The LLM attention mechanism sharpens the diagnosis of breast cancer subtypes

By Zendoric · September 24, 2026. A team from the Islamic University of Kushtia, in Bangladesh, has published in BMC Bioinformatics an artificial intelligence model that simultaneously reads three different layers of a breast tumor's biology —which genes are switched on, which microRNAs regulate them and which chemical methylation marks silence them— in order to classify more accurately the cancer subtype each patient belongs to. The framework, named HCAM-BRCA (Hierarchical Cross-Attention Multi-omics), achieves in its best configuration a Matthews correlation coefficient of 0.7628 —a demanding metric that penalizes easy hits on the majority class— and an area under the precision-recall curve of 0.8925, according to the study's own authors, Md. Neaz Ali and Suman Biswas.

Subtype classification today relies on PAM50, a 50-gene signature that sorts tumors into categories such as Luminal A, Luminal B, HER2-enriched and Basal-like, each with a different prognosis and treatment. The problem, the authors note, is that PAM50 fails precisely at the boundary between Luminal B and HER2-enriched, a confusion that can change the therapy a patient receives. Previous methods for integrating "omics" data —the term for combining several molecular layers of the same tumor— tended to cross at most two of those layers at a time; HCAM-BRCA is among the first to model all three at once and to respect the real hierarchy between them, in which methylation conditions gene expression and microRNAs then modulate it.

The central technical piece is cross-attention: the same mathematical mechanism used by large language models (LLMs, the systems behind conversational assistants such as ChatGPT or Claude) to decide which parts of an input matter most for interpreting another. Applied here, the model first learns which features within each omics layer relate to one another, and then which layers explain the rest. The advantage over a conventional neural network is that the attention weights are legible: instead of a black box, the system delivers a ranked list of the genes that weighed most in each classification, which allowed the authors to verify that those priorities match what biology already knows —genes tied to transcriptional regulation, chromatin organization and immune signaling, a pattern consistent with what would be expected from independent functional analyses based on Gene Ontology and KEGG annotations— and not statistical noise.

The team trained and validated the system with data from The Cancer Genome Atlas (TCGA), the public repository of tumor profiles that has become the standard testbed of computational oncology, and corrected the imbalance between subtypes —there are far fewer HER2-enriched samples than Luminal A, for example— with an oversampling technique called ADASYN. On that basis they tested twelve different machine learning algorithms; CatBoost, a gradient boosting method, was the most accurate, and Extra Trees, an ensemble of decision trees, the most consistent at detecting the minority class. With five-fold cross-validation —the technique that trains and tests the model on different subsets of the data to keep it from memorizing— integrating the three layers outperformed the two-layer combinations across every classifier tested, and the framework performed well precisely at the Luminal B/HER2-enriched boundary that PAM50 handles worst.

The finding with the most clinical mileage comes from a second step: cross-referencing each subtype's priority genes with patient survival. In Basal-like tumors —the most aggressive variant, heavily overlapping with triple-negative breast cancer— the genes CUX1, MYZAP and CEBPA were associated with how the disease progressed; in Luminal B, HOXA10 and the immune gene HLA-DQA2; in HER2-enriched, GDF10 and ZNF879. They are, according to the authors, candidate biomarkers within each subtype, not across all of them at once, which is how precision oncology operates in practice: you do not treat "a breast cancer" but a specific subtype in a specific patient.

It is best not to get ahead of events. The study rests entirely on existing public data —no new patients were recruited—, was published without external funding, and the authors themselves are clear about it: the genes identified are hypotheses that need to be confirmed in independent cohorts and, later, in the clinic. HCAM-BRCA also arrives amid a wave of multi-omics integration methods —such as MOGONET or moBRCA-net— with which it competes for credibility, rather than as an isolated leap.

What strikes us as most significant about this work is not the decimal place of its metric, but the design choice: using attention so that the model explains its own decisions. In medicine, accuracy without explanation has a low ceiling —neither a regulator nor an oncologist signs off on a treatment based on a black box—, so interpretability is not a cosmetic add-on, it is the condition for these tools to reach a hospital one day. It also points to something we have been seeing throughout 2026: the same machinery that trains chatbots is being repurposed, with relatively modest design adaptations, to read biology, and it is being done by a university team with no large budget working solely with public data. That accessibility of the method —a well-known architecture, open TCGA data— is the kind of abundance that underpins our underlying thesis: the more hands that can apply these techniques to biology, the sooner we will stop treating cancer as a single disease and start treating, with growing precision, the exact variant in each patient. The road there is still long —years separate a paper from a clinical trial, and most candidate biomarkers never make it into practice— but it is exactly the kind of small, verifiable brick on which, over time, the eradication of a disease is built.