Enformer Prediction Performance Evaluation Leveraging Deep Learning to Interpret Genetic Variation in Prostate Cancer

Abstract

Prostate cancer is the most commonly diagnosed disease and the main cause of cancer-related deaths among men across the globe. One of the main challenges in the field of prostate cancer biology is that prostate cancer is genetically diverse, with genetic variants are a key factor in defining how the disease develops and progresses. This genetic variation complicates the ability to distinguish between indolent (slow-growing) and aggressive prostate tumors. The challenge arises due to the lack of understanding of the processes responsible for tumor development in prostate cancer. Despite significant advances in genome-wide association studies (GWAS), interpreting the roles of non-coding and coding variants in gene regulation remains a significant obstacle in functional genomics. The difficulty mainly lies in the fact that most variants located in the non-coding regions of the genome, which make up 98% of the DNA. Understanding these variants using traditional methods is both time consuming and necessitates extensive laboratory testing. This situation highlights the urgent need for computational tools that can quickly prioritize variants based on their predicted functional effects. Deep learning models like Enformer present potential solutions for rapid, DNA sequence-based predictions regarding the effects of variants on gene expression and chromatin states. However, the specific applications and effectiveness of these models in the context of prostate cancer have not been thoroughly investigated. This study aimed to assess the capability of the Enformer model to predict the functional impacts of genetic variants linked to prostate cancer and to determine if the model could better understand and interpret human genetic variation and biology in prostate cancer settings. The Enformer model was employed to predict the effects of variants on gene expression and chromatin states, providing numerical scores that reflect the change and magnitude of their impact on gene expression. Well-known prostate genes were chosen for this study. Variants were categorized into coding and non-coding regions, and then stratified based on their link to established prostate cancer genes versus other genes. The research sought to investigate how effectively the model predicted gene expression scores, the “CAGE: Prostate Epithelial Cells” cell line was chosen. The Mann-Whitney U test was utilized to compare coding variants related to prostate genes and variants from other genes, and a similar evaluation was made for non-coding variants. For those variants categorized under “other genes” that received higher predicted scores, their corresponding genes underwent enrichment analysis using various platforms (Enrichr, ShinyGO, g:Profiler) to assess their biological relevance to prostate cancer. High scoring variants from both prostate-associated genes and other genes, along with recurrent variants observed in many patients, were prioritized for presentation. The results indicated that the Enformer model did not assign significantly higher predicted expression scores to prostate gene associated variants compared to variants from other genes, with p-values of 0.98 for coding regions and 0.15 for non-coding regions. However, enrichment analysis of the top-scoring genes within the other genes group showed their association with prostate cancer, and many recurrent variants were identified. This suggests that the Enformer model can potentially provide valuable targets for further research and treatment. Overall, the findings imply that while the Enformer model may not currently be effective for predicting prostate cancer specific outcomes, it can still be useful for discovering new genetic loci associated with the disease. The inability to predict higher scores for known prostate genes indicates a limitation in the model’s ability to learn tissue-specific information. Thus, there is a clear need for training the model with more prostate specific data to improve its predictive capabilities, which could enhance its application in precision oncology.