119P - Predicting the pathogenicity of novel fusion genes and explaining reasons using a large language model: A focused assessment

Background

Fusion genes drive oncogenesis, and accurate pathogenicity assessment is crucial for therapeutic decisions. We have previously developed a system that leverages machine learning on known fusion genes to predict the pathogenicity of novel ones and explain its reasoning using large language model (LLM) (Cancers, accepted). This system achieved an accuracy of 0.98, comparable to the existing technologies, and provides explanations in natural language, a unique feature. However, these evaluations were performed on known data, leaving the accuracy and explanatory power for truly unknown fusion genes unclear. Evaluating explanatory power is challenging and requires multifaceted assessment.

Methods

This study focused on particular set of fusion genes with predictable mechanisms. We used BCR, involved in tumorigenesis, as the 5' partner, and selected 3' partners containing a kinase domain unreported with BCR. We input literature on BCR 5’ fusion genes into an LLM, and had it explain the mechanism of unreported BCR 5’ fusion genes with a 3’ kinase domain.

Results

Among the BCR fusion genes with unknown kinase partners, 30 were predicted as "Pathogenic” with high score. Among the 3' partner genes, receptor tyrosine kinases were the most frequent, followed by cell adhesion and extracellular matrix-related genes. The LLM generated domain-based hypotheses for the pathogenicity mechanisms of BCR fusion genes with unknown partners, considering the mechanisms of BCR::ABL1 and BCR::JAK2.

Conclusions

This study presents a novel approach for predicting and explaining the pathogenicity of unknown fusion genes, demonstrating its potential in certain cases. While the strength of evidence for the inferred mechanisms varied, this research suggests that more reliable explanations can be obtained for specific fusion genes, and further validation is needed for others. Experimental verification is necessary to confirm the validity of the hypotheses.

Clinical trial identification

Editorial acknowledgement

During the preparation of this work the author(s) used ChatGPT and Claude in order to suggest alternative phrasings, and improve clarity and readability. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the publication.

Legal entity responsible for the study

Fujitsu Ltd.

Funding

Fujitsu Ltd.

Disclosure

K. Murakami, S. Tago, S. Takishita, H. Morikawa, R. Kojima, M. Fuji: Financial Interests, Personal, Full or part-time Employment: Fujitsu Ltd.. K. Yokoyama, M. Ogawa, H. Fukushima, H. Takamori, Y. Nannya, S. Imoto: Financial Interests, Institutional, Funding: Fujitsu Ltd.