males, mean age 58.9 years, 38% CASH+) and 75 external (44 males, mean age 62.4 years, 19% CASH+). WholeLiver and ViBE+ models achieved excellent and comparable performance in internal validation (WholeLiver: AUC = 0.857 (95% BS-CI: 0.676-0.954), accuracy = 0.774 (95% BS-CI: 0.613-0.903); ViBE + : AUC = 0.831 (95% BS-CI: 0.549-0.956), accuracy = 0.839 (95% BS-CI: 0.710-0.935); p = 0.814), outperforming ViB model (AUC = 0.724 (95% BS-CI: 0.483-0.896), accuracy = 0.774 (95% BS-CI: 0.613-0.903)). SHAP analysis confirmed liver volume metrics as key predictors. In external validation, the ViBE+ model achieved AUC = 0.745 (95% BS-CI: 0.604-0.886) and accuracy = 0.680 (95% BS-CI: 0.562-0.783). CONCLUSION: Virtual biopsy provides a reproducible, non-invasive radiomic sampling strategy for CASH prediction. Combined with liver volume metrics, it achieves performance comparable to whole-liver analysis, supporting its potential clinical application. KEY POINTS: Question Despite its association with postoperative morbidity, the non-invasive preoperative assessment of CASH remains challenging in patients undergoing liver surgery.
Findings: Radiomic analysis based on a virtual liver biopsy, combined with volumetric measurements, showed excellent performance in predicting CASH, comparable to whole-liver analysis, with liver volume emerging as a key predictor. Critical Relevance Statement A standardized virtual biopsy-based radiomic analysis represents a promising non-invasive approach for evaluating liver parenchymal injury, as demonstrated in preoperative assessment of CASH and identification of high-risk surgical patients.
Journal IF-equivalent: 5.4 (OpenAlex 2-year mean citedness, value as of 2026-11-03, retrieved 2026-11-05; not the official Clarivate IF)Reference: Ammirabile A, Mastroleo F, De Martino R, Pugliese N, Waddle M, Laghi A, et al. Virtual liver biopsy for detecting steatohepatitis in chemotherapy-treated patients: comparison with whole-liver radiomics and external validation. Eur Radiol. 2026 Nov 7 [Epub ahead of print]. doi:10.1007/s00330-026-13004-6.Checked: Abstract only - Abstract
Objectives: Brain metastasis (BM) frequently occurs in advanced cancer patients, and accurate identification of the primary origin is essential for personalized treatment. Conventional imaging often lacks sensitivity and specificity for BM source prediction. This study aimed to develop and validate a multiparametric MRI-based radiomics model emphasizing brain–tumor interface (BTI) features to improve noninvasive BM origin identification. Materials and methods The primary cohort included 429 patients (740 BMs) from two centers, non-small cell lung cancer (NSCLC), small cell lung cancer (SCLC), breast cancer (BC), and gastrointestinal cancer (GIC), and an independent validation cohort of 33 patients (84 BMs) from a third center. All underwent pre-treatment 3.0 T MRI with T1-weighted contrast-enhanced (T1CE) and T2-weighted (T2W) sequences. Radiomic features were extracted from tumor core and BTI regions (2, 4, 6, 8 mm). Least absolute shrinkage and selection operator (LASSO) regression was used for feature selection, and radiomics signatures (RS) were evaluated using receiver operating characteristic (ROC) analysis.
Results: The constructed RS-BTI4 based on the 4-mm BTI region generated the best overall performance, with area under the ROC curve (AUCs) in the training cohort were 0.855 (LC/NLC), 0.910 (SCLC/NSCLC), and 0.890 (BC/GIC); in validation cohort, 0.726, 0.856, and 0.809; and in external validation cohort, 0.739, 0.750, and 0.792, respectively. RS-BTI4 consistently outperformed models based on other BTI radial widths and the whole tumor region in identifying the primary origin of BM.
Conclusion: BTI-derived radiomic features demonstrate strong potential for noninvasive identification of BM origin, with the 4-mm BTI region showing optimal predictive performance. Key Points Question Conventional MRI cannot reliably identify the primary origin of brain metastases, particularly when the primary tumor remains unknown at diagnosis.
Findings: The 4-mm brain–tumor interface model showed the most favorable overall performance, measured by area under the curve, across three binary classification tasks and validation cohorts. Critical relevance statement Integrating brain–tumor interface radiomics with multiparametric MRI may improve noninvasive estimation of BM origin and support more targeted diagnostic evaluation in clinical radiology.
Journal IF-equivalent: 6.5 (OpenAlex 2-year mean citedness, value as of 2026-11-03, retrieved 2026-11-05; not the official Clarivate IF)Reference: Jiang M, Han B, Huang Y, Rong X, Zhao W, Ge X, et al. Radiomics analysis of brain–tumor interface on multiparametric MRI for prediction of brain metastasis primary site. Insights Imaging. 2026 Nov 6;17(1):286. doi:10.1186/s13244-026-02464-5.Checked: Abstract only - Abstract
Protein post-translational modifications (PTMs), particularly phosphorylation, regulate cellular signaling, drug response, and disease-associated pathway activity. However, existing PTM resources are largely static and do not capture how PTM regulation varies across drug, dose, time, cellular context, and modified site. We present DrugPTM-Bench, a curated benchmark derived from decryptM dose-dependent PTM measurements, standardizing site-level drug response across seven cancer cell lines, 27 drugs, 11,167 proteins, six time points, 16 dosage levels, and pEC50 potency values. Reflecting the central role of phosphorylation in drug-responsive signaling, DrugPTM-Bench is phosphorylation-dominant, with 99.5% phosphorylation events. We formulate a protein-disjoint out-of-distribution classification task to predict whether a PTM site is up-regulated, downr-egulated, or unchanged following drug treatment. Our results show that drug-induced PTM regulation is a highly imbalanced and context-dependent prediction problem: baseline machine learning and deep learning models fail to robustly recover rare regulated events, while standard rebalancing strategies improve recall only at the cost of precision and overall F1-score. Beyond classification, DrugPTM-Bench enables pEC50 regression, mechanism-of-action prediction from PTM fingerprints, and drug-specific PTM-site sensitivity ranking.
Journal IF-equivalent: 9.8 (OpenAlex 2-year mean citedness, value as of 2026-11-08, retrieved 2026-11-09; not the official Clarivate IF)Reference: Badkul A, Xie L, Mottaqi M, Xie L. DrugPTM-Bench: A Large-Scale Dataset for Predictive Modeling of Drug-Induced Cell Type-Specific Protein Post-Translational Modifications. Information. 2026 Oct 31;17(11):1079. doi:10.3390/info17111079.Checked: Abstract only - Abstract
Our understanding of protein function and evolution is largely based on the relationship between amino acid sequence and overall fold, now effectively captured by computational models. Yet, predicting how mutations-shaped by epistasis-alter protein behavior, especially in dynamic or structurally ambiguous regions, remains difficult. Here, we present D2D, which combines a self-supervised protein language model with protein-specific evolutionary information, to predict mutational effects using little to no task-specific labeled data. D2D captures long-range epistatic interactions and accurately predicts single and higher-order mutation effects on protein thermostability and binding, without being trained on the task. When fine-tuned, D2D outperforms state-of-the-art methods on latent driver cancer mutations and co-occurring proliferation-enhancing mutations across independent experimental studies. Unlike most existing approaches, D2D avoids biases linked to solvent accessibility or to multiple sequence alignment depth and quality, making it particularly effective for disordered or surface binding regions where structure-based predictors typically falter. Overall, D2D provides a general framework for modeling mutational effects in proteins with limited experimental or structural information.
Journal IF-equivalent: 7.8 (OpenAlex 2-year mean citedness, value as of 2026-11-07, retrieved 2026-11-08; not the official Clarivate IF)Reference: Tzavella K, Olsen C, Vranken W. Evolutionary Constraints Improve Protein Large Language Model Predictions for Protein Stability, Binding Regions, and Epistasis. Comput Struct Biotechnol J. 2026 Sep 18;35(1):0226. doi:10.34133/csbj.0226.Checked: Abstract only