Traditional statistical approaches, such as co-expression network analysis and differential expression analysis, may misrepresent gene importance by overemphasizing statistical significance while overlooking functional relevance. As a result, biologically important genes can be underestimated, whereas statistically significant but less relevant genes may be overestimated.


To address this limitation, we propose a novel interpretable AI framework that integrates the Nucleotide Transformer—a biological foundation model that encodes nucleotide sequences—into probabilistic circuits, enabling theoretically grounded probabilistic interpretability. Using this framework, we trained a classification model to distinguish tumor cells from normal cells by jointly leveraging statistical signals from scRNA-seq data and biological representations derived from promoter sequences.


Importantly, the model provides gene-level probabilistic contributions for each prediction, enabling interpretation of how individual genes influence classification outcomes. Our analysis reveals that 1,524 out of 9,540 genes exhibit contradictory behavior: some genes are statistically more frequent in normal cells but contribute strongly to tumor classification, while others are more frequent in tumor cells yet contribute more to normal classification. This bidirectional discrepancy suggests that incorporating biological representations enables the model to overcome purely statistical biases and capture functionally relevant signals.
Notably, these genes include several well-established cancer-related genes, such as ITGA5, SIGLEC9, NOTUM, and TP73. These findings indicate that our approach can uncover biologically meaningful signals that are overlooked by conventional statistical methods.


Overall, this work goes beyond traditional gene expression–based analyses and provides a new framework for integrating biological knowledge with statistical data, offering deeper insights into gene function in cancer research.

Co-authors: None