AI and Cancer Research ── Cancer biology ── 2026-09-30


Cancer molecular biology and AI
- Abstract
Tissues are complex ecosystems organised in space, and alterations in this organisation underpin multiple diseases. Spatial omics enables molecular profiling of tissue organisation, but linking these patterns to clinical outcomes remains challenging. We present SpaCEy (Spatial Clinical Explainability), an explainable graph neural network that identifies tissue patterns predictive of clinical outcomes in spatial proteomics datasets. SpaCEy models tissues as spatial graphs from molecular marker expression, without using predefined cell-type labels or anatomical regions as model inputs. Its embeddings capture intercellular relationships and molecular dependencies for predicting overall survival and disease progression. An integrated explainer identifies recurring spatial patterns and coordinated marker expression relevant to model predictions. Applied to a spatial proteomic lung cancer cohort, SpaCEy identifies spatial and protein-expression patterns associated with disease progression. Across multiple breast cancer proteomic datasets, it stratifies patients by overall survival, both across and within established clinical subtypes, and highlights protein markers underlying this stratification.
Reference: Rifaioglu AS, Ervin EH, Sarigun A, Germen D, Bodenmiller B, Tanevski J, et al. SpaCEy links spatial tissue patterns to clinical outcomes using explainable graph neural networks. Nat Commun. 2026 Sep 26;17(1):10232. doi:10.1038/s41467-026-77924-z. PMID: 42800816; PMCID: PMC13615965.Checked: Full text checked - Abstract
Predicting drug efficacy across diverse patient contexts remains a major challenge in oncology, as models trained on cancer cell lines often fail to capture patient-specific biology. Emerging biological foundation models and patient-derived technologies offer a promising solution. Here, we present UniCure, a multi-modal model that combines biological and chemical foundation models to predict drug-induced transcriptomic responses across diverse cell and tissue contexts, enabling individualized drug ranking. Trained on 1.9 million transcriptomic perturbation profiles spanning >22,000 compounds, 166 cell types, and 24 tissues, UniCure accurately predicts dose-dependent and combination responses and generalizes across bulk and single-cell data. We further fine-tune UniCure on 345 patient-derived tumor-like cluster (PTC) transcriptomic profiles and validate performance on 396 real-world clinical profiles, demonstrating effective patient-level prediction. The model supports response-based patient stratification and is experimentally validated in cell line and patient-derived models. Overall, UniCure provides a practical framework for translating preclinical data into personalized therapeutic strategies.
Reference: Chen Z, Tian S, Pei J, Gu R, Li Y, Ding S, et al. UniCure: A multi-modal model for predicting personalized cancer therapy response. Cancer Cell. 2026 Sep [Epub ahead of print]. doi:10.1016/j.ccell.2026.07.010. PMID: 42716007.Checked: Abstract only - Abstract
Artificial intelligence-empowered virtual cell models represent an emerging approach for in silico drug discovery 1-3 , yet most existing approaches lack large-scale, time-resolved perturbation proteomics data and interpretable frameworks for predicting therapeutic responses. Here we generated more than 38 million temporal protein-abundance measurements from systematically perturbed breast cancer cell lines, and developed ProteinTalks, a virtual cell model. Central to ProteinTalks is the synergy of this large-scale dynamic proteomic resource and the model architecture, enabling a new pretraining framework that learns transferable dynamical latent representations from temporal proteome trajectories. By modelling how proteins respond conditionally to different perturbations, this approach enables the model to function as an operational tool for diverse drug discovery tasks: predicting drug efficacy and synergy, discovering new drug combinations, probing proteins associated with drug resistance, stratifying patient responses and prioritizing drug candidates for patient organoids. It also shows robust transferability, extending beyond cell lines to patient-derived organoids and clinical biopsies, generally achieving higher performance than the selected benchmark implementations under the evaluated protocols. Together, ProteinTalks shows how scalable pretraining of transferable dynamic representations enables operational, dynamics-aware, proteomics-based virtual cell models to advance in silico drug discovery.
Reference: Sun R, Qian L, Li Y, Liu T, Cheng H, Zhang X, et al. An operational perturbation proteomics-based virtual cell model. Nature. 2026 Sep 9 [Epub ahead of print]. doi:10.1038/s41586-026-11001-9. PMID: 42717098.Checked: Abstract only - Abstract
Major Finding: A pan-cancer foundation model predicts response to immunotherapy from bulk transcriptomic data. Concept: COMPASS takes transcriptomic input, routes through human-readable concepts, and predicts immunotherapy outcomes. Impact: COMPASS’ concept bottleneck architecture substantially improves outcome prediction while providing biological insight.
Reference: [No authors listed]. Immune Concept–Informed AI Improves Immunotherapy Response Prediction. Cancer Discov. 2026 Aug 31;16(9):OF1-OF1. doi:10.1158/2159-8290.cd-rw2026-078. PMID: 42669647.Checked: Abstract only - Abstract
Copy number variations (CNV) are key drivers of cancer progression, yet methods for predicting spatial CNVs directly from haematoxylin and eosin (H&E) images are currently lacking. We introduce HisToSpatialCNV, a multiscale deep-learning framework that integrates histopathological features from H&E-stained images to infer spatial CNV patterns. Combining interpretable feature extraction with graph neural networks and multihead self-attention, our approach captures both local and global tissue contexts. Applied to HER2+ breast cancer, skin cancer and brain cancer datasets, HisToSpatialCNV outperformed existing methods for spatial gene expression inference and showed strong concordance between predicted CNVs and gene expression. In addition, it enabled tumour subclone identification, phylogenetic reconstruction and detection of pathway alterations linked to tumour progression. HisToSpatialCNV generalized across Visium and Xenium spatial transcriptomics platforms, on HER2+ patients. Applying the HisToSpatialCNV HER2+ model to TCGA HER2+ histopathology data identified spatial-molecular subtypes associated with distinct survival outcomes in TCGA HER2+ patients. By connecting histopathology to spatial genomics, HisToSpatialCNV offers a powerful, cost-effective tool for studying intratumour heterogeneity using only routine pathology images.
Reference: Chen T, Unjitwattana T, Guo X, Thakur P, Zhou S, Jing Z, et al. HisToSpatialCNV: an interpretable deep learning method predicting spatial copy number variations from histopathology images. Nat Biomed Eng. 2026 Aug 24 [Epub ahead of print]. doi:10.1038/s41551-026-01754-z. PMID: 42637837.Checked: Abstract only - Abstract
Experimental validation and functional optimization remain bottlenecks in AI-based protein design. We present a scalable workflow for developing AI-designed minibinders against cancer-associated surface proteins. Screening thousands of designs using mammalian cell-surface display identifies several high-affinity PD-L1 minibinders but far fewer for CD276 (B7-H3) and VTCN1 (B7-H4), highlighting substantial target dependence. Interface predicted template modeling (ipTM) scores generated by Chai-1 with ESM embeddings correlate with binding success and capture deleterious effects of interface mutations. Fluorophore-labeled AI-minibinders enable flow-cytometric staining comparable to conventional antibodies. However, when incorporated into chimeric antigen receptors (CAR), some show poor cell-surface trafficking and limited functionality. Redesign through a genetic algorithm-based diversification strategy that preserves the binding interface while changing non-binding surfaces experimentally reveals an isoelectric point (pI) window that improves CAR expression and enhances target-selective tumor cell killing. Our findings identify biochemical optimization beyond the binding interface as a critical requirement for translating AI-minibinders into functional applications.
Reference: Broske B, McEnroe BA, Frechen SC, Kempchen TN, Fandrey CI, Tan E, et al. AI-enabled discovery and biochemical optimization of minibinders targeting cancer cell-surface proteins. Nat Commun. 2026 Aug 20;17(1):8736. doi:10.1038/s41467-026-76760-5. PMID: 42624861; PMCID: PMC13494018.Checked: Full text checked - Abstract
Subcellular spatial transcriptomics can resolve tissue architecture at cellular scale, but sparse gene panels and limited detection sensitivity constrain downstream analysis. Existing enhancement methods often require tissue-matched single-cell RNA sequencing (scRNA-seq) references and dataset-specific retraining. Here we show that stPainter, a conditional generative model pretrained on a pan-cancer scRNA-seq atlas, can enhance spatial transcriptomics data without matched references or retraining. Using a latent diffusion architecture guided by Stochastic Differential Equations (SDE), stPainter reconstructs expanded expression profiles from sparse measurements and produces latent representations for clustering and cell-state analysis. When we apply stPainter upon 6 spatial transcriptomics datasets of different cancer types, we demonstrate that our model empowers downstream biological analyses, including fine-grained subpopulation clustering and pathway enrichment. Comparison with spatially resolved proteomics (CODEX) provided independent support for regional agreement between imputed cellular compositions and protein-level tissue organization. These results establish stPainter as a scalable approach for analyzing tumor microenvironments without auxiliary sequencing data.
Reference: Yang Y, Luo Y, Zhang K, Zhang Z, Peng H, Cao C, et al. Enhancing pan-cancer spatial transcriptomics at single-cell resolution with stPainter. Nat Commun. 2026 Aug 10;17(1):9578. doi:10.1038/s41467-026-76552-x. PMID: 42706255; PMCID: PMC13550648.Checked: Full text checked - Abstract
Spatial proteomics technologies have transformed our understanding of complex tissue architecture in cancer but present unique challenges for computational analysis 1 . Each study uses a different marker panel and protocol, and most methods are tailored to single cohorts, which limits knowledge transfer and robust biomarker discovery. Here we present Virtual Tissues (VirTues), a general-purpose foundation model for spatial proteomics that learns marker-aware, multi-scale representations of proteins, cells, niches and tissues directly from multiplex imaging data. From a single pretrained backbone, VirTues supports marker reconstruction, cell segmentation and typing, niche annotation, spatial biomarker discovery and patient stratification, including zero-shot annotation across heterogeneous panels and datasets. In triple-negative breast cancer, VirTues-derived biomarkers predict anti-PD-L1 chemo-immunotherapy response 2 and stratify disease-free survival in an independent cohort 3 , outperforming state-of-the-art biomarkers derived from the same datasets and current clinical stratification schemes.
Reference: Wenckstern J, Jain E, von Querfurth B, Cheng Y, Vasilev K, Pariset M, et al. The Virtual Tissues foundation model resolves spatial proteomics across scales. Nature. 2026 Aug 5;657(8132):744-54. doi:10.1038/s41586-026-10884-y. PMID: 42557331; PMCID: PMC13581600.Checked: Full text checked - Abstract
Cancer management remains fragmented across its continuum, from late-stage diagnosis and salvage therapies to non-personalized surveillance. Here, we present Oncoformer, a unified multimodal transformer model trained on the China Oncology Multimodal Prediction and Surveillance Study (COMPASS) cohort (3.67 million individuals, 17.7 million clinical visits) and validated on independent external cohorts, including the UK Biobank. Oncoformer integrates longitudinal electronic health records with chest X-ray imaging to address multiple clinical tasks: pan-cancer diagnosis (area under the receiver operating characteristic curve [AUROC] = 0.956), future cancer prediction up to 1 year before diagnosis (AUROC = 0.869), tumor stage inference (mean AUROC > 0.90), patient-specific treatment-response forecasting, and recurrence-free survival stratification across ten cancer types (all p < 0.01). Staging predictions were independently validated against postoperative pathological endpoints and shown to converge on core cancer genomic pathways. By translating routine clinical data into a dynamic view of cancer evolution, Oncoformer provides a framework for risk-informed cancer prediction and treatment stratification using routine clinical data.
Reference: Liu F, Wang K, Xu H, Tang C, Shen X, Wang M, et al. Advancing cancer detection and treatment using longitudinal routine clinical data. Cell. 2026 Sep;189(19):5995-6011.e8. doi:10.1016/j.cell.2026.07.009. PMID: 42508404.Checked: Abstract only - Abstract
Centrosome abnormalities (CA) are a hallmark of epithelial cancers, yet their spatial complexity and phenotypic heterogeneity remain poorly understood due to limitations in conventional image analysis. Here we present CenSegNet (Centrosome Segmentation Network), a modular deep learning framework for high-throughput segmentation of centrosomes and epithelial architecture, enabling accurate and generalisable centrosome phenotyping at spatial and single-cell resolution across imaging modalities and tissue contexts. Applied to tissue microarrays comprising 911 breast cancer cores from 127 patients, CenSegNet enables large-scale, spatially resolved quantification of numerical and structural CA. We show that these CA subtypes are mechanistically uncoupled, exhibiting distinct spatial distributions, age-dependent dynamics, and associations with tumour grade, hormone receptor status, genomic alterations and nodal involvement. Structural CA are associated with overall survival, whereas discordant CA profiles at tumour margins correlate with local tumour aggressiveness and stromal remodelling. These findings establish CenSegNet as a scalable platform for spatially resolved centrosome phenotyping, enabling systematic investigation of centrosome biology and its dysregulation in cancer and other epithelial diseases.
Reference: Cheng J, Fan K, Bailey M, Du X, Jena R, Savva C, et al. CenSegNet: a generalist high-throughput deep learning framework for centrosome phenotyping at spatial and single-cell resolution in heterogeneous tissues. Nat Commun. 2026 Jul 15;17(1):8713. doi:10.1038/s41467-026-75393-y. PMID: 42457714; PMCID: PMC13490574.Checked: Full text checked - Abstract
Current liquid biopsy methods for multi-cancer detection using plasma cell-free RNA (cfRNA, short RNA fragments circulating in blood that can reflect disease states) typically rely on gene annotations, which can overlook signals from unannotated or repetitive genomic regions. We present GeneLLM, a Transformer-based model that directly processes the nucleotide sequences of human-mapped cfRNA reads to identify cancer-indicative signatures. By bypassing gene-level quantification, the model retains signals from transcriptomic dark matter. The model learns latent pseudo-biomarkers (prototype representations from aggregated cfRNA read embeddings) that serve as discriminative features for cancer classification, rather than corresponding to explicit genomic sequences. Here we show that, in a multi-centre cohort, GeneLLM achieves ROC-AUC values ranging from 0.9250 to 0.9962 across several cancers, while maintaining comparable performance at one-sixth of the typical sequencing depth. These results suggest that sequence-level modelling of plasma cfRNA can capture diagnostically relevant information beyond annotation-dependent approaches, enabling more cost-efficient and scalable cancer screening.
Reference: Deng S, Sha L, Jin Y, Zhou T, Wang C, Liu Q, et al. Decoding cancer circulating transcriptomic signatures with language models. Nat Commun. 2026 Jul 9;17(1):8501. doi:10.1038/s41467-026-74411-3. PMID: 42425994; PMCID: PMC13483449.Checked: Full text checked - Abstract
Immune checkpoint inhibitors (ICIs) are a standard treatment across cancers, yet most patients do not respond, and existing biomarkers generalize poorly across tumor types and therapies. Here we present COMPASS, a pan-cancer foundation model that predicts immunotherapy response from bulk tumor transcriptomes using a concept bottleneck transformer. COMPASS encodes gene expression through 44 biologically grounded immune concepts representing immune cell states, tumor-microenvironment interaction and signaling pathways. Trained on 10,184 tumors across 33 cancer types, COMPASS achieves better average performance than 22 methods across 16 clinical cohorts spanning seven cancers and six ICIs, improving accuracy by 8.5% and area under the precision-recall curve by 15.7% on average across cohorts. COMPASS generalizes to cancer types and treatments not represented during fine-tuning and may inform indication selection and patient stratification. In survival analyses, patients classified by COMPASS as responders had longer overall survival (hazard ratio = 4.7, P < 0.0001). Personalized response maps connect gene expression to immune concepts, identifying programs associated with response and resistance; in immune-inflamed non-responders, COMPASS highlights programs including TGFβ signaling, endothelial exclusion, CD4 + T cell dysfunction and B cell deficiency. COMPASS predicts immunotherapy response and provides hypothesis-generating mechanistic insight for trial design and translational studies.
Reference: Shen W, Moon I, Nguyen TH, Li MM, Huang Y, Nair N, et al. Generalizable AI predicts immunotherapy outcomes across cancers and treatments. Nat Med. 2026 Jul 3;32(8):3010-22. doi:10.1038/s41591-026-04502-7. PMID: 42399673; PMCID: PMC13472881.Checked: Full text checked