AI and Cancer Research ── Cancer biology ── 2026-10-20
Cancer molecular biology and AI
- Abstract
Nature Biomedical Engineering, Published online: 19 October 2026; doi:10.1038/s41551-026-01803-7 An interpretable neural network, SPACER, learns the rules of cellular recruitment directly from tissue spatial transcriptomics data. Analyses of multiple cancers reveal a role for the local antigen presentation machinery and tumour mucins in T cell recruitment to tumours, and a mouse model of severe checkpoint inhibitor-induced myocarditis identifies CD4+ T cells as key disease effectors.
Journal IF-equivalent: 15.4 (OpenAlex 2-year mean citedness, value as of 2026-10-19, retrieved 2026-10-20; not the official Clarivate IF)Reference: [No authors listed]. An interpretable AI model decodes the rules of cell recruitment to specific destinations. Nat Biomed Eng. 2026 Oct 19 [Epub ahead of print]. doi:10.1038/s41551-026-01803-7.Checked: Abstract only - Abstract
Genomic association and predictive modeling often operate in high-dimensional regimes where the number of candidate genes far exceeds the number of study subjects, making inference prone to instability and sensitive to missing values, batch effects, and sparse signals. Meanwhile, biology provides rich relational context (protein interactions, pathway co-membership, regulatory targeting, and more) that encodes functional relationships between genes. The recent success of deep learning has demonstrated effective model designs for integrating heterogeneous directed and symmetric relationships into a compact unified representation. Existing methods that leverage graph-based representations of biological knowledge fall along several axes: some use them as input for deep learning models with predictive modules that generate outcome predictions for supervised tasks; others learn general-purpose numerical representation (embeddings) of genes for use in downstream analysis; and a separate line of research constructs regularization priors from the biological sources tailored to a specific application. To the best of our knowledge, no prior work has trained a model to integrate multiple biological sources by learning their relative reliability to build a task-agnostic Laplacian-based smoothness prior for epidemiological inference. We introduce a framework that fuses complementary biological networks into a single symmetric gene-gene coupling matrix via a Heterogeneous Graph Attention Network (HGAT), creating a reusable prior that can regularize any model with gene-indexed parameters. We first perform an intrinsic methodological validation of the approach against multiple benchmark models to assess the embeddings and output graph quality. Through controlled simulation we establish that a faithful prior must capture multi-hop functional coherence rather than direct edges alone, and that the value of learning lies specifically in the adaptive, reliability-weighted fusion of incomplete, heterogeneous sources, where fixed kernels and uniformly-weighted linear baselines fall short. We further validate that the learned prior generalizes by preserving graded biological similarity on regulatory sources held out from training, maintains competitive performance in downstream cancer prediction, and demonstrates a concrete practical advantage in a niche, under-annotated regulatory application where generalization matters. The resulting prior offers transparent biological provenance and proven generalization by integrating fragmented biological knowledge into convex inference without sacrificing predictive performance.
Journal IF-equivalent: 4.8 (OpenAlex 2-year mean citedness, value as of 2026-10-19, retrieved 2026-10-20; not the official Clarivate IF)Reference: Origgi JM, Zeegers M, Mehrkanoon S, von Moltke C, Wesselius A, Nattino G. Genes2Reg: Constructing a biologically-faithful gene coupling prior via deep adaptive fusion of heterogeneous knowledge sources. BMC Bioinformatics. 2026 Oct 17 [Epub ahead of print]. doi:10.1186/s12859-026-06655-1.Checked: Abstract only