SpikeFormer: A Spiking Transformer-Based Cancer Detection Model

Background Cancer remains one of the leading causes of mortality worldwide, with accurate and early diagnosis critical for effective treatment. Advances in machine learning algorithms have prompted the distinguishment ability between cancerous and normal tissues based on gene expression profiling. However, large-scale tumor classification models that can differentiate both among various cancer types and between cancerous and normal samples remain limited. Methods To address this challenge, this study proposes a novel deep learning framework, SpikeFormer, which integrates the attention mechanism of transformers with spiking neuron dynamics to model complex temporal dependencies in gene expression data effectively. SpikeFormer was evaluated on The Cancer Genome Atlas (TCGA) dataset, comprising 10,632 samples across 33 cancer types, demonstrating its robustness in both binary and multi-class classification tasks. Results In binary classification experiments, the model achieved an average accuracy of 99.44% across four TCGA sub-datasets, demonstrating its ability to distinguish between cancerous and normal samples. In multi-class classification, SpikeFormer achieved an accuracy of 96.90% and a Cohen’s Kappa score of 97.00%, outperforming the baseline Transformer model. Conclusions These results highlight the reliability and scalability of the proposed SpikeFormer model for cancer classification, suggesting its potential application in clinical diagnosis and treatment planning. Although SpikeFormer demonstrates promising performance on TCGA RNA-seq datasets, the current evaluation is limited to retrospective public data, and further validation using independent multi-center cohorts and prospective clinical datasets is required before potential clinical translation. This work was supported by the grant from National Natural Science Foundation of China [62561042], the Natural Science Foundation of Inner Mongolia Autonomous Region of China [2024ZD30], and the Fundamental Research Funds for the Central Universities [Y03023206300125113]. The Cancer Genome Atlas (TCGA) data used in this study are available from dbGaP under accession number phs000178.